update readme

choices.py

@@ -0,0 +1,179 @@
+from enum import Enum
+from torch import nn
+
+
+class TrainMode(Enum):
+    # manipulate mode = training the classifier
+    manipulate = 'manipulate'
+    # default training mode!
+    diffusion = 'diffusion'
+    # default latent training mode!
+    # fitting a diffusion model to a given latent
+    latent_diffusion = 'latentdiffusion'
+    
+    def is_manipulate(self):
+        return self in [
+            TrainMode.manipulate,
+        ]
+
+    def is_diffusion(self):
+        return self in [
+            TrainMode.diffusion,
+            TrainMode.latent_diffusion,
+        ]
+
+    def is_autoenc(self):
+        # the network possibly does autoencoding
+        return self in [
+            TrainMode.diffusion,
+        ]
+
+    def is_latent_diffusion(self):
+        return self in [
+            TrainMode.latent_diffusion,
+        ]
+
+    def use_latent_net(self):
+        return self.is_latent_diffusion()
+
+    def require_dataset_infer(self):
+        """
+        whether training in this mode requires the latent variables to be available?
+        """
+        # this will precalculate all the latents before hand
+        # and the dataset will be all the predicted latents
+        return self in [
+            TrainMode.latent_diffusion,
+            TrainMode.manipulate,
+        ]
+
+
+class ManipulateMode(Enum):
+    """
+    how to train the classifier to manipulate
+    """
+    # train on whole celeba attr dataset
+    celebahq_all = 'celebahq_all'
+    # celeba with D2C's crop
+    d2c_fewshot = 'd2cfewshot'
+    d2c_fewshot_allneg = 'd2cfewshotallneg'
+
+    def is_celeba_attr(self):
+        return self in [
+            ManipulateMode.d2c_fewshot,
+            ManipulateMode.d2c_fewshot_allneg,
+            ManipulateMode.celebahq_all,
+        ]
+
+    def is_single_class(self):
+        return self in [
+            ManipulateMode.d2c_fewshot,
+            ManipulateMode.d2c_fewshot_allneg,
+        ]
+
+    def is_fewshot(self):
+        return self in [
+            ManipulateMode.d2c_fewshot,
+            ManipulateMode.d2c_fewshot_allneg,
+        ]
+
+    def is_fewshot_allneg(self):
+        return self in [
+            ManipulateMode.d2c_fewshot_allneg,
+        ]
+
+
+class ModelType(Enum):
+    """
+    Kinds of the backbone models
+    """
+
+    # unconditional ddpm
+    ddpm = 'ddpm'
+    # autoencoding ddpm cannot do unconditional generation
+    autoencoder = 'autoencoder'
+
+    def has_autoenc(self):
+        return self in [
+            ModelType.autoencoder,
+        ]
+
+    def can_sample(self):
+        return self in [ModelType.ddpm]
+
+
+class ModelName(Enum):
+    """
+    List of all supported model classes
+    """
+
+    beatgans_ddpm = 'beatgans_ddpm'
+    beatgans_autoenc = 'beatgans_autoenc'
+
+
+class ModelMeanType(Enum):
+    """
+    Which type of output the model predicts.
+    """
+
+    eps = 'eps'  # the model predicts epsilon
+
+
+class ModelVarType(Enum):
+    """
+    What is used as the model's output variance.
+
+    The LEARNED_RANGE option has been added to allow the model to predict
+    values between FIXED_SMALL and FIXED_LARGE, making its job easier.
+    """
+
+    # posterior beta_t
+    fixed_small = 'fixed_small'
+    # beta_t
+    fixed_large = 'fixed_large'
+
+
+class LossType(Enum):
+    mse = 'mse'  # use raw MSE loss (and KL when learning variances)
+    l1 = 'l1'
+
+
+class GenerativeType(Enum):
+    """
+    How's a sample generated
+    """
+
+    ddpm = 'ddpm'
+    ddim = 'ddim'
+
+
+class OptimizerType(Enum):
+    adam = 'adam'
+    adamw = 'adamw'
+
+
+class Activation(Enum):
+    none = 'none'
+    relu = 'relu'
+    lrelu = 'lrelu'
+    silu = 'silu'
+    tanh = 'tanh'
+
+    def get_act(self):
+        if self == Activation.none:
+            return nn.Identity()
+        elif self == Activation.relu:
+            return nn.ReLU()
+        elif self == Activation.lrelu:
+            return nn.LeakyReLU(negative_slope=0.2)
+        elif self == Activation.silu:
+            return nn.SiLU()
+        elif self == Activation.tanh:
+            return nn.Tanh()
+        else:
+            raise NotImplementedError()
+
+
+class ManipulateLossType(Enum):
+    bce = 'bce'
+    mse = 'mse'

config.py

@@ -0,0 +1,153 @@
+from model.blocks import *
+from diffusion.resample import UniformSampler
+from dataclasses import dataclass
+from diffusion.diffusion import space_timesteps
+from typing import Tuple
+from config_base import BaseConfig
+from diffusion import *
+from diffusion.base import GenerativeType, LossType, ModelMeanType, ModelVarType, get_named_beta_schedule
+from model import *
+from choices import *
+from preprocess import *
+import os
+
+@dataclass
+class TrainConfig(BaseConfig):
+    name: str = ''
+    base_dir: str = './checkpoints/'
+    logdir: str = f'{base_dir}{name}'
+    data_name: str = ''
+    data_val_name: str = ''
+    seq_len: int = 40 # for reconstruction
+    seq_len_future: int = 3 # for prediction
+    in_channels = 9
+    fp16: bool = True
+    lr: float = 1e-4
+    ema_decay: float = 0.9999
+    seed: int = 0 # random seed
+    batch_size: int = 64
+    accum_batches: int = 1
+    batch_size_eval: int = 1024
+    total_epochs: int = 1_000
+    save_every_epochs: int = 10
+    eval_every_epochs: int = 10
+    train_mode: TrainMode = TrainMode.diffusion
+    T: int = 1000
+    T_eval: int = 100
+    diffusion_type: str = 'beatgans'
+    semantic_encoder_type: str = 'gcn'
+    net_beatgans_embed_channels: int = 128
+    beatgans_gen_type: GenerativeType = GenerativeType.ddim
+    beatgans_loss_type: LossType = LossType.mse
+    hand_mse_factor = 1.0
+    head_mse_factor = 1.0
+    beatgans_model_mean_type: ModelMeanType = ModelMeanType.eps
+    beatgans_model_var_type: ModelVarType = ModelVarType.fixed_large
+    beatgans_rescale_timesteps: bool = False
+    beta_scheduler: str = 'linear'    
+    net_ch: int = 64
+    net_ch_mult: Tuple[int, ...]= (1, 2, 4)    
+    net_enc_channel_mult: Tuple[int]  = (1, 2, 4)
+    grad_clip: float = 1
+    optimizer: OptimizerType = OptimizerType.adam
+    weight_decay: float = 0
+    warmup: int = 0
+    model_conf: ModelConfig = None
+    model_name: ModelName = ModelName.beatgans_autoenc
+    model_type: ModelType = None
+    
+    @property
+    def batch_size_effective(self):
+        return self.batch_size*self.accum_batches
+        
+    def _make_diffusion_conf(self, T=None):
+        if self.diffusion_type == 'beatgans':
+            # can use T < self.T for evaluation
+            # follows the guided-diffusion repo conventions
+            # t's are evenly spaced
+            if self.beatgans_gen_type == GenerativeType.ddpm:
+                section_counts = [T]
+            elif self.beatgans_gen_type == GenerativeType.ddim:
+                section_counts = f'ddim{T}'
+            else:
+                raise NotImplementedError()
+
+            return SpacedDiffusionBeatGansConfig(
+                gen_type=self.beatgans_gen_type,
+                model_type=self.model_type,
+                betas=get_named_beta_schedule(self.beta_scheduler, T),
+                model_mean_type=self.beatgans_model_mean_type,
+                model_var_type=self.beatgans_model_var_type,
+                loss_type=self.beatgans_loss_type,
+                rescale_timesteps=self.beatgans_rescale_timesteps,
+                use_timesteps=space_timesteps(num_timesteps=T, section_counts=section_counts),
+                fp16=self.fp16,
+            )
+        else:
+            raise NotImplementedError()
+
+    @property
+    def model_out_channels(self):
+        return self.in_channels
+    
+    @property
+    def model_input_channels(self):
+        return self.in_channels
+        
+    def make_T_sampler(self):        
+        return UniformSampler(self.T)
+
+    def make_diffusion_conf(self):
+        return self._make_diffusion_conf(self.T)
+
+    def make_eval_diffusion_conf(self):
+        return self._make_diffusion_conf(T=self.T_eval)
+
+    def make_model_conf(self):
+        cls = BeatGANsAutoencConfig
+        if self.model_name == ModelName.beatgans_autoenc:
+            self.model_type = ModelType.autoencoder
+        else:
+            raise NotImplementedError()
+            
+        self.model_conf = cls(
+            semantic_encoder_type = self.semantic_encoder_type,
+            channel_mult=self.net_ch_mult,
+            seq_len = self.seq_len,
+            seq_len_future = self.seq_len_future,
+            embed_channels=self.net_beatgans_embed_channels,
+            enc_out_channels=self.net_beatgans_embed_channels,
+            enc_channel_mult=self.net_enc_channel_mult,                
+            in_channels=self.model_input_channels,
+            model_channels=self.net_ch,            
+            out_channels=self.model_out_channels,                
+        )
+        
+        return self.model_conf
+        
+def egobody_autoenc(mode, encoder_type='gcn', hand_mse_factor=1.0, head_mse_factor=1.0, data_sample_rate=1, epoch=130,in_channels=9, seq_len=40):
+    conf = TrainConfig()
+    conf.seq_len = seq_len
+    conf.seq_len_future = 3
+    conf.in_channels = in_channels
+    conf.net_beatgans_embed_channels = 128
+    conf.net_ch = 64
+    conf.net_ch_mult = (1, 1, 1)
+    conf.semantic_encoder_type = encoder_type
+    conf.hand_mse_factor = hand_mse_factor
+    conf.head_mse_factor = head_mse_factor   
+    conf.net_enc_channel_mult = conf.net_ch_mult
+    conf.total_epochs = epoch
+    conf.save_every_epochs = 10
+    conf.eval_every_epochs = 10
+    conf.batch_size = 64
+    conf.batch_size_eval = 1024*4
+    
+    conf.data_dir = "/scratch/hu/pose_forecast/egobody_pose2gaze/"
+    conf.data_sample_rate = data_sample_rate
+    conf.name = 'egobody_autoenc'
+    conf.data_name = 'egobody'
+    
+    conf.mode = mode
+    conf.make_model_conf()
+    return conf

config_base.py

@@ -0,0 +1,72 @@
+import json
+import os
+from copy import deepcopy
+from dataclasses import dataclass
+
+
+@dataclass
+class BaseConfig:
+    def clone(self):
+        return deepcopy(self)
+
+    def inherit(self, another):
+        """inherit common keys from a given config"""
+        common_keys = set(self.__dict__.keys()) & set(another.__dict__.keys())
+        for k in common_keys:
+            setattr(self, k, getattr(another, k))
+
+    def propagate(self):
+        """push down the configuration to all members"""
+        for k, v in self.__dict__.items():
+            if isinstance(v, BaseConfig):
+                v.inherit(self)
+                v.propagate()
+
+    def save(self, save_path):
+        """save config to json file"""
+        dirname = os.path.dirname(save_path)
+        if not os.path.exists(dirname):
+            os.makedirs(dirname)
+        conf = self.as_dict_jsonable()
+        with open(save_path, 'w') as f:
+            json.dump(conf, f)
+
+    def load(self, load_path):
+        """load json config"""
+        with open(load_path) as f:
+            conf = json.load(f)
+        self.from_dict(conf)
+
+    def from_dict(self, dict, strict=False):
+        for k, v in dict.items():
+            if not hasattr(self, k):
+                if strict:
+                    raise ValueError(f"loading extra '{k}'")
+                else:
+                    print(f"loading extra '{k}'")
+                    continue
+            if isinstance(self.__dict__[k], BaseConfig):
+                self.__dict__[k].from_dict(v)
+            else:
+                self.__dict__[k] = v
+
+    def as_dict_jsonable(self):
+        conf = {}
+        for k, v in self.__dict__.items():
+            if isinstance(v, BaseConfig):
+                conf[k] = v.as_dict_jsonable()
+            else:
+                if jsonable(v):
+                    conf[k] = v
+                else:
+                    # ignore not jsonable
+                    pass
+        return conf
+
+
+def jsonable(x):
+    try:
+        json.dumps(x)
+        return True
+    except TypeError:
+        return False

diffusion/__init__.py

@@ -0,0 +1,6 @@
+from typing import Union
+
+from .diffusion import SpacedDiffusionBeatGans, SpacedDiffusionBeatGansConfig
+
+Sampler = Union[SpacedDiffusionBeatGans]
+SamplerConfig = Union[SpacedDiffusionBeatGansConfig]

diffusion/diffusion.py

@@ -0,0 +1,182 @@
+from .base import *
+from dataclasses import dataclass
+
+
+def space_timesteps(num_timesteps, section_counts):
+    """
+    Create a list of timesteps to use from an original diffusion process,
+    given the number of timesteps we want to take from equally-sized portions
+    of the original process.
+
+    For example, if there's 300 timesteps and the section counts are [10,15,20]
+    then the first 100 timesteps are strided to be 10 timesteps, the second 100
+    are strided to be 15 timesteps, and the final 100 are strided to be 20.
+
+    If the stride is a string starting with "ddim", then the fixed striding
+    from the DDIM paper is used, and only one section is allowed.
+
+    :param num_timesteps: the number of diffusion steps in the original
+                          process to divide up.
+    :param section_counts: either a list of numbers, or a string containing
+                           comma-separated numbers, indicating the step count
+                           per section. As a special case, use "ddimN" where N
+                           is a number of steps to use the striding from the
+                           DDIM paper.
+    :return: a set of diffusion steps from the original process to use.
+    """
+    if isinstance(section_counts, str):
+        if section_counts.startswith("ddim"):
+            desired_count = int(section_counts[len("ddim"):])
+            for i in range(1, num_timesteps):
+                if len(range(0, num_timesteps, i)) == desired_count:
+                    return set(range(0, num_timesteps, i))
+            raise ValueError(
+                f"cannot create exactly {num_timesteps} steps with an integer stride"
+            )
+        section_counts = [int(x) for x in section_counts.split(",")]
+    size_per = num_timesteps // len(section_counts)
+    extra = num_timesteps % len(section_counts)
+    start_idx = 0
+    all_steps = []
+    for i, section_count in enumerate(section_counts):
+        size = size_per + (1 if i < extra else 0)
+        if size < section_count:
+            raise ValueError(
+                f"cannot divide section of {size} steps into {section_count}")
+        if section_count <= 1:
+            frac_stride = 1
+        else:
+            frac_stride = (size - 1) / (section_count - 1)
+        cur_idx = 0.0
+        taken_steps = []
+        for _ in range(section_count):
+            taken_steps.append(start_idx + round(cur_idx))
+            cur_idx += frac_stride
+        all_steps += taken_steps
+        start_idx += size
+    return set(all_steps)
+
+
+@dataclass
+class SpacedDiffusionBeatGansConfig(GaussianDiffusionBeatGansConfig):
+    use_timesteps: Tuple[int] = None
+
+    def make_sampler(self):
+        return SpacedDiffusionBeatGans(self)
+
+
+class SpacedDiffusionBeatGans(GaussianDiffusionBeatGans):
+    """
+    A diffusion process which can skip steps in a base diffusion process.
+
+    :param use_timesteps: a collection (sequence or set) of timesteps from the
+                          original diffusion process to retain.
+    :param kwargs: the kwargs to create the base diffusion process.
+    """
+    def __init__(self, conf: SpacedDiffusionBeatGansConfig):
+        self.conf = conf
+        self.use_timesteps = set(conf.use_timesteps)
+        # how the new t's mapped to the old t's
+        self.timestep_map = []
+        self.original_num_steps = len(conf.betas)
+
+        base_diffusion = GaussianDiffusionBeatGans(conf)  # pylint: disable=missing-kwoa
+        last_alpha_cumprod = 1.0
+        new_betas = []
+        for i, alpha_cumprod in enumerate(base_diffusion.alphas_cumprod):
+            if i in self.use_timesteps:
+                # getting the new betas of the new timesteps
+                new_betas.append(1 - alpha_cumprod / last_alpha_cumprod)
+                last_alpha_cumprod = alpha_cumprod
+                self.timestep_map.append(i)
+        conf.betas = np.array(new_betas)
+        super().__init__(conf)
+
+    def p_mean_variance(self, model: Model, *args, **kwargs):  # pylint: disable=signature-differs
+        return super().p_mean_variance(self._wrap_model(model), *args,
+                                       **kwargs)
+
+    def training_losses(self, model: Model, *args, **kwargs):  # pylint: disable=signature-differs
+        return super().training_losses(self._wrap_model(model), *args,
+                                       **kwargs)
+
+    def condition_mean(self, cond_fn, *args, **kwargs):
+        return super().condition_mean(self._wrap_model(cond_fn), *args,
+                                      **kwargs)
+
+    def condition_score(self, cond_fn, *args, **kwargs):
+        return super().condition_score(self._wrap_model(cond_fn), *args,
+                                       **kwargs)
+
+    def _wrap_model(self, model: Model):
+        if isinstance(model, _WrappedModel):
+            return model
+        return _WrappedModel(model, self.timestep_map, self.rescale_timesteps,
+                             self.original_num_steps)
+
+    def _scale_timesteps(self, t):
+        # Scaling is done by the wrapped model.
+        return t
+
+
+class _WrappedModel:
+    """
+    converting the supplied t's to the old t's scales.
+    """
+    def __init__(self, model, timestep_map, rescale_timesteps,
+                 original_num_steps):
+        self.model = model
+        self.timestep_map = timestep_map
+        self.rescale_timesteps = rescale_timesteps
+        self.original_num_steps = original_num_steps
+
+    def forward(self, x, t, t_cond=None, **kwargs):
+        """
+        Args:
+            t: t's with differrent ranges (can be << T due to smaller eval T) need to be converted to the original t's
+            t_cond: the same as t but can be of different values
+        """
+        map_tensor = th.tensor(self.timestep_map,
+                               device=t.device,
+                               dtype=t.dtype)
+
+        def do(t):
+            new_ts = map_tensor[t]
+            if self.rescale_timesteps:
+                new_ts = new_ts.float() * (1000.0 / self.original_num_steps)
+            return new_ts
+
+        if t_cond is not None:
+            # support t_cond
+            t_cond = do(t_cond)
+        ## run_ffhq256.py None
+        return self.model(x=x, t=do(t), t_cond=t_cond, **kwargs) ## self.model.__class__==model.unet_autoenc.BeatGANsAutoencModel
+
+    def __getattr__(self, name):
+        # allow for calling the model's methods
+        if hasattr(self.model, name):
+            func = getattr(self.model, name)
+            return func
+        raise AttributeError(name)
+    
+    # def __call__(self, x, t, t_cond=None, **kwargs):
+    #     """
+    #     Args:
+    #         t: t's with differrent ranges (can be << T due to smaller eval T) need to be converted to the original t's
+    #         t_cond: the same as t but can be of different values
+    #     """
+    #     map_tensor = th.tensor(self.timestep_map,
+    #                            device=t.device,
+    #                            dtype=t.dtype)
+
+    #     def do(t):
+    #         new_ts = map_tensor[t]
+    #         if self.rescale_timesteps:
+    #             new_ts = new_ts.float() * (1000.0 / self.original_num_steps)
+    #         return new_ts
+
+    #     if t_cond is not None:
+    #         # support t_cond
+    #         t_cond = do(t_cond)
+    #     ## run_ffhq256.py None
+    #     return self.model(x=x, t=do(t), t_cond=t_cond, **kwargs)

diffusion/resample.py

@@ -0,0 +1,63 @@
+from abc import ABC, abstractmethod
+
+import numpy as np
+import torch as th
+import torch.distributed as dist
+
+
+def create_named_schedule_sampler(name, diffusion):
+    """
+    Create a ScheduleSampler from a library of pre-defined samplers.
+
+    :param name: the name of the sampler.
+    :param diffusion: the diffusion object to sample for.
+    """
+    if name == "uniform":
+        return UniformSampler(diffusion)
+    else:
+        raise NotImplementedError(f"unknown schedule sampler: {name}")
+
+
+class ScheduleSampler(ABC):
+    """
+    A distribution over timesteps in the diffusion process, intended to reduce
+    variance of the objective.
+
+    By default, samplers perform unbiased importance sampling, in which the
+    objective's mean is unchanged.
+    However, subclasses may override sample() to change how the resampled
+    terms are reweighted, allowing for actual changes in the objective.
+    """
+    @abstractmethod
+    def weights(self):
+        """
+        Get a numpy array of weights, one per diffusion step.
+
+        The weights needn't be normalized, but must be positive.
+        """
+
+    def sample(self, batch_size, device):
+        """
+        Importance-sample timesteps for a batch.
+
+        :param batch_size: the number of timesteps.
+        :param device: the torch device to save to.
+        :return: a tuple (timesteps, weights):
+                 - timesteps: a tensor of timestep indices.
+                 - weights: a tensor of weights to scale the resulting losses.
+        """
+        w = self.weights()
+        p = w / np.sum(w)
+        indices_np = np.random.choice(len(p), size=(batch_size, ), p=p)
+        indices = th.from_numpy(indices_np).long().to(device)
+        weights_np = 1 / (len(p) * p[indices_np])
+        weights = th.from_numpy(weights_np).float().to(device)
+        return indices, weights
+
+        
+class UniformSampler(ScheduleSampler):
+    def __init__(self, num_timesteps): ## all steps are 1
+        self._weights = np.ones([num_timesteps])
+
+    def weights(self):
+        return self._weights

environment/haheae.yml

@@ -0,0 +1,101 @@
+name: haheae
+channels:
+  - defaults
+dependencies:
+  - _libgcc_mutex=0.1=main
+  - _openmp_mutex=5.1=1_gnu
+  - ca-certificates=2023.12.12=h06a4308_0
+  - ld_impl_linux-64=2.38=h1181459_1
+  - libffi=3.4.4=h6a678d5_0
+  - libgcc-ng=11.2.0=h1234567_1
+  - libgomp=11.2.0=h1234567_1
+  - libstdcxx-ng=11.2.0=h1234567_1
+  - ncurses=6.4=h6a678d5_0
+  - openssl=3.0.12=h7f8727e_0
+  - pip=23.3.1=py38h06a4308_0
+  - python=3.8.18=h955ad1f_0
+  - readline=8.2=h5eee18b_0
+  - setuptools=68.2.2=py38h06a4308_0
+  - sqlite=3.41.2=h5eee18b_0
+  - tk=8.6.12=h1ccaba5_0
+  - wheel=0.41.2=py38h06a4308_0
+  - xz=5.4.5=h5eee18b_0
+  - zlib=1.2.13=h5eee18b_0
+  - pip:
+      - absl-py==2.0.0
+      - aiohttp==3.9.1
+      - aiosignal==1.3.1
+      - appdirs==1.4.4
+      - async-timeout==4.0.3
+      - attrs==23.2.0
+      - cachetools==5.3.2
+      - certifi==2023.11.17
+      - charset-normalizer==3.3.2
+      - click==8.1.7
+      - contourpy==1.1.1
+      - cycler==0.12.1
+      - cython==0.29.37
+      - docker-pycreds==0.4.0
+      - fonttools==4.47.2
+      - frozenlist==1.4.1
+      - fsspec==2023.12.2
+      - ftfy==6.1.3
+      - future==0.18.3
+      - gitdb==4.0.11
+      - gitpython==3.1.41
+      - google-auth==2.26.2
+      - google-auth-oauthlib==1.0.0
+      - grpcio==1.60.0
+      - hdbscan==0.8.33
+      - idna==3.6
+      - importlib-metadata==7.0.1
+      - importlib-resources==6.1.1
+      - joblib==1.3.2
+      - kiwisolver==1.4.5
+      - lmdb==1.2.1
+      - lpips==0.1.4
+      - markdown==3.5.2
+      - markupsafe==2.1.3
+      - matplotlib==3.5.3
+      - multidict==6.0.4
+      - numpy==1.24.4
+      - oauthlib==3.2.2
+      - packaging==23.2
+      - pandas==1.5.3
+      - pillow==10.2.0
+      - protobuf==4.25.2
+      - psutil==5.9.8
+      - pyasn1==0.5.1
+      - pyasn1-modules==0.3.0
+      - pydeprecate==0.3.1
+      - pyparsing==3.1.1
+      - python-dateutil==2.8.2
+      - pytorch-fid==0.2.0
+      - pytorch-lightning==1.4.5
+      - pytz==2023.3.post1
+      - pyyaml==6.0.1
+      - regex==2023.12.25
+      - requests==2.31.0
+      - requests-oauthlib==1.3.1
+      - rsa==4.9
+      - scikit-learn==1.3.2
+      - scipy==1.5.4
+      - sentry-sdk==1.39.2
+      - setproctitle==1.3.3
+      - six==1.16.0
+      - smmap==5.0.1
+      - tensorboard==2.14.0
+      - tensorboard-data-server==0.7.2
+      - threadpoolctl==3.2.0
+      - torch==1.8.1
+      - torchmetrics==0.5.0
+      - torchvision==0.9.1
+      - tqdm==4.66.1
+      - typing-extensions==4.9.0
+      - tzdata==2023.4
+      - urllib3==2.1.0
+      - wandb==0.16.2
+      - wcwidth==0.2.13
+      - werkzeug==3.0.1
+      - yarl==1.9.4
+      - zipp==3.17.0

main.py

@@ -0,0 +1,565 @@
+import warnings
+warnings.filterwarnings("ignore")
+import os
+os.nice(5)
+import copy, wandb
+from tqdm import tqdm, trange
+import argparse
+import json
+import re
+import random 
+import math
+import numpy as np
+import pandas as pd
+import pytorch_lightning as pl
+import torch
+from pytorch_lightning import loggers as pl_loggers
+from pytorch_lightning.callbacks import *
+from torch import nn
+import torch.nn.functional as F
+from torch.cuda import amp
+from torch.optim.optimizer import Optimizer
+from config import *
+import time
+import datetime
+
+class LitModel(pl.LightningModule):
+    def __init__(self, conf: TrainConfig):
+        super().__init__()
+
+        ## wandb
+        self.save_hyperparameters({k:v for (k,v) in vars(conf).items() if not callable(v)})
+        
+        if conf.seed is not None:
+            pl.seed_everything(conf.seed)
+            
+        self.save_hyperparameters(conf.as_dict_jsonable())
+        self.conf = conf
+        self.model = conf.make_model_conf().make_model()
+
+        self.ema_model = copy.deepcopy(self.model)
+        self.ema_model.requires_grad_(False)
+        self.ema_model.eval()
+
+        model_size = 0
+        for param in self.model.parameters():
+            model_size += param.data.nelement()
+        print('Model params: %.3f M' % (model_size / 1024 / 1024))
+
+        self.sampler = conf.make_diffusion_conf().make_sampler()
+        self.eval_sampler = conf.make_eval_diffusion_conf().make_sampler()
+        self.T_sampler = conf.make_T_sampler()
+        self.save_every_epochs = conf.save_every_epochs
+        self.eval_every_epochs = conf.eval_every_epochs
+    
+    def setup(self, stage=None) -> None:
+        """
+        make datasets & seeding each worker separately
+        """
+        ##############################################
+        # NEED TO SET THE SEED SEPARATELY HERE
+        if self.conf.seed is not None:
+            seed = self.conf.seed
+            np.random.seed(seed)
+            random.seed(seed)  # Python random module.
+            torch.manual_seed(seed)
+            torch.cuda.manual_seed(seed)
+            torch.cuda.manual_seed_all(seed)
+            torch.backends.cudnn.benchmark = False
+            torch.backends.cudnn.deterministic = True
+            print('seed:', seed)
+        ##############################################
+        
+        ## Load dataset
+        if self.conf.mode == 'train':
+            self.train_data = load_egobody(self.conf.data_dir, self.conf.seq_len+self.conf.seq_len_future, self.conf.data_sample_rate, train=1)
+            self.val_data = load_egobody(self.conf.data_dir, self.conf.seq_len+self.conf.seq_len_future, self.conf.data_sample_rate, train=0)
+            if self.conf.in_channels == 6: # hand only
+                self.train_data = self.train_data[:, :6, :]
+                self.val_data = self.val_data[:, :6, :]
+            if self.conf.in_channels == 3: # head only
+                self.train_data = self.train_data[:, 6:, :]
+                self.val_data = self.val_data[:, 6:, :]
+                        
+    def encode(self, x):
+        assert self.conf.model_type.has_autoenc()
+        cond, pred_hand, pred_head = self.ema_model.encoder.forward(x)
+        return cond, pred_hand, pred_head
+
+    def encode_stochastic(self, x, cond, T=None):
+        if T is None:
+            sampler = self.eval_sampler
+        else: 
+            sampler = self.conf._make_diffusion_conf(T).make_sampler() # get noise at step T
+
+        ## x_0 -> x-T using reverse of inference
+        out = sampler.ddim_reverse_sample_loop(self.ema_model, x, model_kwargs={'cond': cond})
+        ''' 'sample': x_T
+            'sample_t': x_t, t in (1, ..., T)
+            'xstart_t': predicted x_0 at each timestep. "xstart here is a bit different from sampling from T = T-1 to T = 0"
+            'T': (1, ..., T)
+        '''
+        return out['sample']
+        
+    def train_dataloader(self):
+        return torch.utils.data.DataLoader(self.train_data, batch_size=self.conf.batch_size, shuffle=True)
+
+    def val_dataloader(self):
+        return torch.utils.data.DataLoader(self.val_data, batch_size=self.conf.batch_size_eval, shuffle=False)
+        
+    def is_last_accum(self, batch_idx):
+        """
+        is it the last gradient accumulation loop? 
+        used with gradient_accum > 1 and to see if the optimizer will perform "step" in this iteration or not
+        """
+        return (batch_idx + 1) % self.conf.accum_batches == 0
+
+    def training_step(self, batch, batch_idx):
+        """
+        given an input, calculate the loss function
+        no optimization at this stage.
+        """
+        with amp.autocast(False):
+            x_start = batch[:, :, :self.conf.seq_len]            
+            x_future = batch[:, :, self.conf.seq_len:]            
+            
+            if self.conf.train_mode == TrainMode.diffusion:
+                """
+                main training mode!!!
+                """
+                t, weight = self.T_sampler.sample(len(x_start), x_start.device)
+                ''' self.T_sampler: diffusion.resample.UniformSampler (weights for all timesteps are 1)
+                 - t: a tensor of timestep indices.
+                 - weight: a tensor of weights to scale the resulting losses.
+                 ## num_timesteps is self.conf.T == 1000
+                '''
+                losses = self.sampler.training_losses(model=self.model,
+                                                      x_start=x_start,
+                                                      t=t,
+                                                      x_future=x_future,
+                                                      hand_mse_factor = self.conf.hand_mse_factor,
+                                                      head_mse_factor = self.conf.head_mse_factor,                                                      
+                                                      )
+            else:
+                raise NotImplementedError()
+
+            loss = losses['loss'].mean() ## average loss across mini-batches
+            #noise_mse = losses['mse'].mean()
+            #hand_mse = losses['hand_mse'].mean()
+            #head_mse = losses['head_mse'].mean()
+            
+        ## Log loss and metric (wandb)
+        self.log("train_loss", loss, on_epoch=True, prog_bar=True)
+        #self.log("train_noise_mse", noise_mse, on_epoch=True, prog_bar=True)
+        #self.log("train_hand_mse", hand_mse, on_epoch=True, prog_bar=True)
+        #self.log("train_head_mse", head_mse, on_epoch=True, prog_bar=True)
+        return {'loss': loss} 
+        
+    def validation_step(self, batch, batch_idx):
+        if self.conf.in_channels == 9: # both hand and head
+            if((self.current_epoch+1)% self.eval_every_epochs == 0):
+                batch_future = batch[:, :, self.conf.seq_len:]
+                gt_hand_future = batch_future[:, :6, :]
+                gt_head_future = batch_future[:, 6:, :]
+                batch = batch[:, :, :self.conf.seq_len]
+                cond, pred_hand_future, pred_head_future = self.encode(batch)
+                xT = self.encode_stochastic(batch, cond)
+                pred_xstart = self.generate(xT, cond)
+                
+                # hand reconstruction error
+                gt_hand = batch[:, :6, :]
+                pred_hand = pred_xstart[:, :6, :]                        
+                bs, channels, seq_len = gt_hand.shape
+                gt_hand = gt_hand.reshape(bs, 2, 3, seq_len)
+                pred_hand = pred_hand.reshape(bs, 2, 3, seq_len)            
+                hand_traj = torch.mean(torch.norm(gt_hand - pred_hand, dim=2))
+                
+                # hand prediction error
+                bs, channels, seq_len = gt_hand_future.shape
+                gt_hand_future = gt_hand_future.reshape(bs, 2, 3, seq_len)            
+                pred_hand_future = pred_hand_future.reshape(bs, 2, 3, seq_len)
+                baseline_hand_future = gt_hand[:, :, :, -1:].expand(-1, -1, -1, self.conf.seq_len_future).clone()
+                hand_traj_future = torch.mean(torch.norm(gt_hand_future - pred_hand_future, dim=2))
+                hand_traj_future_baseline = torch.mean(torch.norm(gt_hand_future - baseline_hand_future, dim=2))
+                
+                # head reconstruction error
+                gt_head = batch[:, 6:, :]
+                gt_head = F.normalize(gt_head, dim=1) # normalize head orientation to unit vectors
+                pred_head = pred_xstart[:, 6:, :]
+                pred_head = F.normalize(pred_head, dim=1)            
+                head_ang = torch.mean(acos_safe(torch.sum(gt_head*pred_head, 1)))/torch.tensor(math.pi) * 180.0
+                
+                # head prediction error           
+                gt_head_future = F.normalize(gt_head_future, dim=1)
+                pred_head_future = F.normalize(pred_head_future, dim=1)
+                baseline_head_future = gt_head[:, :, -1:].expand(-1, -1, self.conf.seq_len_future).clone()
+                head_ang_future = torch.mean(acos_safe(torch.sum(gt_head_future*pred_head_future, 1)))/torch.tensor(math.pi) * 180.0
+                head_ang_future_baseline = torch.mean(acos_safe(torch.sum(gt_head_future*baseline_head_future, 1)))/torch.tensor(math.pi) * 180.0
+
+                self.log("val_hand_traj", hand_traj, on_epoch=True, prog_bar=True)
+                self.log("val_head_ang", head_ang, on_epoch=True, prog_bar=True)
+                self.log("val_hand_traj_future", hand_traj_future, on_epoch=True, prog_bar=True)
+                self.log("val_head_ang_future", head_ang_future, on_epoch=True, prog_bar=True)
+                self.log("val_hand_traj_future_baseline", hand_traj_future_baseline, on_epoch=True, prog_bar=True)
+                self.log("val_head_ang_future_baseline", head_ang_future_baseline, on_epoch=True, prog_bar=True)
+
+        if self.conf.in_channels == 6: # hand only
+            if((self.current_epoch+1)% self.eval_every_epochs == 0):
+                batch_future = batch[:, :, self.conf.seq_len:]
+                gt_hand_future = batch_future[:, :, :]                
+                batch = batch[:, :, :self.conf.seq_len]
+                cond, pred_hand_future, pred_head_future = self.encode(batch)
+                xT = self.encode_stochastic(batch, cond)
+                pred_xstart = self.generate(xT, cond)
+                
+                # hand reconstruction error
+                gt_hand = batch[:, :, :]
+                pred_hand = pred_xstart[:, :, :]                        
+                bs, channels, seq_len = gt_hand.shape
+                gt_hand = gt_hand.reshape(bs, 2, 3, seq_len)
+                pred_hand = pred_hand.reshape(bs, 2, 3, seq_len)            
+                hand_traj = torch.mean(torch.norm(gt_hand - pred_hand, dim=2))
+                
+                # hand prediction error
+                bs, channels, seq_len = gt_hand_future.shape
+                gt_hand_future = gt_hand_future.reshape(bs, 2, 3, seq_len)            
+                pred_hand_future = pred_hand_future.reshape(bs, 2, 3, seq_len)
+                baseline_hand_future = gt_hand[:, :, :, -1:].expand(-1, -1, -1, self.conf.seq_len_future).clone()
+                hand_traj_future = torch.mean(torch.norm(gt_hand_future - pred_hand_future, dim=2))
+                hand_traj_future_baseline = torch.mean(torch.norm(gt_hand_future - baseline_hand_future, dim=2))
+                
+                self.log("val_hand_traj", hand_traj, on_epoch=True, prog_bar=True)
+                self.log("val_hand_traj_future", hand_traj_future, on_epoch=True, prog_bar=True)
+                self.log("val_hand_traj_future_baseline", hand_traj_future_baseline, on_epoch=True, prog_bar=True)
+
+        if self.conf.in_channels == 3: # head only
+            if((self.current_epoch+1)% self.eval_every_epochs == 0):
+                batch_future = batch[:, :, self.conf.seq_len:]
+                gt_head_future = batch_future[:, :, :]
+                batch = batch[:, :, :self.conf.seq_len]
+                cond, pred_hand_future, pred_head_future = self.encode(batch)
+                xT = self.encode_stochastic(batch, cond)
+                pred_xstart = self.generate(xT, cond)
+                                               
+                # head reconstruction error
+                gt_head = batch[:, :, :]
+                gt_head = F.normalize(gt_head, dim=1) # normalize head orientation to unit vectors
+                pred_head = pred_xstart[:, :, :]
+                pred_head = F.normalize(pred_head, dim=1)
+                head_ang = torch.mean(acos_safe(torch.sum(gt_head*pred_head, 1)))/torch.tensor(math.pi) * 180.0
+                
+                # head prediction error           
+                gt_head_future = F.normalize(gt_head_future, dim=1)
+                pred_head_future = F.normalize(pred_head_future, dim=1)
+                baseline_head_future = gt_head[:, :, -1:].expand(-1, -1, self.conf.seq_len_future).clone()
+                head_ang_future = torch.mean(acos_safe(torch.sum(gt_head_future*pred_head_future, 1)))/torch.tensor(math.pi) * 180.0
+                head_ang_future_baseline = torch.mean(acos_safe(torch.sum(gt_head_future*baseline_head_future, 1)))/torch.tensor(math.pi) * 180.0
+
+                self.log("val_head_ang", head_ang, on_epoch=True, prog_bar=True)
+                self.log("val_head_ang_future", head_ang_future, on_epoch=True, prog_bar=True)
+                self.log("val_head_ang_future_baseline", head_ang_future_baseline, on_epoch=True, prog_bar=True)
+
+            
+    def test_step(self, batch, batch_idx):
+        batch_future = batch[:, :, self.conf.seq_len:]
+        gt_hand_future = batch_future[:, :6, :]
+        gt_head_future = batch_future[:, 6:, :]
+        batch = batch[:, :, :self.conf.seq_len]
+        cond, pred_hand_future, pred_head_future = self.encode(batch)
+        xT = self.encode_stochastic(batch, cond)
+        pred_xstart = self.generate(xT, cond)
+        
+        # hand reconstruction error
+        gt_hand = batch[:, :6, :]
+        pred_hand = pred_xstart[:, :6, :]                        
+        bs, channels, seq_len = gt_hand.shape
+        gt_hand = gt_hand.reshape(bs, 2, 3, seq_len)
+        pred_hand = pred_hand.reshape(bs, 2, 3, seq_len)            
+        hand_traj = torch.mean(torch.norm(gt_hand - pred_hand, dim=2))
+        
+        # hand prediction error
+        bs, channels, seq_len = gt_hand_future.shape
+        gt_hand_future = gt_hand_future.reshape(bs, 2, 3, seq_len)            
+        pred_hand_future = pred_hand_future.reshape(bs, 2, 3, seq_len)
+        baseline_hand_future = gt_hand[:, :, :, -1:].expand(-1, -1, -1, self.conf.seq_len_future).clone()
+        hand_traj_future = torch.mean(torch.norm(gt_hand_future - pred_hand_future, dim=2))
+        hand_traj_future_baseline = torch.mean(torch.norm(gt_hand_future - baseline_hand_future, dim=2))
+        
+        # head reconstruction error
+        gt_head = batch[:, 6:, :]
+        gt_head = F.normalize(gt_head, dim=1) # normalize head orientation to unit vectors
+        pred_head = pred_xstart[:, 6:, :]
+        pred_head = F.normalize(pred_head, dim=1)            
+        head_ang = torch.mean(acos_safe(torch.sum(gt_head*pred_head, 1)))/torch.tensor(math.pi) * 180.0
+        
+        # head prediction error           
+        gt_head_future = F.normalize(gt_head_future, dim=1)
+        pred_head_future = F.normalize(pred_head_future, dim=1)
+        baseline_head_future = gt_head[:, :, -1:].expand(-1, -1, self.conf.seq_len_future).clone()
+        head_ang_future = torch.mean(acos_safe(torch.sum(gt_head_future*pred_head_future, 1)))/torch.tensor(math.pi) * 180.0
+        head_ang_future_baseline = torch.mean(acos_safe(torch.sum(gt_head_future*baseline_head_future, 1)))/torch.tensor(math.pi) * 180.0
+
+        self.log("test_hand_traj", hand_traj, on_epoch=True, prog_bar=True)
+        self.log("test_head_ang", head_ang, on_epoch=True, prog_bar=True)                    
+        self.log("test_hand_traj_future", hand_traj_future, on_epoch=True, prog_bar=True)
+        self.log("test_head_ang_future", head_ang_future, on_epoch=True, prog_bar=True)        
+        self.log("test_hand_traj_future_baseline", hand_traj_future_baseline, on_epoch=True, prog_bar=True)            
+        self.log("test_head_ang_future_baseline", head_ang_future_baseline, on_epoch=True, prog_bar=True)
+       
+        
+    def generate(self, noise, cond=None, ema=True, T=None):
+        if T is None:
+            sampler = self.eval_sampler
+        else:
+            sampler = self.conf._make_diffusion_conf(T).make_sampler()
+            
+        if ema:
+            model = self.ema_model
+        else:
+            model = self.model
+                                  
+        gen = sampler.sample(model=model, noise=noise, model_kwargs={'cond': cond})
+        return gen
+        
+    def on_train_batch_end(self, outputs, batch, batch_idx: int,
+                           dataloader_idx: int) -> None:
+        """
+        after each training step ...
+        """
+        if self.is_last_accum(batch_idx):
+            # only apply ema on the last gradient accumulation step,
+            # if it is the iteration that has optimizer.step()
+            
+            ema(self.model, self.ema_model, self.conf.ema_decay)
+            
+            if (batch_idx==len(self.train_dataloader())-1) and ((self.current_epoch+1) % self.save_every_epochs == 0):
+                save_path = os.path.join(self.conf.logdir, 'epoch_%d.ckpt' % (self.current_epoch+1))
+                torch.save({
+                    'state_dict': self.state_dict(),
+                    'global_step': self.global_step,
+                    'loss': outputs['loss'],
+                }, save_path)
+                
+    def on_before_optimizer_step(self, optimizer: Optimizer,
+                                 optimizer_idx: int) -> None:
+        # fix the fp16 + clip grad norm problem with pytorch lightinng
+        # this is the currently correct way to do it
+        if self.conf.grad_clip > 0: 
+            # from trainer.params_grads import grads_norm, iter_opt_params
+            params = [
+                p for group in optimizer.param_groups for p in group['params']
+            ]
+            # print('before:', grads_norm(iter_opt_params(optimizer)))
+            torch.nn.utils.clip_grad_norm_(params, max_norm=self.conf.grad_clip)
+            # print('after:', grads_norm(iter_opt_params(optimizer)))
+
+    def configure_optimizers(self):
+        out = {}
+        if self.conf.optimizer == OptimizerType.adam:
+            optim = torch.optim.Adam(self.model.parameters(),
+                                     lr=self.conf.lr,
+                                     weight_decay=self.conf.weight_decay)
+        elif self.conf.optimizer == OptimizerType.adamw:
+            optim = torch.optim.AdamW(self.model.parameters(),
+                                      lr=self.conf.lr,
+                                      weight_decay=self.conf.weight_decay)
+        else:
+            raise NotImplementedError()
+        out['optimizer'] = optim
+        if self.conf.warmup > 0:
+            sched = torch.optim.lr_scheduler.LambdaLR(optim,
+                                                      lr_lambda=WarmupLR(
+                                                          self.conf.warmup))
+            out['lr_scheduler'] = {
+                'scheduler': sched,
+                'interval': 'step',
+            }
+        return out
+        
+        
+def ema(source, target, decay):
+    source_dict = source.state_dict()
+    target_dict = target.state_dict()
+    for key in source_dict.keys():
+        target_dict[key].data.copy_(target_dict[key].data * decay +
+                                    source_dict[key].data * (1 - decay))
+
+
+class WarmupLR:
+    def __init__(self, warmup) -> None:
+        self.warmup = warmup
+
+    def __call__(self, step):
+        return min(step, self.warmup) / self.warmup
+
+        
+def train(conf: TrainConfig, model: LitModel, gpus):
+    checkpoint = ModelCheckpoint(dirpath=conf.logdir,
+                                 filename='last',
+                                 save_last=True,
+                                 save_top_k=1,
+                                 every_n_epochs=conf.save_every_epochs,
+                                 )
+    checkpoint_path = f'{conf.logdir}last.ckpt'
+    if os.path.exists(checkpoint_path):
+        resume = checkpoint_path
+        if conf.mode == 'train':
+            print('ckpt path:', checkpoint_path)
+    else:
+        print('checkpoint not found!')
+        resume = None
+            
+    wandb_logger = pl_loggers.WandbLogger(project='haheae', 
+                        name='%s_%s'%(model.conf.data_name, conf.logdir.split('/')[-2]),
+                        log_model=True,
+                        save_dir=conf.logdir, 
+                        dir = conf.logdir,
+                        config=vars(model.conf), 
+                        save_code=True,
+                        settings=wandb.Settings(code_dir="."))
+                        
+    trainer = pl.Trainer(
+        max_epochs=conf.total_epochs,
+        resume_from_checkpoint=resume,
+        gpus=gpus,
+        precision=16 if conf.fp16 else 32,
+        callbacks=[
+            checkpoint,
+            LearningRateMonitor(),
+        ],
+        logger= wandb_logger,
+        accumulate_grad_batches=conf.accum_batches,
+        progress_bar_refresh_rate=4,
+    )
+    
+    if conf.mode == 'train':
+        trainer.fit(model)
+    elif conf.mode == 'eval':
+        checkpoint_path = f'{conf.logdir}last.ckpt'
+        # load the latest checkpoint
+        print('loading from:', checkpoint_path)
+        state = torch.load(checkpoint_path)
+        model.load_state_dict(state['state_dict'])
+        
+        test_datasets = ['egobody', 'adt', 'gimo']
+        for dataset_name in test_datasets:
+            if dataset_name == 'egobody':
+                data_dir = "/scratch/hu/pose_forecast/egobody_pose2gaze/"
+                test_data = load_egobody(data_dir, conf.seq_len+conf.seq_len_future, 1, train=0) # use the test set
+            elif dataset_name == 'adt':
+                data_dir = "/scratch/hu/pose_forecast/adt_pose2gaze/"
+                test_data = load_adt(data_dir, conf.seq_len+conf.seq_len_future, 1, train=2) # use the train+test set
+            elif dataset_name == 'gimo':
+                data_dir = "/scratch/hu/pose_forecast/gimo_pose2gaze/"
+                test_data = load_gimo(data_dir, conf.seq_len+conf.seq_len_future, 1, train=2) # use the train+test set
+                
+            test_dataloader = torch.utils.data.DataLoader(test_data, batch_size=conf.batch_size_eval, shuffle=False)        
+            results = trainer.test(model, dataloaders=test_dataloader, verbose=False)
+            print("\n\nTest on {}, dataset size: {}".format(dataset_name, test_data.shape))            
+            print("test_hand_traj: {:.3f} cm".format(results[0]['test_hand_traj']*100))
+            print("test_head_ang: {:.3f} deg".format(results[0]['test_head_ang']))
+            print("test_hand_traj_future: {:.3f} cm".format(results[0]['test_hand_traj_future']*100))
+            print("test_head_ang_future: {:.3f} deg".format(results[0]['test_head_ang_future']))
+            print("test_hand_traj_future_baseline: {:.3f} cm".format(results[0]['test_hand_traj_future_baseline']*100))
+            print("test_head_ang_future_baseline: {:.3f} deg\n\n".format(results[0]['test_head_ang_future_baseline']))
+            
+    wandb.finish()
+
+    
+def acos_safe(x, eps=1e-6):
+    slope = np.arccos(1-eps) / eps
+    buf = torch.empty_like(x)
+    good = abs(x) <= 1-eps
+    bad = ~good
+    sign = torch.sign(x[bad])
+    buf[good] = torch.acos(x[good])
+    buf[bad] = torch.acos(sign * (1 - eps)) - slope*sign*(abs(x[bad]) - 1 + eps)
+    return buf
+
+    
+def get_representation(model, dataset, conf, device='cuda'):
+    model = model.to(device)
+    model.eval()
+    dataloader = torch.utils.data.DataLoader(dataset, batch_size=conf.batch_size_eval, shuffle=False)
+    with torch.no_grad():
+        conds = [] # semantic representation
+        xTs = [] # stochastic representation
+        for batch in tqdm(dataloader, total=len(dataloader), desc='infer'):
+            batch = batch.to(device)
+            cond, _, _ = model.encode(batch)
+            xT = model.encode_stochastic(batch, cond)            
+            cond_cpu = cond.cpu().data.numpy()
+            xT_cpu = xT.cpu().data.numpy()
+            if len(conds) == 0:
+                conds = cond_cpu
+                xTs = xT_cpu
+            else:
+                conds = np.concatenate((conds, cond_cpu), axis=0)
+                xTs = np.concatenate((xTs, xT_cpu), axis=0)                
+    return conds, xTs
+
+    
+def generate_from_representation(model, conds, xTs, device='cuda'):
+    model = model.to(device)
+    model.eval()
+    conds = torch.from_numpy(conds).to(device)
+    xTs = torch.from_numpy(xTs).to(device)
+    rec = model.generate(xTs, conds)
+    rec = rec.cpu().data.numpy()    
+    return rec
+
+
+def evaluate_reconstruction(gt, rec):
+    # hand reconstruction error (cm)
+    gt_hand = gt[:, :6, :]
+    rec_hand = rec[:, :6, :]                        
+    bs, channels, seq_len = gt_hand.shape
+    gt_hand = gt_hand.reshape(bs, 2, 3, seq_len)
+    rec_hand = rec_hand.reshape(bs, 2, 3, seq_len)            
+    hand_traj_errors = np.mean(np.mean(np.linalg.norm(gt_hand - rec_hand, axis=2), axis=1), axis=1)*100
+    
+    # head reconstruction error (deg)
+    gt_head = gt[:, 6:, :]
+    gt_head_norm = np.linalg.norm(gt_head, axis=1, keepdims=True)
+    gt_head = gt_head/gt_head_norm
+    rec_head = rec[:, 6:, :]
+    rec_head_norm = np.linalg.norm(rec_head, axis=1, keepdims=True)
+    rec_head = rec_head/rec_head_norm
+    dot_sum = np.clip(np.sum(gt_head*rec_head, axis=1), -1, 1)
+    head_ang_errors = np.mean(np.arccos(dot_sum), axis=1)/np.pi * 180.0
+    return hand_traj_errors, head_ang_errors
+
+    
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--gpus', default=7, type=int)
+    parser.add_argument('--mode', default='eval', type=str)
+    parser.add_argument('--encoder_type', default='gcn', type=str)
+    parser.add_argument('--model_name', default='haheae', type=str)
+    parser.add_argument('--hand_mse_factor', default=1.0, type=float)
+    parser.add_argument('--head_mse_factor', default=1.0, type=float)
+    parser.add_argument('--data_sample_rate', default=1, type=int)
+    parser.add_argument('--epoch', default=130, type=int)
+    parser.add_argument('--in_channels', default=9, type=int)    
+    args = parser.parse_args()
+    
+    conf = egobody_autoenc(args.mode, args.encoder_type, args.hand_mse_factor, args.head_mse_factor, args.data_sample_rate, args.epoch, args.in_channels)
+    model = LitModel(conf)
+    conf.logdir = f'{conf.logdir}{args.model_name}/'
+    print('log dir: {}'.format(conf.logdir))
+    MakeDir(conf.logdir)
+    
+    if conf.mode == 'train' or conf.mode == 'eval': # train or evaluate the model
+        os.environ['WANDB_CACHE_DIR'] = conf.logdir
+        os.environ['WANDB_DATA_DIR'] = conf.logdir
+        # set wandb to not upload checkpoints, but all the others
+        os.environ['WANDB_IGNORE_GLOBS'] = '*.ckpt'
+        local_time = time.asctime(time.localtime(time.time()))
+        print('\n{} starts at {}'.format(conf.mode, local_time))
+        start_time = datetime.datetime.now()
+        train(conf, model, gpus=[args.gpus])
+        end_time = datetime.datetime.now()
+        total_time = (end_time - start_time).seconds/60
+        print('\nTotal time: {:.3f} min'.format(total_time))
+        local_time = time.asctime(time.localtime(time.time()))
+        print('\n{} ends at {}'.format(conf.mode, local_time))

model/__init__.py

@@ -0,0 +1,6 @@
+from typing import Union
+from .unet import BeatGANsUNetModel, BeatGANsUNetConfig, GCNUNetModel, GCNUNetConfig
+from .unet_autoenc import BeatGANsAutoencConfig, BeatGANsAutoencModel, GCNAutoencConfig, GCNAutoencModel
+
+Model = Union[BeatGANsUNetModel, BeatGANsAutoencModel, GCNUNetModel, GCNAutoencModel]
+ModelConfig = Union[BeatGANsUNetConfig, BeatGANsAutoencConfig, GCNUNetConfig,GCNAutoencConfig]

model/blocks.py

@@ -0,0 +1,579 @@
+import math, pdb
+from abc import abstractmethod
+from dataclasses import dataclass
+from numbers import Number
+
+import torch as th
+import torch.nn.functional as F
+from choices import *
+from config_base import BaseConfig
+from torch import nn
+import numpy as np
+from .nn import (avg_pool_nd, conv_nd, linear, normalization,
+                 timestep_embedding, zero_module)
+                                                                  
+                                                  
+class TimestepBlock(nn.Module):
+    """
+    Any module where forward() takes timestep embeddings as a second argument.
+    """
+    @abstractmethod
+    def forward(self, x, emb=None, cond=None, lateral=None):
+        """
+        Apply the module to `x` given `emb` timestep embeddings.
+        """
+
+        
+class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
+    """
+    A sequential module that passes timestep embeddings to the children that
+    support it as an extra input.
+    """
+    def forward(self, x, emb=None, cond=None, lateral=None):
+        for layer in self:
+            if isinstance(layer, TimestepBlock):
+                '''if layer(x, emb=emb, cond=cond, lateral=lateral).shape[-1]==10:
+                    pdb.set_trace()'''
+                x = layer(x, emb=emb, cond=cond, lateral=lateral)
+            else:
+                '''if layer(x).shape[-1]==10:
+                    pdb.set_trace()'''
+                x = layer(x)
+        return x
+
+
+@dataclass
+class ResBlockConfig(BaseConfig):
+    channels: int
+    emb_channels: int
+    dropout: float
+    out_channels: int = None
+    # condition the resblock with time (and encoder's output)
+    use_condition: bool = True
+    # whether to use 3x3 conv for skip path when the channels aren't matched
+    use_conv: bool = False
+    # dimension of conv (always 1 = 1d)
+    dims: int = 1
+    up: bool = False
+    down: bool = False
+    # whether to condition with both time & encoder's output
+    two_cond: bool = False
+    # number of encoders' output channels
+    cond_emb_channels: int = None
+    # suggest: False
+    has_lateral: bool = False
+    lateral_channels: int = None
+    # whether to init the convolution with zero weights
+    # this is default from BeatGANs and seems to help learning
+    use_zero_module: bool = True
+
+    def __post_init__(self):
+        self.out_channels = self.out_channels or self.channels
+        self.cond_emb_channels = self.cond_emb_channels or self.emb_channels
+
+    def make_model(self):
+        return ResBlock(self)
+      
+      
+class ResBlock(TimestepBlock):
+    """
+    A residual block that can optionally change the number of channels.
+
+    total layers:
+        in_layers
+        - norm
+        - act
+        - conv
+        out_layers
+        - norm
+        - (modulation)
+        - act
+        - conv
+    """
+    def __init__(self, conf: ResBlockConfig):
+        super().__init__()
+        self.conf = conf
+        
+        #############################
+        # IN LAYERS
+        #############################
+        assert conf.lateral_channels is None
+        layers = [
+            normalization(conf.channels),
+            nn.SiLU(),
+            conv_nd(conf.dims, conf.channels, conf.out_channels, 3, padding=1) ## 3 is kernel size
+        ]
+        self.in_layers = nn.Sequential(*layers)
+
+        self.updown = conf.up or conf.down
+
+        if conf.up:
+            self.h_upd = Upsample(conf.channels, False, conf.dims)
+            self.x_upd = Upsample(conf.channels, False, conf.dims)
+        elif conf.down:
+            self.h_upd = Downsample(conf.channels, False, conf.dims)
+            self.x_upd = Downsample(conf.channels, False, conf.dims)
+        else:
+            self.h_upd = self.x_upd = nn.Identity()
+
+        #############################
+        # OUT LAYERS CONDITIONS
+        #############################
+        if conf.use_condition:
+            # condition layers for the out_layers
+            self.emb_layers = nn.Sequential(
+                nn.SiLU(),
+                linear(conf.emb_channels, 2 * conf.out_channels),
+            )
+
+            if conf.two_cond:
+                self.cond_emb_layers = nn.Sequential(
+                    nn.SiLU(),
+                    linear(conf.cond_emb_channels, conf.out_channels),
+                )
+            #############################
+            # OUT LAYERS (ignored when there is no condition)
+            #############################
+            # original version
+            conv = conv_nd(conf.dims,
+                           conf.out_channels,
+                           conf.out_channels,
+                           3,
+                           padding=1)
+            if conf.use_zero_module:
+                # zere out the weights
+                # it seems to help training
+                conv = zero_module(conv)
+
+            # construct the layers
+            # - norm
+            # - (modulation)
+            # - act
+            # - dropout
+            # - conv
+            layers = []
+            layers += [
+                normalization(conf.out_channels),
+                nn.SiLU(),
+                nn.Dropout(p=conf.dropout),
+                conv,
+            ]
+            self.out_layers = nn.Sequential(*layers)
+
+        #############################
+        # SKIP LAYERS
+        #############################
+        if conf.out_channels == conf.channels:
+            # cannot be used with gatedconv, also gatedconv is alsways used as the first block
+            self.skip_connection = nn.Identity()
+        else:
+            if conf.use_conv:
+                kernel_size = 3
+                padding = 1
+            else:
+                kernel_size = 1
+                padding = 0
+
+            self.skip_connection = conv_nd(conf.dims,
+                                           conf.channels,
+                                           conf.out_channels,
+                                           kernel_size,
+                                           padding=padding)
+
+    def forward(self, x, emb=None, cond=None, lateral=None):
+        """
+        Apply the block to a Tensor, conditioned on a timestep embedding.
+
+        Args:
+            x: input
+            lateral: lateral connection from the encoder
+        """
+        return self._forward(x, emb, cond, lateral)
+
+    def _forward(
+        self,
+        x,
+        emb=None,
+        cond=None,
+        lateral=None,
+    ):
+        """
+        Args:
+            lateral: required if "has_lateral" and non-gated, with gated, it can be supplied optionally    
+        """
+        if self.conf.has_lateral:
+            # lateral may be supplied even if it doesn't require
+            # the model will take the lateral only if "has_lateral"
+            assert lateral is not None
+            # x = F.interpolate(x, size=(lateral.size(2)), mode='linear' )
+            x = th.cat([x, lateral], dim=1)
+
+        if self.updown:
+            in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
+            h = in_rest(x)
+            h = self.h_upd(h)
+            x = self.x_upd(x)
+            h = in_conv(h)
+        else:
+            h = self.in_layers(x)
+
+        if self.conf.use_condition:
+            # it's possible that the network may not receieve the time emb
+            # this happens with autoenc and setting the time_at
+            if emb is not None:
+                emb_out = self.emb_layers(emb).type(h.dtype)
+            else:
+                emb_out = None
+                
+            if self.conf.two_cond:
+                # it's possible that the network is two_cond
+                # but it doesn't get the second condition
+                # in which case, we ignore the second condition
+                # and treat as if the network has one condition
+                if cond is None:
+                    cond_out = None
+                else:
+                    if not isinstance(cond, th.Tensor):
+                        assert isinstance(cond, dict)
+                        cond = cond['cond']                    
+                    cond_out = self.cond_emb_layers(cond).type(h.dtype)
+                if cond_out is not None:
+                    while len(cond_out.shape) < len(h.shape):
+                        cond_out = cond_out[..., None]
+            else:
+                cond_out = None
+
+            # this is the new refactored code
+            h = apply_conditions(
+                h=h,
+                emb=emb_out,
+                cond=cond_out,
+                layers=self.out_layers,
+                scale_bias=1,
+                in_channels=self.conf.out_channels,
+                up_down_layer=None,
+            )
+
+        return self.skip_connection(x) + h
+
+
+def apply_conditions(
+    h,
+    emb=None,
+    cond=None,
+    layers: nn.Sequential = None,
+    scale_bias: float = 1,
+    in_channels: int = 512,
+    up_down_layer: nn.Module = None,
+):
+    """
+    apply conditions on the feature maps
+
+    Args:
+        emb: time conditional (ready to scale + shift)
+        cond: encoder's conditional (ready to scale + shift)
+    """
+    two_cond = emb is not None and cond is not None
+
+    if emb is not None:
+        # adjusting shapes
+        while len(emb.shape) < len(h.shape):            
+            emb = emb[..., None]            
+            
+    if two_cond:
+        # adjusting shapes        
+        while len(cond.shape) < len(h.shape):            
+            cond = cond[..., None]            
+        # time first
+        scale_shifts = [emb, cond]
+    else:
+        # "cond" is not used with single cond mode
+        scale_shifts = [emb]
+
+    # support scale, shift or shift only
+    for i, each in enumerate(scale_shifts):
+        if each is None:
+            # special case: the condition is not provided
+            a = None
+            b = None
+        else:
+            if each.shape[1] == in_channels * 2:
+                a, b = th.chunk(each, 2, dim=1)
+            else:
+                a = each
+                b = None
+        scale_shifts[i] = (a, b)
+
+    # condition scale bias could be a list
+    if isinstance(scale_bias, Number):
+        biases = [scale_bias] * len(scale_shifts)
+    else:
+        # a list
+        biases = scale_bias
+
+    # default, the scale & shift are applied after the group norm but BEFORE SiLU
+    pre_layers, post_layers = layers[0], layers[1:]
+
+    # spilt the post layer to be able to scale up or down before conv
+    # post layers will contain only the conv
+    mid_layers, post_layers = post_layers[:-2], post_layers[-2:]
+
+    h = pre_layers(h)
+    # scale and shift for each condition
+    for i, (scale, shift) in enumerate(scale_shifts):
+        # if scale is None, it indicates that the condition is not provided
+        if scale is not None:                        
+            h = h * (biases[i] + scale)
+            if shift is not None:
+                h = h + shift
+    h = mid_layers(h)
+
+    # upscale or downscale if any just before the last conv
+    if up_down_layer is not None:
+        h = up_down_layer(h)
+    h = post_layers(h)
+    return h
+
+
+class Upsample(nn.Module):
+    """
+    An upsampling layer with an optional convolution.
+
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 upsampling occurs in the inner-two dimensions.
+    """
+    def __init__(self, channels, use_conv, dims=2, out_channels=None):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        if use_conv:
+            self.conv = conv_nd(dims,
+                                self.channels,
+                                self.out_channels,
+                                3,
+                                padding=1)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        if self.dims == 3:
+            x = F.interpolate(x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2),
+                              mode="nearest")
+        else:
+            # if x.shape[2] == 4:
+            #     feature = 9
+            #     x = F.interpolate(x, size=(feature), mode="nearest")
+            # if x.shape[2] == 8:
+            #     feature = 9
+            #     x = F.interpolate(x, size=(feature), mode="nearest")
+            # else:
+            x = F.interpolate(x, scale_factor=2, mode="nearest")
+        if self.use_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample(nn.Module):
+    """
+    A downsampling layer with an optional convolution.
+
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 downsampling occurs in the inner-two dimensions.
+    """
+    def __init__(self, channels, use_conv, dims=2, out_channels=None):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        self.stride = 2 if dims != 3 else (1, 2, 2)
+        if use_conv:
+            self.op = conv_nd(dims,
+                              self.channels,
+                              self.out_channels,
+                              3,
+                              stride=self.stride,
+                              padding=1)
+        else:
+            assert self.channels == self.out_channels
+            self.op = avg_pool_nd(dims, kernel_size=self.stride, stride=self.stride)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        # if x.shape[2] % 2 != 0:
+        #     op = avg_pool_nd(self.dims, kernel_size=3, stride=2)
+        #     return op(x)
+        # if x.shape[2] % 2 != 0:
+        #     op = avg_pool_nd(self.dims, kernel_size=2, stride=1)
+        #     return op(x)
+        # else: 
+        return self.op(x)
+
+        
+class AttentionBlock(nn.Module):
+    """
+    An attention block that allows spatial positions to attend to each other.
+
+    Originally ported from here, but adapted to the N-d case.
+    https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66.
+    """
+    def __init__(
+        self,
+        channels,
+        num_heads=1,
+        num_head_channels=-1,
+        use_new_attention_order=False,
+    ):
+        super().__init__()
+        self.channels = channels
+        if num_head_channels == -1:
+            self.num_heads = num_heads
+        else:
+            assert (
+                channels % num_head_channels == 0
+            ), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}"
+            self.num_heads = channels // num_head_channels
+        self.norm = normalization(channels)
+        self.qkv = conv_nd(1, channels, channels * 3, 1)
+        if use_new_attention_order:
+            # split qkv before split heads
+            self.attention = QKVAttention(self.num_heads)
+        else:
+            # split heads before split qkv
+            self.attention = QKVAttentionLegacy(self.num_heads)
+
+        self.proj_out = zero_module(conv_nd(1, channels, channels, 1))
+
+    def forward(self, x):
+        return self._forward(x)
+
+    def _forward(self, x):
+        b, c, *spatial = x.shape
+        x = x.reshape(b, c, -1)
+        qkv = self.qkv(self.norm(x))
+        h = self.attention(qkv)
+        h = self.proj_out(h)
+        return (x + h).reshape(b, c, *spatial)
+
+
+def count_flops_attn(model, _x, y):
+    """
+    A counter for the `thop` package to count the operations in an
+    attention operation.
+    Meant to be used like:
+        macs, params = thop.profile(
+            model,
+            inputs=(inputs, timestamps),
+            custom_ops={QKVAttention: QKVAttention.count_flops},
+        )
+    """
+    b, c, *spatial = y[0].shape
+    num_spatial = int(np.prod(spatial))
+    # We perform two matmuls with the same number of ops.
+    # The first computes the weight matrix, the second computes
+    # the combination of the value vectors.
+    matmul_ops = 2 * b * (num_spatial**2) * c
+    model.total_ops += th.DoubleTensor([matmul_ops])
+
+
+class QKVAttentionLegacy(nn.Module):
+    """
+    A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping
+    """
+    def __init__(self, n_heads):
+        super().__init__()
+        self.n_heads = n_heads
+
+    def forward(self, qkv):
+        """
+        Apply QKV attention.
+
+        :param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs.
+        :return: an [N x (H * C) x T] tensor after attention.
+        """
+        bs, width, length = qkv.shape
+        assert width % (3 * self.n_heads) == 0
+        ch = width // (3 * self.n_heads)
+        q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1)
+        scale = 1 / math.sqrt(math.sqrt(ch))
+        weight = th.einsum(
+            "bct,bcs->bts", q * scale,
+            k * scale)  # More stable with f16 than dividing afterwards
+        weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
+        a = th.einsum("bts,bcs->bct", weight, v)
+        return a.reshape(bs, -1, length)
+
+    @staticmethod
+    def count_flops(model, _x, y):
+        return count_flops_attn(model, _x, y)
+
+
+class QKVAttention(nn.Module):
+    """
+    A module which performs QKV attention and splits in a different order.
+    """
+    def __init__(self, n_heads):
+        super().__init__()
+        self.n_heads = n_heads
+
+    def forward(self, qkv):
+        """
+        Apply QKV attention.
+
+        :param qkv: an [N x (3 * H * C) x T] tensor of Qs, Ks, and Vs.
+        :return: an [N x (H * C) x T] tensor after attention.
+        """
+        bs, width, length = qkv.shape
+        pdb.set_trace()
+        assert width % (3 * self.n_heads) == 0
+        ch = width // (3 * self.n_heads)
+        q, k, v = qkv.chunk(3, dim=1)
+        scale = 1 / math.sqrt(math.sqrt(ch))
+        weight = th.einsum(
+            "bct,bcs->bts",
+            (q * scale).view(bs * self.n_heads, ch, length),
+            (k * scale).view(bs * self.n_heads, ch, length),
+        )  # More stable with f16 than dividing afterwards
+        weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
+        a = th.einsum("bts,bcs->bct", weight,
+                      v.reshape(bs * self.n_heads, ch, length))
+        return a.reshape(bs, -1, length)
+
+    @staticmethod
+    def count_flops(model, _x, y):
+        return count_flops_attn(model, _x, y)
+        
+        
+class AttentionPool2d(nn.Module):
+    """
+    Adapted from CLIP: https://github.com/openai/CLIP/blob/main/clip/model.py
+    """
+    def __init__(
+        self,
+        spacial_dim: int,
+        embed_dim: int,
+        num_heads_channels: int,
+        output_dim: int = None,
+    ):
+        super().__init__()
+        self.positional_embedding = nn.Parameter(
+            th.randn(embed_dim, spacial_dim**2 + 1) / embed_dim**0.5)
+        self.qkv_proj = conv_nd(1, embed_dim, 3 * embed_dim, 1)
+        self.c_proj = conv_nd(1, embed_dim, output_dim or embed_dim, 1)
+        self.num_heads = embed_dim // num_heads_channels
+        self.attention = QKVAttention(self.num_heads)
+
+    def forward(self, x):
+        b, c, *_spatial = x.shape
+        x = x.reshape(b, c, -1)  # NC(HW)
+        x = th.cat([x.mean(dim=-1, keepdim=True), x], dim=-1)  # NC(HW+1)
+        x = x + self.positional_embedding[None, :, :].to(x.dtype)  # NC(HW+1)
+        x = self.qkv_proj(x)
+        x = self.attention(x)
+        x = self.c_proj(x)
+        return x[:, :, 0]

model/graph_convolution_network.py

@@ -0,0 +1,173 @@
+import torch.nn as nn
+import torch
+from dataclasses import dataclass
+from torch.nn.parameter import Parameter
+from numbers import Number
+import torch.nn.functional as F
+from .blocks import *
+import math
+
+
+class graph_convolution(nn.Module):
+    def __init__(self, in_features, out_features, node_n = 3, seq_len = 80, bias=True):
+        super(graph_convolution, self).__init__()
+        
+        self.temporal_graph_weights = Parameter(torch.FloatTensor(seq_len, seq_len))
+        self.feature_weights = Parameter(torch.FloatTensor(in_features, out_features))
+        self.spatial_graph_weights = Parameter(torch.FloatTensor(node_n, node_n))
+        
+        if bias:
+            self.bias = Parameter(torch.FloatTensor(seq_len))
+            
+        self.reset_parameters()
+        
+    def reset_parameters(self):
+        stdv = 1. / math.sqrt(self.spatial_graph_weights.size(1))
+        self.feature_weights.data.uniform_(-stdv, stdv)
+        self.temporal_graph_weights.data.uniform_(-stdv, stdv)
+        self.spatial_graph_weights.data.uniform_(-stdv, stdv)
+        if self.bias is not None:
+            self.bias.data.uniform_(-stdv, stdv)
+            
+    def forward(self, x):
+        y = torch.matmul(x, self.temporal_graph_weights)
+        y = torch.matmul(y.permute(0, 3, 2, 1), self.feature_weights)    
+        y = torch.matmul(self.spatial_graph_weights, y).permute(0, 3, 2, 1).contiguous()
+        
+        if self.bias is not None:
+            return (y + self.bias)
+        else:
+            return y
+
+            
+@dataclass
+class residual_graph_convolution_config():
+    in_features: int
+    seq_len: int
+    emb_channels: int
+    dropout: float
+    out_features: int = None
+    node_n: int = 3
+    # condition the block with time (and encoder's output)
+    use_condition: bool = True    
+    # whether to condition with both time & encoder's output
+    two_cond: bool = False
+    # number of encoders' output channels
+    cond_emb_channels: int = None    
+    has_lateral: bool = False
+    graph_convolution_bias: bool = True
+    scale_bias: float = 1
+    
+    def __post_init__(self):
+        self.out_features = self.out_features or self.in_features
+        self.cond_emb_channels = self.cond_emb_channels or self.emb_channels
+
+    def make_model(self):
+        return residual_graph_convolution(self)
+        
+        
+class residual_graph_convolution(TimestepBlock):
+    def __init__(self, conf: residual_graph_convolution_config):
+        super(residual_graph_convolution, self).__init__()
+        self.conf = conf        
+        
+        self.gcn = graph_convolution(conf.in_features, conf.out_features, node_n=conf.node_n, seq_len=conf.seq_len, bias=conf.graph_convolution_bias)        
+        self.ln = nn.LayerNorm([conf.out_features, conf.node_n, conf.seq_len])
+        self.act_f = nn.Tanh()
+        self.dropout = nn.Dropout(conf.dropout)
+        
+        if conf.use_condition:
+            # condition layers for the out_layers
+            self.emb_layers = nn.Sequential(
+                nn.SiLU(),
+                nn.Linear(conf.emb_channels, conf.out_features),
+            )
+
+            if conf.two_cond:
+                self.cond_emb_layers = nn.Sequential(
+                    nn.SiLU(),
+                    nn.Linear(conf.cond_emb_channels, conf.out_features),
+                )
+        
+        if conf.in_features == conf.out_features:
+            self.skip_connection = nn.Identity()
+        else:
+            self.skip_connection = nn.Sequential(
+                    graph_convolution(conf.in_features, conf.out_features, node_n=conf.node_n, seq_len=conf.seq_len, bias=conf.graph_convolution_bias),
+                    nn.Tanh(),
+                )
+                                          
+    def forward(self, x, emb=None, cond=None, lateral=None):
+        if self.conf.has_lateral:
+            # lateral may be supplied even if it doesn't require
+            # the model will take the lateral only if "has_lateral"
+            assert lateral is not None
+            x = torch.cat((x, lateral), dim =1)
+
+        y = self.gcn(x)
+        y = self.ln(y)
+        
+        if self.conf.use_condition:
+            if emb is not None:
+                emb = self.emb_layers(emb).type(x.dtype)
+                # adjusting shapes
+                while len(emb.shape) < len(y.shape):            
+                    emb = emb[..., None]
+            
+            if self.conf.two_cond or True:
+                if cond is not None:
+                    if not isinstance(cond, torch.Tensor):
+                        assert isinstance(cond, dict)
+                        cond = cond['cond']
+                    cond = self.cond_emb_layers(cond).type(x.dtype)
+                    while len(cond.shape) < len(y.shape):
+                        cond = cond[..., None]
+                scales = [emb, cond]
+            else:
+                scales = [emb]
+                
+            # condition scale bias could be a list
+            if isinstance(self.conf.scale_bias, Number):
+                biases = [self.conf.scale_bias] * len(scales)
+            else:
+                # a list
+                biases = self.conf.scale_bias
+                
+            # scale for each condition
+            for i, scale in enumerate(scales):
+                # if scale is None, it indicates that the condition is not provided
+                if scale is not None:                        
+                    y = y*(biases[i] + scale)
+                                
+        y = self.act_f(y)
+        y = self.dropout(y)        
+        return self.skip_connection(x) + y
+
+        
+class graph_downsample(nn.Module):
+    """
+    A downsampling layer   
+    """
+    def __init__(self, kernel_size = 2):
+        super().__init__()
+        self.downsample = nn.AvgPool1d(kernel_size = kernel_size)
+        
+    def forward(self, x):               
+        bs, features, node_n, seq_len = x.shape
+        x = x.reshape(bs, features*node_n, seq_len)
+        x = self.downsample(x)
+        x = x.reshape(bs, features, node_n, -1)
+        return x
+
+        
+class graph_upsample(nn.Module):
+    """
+    An upsampling layer
+    """
+    def __init__(self, scale_factor=2):
+        super().__init__()
+        self.scale_factor = scale_factor
+        
+    def forward(self, x):           
+        x = F.interpolate(x, (x.shape[2], x.shape[3]*self.scale_factor), mode="nearest")
+        return x

model/nn.py

@@ -0,0 +1,141 @@
+"""
+Various utilities for neural networks.
+"""
+
+from enum import Enum
+import math, pdb
+from typing import Optional
+
+import torch as th
+import torch.nn as nn
+import torch.utils.checkpoint
+
+import torch.nn.functional as F
+
+
+# PyTorch 1.7 has SiLU, but we support PyTorch 1.5.
+class SiLU(nn.Module):
+    # @th.jit.script
+    def forward(self, x):
+        return x * th.sigmoid(x)
+
+
+class GroupNorm32(nn.GroupNorm):
+    def forward(self, x):
+        y = super().forward(x.float()).type(x.dtype)
+        return y
+
+
+def conv_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D convolution module.
+    """
+    assert dims==1
+    if dims == 1:
+        return nn.Conv1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.Conv2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.Conv3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+def linear(*args, **kwargs):
+    """
+    Create a linear module.
+    """
+    return nn.Linear(*args, **kwargs)
+
+
+def avg_pool_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D average pooling module.
+    """
+    assert dims==1
+    if dims == 1:
+        return nn.AvgPool1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.AvgPool2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.AvgPool3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+def update_ema(target_params, source_params, rate=0.99):
+    """
+    Update target parameters to be closer to those of source parameters using
+    an exponential moving average.
+
+    :param target_params: the target parameter sequence.
+    :param source_params: the source parameter sequence.
+    :param rate: the EMA rate (closer to 1 means slower).
+    """
+    for targ, src in zip(target_params, source_params):
+        targ.detach().mul_(rate).add_(src, alpha=1 - rate)
+
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+def scale_module(module, scale):
+    """
+    Scale the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().mul_(scale)
+    return module
+
+
+def mean_flat(tensor):
+    """
+    Take the mean over all non-batch dimensions.
+    """
+    return tensor.mean(dim=list(range(1, len(tensor.shape))))
+
+
+def normalization(channels):
+    """
+    Make a standard normalization layer.
+
+    :param channels: number of input channels.
+    :return: an nn.Module for normalization.
+    """
+    # return GroupNorm32(channels, channels)
+    return GroupNorm32(min(4, channels), channels)
+
+
+def timestep_embedding(timesteps, dim, max_period=10000):
+    """
+    Create sinusoidal timestep embeddings.
+
+    :param timesteps: a 1-D Tensor of N indices, one per batch element.
+                      These may be fractional.
+    :param dim: the dimension of the output.
+    :param max_period: controls the minimum frequency of the embeddings.
+    :return: an [N x dim] Tensor of positional embeddings.
+    """
+    half = dim // 2
+    freqs = th.exp(-math.log(max_period) *
+                   th.arange(start=0, end=half, dtype=th.float32) /
+                   half).to(device=timesteps.device)
+    args = timesteps[:, None].float() * freqs[None]
+    embedding = th.cat([th.cos(args), th.sin(args)], dim=-1)
+    if dim % 2:
+        embedding = th.cat(
+            [embedding, th.zeros_like(embedding[:, :1])], dim=-1)
+    return embedding
+    
+    
+def torch_checkpoint(func, args, flag, preserve_rng_state=False):
+    # torch's gradient checkpoint works with automatic mixed precision, given torch >= 1.8
+    if flag:
+        return torch.utils.checkpoint.checkpoint(
+            func, *args, preserve_rng_state=preserve_rng_state)
+    else:
+        return func(*args)

model/unet.py

@@ -0,0 +1,954 @@
+import math
+from dataclasses import dataclass
+from numbers import Number
+from typing import NamedTuple, Tuple, Union
+
+import numpy as np
+import torch as th
+from torch import nn
+import torch.nn.functional as F
+from choices import *
+from config_base import BaseConfig
+from .blocks import *
+from .graph_convolution_network import *
+from .nn import (conv_nd, linear, normalization, timestep_embedding, zero_module)
+
+
+@dataclass
+class BeatGANsUNetConfig(BaseConfig):
+    seq_len: int = 80
+    in_channels: int = 9
+    # base channels, will be multiplied
+    model_channels: int = 64
+    # output of the unet
+    out_channels: int = 9
+    # how many repeating resblocks per resolution
+    # the decoding side would have "one more" resblock
+    # default: 2
+    num_res_blocks: int = 2
+    # number of time embed channels and style channels
+    embed_channels: int = 256
+    # at what resolutions you want to do self-attention of the feature maps
+    # attentions generally improve performance
+    attention_resolutions: Tuple[int] = (0, )
+    # dropout applies to the resblocks (on feature maps)
+    dropout: float = 0.1
+    channel_mult: Tuple[int] = (1, 2, 4)    
+    conv_resample: bool = True
+    # 1 = 1d conv
+    dims: int = 1    
+    # number of attention heads
+    num_heads: int = 1
+    # or specify the number of channels per attention head
+    num_head_channels: int = -1
+    # use resblock for upscale/downscale blocks (expensive)
+    # default: True (BeatGANs)
+    resblock_updown: bool = True
+    use_new_attention_order: bool = False
+    resnet_two_cond: bool = True
+    resnet_cond_channels: int = None
+    # init the decoding conv layers with zero weights, this speeds up training
+    # default: True (BeatGANs)
+    resnet_use_zero_module: bool = True
+
+    def make_model(self):
+        return BeatGANsUNetModel(self)
+
+        
+class BeatGANsUNetModel(nn.Module):
+    def __init__(self, conf: BeatGANsUNetConfig):
+        super().__init__()
+        self.conf = conf
+
+        self.dtype = th.float32
+
+        self.time_emb_channels = conf.model_channels
+        self.time_embed = nn.Sequential(
+            linear(self.time_emb_channels, conf.embed_channels),
+            nn.SiLU(),
+            linear(conf.embed_channels, conf.embed_channels),
+        )
+
+        ch = input_ch = int(conf.channel_mult[0] * conf.model_channels)
+        self.input_blocks = nn.ModuleList([
+            TimestepEmbedSequential(
+                conv_nd(conf.dims, conf.in_channels, ch, 3, padding=1)),
+        ])
+
+        kwargs = dict(
+            use_condition=True,
+            two_cond=conf.resnet_two_cond,
+            use_zero_module=conf.resnet_use_zero_module,
+            # style channels for the resnet block
+            cond_emb_channels=conf.resnet_cond_channels,
+        )
+
+        self._feature_size = ch
+
+        # input_block_chans = [ch]
+        input_block_chans = [[] for _ in range(len(conf.channel_mult))]
+        input_block_chans[0].append(ch)
+
+        # number of blocks at each resolution
+        self.input_num_blocks = [0 for _ in range(len(conf.channel_mult))]
+        self.input_num_blocks[0] = 1
+        self.output_num_blocks = [0 for _ in range(len(conf.channel_mult))]
+
+        ds = 1
+        resolution = conf.seq_len
+        for level, mult in enumerate(conf.channel_mult):
+            for _ in range(conf.num_res_blocks):
+                layers = [
+                    ResBlockConfig(
+                        ch,
+                        conf.embed_channels,
+                        conf.dropout,
+                        out_channels=int(mult * conf.model_channels),
+                        dims=conf.dims,                        
+                        **kwargs,
+                    ).make_model()
+                ]
+                ch = int(mult * conf.model_channels)                
+                if resolution in conf.attention_resolutions:
+                    layers.append(
+                        AttentionBlock(
+                            ch,                            
+                            num_heads=conf.num_heads,
+                            num_head_channels=conf.num_head_channels,
+                            use_new_attention_order=conf.
+                            use_new_attention_order,
+                        ))
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+                # input_block_chans.append(ch)
+                input_block_chans[level].append(ch)
+                self.input_num_blocks[level] += 1
+                # print(input_block_chans)
+            if level != len(conf.channel_mult) - 1:
+                resolution //= 2
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        ResBlockConfig(
+                            ch,
+                            conf.embed_channels,
+                            conf.dropout,
+                            out_channels=out_ch,
+                            dims=conf.dims,                            
+                            down=True,
+                            **kwargs,
+                        ).make_model() if conf.
+                        resblock_updown else Downsample(ch,
+                                                        conf.conv_resample,
+                                                        dims=conf.dims,
+                                                        out_channels=out_ch)))
+                ch = out_ch
+                # input_block_chans.append(ch)
+                input_block_chans[level + 1].append(ch)
+                self.input_num_blocks[level + 1] += 1
+                ds *= 2
+                self._feature_size += ch
+                
+        self.middle_block = TimestepEmbedSequential(
+            ResBlockConfig(
+                ch,
+                conf.embed_channels,
+                conf.dropout,
+                dims=conf.dims,                
+                **kwargs,
+            ).make_model(),
+            #AttentionBlock(
+            #    ch,                
+            #    num_heads=conf.num_heads,
+            #    num_head_channels=conf.num_head_channels,
+            #    use_new_attention_order=conf.use_new_attention_order,
+            #),
+            ResBlockConfig(
+                ch,
+                conf.embed_channels,
+                conf.dropout,
+                dims=conf.dims,                
+                **kwargs,
+            ).make_model(),
+        )
+        self._feature_size += ch
+
+        self.output_blocks = nn.ModuleList([])
+        for level, mult in list(enumerate(conf.channel_mult))[::-1]:
+            for i in range(conf.num_res_blocks + 1):
+                # print(input_block_chans)
+                # ich = input_block_chans.pop()
+                try:
+                    ich = input_block_chans[level].pop()
+                except IndexError:
+                    # this happens only when num_res_block > num_enc_res_block
+                    # we will not have enough lateral (skip) connecions for all decoder blocks
+                    ich = 0
+                # print('pop:', ich)
+                layers = [
+                    ResBlockConfig(
+                        # only direct channels when gated
+                        channels=ch + ich,
+                        emb_channels=conf.embed_channels,
+                        dropout=conf.dropout,
+                        out_channels=int(conf.model_channels * mult),
+                        dims=conf.dims,                        
+                        # lateral channels are described here when gated
+                        has_lateral=True if ich > 0 else False,
+                        lateral_channels=None,
+                        **kwargs,
+                    ).make_model()
+                ]
+                ch = int(conf.model_channels * mult)
+                if resolution in conf.attention_resolutions:
+                    layers.append(
+                        AttentionBlock(
+                            ch,                            
+                            num_heads=conf.num_heads,
+                            num_head_channels=conf.num_head_channels,
+                            use_new_attention_order=conf.
+                            use_new_attention_order,
+                        ))
+                if level and i == conf.num_res_blocks:
+                    resolution *= 2
+                    out_ch = ch
+                    layers.append(
+                        ResBlockConfig(
+                            ch,
+                            conf.embed_channels,
+                            conf.dropout,
+                            out_channels=out_ch,
+                            dims=conf.dims,                            
+                            up=True,
+                            **kwargs,
+                        ).make_model() if (
+                            conf.resblock_updown
+                        ) else Upsample(ch,
+                                        conf.conv_resample,
+                                        dims=conf.dims,
+                                        out_channels=out_ch))
+                    ds //= 2
+                self.output_blocks.append(TimestepEmbedSequential(*layers))
+                self.output_num_blocks[level] += 1
+                self._feature_size += ch
+
+        # print(input_block_chans)
+        # print('inputs:', self.input_num_blocks)
+        # print('outputs:', self.output_num_blocks)
+
+        if conf.resnet_use_zero_module:
+            self.out = nn.Sequential(
+                normalization(ch),
+                nn.SiLU(),
+                zero_module(
+                    conv_nd(conf.dims,
+                            input_ch,
+                            conf.out_channels,
+                            3,  ## kernel size
+                            padding=1)),
+            )
+        else:
+            self.out = nn.Sequential(
+                normalization(ch),
+                nn.SiLU(),
+                conv_nd(conf.dims, input_ch, conf.out_channels, 3, padding=1),
+            )
+
+    def forward(self, x, t, **kwargs):
+        """
+        Apply the model to an input batch.
+
+        :param x: an [N x C x ...] Tensor of inputs.
+        :param timesteps: a 1-D batch of timesteps.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+
+        # hs = []
+        hs = [[] for _ in range(len(self.conf.channel_mult))]
+        emb = self.time_embed(timestep_embedding(t, self.time_emb_channels))
+
+        # new code supports input_num_blocks != output_num_blocks
+        h = x.type(self.dtype)
+        k = 0
+        for i in range(len(self.input_num_blocks)):
+            for j in range(self.input_num_blocks[i]):
+                h = self.input_blocks[k](h, emb=emb)
+                # print(i, j, h.shape)
+                hs[i].append(h) ## Get output from each layer
+                k += 1
+        assert k == len(self.input_blocks)
+
+        # middle blocks
+        h = self.middle_block(h, emb=emb)
+
+        # output blocks
+        k = 0
+        for i in range(len(self.output_num_blocks)):
+            for j in range(self.output_num_blocks[i]):
+                # take the lateral connection from the same layer (in reserve)
+                # until there is no more, use None
+                try:
+                    lateral = hs[-i - 1].pop()
+                    # print(i, j, lateral.shape)
+                except IndexError:
+                    lateral = None
+                    # print(i, j, lateral)
+                h = self.output_blocks[k](h, emb=emb, lateral=lateral)
+                k += 1
+
+        h = h.type(x.dtype)
+        pred = self.out(h)
+        return Return(pred=pred)
+
+
+class Return(NamedTuple):
+    pred: th.Tensor
+    
+    
+@dataclass
+class BeatGANsEncoderConfig(BaseConfig):    
+    in_channels: int
+    seq_len: int = 80
+    num_res_blocks: int = 2
+    attention_resolutions: Tuple[int] = (0, )
+    model_channels: int = 32
+    out_channels: int = 256
+    dropout: float = 0.1
+    channel_mult: Tuple[int] = (1, 2, 4)
+    use_time_condition: bool = False
+    conv_resample: bool = True
+    dims: int = 1    
+    num_heads: int = 1
+    num_head_channels: int = -1
+    resblock_updown: bool = True
+    use_new_attention_order: bool = False
+    pool: str = 'adaptivenonzero'
+
+    def make_model(self):
+        return BeatGANsEncoderModel(self)
+
+        
+class BeatGANsEncoderModel(nn.Module):
+    """
+    The half UNet model with attention and timestep embedding.
+    
+    For usage, see UNet.
+    """
+    def __init__(self, conf: BeatGANsEncoderConfig):
+        super().__init__()
+        self.conf = conf
+        self.dtype = th.float32
+
+        if conf.use_time_condition: 
+            time_embed_dim = conf.model_channels
+            self.time_embed = nn.Sequential(
+                linear(conf.model_channels, time_embed_dim),
+                nn.SiLU(),
+                linear(time_embed_dim, time_embed_dim),
+            )
+        else:
+            time_embed_dim = None
+
+        ch = int(conf.channel_mult[0] * conf.model_channels)
+        self.input_blocks = nn.ModuleList([
+            TimestepEmbedSequential(
+                conv_nd(conf.dims, conf.in_channels, ch, 3, padding=1),
+                )
+        ])
+        self._feature_size = ch
+        input_block_chans = [ch]
+        ds = 1
+        resolution = conf.seq_len
+        for level, mult in enumerate(conf.channel_mult):
+            for _ in range(conf.num_res_blocks):
+                layers = [
+                    ResBlockConfig(
+                        ch,
+                        time_embed_dim,
+                        conf.dropout,
+                        out_channels=int(mult * conf.model_channels),
+                        dims=conf.dims,
+                        use_condition=conf.use_time_condition,                        
+                    ).make_model()
+                ]
+                ch = int(mult * conf.model_channels)               
+                if resolution in conf.attention_resolutions:
+                    layers.append(
+                        AttentionBlock(
+                            ch,                            
+                            num_heads=conf.num_heads,
+                            num_head_channels=conf.num_head_channels,
+                            use_new_attention_order=conf.
+                            use_new_attention_order,
+                        ))
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+                input_block_chans.append(ch)
+            if level != len(conf.channel_mult) - 1:
+                resolution //= 2
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        ResBlockConfig(
+                            ch,
+                            time_embed_dim,
+                            conf.dropout,
+                            out_channels=out_ch,
+                            dims=conf.dims,
+                            use_condition=conf.use_time_condition,
+                            down=True,
+                        ).make_model() if (
+                            conf.resblock_updown
+                        ) else Downsample(ch,
+                                          conf.conv_resample,
+                                          dims=conf.dims,
+                                          out_channels=out_ch)))
+                ch = out_ch
+                input_block_chans.append(ch)
+                ds *= 2
+                self._feature_size += ch
+
+        self.middle_block = TimestepEmbedSequential(
+            ResBlockConfig(
+                ch,
+                time_embed_dim,
+                conf.dropout,
+                dims=conf.dims,
+                use_condition=conf.use_time_condition,                
+            ).make_model(),
+            AttentionBlock(
+                ch,                
+                num_heads=conf.num_heads,
+                num_head_channels=conf.num_head_channels,
+                use_new_attention_order=conf.use_new_attention_order,
+            ),
+            ResBlockConfig(
+                ch,
+                time_embed_dim,
+                conf.dropout,
+                dims=conf.dims,
+                use_condition=conf.use_time_condition,                
+            ).make_model(),
+        )
+        self._feature_size += ch
+        if conf.pool == "adaptivenonzero":
+            self.out = nn.Sequential(
+                normalization(ch),
+                nn.SiLU(),
+                ## nn.AdaptiveAvgPool2d((1, 1)),
+                nn.AdaptiveAvgPool1d(1),
+                conv_nd(conf.dims, ch, conf.out_channels, 1),
+                nn.Flatten(),
+            )
+        else:
+            raise NotImplementedError(f"Unexpected {conf.pool} pooling")
+
+    def forward(self, x, t=None):
+        """
+        Apply the model to an input batch.
+
+        :param x: an [N x C x ...] Tensor of inputs.
+        :param timesteps: a 1-D batch of timesteps.
+        :return: an [N x K] Tensor of outputs.
+        """
+        if self.conf.use_time_condition:
+            emb = self.time_embed(timestep_embedding(t, self.model_channels))
+        else: ## autoencoding.py
+            emb = None
+
+        results = []
+        h = x.type(self.dtype)        
+        for module in self.input_blocks: ## flow input x over all the input blocks
+            h = module(h, emb=emb)                        
+            if self.conf.pool.startswith("spatial"):
+                results.append(h.type(x.dtype).mean(dim=(2, 3)))                
+        h = self.middle_block(h, emb=emb) ## TimestepEmbedSequential(...)
+        if self.conf.pool.startswith("spatial"):
+            results.append(h.type(x.dtype).mean(dim=(2, 3)))
+            h = th.cat(results, axis=-1)
+        else: ## autoencoder.py
+            h = h.type(x.dtype)
+
+        h = h.float()
+        h = self.out(h)
+
+        return h
+
+        
+@dataclass
+class GCNUNetConfig(BaseConfig):
+    in_channels: int = 9
+    node_n: int = 3
+    seq_len: int = 80    
+    # base channels, will be multiplied
+    model_channels: int = 32
+    # output of the unet
+    out_channels: int = 9
+    # how many repeating resblocks per resolution
+    num_res_blocks: int = 8
+    # number of time embed channels and style channels
+    embed_channels: int = 256
+    # dropout applies to the resblocks
+    dropout: float = 0.1
+    channel_mult: Tuple[int] = (1, 2, 4)    
+    resnet_two_cond: bool = True
+    
+    def make_model(self):
+        return GCNUNetModel(self)
+
+        
+class GCNUNetModel(nn.Module):
+    def __init__(self, conf: GCNUNetConfig):
+        super().__init__()
+        self.conf = conf
+        self.dtype = th.float32
+        assert conf.in_channels%conf.node_n == 0                
+        self.in_features = conf.in_channels//conf.node_n
+
+        self.time_emb_channels = conf.model_channels*4
+        self.time_embed = nn.Sequential(
+            linear(self.time_emb_channels, conf.embed_channels),
+            nn.SiLU(),
+            linear(conf.embed_channels, conf.embed_channels),
+        )
+        
+        ch = int(conf.channel_mult[0] * conf.model_channels)
+        self.input_blocks = nn.ModuleList([
+            TimestepEmbedSequential(
+                graph_convolution(in_features=self.in_features, out_features=ch, node_n=conf.node_n, seq_len=conf.seq_len)),
+        ])
+
+        kwargs = dict(
+            use_condition=True,
+            two_cond=conf.resnet_two_cond,
+        )
+
+        input_block_chans = [[] for _ in range(len(conf.channel_mult))]
+        input_block_chans[0].append(ch)
+
+        # number of blocks at each resolution
+        self.input_num_blocks = [0 for _ in range(len(conf.channel_mult))]
+        self.input_num_blocks[0] = 1
+        self.output_num_blocks = [0 for _ in range(len(conf.channel_mult))]
+
+        ds = 1
+        resolution = conf.seq_len
+        for level, mult in enumerate(conf.channel_mult):
+            for _ in range(conf.num_res_blocks):
+                layers = [
+                    residual_graph_convolution_config(
+                    in_features=ch,
+                    seq_len=resolution,
+                    emb_channels = conf.embed_channels,
+                    dropout=conf.dropout,
+                    out_features=int(mult * conf.model_channels),
+                    node_n=conf.node_n,
+                    **kwargs,
+                    ).make_model()                
+                ]
+                ch = int(mult * conf.model_channels)                
+                self.input_blocks.append(*layers)
+                input_block_chans[level].append(ch)
+                self.input_num_blocks[level] += 1
+            if level != len(conf.channel_mult) - 1:
+                resolution //= 2
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        graph_downsample()))
+                ch = out_ch
+                input_block_chans[level + 1].append(ch)
+                self.input_num_blocks[level + 1] += 1
+                ds *= 2
+                
+        self.output_blocks = nn.ModuleList([])
+        for level, mult in list(enumerate(conf.channel_mult))[::-1]:
+            for i in range(conf.num_res_blocks + 1):
+                try:
+                    ich = input_block_chans[level].pop()
+                except IndexError:
+                    # this happens only when num_res_block > num_enc_res_block
+                    # we will not have enough lateral (skip) connecions for all decoder blocks
+                    ich = 0
+                layers = [
+                    residual_graph_convolution_config(
+                    in_features=ch + ich,
+                    seq_len=resolution,
+                    emb_channels = conf.embed_channels,
+                    dropout=conf.dropout,
+                    out_features=int(mult * conf.model_channels),
+                    node_n=conf.node_n,
+                    has_lateral=True if ich > 0 else False,
+                    **kwargs,
+                    ).make_model()
+                ]
+                ch = int(mult*conf.model_channels)
+                if level and i == conf.num_res_blocks:
+                    resolution *= 2
+                    out_ch = ch
+                    layers.append(graph_upsample())
+                    ds //= 2
+                self.output_blocks.append(TimestepEmbedSequential(*layers))
+                self.output_num_blocks[level] += 1
+                
+        self.out = nn.Sequential(
+        graph_convolution(in_features=ch, out_features=self.in_features, node_n=conf.node_n, seq_len=conf.seq_len),
+        nn.Tanh(),
+        )
+        
+    def forward(self, x, t, **kwargs):
+        """
+        Apply the model to an input batch.
+        
+        :param x: an [N x C x ...] Tensor of inputs.
+        :param timesteps: a 1-D batch of timesteps.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        bs, channels, seq_len = x.shape
+        x = x.reshape(bs, self.conf.node_n, self.in_features, seq_len).permute(0, 2, 1, 3)
+        
+        hs = [[] for _ in range(len(self.conf.channel_mult))]
+        emb = self.time_embed(timestep_embedding(t, self.time_emb_channels))
+        
+        # new code supports input_num_blocks != output_num_blocks
+        h = x.type(self.dtype)
+        k = 0
+        for i in range(len(self.input_num_blocks)):
+            for j in range(self.input_num_blocks[i]):
+                h = self.input_blocks[k](h, emb=emb)
+                # print(i, j, h.shape)
+                hs[i].append(h) ## Get output from each layer
+                k += 1
+        assert k == len(self.input_blocks)
+
+        # output blocks
+        k = 0
+        for i in range(len(self.output_num_blocks)):
+            for j in range(self.output_num_blocks[i]):
+                # take the lateral connection from the same layer (in reserve)
+                # until there is no more, use None
+                try:
+                    lateral = hs[-i - 1].pop()
+                    # print(i, j, lateral.shape)
+                except IndexError:
+                    lateral = None
+                    # print(i, j, lateral)
+                h = self.output_blocks[k](h, emb=emb, lateral=lateral)
+                k += 1
+
+        h = h.type(x.dtype)
+        pred = self.out(h)
+        pred = pred.permute(0, 2, 1, 3).reshape(bs, -1, seq_len)
+        
+        return Return(pred=pred)
+
+        
+@dataclass
+class GCNEncoderConfig(BaseConfig):    
+    in_channels: int
+    in_features = 3 # features for one node        
+    seq_len: int = 40
+    seq_len_future: int = 3
+    num_res_blocks: int = 2
+    model_channels: int = 32
+    out_channels: int = 32
+    dropout: float = 0.1
+    channel_mult: Tuple[int] = (1, 2, 4)
+    use_time_condition: bool = False
+    
+    def make_model(self):
+        return GCNEncoderModel(self)
+
+        
+class GCNEncoderModel(nn.Module):
+    def __init__(self, conf: GCNEncoderConfig):
+        super().__init__()
+        self.conf = conf
+        self.dtype = th.float32
+        assert conf.in_channels%conf.in_features == 0                
+        self.in_features = conf.in_features
+        self.node_n = conf.in_channels//conf.in_features
+                        
+        ch = int(conf.channel_mult[0] * conf.model_channels)
+        self.input_blocks = nn.ModuleList([
+            graph_convolution(in_features=self.in_features, out_features=ch, node_n=self.node_n, seq_len=conf.seq_len),
+        ])
+        input_block_chans = [ch]
+        ds = 1
+        resolution = conf.seq_len
+        for level, mult in enumerate(conf.channel_mult):
+            for _ in range(conf.num_res_blocks):
+                layers = [
+                    residual_graph_convolution_config(
+                    in_features=ch,
+                    seq_len=resolution,
+                    emb_channels = None,
+                    dropout=conf.dropout,
+                    out_features=int(mult * conf.model_channels),
+                    node_n=self.node_n,
+                    use_condition=conf.use_time_condition,
+                    ).make_model()
+                ]                
+                ch = int(mult * conf.model_channels)
+                self.input_blocks.append(*layers)
+                input_block_chans.append(ch)
+            if level != len(conf.channel_mult) - 1:
+                resolution //= 2
+                out_ch = ch
+                self.input_blocks.append(
+                    graph_downsample())
+                ch = out_ch
+                input_block_chans.append(ch)
+                ds *= 2
+
+        self.hand_prediction = nn.Sequential(
+            conv_nd(1, ch*2, ch*2, 3, padding=1),
+            nn.LayerNorm([ch*2, conf.seq_len_future]),
+            nn.Tanh(),
+            conv_nd(1, ch*2, self.in_features*2, 1),
+            nn.Tanh(),
+        )
+        
+        self.head_prediction = nn.Sequential(
+            conv_nd(1, ch, ch, 3, padding=1),
+            nn.LayerNorm([ch, conf.seq_len_future]),
+            nn.Tanh(),            
+            conv_nd(1, ch, self.in_features, 1),
+            nn.Tanh(),
+        )
+        
+        self.out = nn.Sequential(
+            nn.AdaptiveAvgPool1d(1),
+            conv_nd(1, ch*self.node_n, conf.out_channels, 1),
+            nn.Flatten(),
+        )
+
+        
+    def forward(self, x, t=None):
+        bs, channels, seq_len = x.shape
+
+        if self.node_n == 3: # both hand and head
+            hand_last = x[:, :6, -1:].expand(-1, -1, self.conf.seq_len_future).clone() #last observed hand position
+            head_last = x[:, 6:, -1:].expand(-1, -1, self.conf.seq_len_future).clone()# last observed head orientation
+
+        if self.node_n == 2: # hand only
+            hand_last = x[:, :, -1:].expand(-1, -1, self.conf.seq_len_future).clone() #last observed hand position
+            
+        if self.node_n == 1: # head only
+            head_last = x[:, :, -1:].expand(-1, -1, self.conf.seq_len_future).clone()# last observed head orientation
+
+        
+        x = x.reshape(bs, self.node_n, self.in_features, seq_len).permute(0, 2, 1, 3)
+        
+        h = x.type(self.dtype)
+        for module in self.input_blocks:
+            h = module(h)
+                    
+        h = h.type(x.dtype)
+        h = h.float()
+        bs, features, node_n, seq_len = h.shape
+        
+        if self.node_n == 3: # both hand and head
+            hand_features = h[:, :, :2, -self.conf.seq_len_future:].reshape(bs, features*2, -1)        
+            head_features = h[:, :, 2:, -self.conf.seq_len_future:].reshape(bs, features, -1)
+            
+            pred_hand = self.hand_prediction(hand_features) + hand_last
+            pred_head = self.head_prediction(head_features) + head_last
+            pred_head = F.normalize(pred_head, dim=1)# normalize head orientation to unit vectors
+            
+        if self.node_n == 2: # hand only
+            hand_features = h[:, :, :, -self.conf.seq_len_future:].reshape(bs, features*2, -1)            
+            pred_hand = self.hand_prediction(hand_features) + hand_last
+            pred_head = None
+        
+        if self.node_n == 1: # head only        
+            head_features = h[:, :, :, -self.conf.seq_len_future:].reshape(bs, features, -1)                        
+            pred_head = self.head_prediction(head_features) + head_last
+            pred_head = F.normalize(pred_head, dim=1)# normalize head orientation to unit vectors
+            pred_hand = None
+        
+        h = h.reshape(bs, features*node_n, seq_len)
+        h = self.out(h)
+        
+        return h, pred_hand, pred_head
+
+        
+@dataclass
+class CNNEncoderConfig(BaseConfig):    
+    in_channels: int    
+    seq_len: int = 40
+    seq_len_future: int = 3    
+    out_channels: int = 128
+        
+    def make_model(self):
+        return CNNEncoderModel(self)
+
+        
+class CNNEncoderModel(nn.Module):
+    def __init__(self, conf: CNNEncoderConfig):
+        super().__init__()
+        self.conf = conf
+        self.dtype = th.float32
+        input_dim = conf.in_channels
+        length = conf.seq_len
+        out_channels = conf.out_channels
+        
+        self.encoder = nn.Sequential(
+            nn.Conv1d(input_dim, 32, kernel_size=3, padding=1),
+            nn.LayerNorm([32, length]),
+            nn.ReLU(inplace=True),
+            nn.Conv1d(32, 32, kernel_size=3, padding=1),
+            nn.LayerNorm([32, length]),
+            nn.ReLU(inplace=True),
+            nn.Conv1d(32, 32, kernel_size=3, padding=1),
+            nn.LayerNorm([32, length]),
+            nn.ReLU(inplace=True)            
+        )
+        
+        self.out = nn.Linear(32 * length, out_channels)
+        
+    def forward(self, x, t=None):
+        bs, channels, seq_len = x.shape
+        hand_last = x[:, :6, -1:].expand(-1, -1, self.conf.seq_len_future).clone() #last observed hand position
+        head_last = x[:, 6:, -1:].expand(-1, -1, self.conf.seq_len_future).clone()# last observed head orientation
+        
+        h = x.type(self.dtype)
+        h = self.encoder(h)
+        h = h.view(h.shape[0], -1)
+        
+        h = h.type(x.dtype)
+        h = h.float()
+        
+        h = self.out(h)
+        return h, hand_last, head_last
+
+        
+@dataclass
+class GRUEncoderConfig(BaseConfig):    
+    in_channels: int    
+    seq_len: int = 40
+    seq_len_future: int = 3    
+    out_channels: int = 128
+        
+    def make_model(self):
+        return GRUEncoderModel(self)
+
+        
+class GRUEncoderModel(nn.Module):
+    def __init__(self, conf: GRUEncoderConfig):
+        super().__init__()
+        self.conf = conf
+        self.dtype = th.float32
+        input_dim = conf.in_channels
+        length = conf.seq_len
+        feature_channels = 32
+        out_channels = conf.out_channels
+        
+        self.encoder = nn.GRU(input_dim, feature_channels, 1, batch_first=True)
+        
+        self.out = nn.Linear(feature_channels * length, out_channels)
+        
+    def forward(self, x, t=None):
+        bs, channels, seq_len = x.shape
+        hand_last = x[:, :6, -1:].expand(-1, -1, self.conf.seq_len_future).clone() #last observed hand position
+        head_last = x[:, 6:, -1:].expand(-1, -1, self.conf.seq_len_future).clone()# last observed head orientation
+        
+        h = x.type(self.dtype)
+        h, _ = self.encoder(h.permute(0, 2, 1))
+        h = h.reshape(h.shape[0], -1)
+        
+        h = h.type(x.dtype)
+        h = h.float()
+        
+        h = self.out(h)
+        return h, hand_last, head_last
+
+        
+@dataclass
+class LSTMEncoderConfig(BaseConfig):    
+    in_channels: int    
+    seq_len: int = 40
+    seq_len_future: int = 3    
+    out_channels: int = 128
+        
+    def make_model(self):
+        return LSTMEncoderModel(self)
+
+        
+class LSTMEncoderModel(nn.Module):
+    def __init__(self, conf: LSTMEncoderConfig):
+        super().__init__()
+        self.conf = conf
+        self.dtype = th.float32
+        input_dim = conf.in_channels
+        length = conf.seq_len
+        feature_channels = 32
+        out_channels = conf.out_channels        
+        
+        self.encoder = nn.LSTM(input_dim, feature_channels, 1, batch_first=True)
+                
+        self.out = nn.Linear(feature_channels * length, out_channels)
+        
+    def forward(self, x, t=None):
+        bs, channels, seq_len = x.shape
+        hand_last = x[:, :6, -1:].expand(-1, -1, self.conf.seq_len_future).clone() #last observed hand position
+        head_last = x[:, 6:, -1:].expand(-1, -1, self.conf.seq_len_future).clone()# last observed head orientation
+        
+        h = x.type(self.dtype)
+        h, _ = self.encoder(h.permute(0, 2, 1))
+        h = h.reshape(h.shape[0], -1)
+                       
+        h = h.type(x.dtype)
+        h = h.float()
+        
+        h = self.out(h)
+        return h, hand_last, head_last                
+
+        
+@dataclass
+class MLPEncoderConfig(BaseConfig):    
+    in_channels: int    
+    seq_len: int = 40
+    seq_len_future: int = 3    
+    out_channels: int = 128
+        
+    def make_model(self):
+        return MLPEncoderModel(self)
+
+                
+class MLPEncoderModel(nn.Module):
+    def __init__(self, conf: MLPEncoderConfig):
+        super().__init__()
+        self.conf = conf
+        self.dtype = th.float32
+        input_dim = conf.in_channels
+        length = conf.seq_len
+        out_channels = conf.out_channels
+        
+        linear_size = 128
+        self.encoder = nn.Sequential(
+            nn.Linear(length*input_dim, linear_size),
+            nn.LayerNorm([linear_size]),
+            nn.ReLU(inplace=True),
+            nn.Linear(linear_size, linear_size),
+            nn.LayerNorm([linear_size]),
+            nn.ReLU(inplace=True),
+        )
+                        
+        self.out = nn.Linear(linear_size, out_channels)
+        
+    def forward(self, x, t=None):
+        bs, channels, seq_len = x.shape
+        hand_last = x[:, :6, -1:].expand(-1, -1, self.conf.seq_len_future).clone() #last observed hand position
+        head_last = x[:, 6:, -1:].expand(-1, -1, self.conf.seq_len_future).clone()# last observed head orientation
+        
+        h = x.type(self.dtype)
+        
+        h = h.view(h.shape[0], -1)
+        h = self.encoder(h)
+                     
+        h = h.type(x.dtype)
+        h = h.float()
+        
+        h = self.out(h)
+        return h, hand_last, head_last

model/unet_autoenc.py

@@ -0,0 +1,418 @@
+from enum import Enum
+
+import torch, pdb
+import os
+from torch import Tensor
+from torch.nn.functional import silu
+from .unet import *
+from choices import *
+
+
+@dataclass
+class BeatGANsAutoencConfig(BeatGANsUNetConfig):
+    seq_len_future: int = 3
+    enc_out_channels: int = 128
+    semantic_encoder_type: str = 'gcn'
+    enc_channel_mult: Tuple[int] = None    
+    def make_model(self):
+        return BeatGANsAutoencModel(self)
+        
+class BeatGANsAutoencModel(BeatGANsUNetModel):
+    def __init__(self, conf: BeatGANsAutoencConfig):
+        super().__init__(conf)
+        self.conf = conf
+        
+        # having only time, cond
+        self.time_embed = TimeStyleSeperateEmbed(
+            time_channels=conf.model_channels,
+            time_out_channels=conf.embed_channels,
+        )
+        
+        if conf.semantic_encoder_type == 'gcn':
+            self.encoder = GCNEncoderConfig(
+                seq_len=conf.seq_len,
+                seq_len_future=conf.seq_len_future,
+                in_channels=conf.in_channels,
+                model_channels=16,
+                out_channels=conf.enc_out_channels,            
+                channel_mult=conf.enc_channel_mult or conf.channel_mult,
+            ).make_model()
+        elif conf.semantic_encoder_type == '1dcnn':
+            self.encoder = CNNEncoderConfig(
+                seq_len=conf.seq_len,
+                seq_len_future=conf.seq_len_future,
+                in_channels=conf.in_channels,
+                out_channels=conf.enc_out_channels,
+            ).make_model()
+        elif conf.semantic_encoder_type == 'gru':
+            # ensure deterministic behavior of RNNs
+            os.environ["CUBLAS_WORKSPACE_CONFIG"]=":4096:2"
+            self.encoder = GRUEncoderConfig(
+                seq_len=conf.seq_len,
+                seq_len_future=conf.seq_len_future,
+                in_channels=conf.in_channels,
+                out_channels=conf.enc_out_channels,
+            ).make_model()            
+        elif conf.semantic_encoder_type == 'lstm':
+            # ensure deterministic behavior of RNNs
+            os.environ["CUBLAS_WORKSPACE_CONFIG"]=":4096:2"        
+            self.encoder = LSTMEncoderConfig(
+                seq_len=conf.seq_len,
+                seq_len_future=conf.seq_len_future,
+                in_channels=conf.in_channels,
+                out_channels=conf.enc_out_channels,
+            ).make_model()            
+        elif conf.semantic_encoder_type == 'mlp':
+            self.encoder = MLPEncoderConfig(
+                seq_len=conf.seq_len,
+                seq_len_future=conf.seq_len_future,
+                in_channels=conf.in_channels,
+                out_channels=conf.enc_out_channels,
+            ).make_model()
+        else:
+            raise NotImplementedError()
+
+    def reparameterize(self, mu: Tensor, logvar: Tensor) -> Tensor:
+        """
+        Reparameterization trick to sample from N(mu, var) from
+        N(0,1).
+        :param mu: (Tensor) Mean of the latent Gaussian [B x D]
+        :param logvar: (Tensor) Standard deviation of the latent Gaussian [B x D]
+        :return: (Tensor) [B x D]
+        """
+        assert self.conf.is_stochastic
+        std = torch.exp(0.5 * logvar)
+        eps = torch.randn_like(std)
+        return eps * std + mu
+
+    def sample_z(self, n: int, device):
+        assert self.conf.is_stochastic
+        return torch.randn(n, self.conf.enc_out_channels, device=device)
+
+    def noise_to_cond(self, noise: Tensor):
+        raise NotImplementedError()
+        assert self.conf.noise_net_conf is not None
+        return self.noise_net.forward(noise)
+
+    def encode(self, x):
+        cond, pred_hand, pred_head = self.encoder.forward(x)
+        return cond, pred_hand, pred_head
+
+    @property
+    def stylespace_sizes(self):
+        modules = list(self.input_blocks.modules()) + list(
+            self.middle_block.modules()) + list(self.output_blocks.modules())
+        sizes = []
+        for module in modules:
+            if isinstance(module, ResBlock):
+                linear = module.cond_emb_layers[-1]
+                sizes.append(linear.weight.shape[0])
+        return sizes
+
+    def encode_stylespace(self, x, return_vector: bool = True):
+        """
+        encode to style space
+        """
+        modules = list(self.input_blocks.modules()) + list(
+            self.middle_block.modules()) + list(self.output_blocks.modules())
+        # (n, c)
+        cond = self.encoder.forward(x)
+        S = []
+        for module in modules:
+            if isinstance(module, ResBlock):
+                # (n, c')
+                s = module.cond_emb_layers.forward(cond)
+                S.append(s)
+
+        if return_vector:
+            # (n, sum_c)
+            return torch.cat(S, dim=1)
+        else:
+            return S
+            
+    def forward(self,
+                x,
+                t,                
+                x_start=None,
+                cond=None,
+                style=None,
+                noise=None,
+                t_cond=None,
+                **kwargs):
+        """
+        Apply the model to an input batch.
+
+        Args:
+            x_start: the original image to encode
+            cond: output of the encoder
+            noise: random noise (to predict the cond)
+        """
+        if t_cond is None:
+            t_cond = t ## randomly sampled timestep with the size of [batch_size]
+
+        if noise is not None: 
+            # if the noise is given, we predict the cond from noise
+            cond = self.noise_to_cond(noise)
+
+        cond_given = True
+        if cond is None:
+            cond_given = False
+            if x is not None:
+                assert len(x) == len(x_start), f'{len(x)} != {len(x_start)}'
+
+            cond, pred_hand, pred_head = self.encode(x_start)
+            
+        if t is not None: ## t==t_cond
+            _t_emb = timestep_embedding(t, self.conf.model_channels)
+            #print("t: {}, _t_emb:{}".format(t, _t_emb))
+            _t_cond_emb = timestep_embedding(t_cond, self.conf.model_channels)
+            #print("t_cond: {}, _t_cond_emb:{}".format(t, _t_cond_emb))
+        else:
+            # this happens when training only autoenc
+            _t_emb = None
+            _t_cond_emb = None
+
+        if self.conf.resnet_two_cond:
+            res = self.time_embed.forward( ## self.time_embed is an MLP
+                time_emb=_t_emb,
+                cond=cond,
+                time_cond_emb=_t_cond_emb,
+            )
+        else:
+            raise NotImplementedError()
+
+        if self.conf.resnet_two_cond:
+            # two cond: first = time emb, second = cond_emb
+            emb = res.time_emb
+            cond_emb = res.emb
+        else:
+            # one cond = combined of both time and cond
+            emb = res.emb
+            cond_emb = None
+
+        # override the style if given
+        style = style or res.style ## style==None, res.style: cond, torch.Size([64, 512])
+        
+
+        # where in the model to supply time conditions
+        enc_time_emb = emb ## time embeddings
+        mid_time_emb = emb
+        dec_time_emb = emb
+        # where in the model to supply style conditions
+        enc_cond_emb = cond_emb ## z_sem embeddings
+        mid_cond_emb = cond_emb
+        dec_cond_emb = cond_emb        
+
+        # hs = []
+        hs = [[] for _ in range(len(self.conf.channel_mult))]
+
+        if x is not None:
+            h = x.type(self.dtype)
+            # input blocks
+            k = 0
+            for i in range(len(self.input_num_blocks)):
+                for j in range(self.input_num_blocks[i]):
+                    h = self.input_blocks[k](h,
+                                             emb=enc_time_emb,
+                                             cond=enc_cond_emb)
+                    # print(i, j, h.shape)
+                    '''if h.shape[-1]%2==1:
+                        pdb.set_trace()'''
+                    hs[i].append(h)
+                    k += 1
+            assert k == len(self.input_blocks)
+
+            # middle blocks
+            h = self.middle_block(h, emb=mid_time_emb, cond=mid_cond_emb)
+        else:
+            # no lateral connections
+            # happens when training only the autonecoder
+            h = None
+            hs = [[] for _ in range(len(self.conf.channel_mult))]
+            pdb.set_trace()
+        
+        # output blocks
+        k = 0
+        for i in range(len(self.output_num_blocks)):
+            for j in range(self.output_num_blocks[i]):
+                # take the lateral connection from the same layer (in reserve)
+                # until there is no more, use None
+                try:
+                    lateral = hs[-i - 1].pop() ## in the reverse order (symmetric)
+                except IndexError:
+                    lateral = None
+                '''print(i, j, lateral.shape, h.shape)
+                if lateral.shape[-1]!=h.shape[-1]:
+                    pdb.set_trace()'''
+                # print("h is", h.size())
+                # print("lateral is", lateral.size())
+                h = self.output_blocks[k](h,
+                                          emb=dec_time_emb,
+                                          cond=dec_cond_emb,
+                                          lateral=lateral)
+                k += 1
+                
+        pred = self.out(h)
+        # print("h:", h.shape)
+        # print("pred:", pred.shape)
+        
+        if cond_given == True:
+            return AutoencReturn(pred=pred, cond=cond)
+        else:
+            return AutoencReturn(pred=pred, cond=cond, pred_hand=pred_hand, pred_head=pred_head)
+
+        
+@dataclass
+class GCNAutoencConfig(GCNUNetConfig):
+    # number of style channels
+    enc_out_channels: int = 256
+    enc_channel_mult: Tuple[int] = None
+    def make_model(self):
+        return GCNAutoencModel(self)
+        
+        
+class GCNAutoencModel(GCNUNetModel):
+    def __init__(self, conf: GCNAutoencConfig):
+        super().__init__(conf)
+        self.conf = conf
+        
+        # having only time, cond
+        self.time_emb_channels = conf.model_channels
+        self.time_embed = TimeStyleSeperateEmbed(
+            time_channels=self.time_emb_channels,
+            time_out_channels=conf.embed_channels,
+        )
+        
+        self.encoder = GCNEncoderConfig(
+            seq_len=conf.seq_len,
+            in_channels=conf.in_channels,
+            model_channels=32,
+            out_channels=conf.enc_out_channels,            
+            channel_mult=conf.enc_channel_mult or conf.channel_mult,
+        ).make_model()
+
+    def encode(self, x):
+        cond = self.encoder.forward(x)
+        return {'cond': cond}
+        
+    def forward(self,
+                x,
+                t,
+                x_start=None,
+                cond=None,                
+                **kwargs):
+        """
+        Apply the model to an input batch.
+
+        Args:
+            x_start: the original input to encode
+            cond: output of the encoder
+        """
+        
+        if cond is None:
+            if x is not None:
+                assert len(x) == len(x_start), f'{len(x)} != {len(x_start)}'
+
+            tmp = self.encode(x_start) 
+            cond = tmp['cond'] 
+            
+        if t is not None: 
+            _t_emb = timestep_embedding(t, self.time_emb_channels)
+        else:
+            # this happens when training only autoenc
+            _t_emb = None
+
+        if self.conf.resnet_two_cond:
+            res = self.time_embed.forward( ## self.time_embed is an MLP
+                time_emb=_t_emb,
+                cond=cond,
+            )
+            # two cond: first = time emb, second = cond_emb
+            emb = res.time_emb
+            cond_emb = res.emb            
+        else:
+            raise NotImplementedError()
+       
+        # where in the model to supply time conditions
+        enc_time_emb = emb ## time embeddings
+        mid_time_emb = emb
+        dec_time_emb = emb
+        enc_cond_emb = cond_emb ## z_sem embeddings
+        mid_cond_emb = cond_emb
+        dec_cond_emb = cond_emb
+        
+        
+        bs, channels, seq_len = x.shape
+        x = x.reshape(bs, self.conf.node_n, self.in_features, seq_len).permute(0, 2, 1, 3)        
+        hs = [[] for _ in range(len(self.conf.channel_mult))]
+        h = x.type(self.dtype)
+        
+        # input blocks
+        k = 0
+        for i in range(len(self.input_num_blocks)):
+            for j in range(self.input_num_blocks[i]):                
+                h = self.input_blocks[k](h,
+                                         emb=enc_time_emb,
+                                         cond=enc_cond_emb)
+                hs[i].append(h)
+                k += 1
+        assert k == len(self.input_blocks)
+                
+        # output blocks
+        k = 0
+        for i in range(len(self.output_num_blocks)):
+            for j in range(self.output_num_blocks[i]):
+                # take the lateral connection from the same layer (in reserve)
+                # until there is no more, use None
+                try:
+                    lateral = hs[-i - 1].pop() ## in the reverse order (symmetric)
+                except IndexError:
+                    lateral = None
+                h = self.output_blocks[k](h,
+                                          emb=dec_time_emb,
+                                          cond=dec_cond_emb,
+                                          lateral=lateral)
+                k += 1
+
+
+        pred = self.out(h)
+        pred = pred.permute(0, 2, 1, 3).reshape(bs, -1, seq_len)
+
+        return AutoencReturn(pred=pred, cond=cond)
+
+
+class AutoencReturn(NamedTuple):
+    pred: Tensor
+    cond: Tensor = None
+    pred_hand: Tensor = None
+    pred_head: Tensor = None
+
+    
+class EmbedReturn(NamedTuple):
+    # style and time
+    emb: Tensor = None
+    # time only
+    time_emb: Tensor = None
+    # style only (but could depend on time)
+    style: Tensor = None
+
+
+class TimeStyleSeperateEmbed(nn.Module):
+    # embed only style
+    def __init__(self, time_channels, time_out_channels):
+        super().__init__()
+        self.time_embed = nn.Sequential(
+            linear(time_channels, time_out_channels),
+            nn.SiLU(),
+            linear(time_out_channels, time_out_channels),
+        )
+        self.style = nn.Identity()
+
+    def forward(self, time_emb=None, cond=None, **kwargs):
+        if time_emb is None:
+            # happens with autoenc training mode
+            time_emb = None
+        else:
+            time_emb = self.time_embed(time_emb)
+        style = self.style(cond) ## style==cond
+        return EmbedReturn(emb=style, time_emb=time_emb, style=style)

preprocess.py

@@ -0,0 +1,193 @@
+import os, random, math, copy
+import pandas as pd
+import numpy as np
+import pickle as pkl
+import logging, sys
+from torch.utils.data import DataLoader,Dataset
+import multiprocessing as mp
+import json
+import matplotlib.pyplot as plt
+
+
+def MakeDir(dirpath):
+    if not os.path.exists(dirpath):
+        os.makedirs(dirpath)
+
+        
+def load_egobody(data_dir, seq_len, sample_rate=1, train=1):
+        data_dir_train = data_dir + 'train/'
+        data_dir_test = data_dir + 'test/'
+        if train == 0:
+            data_dirs = [data_dir_test] # test
+        elif train == 1:
+            data_dirs = [data_dir_train] # train
+        elif train == 2:
+            data_dirs = [data_dir_train, data_dir_test] # train + test
+        
+        hand_head = []
+        for data_dir in data_dirs:
+            file_paths = sorted(os.listdir(data_dir))
+            pose_xyz_file_paths = []
+            head_file_paths = []
+            for path in file_paths:
+                path_split = path.split('_')
+                data_type = path_split[-1][:-4]
+                if(data_type == 'xyz'):
+                    pose_xyz_file_paths.append(path)
+                if(data_type == 'head'):
+                    head_file_paths.append(path)    
+                
+            file_num = len(pose_xyz_file_paths)        
+            for i in range(file_num):
+                pose_data = np.load(data_dir + pose_xyz_file_paths[i])
+                head_data = np.load(data_dir + head_file_paths[i])
+                num_frames = pose_data.shape[0]
+                if num_frames < seq_len:
+                    continue
+                
+                head_pos = pose_data[:, 15*3:16*3]
+                left_hand_pos = pose_data[:, 20*3:21*3]
+                right_hand_pos = pose_data[:, 21*3:22*3]
+                head_ori = head_data
+                left_hand_pos -= head_pos # convert hand positions to head coordinate system
+                right_hand_pos -= head_pos
+                hand_head_data = left_hand_pos
+                hand_head_data = np.concatenate((hand_head_data, right_hand_pos), axis=1)
+                hand_head_data = np.concatenate((hand_head_data, head_ori), axis=1)
+                            
+                fs = np.arange(0, num_frames - seq_len + 1)
+                fs_sel = fs
+                for i in np.arange(seq_len - 1):
+                    fs_sel = np.vstack((fs_sel, fs + i + 1))
+                fs_sel = fs_sel.transpose()
+                seq_sel = hand_head_data[fs_sel, :]
+                seq_sel = seq_sel[0::sample_rate, :, :]                    
+                if len(hand_head) == 0:
+                    hand_head = seq_sel
+                else:
+                    hand_head = np.concatenate((hand_head, seq_sel), axis=0)
+        
+        hand_head = np.transpose(hand_head, (0, 2, 1))
+        return hand_head
+
+        
+def load_adt(data_dir, seq_len, sample_rate=1, train=1):
+        data_dir_train = data_dir + 'train/'
+        data_dir_test = data_dir + 'test/'
+        if train == 0:
+            data_dirs = [data_dir_test] # test
+        elif train == 1:
+            data_dirs = [data_dir_train] # train
+        elif train == 2:
+            data_dirs = [data_dir_train, data_dir_test] # train + test
+            
+        hand_head = []
+        for data_dir in data_dirs:
+            file_paths = sorted(os.listdir(data_dir))
+            pose_xyz_file_paths = []
+            head_file_paths = []
+            for path in file_paths:
+                path_split = path.split('_')
+                data_type = path_split[-1][:-4]
+                if(data_type == 'xyz'):
+                    pose_xyz_file_paths.append(path)
+                if(data_type == 'head'):
+                    head_file_paths.append(path)    
+                
+            file_num = len(pose_xyz_file_paths)        
+            for i in range(file_num):
+                pose_data = np.load(data_dir + pose_xyz_file_paths[i])
+                head_data = np.load(data_dir + head_file_paths[i])
+                num_frames = pose_data.shape[0]
+                if num_frames < seq_len:
+                    continue
+                
+                head_pos = pose_data[:, 4*3:5*3]
+                left_hand_pos = pose_data[:, 8*3:9*3]
+                right_hand_pos = pose_data[:, 12*3:13*3]
+                head_ori = head_data
+                left_hand_pos -= head_pos # convert hand positions to head coordinate system
+                right_hand_pos -= head_pos
+                hand_head_data = left_hand_pos
+                hand_head_data = np.concatenate((hand_head_data, right_hand_pos), axis=1)
+                hand_head_data = np.concatenate((hand_head_data, head_ori), axis=1)
+                            
+                fs = np.arange(0, num_frames - seq_len + 1)
+                fs_sel = fs
+                for i in np.arange(seq_len - 1):
+                    fs_sel = np.vstack((fs_sel, fs + i + 1))
+                fs_sel = fs_sel.transpose()
+                seq_sel = hand_head_data[fs_sel, :]
+                seq_sel = seq_sel[0::sample_rate, :, :]                    
+                if len(hand_head) == 0:
+                    hand_head = seq_sel
+                else:
+                    hand_head = np.concatenate((hand_head, seq_sel), axis=0)            
+        
+        hand_head = np.transpose(hand_head, (0, 2, 1))
+        return hand_head
+       
+        
+def load_gimo(data_dir, seq_len, sample_rate=1, train=1):
+        data_dir_train = data_dir + 'train/'
+        data_dir_test = data_dir + 'test/'
+        if train == 0:
+            data_dirs = [data_dir_test] # test
+        elif train == 1:
+            data_dirs = [data_dir_train] # train
+        elif train == 2:
+            data_dirs = [data_dir_train, data_dir_test] # train + test
+            
+        hand_head = []
+        for data_dir in data_dirs:
+            file_paths = sorted(os.listdir(data_dir))
+            pose_xyz_file_paths = []
+            head_file_paths = []
+            for path in file_paths:
+                path_split = path.split('_')
+                data_type = path_split[-1][:-4]
+                if(data_type == 'xyz'):
+                    pose_xyz_file_paths.append(path)
+                if(data_type == 'head'):
+                    head_file_paths.append(path)    
+            
+            file_num = len(pose_xyz_file_paths)        
+            for i in range(file_num):
+                pose_data = np.load(data_dir + pose_xyz_file_paths[i])
+                head_data = np.load(data_dir + head_file_paths[i])
+                num_frames = pose_data.shape[0]
+                if num_frames < seq_len:
+                    continue
+                
+                head_pos = pose_data[:, 15*3:16*3]
+                left_hand_pos = pose_data[:, 20*3:21*3]
+                right_hand_pos = pose_data[:, 21*3:22*3]
+                head_ori = head_data
+                left_hand_pos -= head_pos # convert hand positions to head coordinate system
+                right_hand_pos -= head_pos
+                hand_head_data = left_hand_pos
+                hand_head_data = np.concatenate((hand_head_data, right_hand_pos), axis=1)
+                hand_head_data = np.concatenate((hand_head_data, head_ori), axis=1)
+                            
+                fs = np.arange(0, num_frames - seq_len + 1)
+                fs_sel = fs
+                for i in np.arange(seq_len - 1):
+                    fs_sel = np.vstack((fs_sel, fs + i + 1))
+                fs_sel = fs_sel.transpose()
+                seq_sel = hand_head_data[fs_sel, :]
+                seq_sel = seq_sel[0::sample_rate, :, :]                    
+                if len(hand_head) == 0:
+                    hand_head = seq_sel
+                else:
+                    hand_head = np.concatenate((hand_head, seq_sel), axis=0)            
+        
+        hand_head = np.transpose(hand_head, (0, 2, 1))
+        return hand_head
+
+        
+if __name__ == "__main__":
+    data_dir = "/scratch/hu/pose_forecast/egobody_pose2gaze/"
+    seq_len = 40
+    
+    test_data = load_egobody(data_dir, seq_len, sample_rate=10, train=0)
+    print("\ndataset size: {}".format(test_data.shape))

train.sh

@@ -0,0 +1,3 @@
+#python main.py --gpus 7 --mode 'train' --model_name 'haheae';
+
+python main.py --gpus 7 --mode 'eval' --model_name 'haheae';