113 KiB
113 KiB
In [1]:
import numpy as np import keras import sys import os, glob from keras.models import Model import tensorflow as tf from keras.utils import Sequence from keras.optimizers import Adam from keras.callbacks import ModelCheckpoint, LearningRateScheduler #from IPython.display import clear_output import tqdm import math sys.path.append('../src') from util import get_model_by_name from sal_imp_utilities import * from cb import InteractivePlot config = tf.ConfigProto() config.gpu_options.allow_growth = True sess = tf.Session(config=config) %load_ext autoreload %autoreload 2
Using TensorFlow backend.
/netpool/homes/wangyo/.conda/envs/tf-cuda9/lib/python3.7/site-packages/tensorflow/python/framework/dtypes.py:516: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
_np_qint8 = np.dtype([("qint8", np.int8, 1)])
/netpool/homes/wangyo/.conda/envs/tf-cuda9/lib/python3.7/site-packages/tensorflow/python/framework/dtypes.py:517: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
_np_quint8 = np.dtype([("quint8", np.uint8, 1)])
/netpool/homes/wangyo/.conda/envs/tf-cuda9/lib/python3.7/site-packages/tensorflow/python/framework/dtypes.py:518: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
_np_qint16 = np.dtype([("qint16", np.int16, 1)])
/netpool/homes/wangyo/.conda/envs/tf-cuda9/lib/python3.7/site-packages/tensorflow/python/framework/dtypes.py:519: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
_np_quint16 = np.dtype([("quint16", np.uint16, 1)])
/netpool/homes/wangyo/.conda/envs/tf-cuda9/lib/python3.7/site-packages/tensorflow/python/framework/dtypes.py:520: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
_np_qint32 = np.dtype([("qint32", np.int32, 1)])
/netpool/homes/wangyo/.conda/envs/tf-cuda9/lib/python3.7/site-packages/tensorflow/python/framework/dtypes.py:525: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
np_resource = np.dtype([("resource", np.ubyte, 1)])
/netpool/homes/wangyo/.conda/envs/tf-cuda9/lib/python3.7/site-packages/tensorboard/compat/tensorflow_stub/dtypes.py:541: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
_np_qint8 = np.dtype([("qint8", np.int8, 1)])
/netpool/homes/wangyo/.conda/envs/tf-cuda9/lib/python3.7/site-packages/tensorboard/compat/tensorflow_stub/dtypes.py:542: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
_np_quint8 = np.dtype([("quint8", np.uint8, 1)])
/netpool/homes/wangyo/.conda/envs/tf-cuda9/lib/python3.7/site-packages/tensorboard/compat/tensorflow_stub/dtypes.py:543: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
_np_qint16 = np.dtype([("qint16", np.int16, 1)])
/netpool/homes/wangyo/.conda/envs/tf-cuda9/lib/python3.7/site-packages/tensorboard/compat/tensorflow_stub/dtypes.py:544: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
_np_quint16 = np.dtype([("quint16", np.uint16, 1)])
/netpool/homes/wangyo/.conda/envs/tf-cuda9/lib/python3.7/site-packages/tensorboard/compat/tensorflow_stub/dtypes.py:545: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
_np_qint32 = np.dtype([("qint32", np.int32, 1)])
/netpool/homes/wangyo/.conda/envs/tf-cuda9/lib/python3.7/site-packages/tensorboard/compat/tensorflow_stub/dtypes.py:550: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
np_resource = np.dtype([("resource", np.ubyte, 1)])
In [2]:
%%bash nvidia-smi
Mon May 9 14:23:41 2022
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 470.103.01 Driver Version: 470.103.01 CUDA Version: 11.4 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 NVIDIA GeForce ... Off | 00000000:29:00.0 On | N/A |
| 30% 41C P2 38W / 184W | 975MiB / 7959MiB | 5% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| 0 N/A N/A 1377 G /usr/lib/xorg/Xorg 95MiB |
| 0 N/A N/A 2243 G /usr/lib/xorg/Xorg 432MiB |
| 0 N/A N/A 2378 G /usr/bin/gnome-shell 72MiB |
| 0 N/A N/A 3521 G ...AAAAAAAAA= --shared-files 16MiB |
| 0 N/A N/A 5912 G ...RendererForSitePerProcess 103MiB |
| 0 N/A N/A 18975 G ...sktop/bin/mendeleydesktop 42MiB |
| 0 N/A N/A 168057 G ...178058406726361824,131072 78MiB |
| 0 N/A N/A 178813 C .../envs/tf-cuda9/bin/python 87MiB |
+-----------------------------------------------------------------------------+
In [3]:
os.environ["CUDA_VISIBLE_DEVICES"] = '0' os.environ["CUDA_VISIBLE_DEVICES"]
Out[3]:
'0'
ENV¶
In [4]:
data_path = '/your/path/VisRecall/' # 0: T, 1: FE, 2: F, 3: RV, 4: U TYPE = 0 split = 0
Load data¶
In [5]:
TYPE_Q = ['T','FE','F','RV','U'] bp_imp = data_path + 'merged/src/' training_set = np.load(data_path + 'training_data/%s-question/train_split%d.npy'%(TYPE_Q[TYPE], split),allow_pickle=True) val_set = np.load(data_path + 'training_data/%s-question/val_split%d.npy'%(TYPE_Q[TYPE], split),allow_pickle=True)
In [6]:
train_filename = [] train_label = [] train_mean = [] train_type = [] for data in training_set: one_hot = [0,0,0,0,0,0] one_hot[data['vistype']] = 1 train_filename.append(bp_imp+data['name']) train_label.append(one_hot) train_mean.append(data['norm_mean_acc_withD']) train_type.append(data['norm_%d_withD'%(TYPE)])
In [7]:
val_filename = [] val_label = [] val_mean = [] val_type = [] for data in val_set: one_hot = [0,0,0,0,0,0] one_hot[data['vistype']] = 1 val_filename.append(bp_imp+data['name']) val_label.append(one_hot) val_mean.append(data['norm_mean_acc_withD']) val_type.append(data['norm_%d_withD'%(TYPE)]) #val_filename
Model and training params¶
In [8]:
# FILL THESE IN: set training parameters ckpt_savedir = "ckpt" load_weights = False weightspath = "" batch_size = 4 init_lr = 0.002 lr_reduce_by = .1 reduce_at_epoch = 3 n_epochs = 15 opt = Adam(lr=init_lr) model_name = "RecallNet_xception" model_inp_size = (240, 320)
In [9]:
input_shape = model_inp_size + (3,)
In [10]:
# get model model_params = { 'input_shape': input_shape, #'n_outs': len(losses), 'n_outs': 2 } model_func, mode = get_model_by_name(model_name) #assert mode == "simple", "%s is a multi-duration model! Please use the multi-duration notebook to train." % model_name model = model_func(**model_params) if load_weights: model.load_weights(weightspath) print("load")
WARNING:tensorflow:From /netpool/homes/wangyo/.conda/envs/tf-cuda9/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py:4070: The name tf.nn.max_pool is deprecated. Please use tf.nn.max_pool2d instead.
xception output shapes: (?, 30, 40, 2048)
Model: "model_1"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) (None, 240, 320, 3) 0
__________________________________________________________________________________________________
block1_conv1 (Conv2D) (None, 119, 159, 32) 864 input_1[0][0]
__________________________________________________________________________________________________
block1_conv1_bn (BatchNormaliza (None, 119, 159, 32) 128 block1_conv1[0][0]
__________________________________________________________________________________________________
block1_conv1_act (Activation) (None, 119, 159, 32) 0 block1_conv1_bn[0][0]
__________________________________________________________________________________________________
block1_conv2 (Conv2D) (None, 117, 157, 64) 18432 block1_conv1_act[0][0]
__________________________________________________________________________________________________
block1_conv2_bn (BatchNormaliza (None, 117, 157, 64) 256 block1_conv2[0][0]
__________________________________________________________________________________________________
block1_conv2_act (Activation) (None, 117, 157, 64) 0 block1_conv2_bn[0][0]
__________________________________________________________________________________________________
block2_sepconv1 (SeparableConv2 (None, 117, 157, 128 8768 block1_conv2_act[0][0]
__________________________________________________________________________________________________
block2_sepconv1_bn (BatchNormal (None, 117, 157, 128 512 block2_sepconv1[0][0]
__________________________________________________________________________________________________
block2_sepconv2_act (Activation (None, 117, 157, 128 0 block2_sepconv1_bn[0][0]
__________________________________________________________________________________________________
block2_sepconv2 (SeparableConv2 (None, 117, 157, 128 17536 block2_sepconv2_act[0][0]
__________________________________________________________________________________________________
block2_sepconv2_bn (BatchNormal (None, 117, 157, 128 512 block2_sepconv2[0][0]
__________________________________________________________________________________________________
conv2d_1 (Conv2D) (None, 59, 79, 128) 8192 block1_conv2_act[0][0]
__________________________________________________________________________________________________
block2_pool (MaxPooling2D) (None, 59, 79, 128) 0 block2_sepconv2_bn[0][0]
__________________________________________________________________________________________________
batch_normalization_1 (BatchNor (None, 59, 79, 128) 512 conv2d_1[0][0]
__________________________________________________________________________________________________
add_1 (Add) (None, 59, 79, 128) 0 block2_pool[0][0]
batch_normalization_1[0][0]
__________________________________________________________________________________________________
block3_sepconv1_act (Activation (None, 59, 79, 128) 0 add_1[0][0]
__________________________________________________________________________________________________
block3_sepconv1 (SeparableConv2 (None, 59, 79, 256) 33920 block3_sepconv1_act[0][0]
__________________________________________________________________________________________________
block3_sepconv1_bn (BatchNormal (None, 59, 79, 256) 1024 block3_sepconv1[0][0]
__________________________________________________________________________________________________
block3_sepconv2_act (Activation (None, 59, 79, 256) 0 block3_sepconv1_bn[0][0]
__________________________________________________________________________________________________
block3_sepconv2 (SeparableConv2 (None, 59, 79, 256) 67840 block3_sepconv2_act[0][0]
__________________________________________________________________________________________________
block3_sepconv2_bn (BatchNormal (None, 59, 79, 256) 1024 block3_sepconv2[0][0]
__________________________________________________________________________________________________
conv2d_2 (Conv2D) (None, 30, 40, 256) 32768 add_1[0][0]
__________________________________________________________________________________________________
block3_pool (MaxPooling2D) (None, 30, 40, 256) 0 block3_sepconv2_bn[0][0]
__________________________________________________________________________________________________
batch_normalization_2 (BatchNor (None, 30, 40, 256) 1024 conv2d_2[0][0]
__________________________________________________________________________________________________
add_2 (Add) (None, 30, 40, 256) 0 block3_pool[0][0]
batch_normalization_2[0][0]
__________________________________________________________________________________________________
block4_sepconv1_act (Activation (None, 30, 40, 256) 0 add_2[0][0]
__________________________________________________________________________________________________
block4_sepconv1 (SeparableConv2 (None, 30, 40, 728) 188672 block4_sepconv1_act[0][0]
__________________________________________________________________________________________________
block4_sepconv1_bn (BatchNormal (None, 30, 40, 728) 2912 block4_sepconv1[0][0]
__________________________________________________________________________________________________
block4_sepconv2_act (Activation (None, 30, 40, 728) 0 block4_sepconv1_bn[0][0]
__________________________________________________________________________________________________
block4_sepconv2 (SeparableConv2 (None, 30, 40, 728) 536536 block4_sepconv2_act[0][0]
__________________________________________________________________________________________________
block4_sepconv2_bn (BatchNormal (None, 30, 40, 728) 2912 block4_sepconv2[0][0]
__________________________________________________________________________________________________
conv2d_3 (Conv2D) (None, 30, 40, 728) 186368 add_2[0][0]
__________________________________________________________________________________________________
block4_pool (MaxPooling2D) (None, 30, 40, 728) 0 block4_sepconv2_bn[0][0]
__________________________________________________________________________________________________
batch_normalization_3 (BatchNor (None, 30, 40, 728) 2912 conv2d_3[0][0]
__________________________________________________________________________________________________
add_3 (Add) (None, 30, 40, 728) 0 block4_pool[0][0]
batch_normalization_3[0][0]
__________________________________________________________________________________________________
block5_sepconv1_act (Activation (None, 30, 40, 728) 0 add_3[0][0]
__________________________________________________________________________________________________
block5_sepconv1 (SeparableConv2 (None, 30, 40, 728) 536536 block5_sepconv1_act[0][0]
__________________________________________________________________________________________________
block5_sepconv1_bn (BatchNormal (None, 30, 40, 728) 2912 block5_sepconv1[0][0]
__________________________________________________________________________________________________
block5_sepconv2_act (Activation (None, 30, 40, 728) 0 block5_sepconv1_bn[0][0]
__________________________________________________________________________________________________
block5_sepconv2 (SeparableConv2 (None, 30, 40, 728) 536536 block5_sepconv2_act[0][0]
__________________________________________________________________________________________________
block5_sepconv2_bn (BatchNormal (None, 30, 40, 728) 2912 block5_sepconv2[0][0]
__________________________________________________________________________________________________
block5_sepconv3_act (Activation (None, 30, 40, 728) 0 block5_sepconv2_bn[0][0]
__________________________________________________________________________________________________
block5_sepconv3 (SeparableConv2 (None, 30, 40, 728) 536536 block5_sepconv3_act[0][0]
__________________________________________________________________________________________________
block5_sepconv3_bn (BatchNormal (None, 30, 40, 728) 2912 block5_sepconv3[0][0]
__________________________________________________________________________________________________
add_4 (Add) (None, 30, 40, 728) 0 block5_sepconv3_bn[0][0]
add_3[0][0]
__________________________________________________________________________________________________
block6_sepconv1_act (Activation (None, 30, 40, 728) 0 add_4[0][0]
__________________________________________________________________________________________________
block6_sepconv1 (SeparableConv2 (None, 30, 40, 728) 536536 block6_sepconv1_act[0][0]
__________________________________________________________________________________________________
block6_sepconv1_bn (BatchNormal (None, 30, 40, 728) 2912 block6_sepconv1[0][0]
__________________________________________________________________________________________________
block6_sepconv2_act (Activation (None, 30, 40, 728) 0 block6_sepconv1_bn[0][0]
__________________________________________________________________________________________________
block6_sepconv2 (SeparableConv2 (None, 30, 40, 728) 536536 block6_sepconv2_act[0][0]
__________________________________________________________________________________________________
block6_sepconv2_bn (BatchNormal (None, 30, 40, 728) 2912 block6_sepconv2[0][0]
__________________________________________________________________________________________________
block6_sepconv3_act (Activation (None, 30, 40, 728) 0 block6_sepconv2_bn[0][0]
__________________________________________________________________________________________________
block6_sepconv3 (SeparableConv2 (None, 30, 40, 728) 536536 block6_sepconv3_act[0][0]
__________________________________________________________________________________________________
block6_sepconv3_bn (BatchNormal (None, 30, 40, 728) 2912 block6_sepconv3[0][0]
__________________________________________________________________________________________________
add_5 (Add) (None, 30, 40, 728) 0 block6_sepconv3_bn[0][0]
add_4[0][0]
__________________________________________________________________________________________________
block7_sepconv1_act (Activation (None, 30, 40, 728) 0 add_5[0][0]
__________________________________________________________________________________________________
block7_sepconv1 (SeparableConv2 (None, 30, 40, 728) 536536 block7_sepconv1_act[0][0]
__________________________________________________________________________________________________
block7_sepconv1_bn (BatchNormal (None, 30, 40, 728) 2912 block7_sepconv1[0][0]
__________________________________________________________________________________________________
block7_sepconv2_act (Activation (None, 30, 40, 728) 0 block7_sepconv1_bn[0][0]
__________________________________________________________________________________________________
block7_sepconv2 (SeparableConv2 (None, 30, 40, 728) 536536 block7_sepconv2_act[0][0]
__________________________________________________________________________________________________
block7_sepconv2_bn (BatchNormal (None, 30, 40, 728) 2912 block7_sepconv2[0][0]
__________________________________________________________________________________________________
block7_sepconv3_act (Activation (None, 30, 40, 728) 0 block7_sepconv2_bn[0][0]
__________________________________________________________________________________________________
block7_sepconv3 (SeparableConv2 (None, 30, 40, 728) 536536 block7_sepconv3_act[0][0]
__________________________________________________________________________________________________
block7_sepconv3_bn (BatchNormal (None, 30, 40, 728) 2912 block7_sepconv3[0][0]
__________________________________________________________________________________________________
add_6 (Add) (None, 30, 40, 728) 0 block7_sepconv3_bn[0][0]
add_5[0][0]
__________________________________________________________________________________________________
block8_sepconv1_act (Activation (None, 30, 40, 728) 0 add_6[0][0]
__________________________________________________________________________________________________
block8_sepconv1 (SeparableConv2 (None, 30, 40, 728) 536536 block8_sepconv1_act[0][0]
__________________________________________________________________________________________________
block8_sepconv1_bn (BatchNormal (None, 30, 40, 728) 2912 block8_sepconv1[0][0]
__________________________________________________________________________________________________
block8_sepconv2_act (Activation (None, 30, 40, 728) 0 block8_sepconv1_bn[0][0]
__________________________________________________________________________________________________
block8_sepconv2 (SeparableConv2 (None, 30, 40, 728) 536536 block8_sepconv2_act[0][0]
__________________________________________________________________________________________________
block8_sepconv2_bn (BatchNormal (None, 30, 40, 728) 2912 block8_sepconv2[0][0]
__________________________________________________________________________________________________
block8_sepconv3_act (Activation (None, 30, 40, 728) 0 block8_sepconv2_bn[0][0]
__________________________________________________________________________________________________
block8_sepconv3 (SeparableConv2 (None, 30, 40, 728) 536536 block8_sepconv3_act[0][0]
__________________________________________________________________________________________________
block8_sepconv3_bn (BatchNormal (None, 30, 40, 728) 2912 block8_sepconv3[0][0]
__________________________________________________________________________________________________
add_7 (Add) (None, 30, 40, 728) 0 block8_sepconv3_bn[0][0]
add_6[0][0]
__________________________________________________________________________________________________
block9_sepconv1_act (Activation (None, 30, 40, 728) 0 add_7[0][0]
__________________________________________________________________________________________________
block9_sepconv1 (SeparableConv2 (None, 30, 40, 728) 536536 block9_sepconv1_act[0][0]
__________________________________________________________________________________________________
block9_sepconv1_bn (BatchNormal (None, 30, 40, 728) 2912 block9_sepconv1[0][0]
__________________________________________________________________________________________________
block9_sepconv2_act (Activation (None, 30, 40, 728) 0 block9_sepconv1_bn[0][0]
__________________________________________________________________________________________________
block9_sepconv2 (SeparableConv2 (None, 30, 40, 728) 536536 block9_sepconv2_act[0][0]
__________________________________________________________________________________________________
block9_sepconv2_bn (BatchNormal (None, 30, 40, 728) 2912 block9_sepconv2[0][0]
__________________________________________________________________________________________________
block9_sepconv3_act (Activation (None, 30, 40, 728) 0 block9_sepconv2_bn[0][0]
__________________________________________________________________________________________________
block9_sepconv3 (SeparableConv2 (None, 30, 40, 728) 536536 block9_sepconv3_act[0][0]
__________________________________________________________________________________________________
block9_sepconv3_bn (BatchNormal (None, 30, 40, 728) 2912 block9_sepconv3[0][0]
__________________________________________________________________________________________________
add_8 (Add) (None, 30, 40, 728) 0 block9_sepconv3_bn[0][0]
add_7[0][0]
__________________________________________________________________________________________________
block10_sepconv1_act (Activatio (None, 30, 40, 728) 0 add_8[0][0]
__________________________________________________________________________________________________
block10_sepconv1 (SeparableConv (None, 30, 40, 728) 536536 block10_sepconv1_act[0][0]
__________________________________________________________________________________________________
block10_sepconv1_bn (BatchNorma (None, 30, 40, 728) 2912 block10_sepconv1[0][0]
__________________________________________________________________________________________________
block10_sepconv2_act (Activatio (None, 30, 40, 728) 0 block10_sepconv1_bn[0][0]
__________________________________________________________________________________________________
block10_sepconv2 (SeparableConv (None, 30, 40, 728) 536536 block10_sepconv2_act[0][0]
__________________________________________________________________________________________________
block10_sepconv2_bn (BatchNorma (None, 30, 40, 728) 2912 block10_sepconv2[0][0]
__________________________________________________________________________________________________
block10_sepconv3_act (Activatio (None, 30, 40, 728) 0 block10_sepconv2_bn[0][0]
__________________________________________________________________________________________________
block10_sepconv3 (SeparableConv (None, 30, 40, 728) 536536 block10_sepconv3_act[0][0]
__________________________________________________________________________________________________
block10_sepconv3_bn (BatchNorma (None, 30, 40, 728) 2912 block10_sepconv3[0][0]
__________________________________________________________________________________________________
add_9 (Add) (None, 30, 40, 728) 0 block10_sepconv3_bn[0][0]
add_8[0][0]
__________________________________________________________________________________________________
block11_sepconv1_act (Activatio (None, 30, 40, 728) 0 add_9[0][0]
__________________________________________________________________________________________________
block11_sepconv1 (SeparableConv (None, 30, 40, 728) 536536 block11_sepconv1_act[0][0]
__________________________________________________________________________________________________
block11_sepconv1_bn (BatchNorma (None, 30, 40, 728) 2912 block11_sepconv1[0][0]
__________________________________________________________________________________________________
block11_sepconv2_act (Activatio (None, 30, 40, 728) 0 block11_sepconv1_bn[0][0]
__________________________________________________________________________________________________
block11_sepconv2 (SeparableConv (None, 30, 40, 728) 536536 block11_sepconv2_act[0][0]
__________________________________________________________________________________________________
block11_sepconv2_bn (BatchNorma (None, 30, 40, 728) 2912 block11_sepconv2[0][0]
__________________________________________________________________________________________________
block11_sepconv3_act (Activatio (None, 30, 40, 728) 0 block11_sepconv2_bn[0][0]
__________________________________________________________________________________________________
block11_sepconv3 (SeparableConv (None, 30, 40, 728) 536536 block11_sepconv3_act[0][0]
__________________________________________________________________________________________________
block11_sepconv3_bn (BatchNorma (None, 30, 40, 728) 2912 block11_sepconv3[0][0]
__________________________________________________________________________________________________
add_10 (Add) (None, 30, 40, 728) 0 block11_sepconv3_bn[0][0]
add_9[0][0]
__________________________________________________________________________________________________
block12_sepconv1_act (Activatio (None, 30, 40, 728) 0 add_10[0][0]
__________________________________________________________________________________________________
block12_sepconv1 (SeparableConv (None, 30, 40, 728) 536536 block12_sepconv1_act[0][0]
__________________________________________________________________________________________________
block12_sepconv1_bn (BatchNorma (None, 30, 40, 728) 2912 block12_sepconv1[0][0]
__________________________________________________________________________________________________
block12_sepconv2_act (Activatio (None, 30, 40, 728) 0 block12_sepconv1_bn[0][0]
__________________________________________________________________________________________________
block12_sepconv2 (SeparableConv (None, 30, 40, 728) 536536 block12_sepconv2_act[0][0]
__________________________________________________________________________________________________
block12_sepconv2_bn (BatchNorma (None, 30, 40, 728) 2912 block12_sepconv2[0][0]
__________________________________________________________________________________________________
block12_sepconv3_act (Activatio (None, 30, 40, 728) 0 block12_sepconv2_bn[0][0]
__________________________________________________________________________________________________
block12_sepconv3 (SeparableConv (None, 30, 40, 728) 536536 block12_sepconv3_act[0][0]
__________________________________________________________________________________________________
block12_sepconv3_bn (BatchNorma (None, 30, 40, 728) 2912 block12_sepconv3[0][0]
__________________________________________________________________________________________________
add_11 (Add) (None, 30, 40, 728) 0 block12_sepconv3_bn[0][0]
add_10[0][0]
__________________________________________________________________________________________________
block13_sepconv1_act (Activatio (None, 30, 40, 728) 0 add_11[0][0]
__________________________________________________________________________________________________
block13_sepconv1 (SeparableConv (None, 30, 40, 728) 536536 block13_sepconv1_act[0][0]
__________________________________________________________________________________________________
block13_sepconv1_bn (BatchNorma (None, 30, 40, 728) 2912 block13_sepconv1[0][0]
__________________________________________________________________________________________________
block13_sepconv2_act (Activatio (None, 30, 40, 728) 0 block13_sepconv1_bn[0][0]
__________________________________________________________________________________________________
block13_sepconv2 (SeparableConv (None, 30, 40, 1024) 752024 block13_sepconv2_act[0][0]
__________________________________________________________________________________________________
block13_sepconv2_bn (BatchNorma (None, 30, 40, 1024) 4096 block13_sepconv2[0][0]
__________________________________________________________________________________________________
conv2d_4 (Conv2D) (None, 30, 40, 1024) 745472 add_11[0][0]
__________________________________________________________________________________________________
block13_pool (MaxPooling2D) (None, 30, 40, 1024) 0 block13_sepconv2_bn[0][0]
__________________________________________________________________________________________________
batch_normalization_4 (BatchNor (None, 30, 40, 1024) 4096 conv2d_4[0][0]
__________________________________________________________________________________________________
add_12 (Add) (None, 30, 40, 1024) 0 block13_pool[0][0]
batch_normalization_4[0][0]
__________________________________________________________________________________________________
block14_sepconv1 (SeparableConv (None, 30, 40, 1536) 1582080 add_12[0][0]
__________________________________________________________________________________________________
block14_sepconv1_bn (BatchNorma (None, 30, 40, 1536) 6144 block14_sepconv1[0][0]
__________________________________________________________________________________________________
block14_sepconv1_act (Activatio (None, 30, 40, 1536) 0 block14_sepconv1_bn[0][0]
__________________________________________________________________________________________________
block14_sepconv2 (SeparableConv (None, 30, 40, 2048) 3159552 block14_sepconv1_act[0][0]
__________________________________________________________________________________________________
block14_sepconv2_bn (BatchNorma (None, 30, 40, 2048) 8192 block14_sepconv2[0][0]
__________________________________________________________________________________________________
block14_sepconv2_act (Activatio (None, 30, 40, 2048) 0 block14_sepconv2_bn[0][0]
__________________________________________________________________________________________________
global_conv (Conv2D) (None, 10, 14, 256) 4718592 block14_sepconv2_act[0][0]
__________________________________________________________________________________________________
global_BN (BatchNormalization) (None, 10, 14, 256) 1024 global_conv[0][0]
__________________________________________________________________________________________________
activation_1 (Activation) (None, 10, 14, 256) 0 global_BN[0][0]
__________________________________________________________________________________________________
dropout_1 (Dropout) (None, 10, 14, 256) 0 activation_1[0][0]
__________________________________________________________________________________________________
global_average_pooling2d_1 (Glo (None, 256) 0 dropout_1[0][0]
__________________________________________________________________________________________________
global_dense (Dense) (None, 256) 65792 global_average_pooling2d_1[0][0]
__________________________________________________________________________________________________
dropout_2 (Dropout) (None, 256) 0 global_dense[0][0]
__________________________________________________________________________________________________
out_mean_acc (Dense) (None, 1) 257 dropout_2[0][0]
__________________________________________________________________________________________________
out_type0_acc (Dense) (None, 1) 257 dropout_2[0][0]
==================================================================================================
Total params: 25,647,402
Trainable params: 25,592,362
Non-trainable params: 55,040
__________________________________________________________________________________________________
In [11]:
# set up data generation and checkpoints if not os.path.exists(ckpt_savedir): os.makedirs(ckpt_savedir) # Generators gen_train = RecallNet_Generator(train_filename, train_label, train_mean, train_type, batch_size = 4) gen_val = RecallNet_Generator(val_filename, val_label, val_mean, val_type, 1) # Callbacks # where to save checkpoints #filepath = os.path.join(ckpt_savedir, dataset_sal + "_" + l_str + '_ep{epoch:02d}_valloss{val_loss:.4f}.hdf5') filepath = os.path.join(ckpt_savedir, model_name, model_name+'_ep{epoch:02d}_valloss{val_loss:.4f}.hdf5') print("Checkpoints will be saved with format %s" % filepath) cb_chk = ModelCheckpoint(filepath, monitor='val_loss', verbose=1, save_weights_only=False, period=1) cb_plot = InteractivePlot() def step_decay(epoch): lrate = init_lr * math.pow(lr_reduce_by, math.floor((1+epoch)/reduce_at_epoch)) if epoch%reduce_at_epoch: print('Reducing lr. New lr is:', lrate) return lrate cb_sched = LearningRateScheduler(step_decay) cbs = [cb_chk, cb_sched, cb_plot]
Checkpoints will be saved with format ckpt/RecallNet_xception/RecallNet_xception_ep{epoch:02d}_valloss{val_loss:.4f}.hdf5
In [12]:
#test the generator batch_img, out= gen_train.__getitem__(1) print(out)
[array([0.74688797, 0.84647303, 0.93360996, 0.8340249 ]), array([0.7826087 , 0.95652174, 1. , 0.86956522])]
Train¶
In [14]:
#model.compile(optimizer=opt, loss={'out_mean_acc': 'mean_squared_error', 'out_type0_acc':'mean_squared_error','out_classif':'binary_crossentropy'}, loss_weights={'out_mean_acc': 3, 'out_type0_acc': 3, 'out_classif': 1}) model.compile(optimizer=opt, loss={'out_mean_acc': 'mean_squared_error', 'out_type0_acc':'mean_squared_error'}, loss_weights={'out_mean_acc': 1, 'out_type0_acc': 1}) print('Ready to train') model.fit_generator(gen_train, epochs=n_epochs, verbose=1, callbacks=cbs, validation_data=gen_val, max_queue_size=10, workers=5)
Out[14]:
<keras.callbacks.callbacks.History at 0x7ff9ed6f3950>
validation¶
In [77]:
data_path = '/your/path/VisRecall/' if True: W = "./ckpt/RecallNet/U-question/Fold4/RecallNet_xception_ep10_valloss0.0313.hdf5" model.load_weights(W) print('load')
load
In [78]:
from sklearn.metrics import mean_squared_error mean_list = [] type_list = [] num_t = 0.0 for i in range(len(val_filename)): img = preprocess_images([val_filename[i]], 240, 320) preds = model.predict(img) #if(np.argmax(preds[2][0]) == np.argmax(val_label[i])): # num_t += 1 mean_acc = preds[0][0] type_acc = preds[1][0] mean_list.append(mean_squared_error([val_mean[i]], mean_acc)) type_list.append(mean_squared_error([val_type[i]], type_acc)) #print(val_label[i]) #print('classificaion: ', num_t/len(val_filename)) print('avg, type: ', np.mean(mean_list),np.mean(type_list))
avg, type: 0.02553353417062307 0.02617591625785007
In [13]:
from sklearn.metrics import mean_squared_error
In [24]:
# 0: T, 1: FE, 2: F, 3: RV, 4: U TYPE = 4 final = [] for split in range(5): TYPE_Q = ['T','FE','F','RV','U'] bp_imp = data_path + 'merged/src/' training_set = np.load(data_path + 'training_data/%s-question/train_split%d.npy'%(TYPE_Q[TYPE], split),allow_pickle=True) val_set = np.load(data_path + 'training_data/%s-question/val_split%d.npy'%(TYPE_Q[TYPE], split),allow_pickle=True) train_filename = [] train_label = [] train_mean = [] train_type = [] for data in training_set: one_hot = [0,0,0,0,0,0] one_hot[data['vistype']] = 1 train_filename.append(bp_imp+data['name']) train_label.append(one_hot) train_mean.append(data['norm_mean_acc_withD']) train_type.append(data['norm_%d_withD'%(TYPE)]) val_filename = [] val_label = [] val_mean = [] val_type = [] for data in val_set: one_hot = [0,0,0,0,0,0] one_hot[data['vistype']] = 1 val_filename.append(bp_imp+data['name']) val_label.append(one_hot) val_mean.append(data['norm_mean_acc_withD']) val_type.append(data['norm_%d_withD'%(TYPE)]) mean_list = [] type_list = [] num_t = 0.0 if True: W = "./ckpt/RecallNet_xception/U-question/Fold"+str(split)+"/RecallNet_xception_ep10.hdf5" model.load_weights(W) print('load') for i in range(len(val_filename)): img = preprocess_images([val_filename[i]], 240, 320) preds = model.predict(img) #if(np.argmax(preds[2][0]) == np.argmax(val_label[i])): # num_t += 1 mean_acc = preds[0][0] type_acc = preds[1][0] mean_list.append(mean_squared_error([val_mean[i]], mean_acc)) type_list.append(mean_squared_error([val_type[i]], type_acc)) final.append((np.mean(mean_list),np.mean(type_list))) print(final)
load [(0.1154690275375791, 0.07620564377135046)] load [(0.1154690275375791, 0.07620564377135046), (0.10878741427577963, 0.2667425108749781)] load [(0.1154690275375791, 0.07620564377135046), (0.10878741427577963, 0.2667425108749781), (0.5932855723937102, 0.38129302771926993)] load [(0.1154690275375791, 0.07620564377135046), (0.10878741427577963, 0.2667425108749781), (0.5932855723937102, 0.38129302771926993), (0.015003968232081454, 0.06507362662601399)] load [(0.1154690275375791, 0.07620564377135046), (0.10878741427577963, 0.2667425108749781), (0.5932855723937102, 0.38129302771926993), (0.015003968232081454, 0.06507362662601399), (0.02553353417062307, 0.02617591625785007)]
In [25]:
np.mean(final ,axis=0),np.std(final ,axis=0)
Out[25]:
(array([0.1716159 , 0.16309815]), array([0.21483601, 0.13730111]))
Test¶
In [36]:
#W_T = "./ckpt/VMQA_UMSI/T-question/UMSI_weight1:1/Fold0/VMQA_UMSI_ep10_valloss0.6198.hdf5" W_T = "./ckpt/RecallNet_xception/T-question/Fold4/RecallNet_xception_ep10.hdf5" model.load_weights(W_T) image_path = "/your/path/to/src" list_img_targets = glob.glob(image_path + '/*.png') print(list_img_targets) images = preprocess_images(list_img_targets, model_inp_size[0], model_inp_size[1])
['/netpool/homes/wangyo/Projects/2020_luwei_VisQA/Dataset/application/src/bar4.png', '/netpool/homes/wangyo/Projects/2020_luwei_VisQA/Dataset/application/src/pie4.png', '/netpool/homes/wangyo/Projects/2020_luwei_VisQA/Dataset/application/src/bar1.png', '/netpool/homes/wangyo/Projects/2020_luwei_VisQA/Dataset/application/src/pie2.png', '/netpool/homes/wangyo/Projects/2020_luwei_VisQA/Dataset/application/src/bar3.png', '/netpool/homes/wangyo/Projects/2020_luwei_VisQA/Dataset/application/src/line1.png', '/netpool/homes/wangyo/Projects/2020_luwei_VisQA/Dataset/application/src/bar2.png', '/netpool/homes/wangyo/Projects/2020_luwei_VisQA/Dataset/application/src/pie1.png', '/netpool/homes/wangyo/Projects/2020_luwei_VisQA/Dataset/application/src/line2.png', '/netpool/homes/wangyo/Projects/2020_luwei_VisQA/Dataset/application/src/line4.png', '/netpool/homes/wangyo/Projects/2020_luwei_VisQA/Dataset/application/src/line3.png', '/netpool/homes/wangyo/Projects/2020_luwei_VisQA/Dataset/application/src/pie3.png']
In [37]:
preds_recallNet = model.predict(images)
In [38]:
print(preds_recallNet[0].shape, preds_recallNet)
(12, 1) [array([[0.6729137 ],
[0.62224036],
[0.6644782 ],
[0.63536835],
[0.6595377 ],
[0.6628296 ],
[0.600067 ],
[0.64756936],
[0.600191 ],
[0.6459451 ],
[0.6242438 ],
[0.60450727]], dtype=float32), array([[0.7803583 ],
[0.67262685],
[0.75816834],
[0.718869 ],
[0.7442306 ],
[0.76404536],
[0.8003189 ],
[0.7172563 ],
[0.7985541 ],
[0.81216127],
[0.6953638 ],
[0.6564585 ]], dtype=float32)]