add code

.gitignore

@@ -0,0 +1,2 @@
+.DS_STORE
+.pyc

Code/dataset_new.py

@@ -0,0 +1,18 @@
+import torch
+from torch.utils.data import Dataset
+import numpy as np
+
+class ImagesWithSaliency(Dataset):
+    def __init__(self, npy_path, dtype=None):
+        self.dtype = dtype
+        self.datas = np.load(npy_path, allow_pickle = True)
+
+    def __len__(self):
+        return len(self.datas)
+
+    def __getitem__(self, idx):
+        if self.dtype:
+            self.datas[idx][0] = self.datas[idx][0].type(self.dtype)
+            self.datas[idx][3] = self.datas[idx][3].type(self.dtype)
+
+        return self.datas[idx]

Code/env.py

@@ -0,0 +1,6 @@
+
+import os
+os.environ['TORCH_HOME'] = '/projects/wang/.cache/torch'
+os.environ['TRANSFORMERS_CACHE'] = '/projects/wang/.cache'
+
+my_variable = os.environ.get('TORCH_HOME')

Code/evaluation.py

@@ -0,0 +1,110 @@
+import torch
+from torch.utils.data import DataLoader
+from env import *
+
+import argparse
+from dataset_new import ImagesWithSaliency
+from torchvision.utils import save_image
+from transformers import SwinModel
+from pathlib import Path
+
+def evaluation(Model:str, ckpt: str, device, batch_size:int):
+    eps=1e-10
+
+    if Model == 'llama':
+        from model_llama import SalFormer
+        from transformers import LlamaModel
+        from tokenizer_llama import padding_fn
+        # llm = LlamaModel.from_pretrained("Enoch/llama-7b-hf", low_cpu_mem_usage=True)
+        llm = LlamaModel.from_pretrained("daryl149/Llama-2-7b-chat-hf", low_cpu_mem_usage=True)
+        neuron_n = 4096
+        print("llama loaded")
+    elif Model == 'bloom':
+        from model_llama import SalFormer
+        from transformers import BloomModel
+        from tokenizer_bloom import padding_fn
+        llm = BloomModel.from_pretrained("bigscience/bloom-3b")
+        neuron_n = 2560
+        print('BloomModel loaded')
+    elif Model == 'bert':
+        from model_swin import SalFormer
+        from transformers import BertModel
+        from tokenizer_bert import padding_fn
+        llm = BertModel.from_pretrained("bert-base-uncased")
+        print('BertModel loaded')
+    else:
+        print('model not available, possiblilities: llama, bloom, bert')
+        return
+
+    test_set = ImagesWithSaliency("data/test.npy")
+
+    Path('./eval_results').mkdir(parents=True, exist_ok=True)
+
+    # vit = ViTModel.from_pretrained("google/vit-base-patch16-224-in21k")
+    vit = SwinModel.from_pretrained("microsoft/swin-tiny-patch4-window7-224")
+    # vit = timm.create_model('xception41p.ra3_in1k', pretrained=True)
+
+    if Model == 'bert':
+        model = SalFormer(vit, llm).to(device)
+    else:
+        model = SalFormer(vit, llm, neuron_n = neuron_n).to(device)
+
+    checkpoint = torch.load(ckpt)
+    model.load_state_dict(checkpoint['model_state_dict'])
+    model.eval()
+
+
+    test_dataloader = DataLoader(test_set, batch_size=batch_size, shuffle=False, collate_fn=padding_fn, num_workers=8)
+    kl_loss = torch.nn.KLDivLoss(reduction="batchmean", log_target=True)
+
+    test_kl, test_cc, test_nss = 0,0,0 
+    for batch, (img, input_ids, fix, hm, name) in enumerate(test_dataloader):
+        img = img.to(device)
+        input_ids = input_ids.to(device)
+        fix = fix.to(device)
+        hm = hm.to(device)
+
+        y = model(img, input_ids)
+        
+        y_sum = y.view(y.shape[0], -1).sum(1, keepdim=True)
+        y_distribution = y / (y_sum[:, :, None, None] + eps)
+
+        hm_sum = hm.view(y.shape[0], -1).sum(1, keepdim=True)
+        hm_distribution = hm / (hm_sum[:, :, None, None] + eps)
+        hm_distribution = hm_distribution + eps
+        hm_distribution = hm_distribution / (1+eps)
+
+        if fix.sum() != 0:
+            normal_y = (y-y.mean())/y.std()
+            nss = torch.sum(normal_y*fix)/fix.sum()
+        else:
+            nss = torch.Tensor([0.0]).to(device)
+        kl = kl_loss(torch.log(y_distribution), torch.log(hm_distribution))
+
+        vy = y - torch.mean(y)
+        vhm = hm - torch.mean(hm)  
+
+        if (torch.sqrt(torch.sum(vy ** 2)) * torch.sqrt(torch.sum(vhm ** 2))) != 0:
+            cc = torch.sum(vy * vhm) / (torch.sqrt(torch.sum(vy ** 2)) * torch.sqrt(torch.sum(vhm ** 2)))
+        else: 
+            cc = torch.Tensor([0.0]).to(device)
+        
+        test_kl += kl.item()/len(test_dataloader)
+        test_cc += cc.item()/len(test_dataloader)
+        test_nss += nss.item()/len(test_dataloader)
+
+        for i in range(0, y.shape[0]):
+            save_image(y[i], f"./eval_results/{name[i]}")
+
+    print("kl:", test_kl, "cc", test_cc, "nss", test_nss)
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model", type=str, default='bert')
+    parser.add_argument("--device", type=str, default='cuda')
+    parser.add_argument("--batch_size", type=int, default=16)
+    parser.add_argument("--ckpt", type=str, default='./ckpt/model_bert_freeze_10kl_5cc_2nss.tar')
+    args = vars(parser.parse_args())
+
+    evaluation(Model = args['model'], device = args['device'], ckpt = args['ckpt'], batch_size = args['batch_size'])

Code/evaluation.sh

@@ -0,0 +1 @@
+python evaluation.py --model 'bert' --ckpt './VisSalFormer_weights.tar' --device 'cuda'

Code/model_swin.py

@@ -0,0 +1,116 @@
+import torch
+
+class SalFormer(torch.nn.Module):
+    def __init__(self, vision_encoder, bert):
+        """
+        In the constructor we instantiate four parameters and assign them as
+        member parameters.
+        """
+        super().__init__()
+
+        self.vit = vision_encoder
+        self.feature_dim = 768
+        self.bert = bert
+
+        self.vision_head = torch.nn.Sequential(
+            torch.nn.Linear(self.feature_dim, self.feature_dim),
+            torch.nn.GELU(),
+            torch.nn.Linear(self.feature_dim, self.feature_dim),
+            torch.nn.GELU()
+        )
+
+        self.text_head = torch.nn.Sequential(
+            torch.nn.Linear(self.feature_dim, self.feature_dim),
+            torch.nn.GELU(),
+            torch.nn.Linear(self.feature_dim, self.feature_dim),
+            torch.nn.GELU()
+        )
+
+        self.cross_attention = torch.nn.MultiheadAttention(self.feature_dim, 16, kdim=self.feature_dim, vdim=self.feature_dim, batch_first=True)
+        self.cross_attention1 = torch.nn.MultiheadAttention(self.feature_dim, 16, kdim=self.feature_dim, vdim=self.feature_dim, batch_first=True)
+
+        self.ln1 = torch.nn.LayerNorm(self.feature_dim)
+        self.ln2 = torch.nn.LayerNorm(self.feature_dim)
+
+        self.self_attetion = torch.nn.MultiheadAttention(self.feature_dim, 16, batch_first=True)
+
+        self.text_feature_query = torch.nn.Parameter(torch.randn(10, self.feature_dim).unsqueeze(0)/2)
+        self.img_positional_embedding = torch.nn.Parameter(torch.zeros(49, self.feature_dim))
+        self.text_positional_embedding = torch.nn.Parameter(torch.zeros(10, self.feature_dim))
+
+
+        self.dense1 = torch.nn.Linear(self.feature_dim, self.feature_dim)
+        self.relu1 = torch.nn.ReLU()
+
+        self.decoder = torch.nn.Sequential(
+            torch.nn.Conv2d(self.feature_dim, 512, 3),
+            torch.nn.BatchNorm2d(512),
+            torch.nn.ReLU(),
+            torch.nn.Dropout(p=0.1),
+            torch.nn.Conv2d(512, 512, 3),
+            torch.nn.BatchNorm2d(512),
+            torch.nn.ReLU(),
+            torch.nn.Dropout(p=0.1),
+            torch.nn.Upsample((16,16), mode='bilinear'),
+            torch.nn.Conv2d(512, 256, 3),
+            torch.nn.BatchNorm2d(256),
+            torch.nn.ReLU(),
+            torch.nn.Dropout(p=0.1),
+            torch.nn.Conv2d(256, 256, 3),
+            torch.nn.BatchNorm2d(256),
+            torch.nn.ReLU(),
+            torch.nn.Dropout(p=0.1),
+            torch.nn.Upsample((32,32), mode='bilinear'),
+            torch.nn.Conv2d(256, 128, 3),
+            torch.nn.BatchNorm2d(128),
+            torch.nn.ReLU(),
+            torch.nn.Dropout(p=0.1),
+            torch.nn.Conv2d(128, 128, 3),
+            torch.nn.BatchNorm2d(128),
+            torch.nn.ReLU(),
+            torch.nn.Dropout(p=0.1),
+            torch.nn.Upsample((130,130), mode='bilinear'),
+            torch.nn.Conv2d(128, 1, 3),
+            torch.nn.BatchNorm2d(1),
+            torch.nn.Sigmoid(),
+        )
+
+        self.vit.eval()
+        self.bert.eval()
+        self.train(True)
+
+    # def eval(self):
+    #     super().eval()
+    #     self.vit.eval()
+    #     self.bert.eval()
+
+    # def train(self, mode=True):
+    #     super().train(mode)
+    #     self.vit.train(mode)
+    #     self.bert.train(mode)
+
+    def forward(self, img, q_inputs):
+
+        img_features =  self.vit.forward(img, return_dict =True)["last_hidden_state"]
+        with torch.no_grad():
+            text_features =  self.bert(**q_inputs)["last_hidden_state"]
+            # text_features =  torch.unsqueeze(bert_output["last_hidden_state"][:,0,:], 1)
+        text_features = self.cross_attention.forward(self.text_feature_query.repeat([text_features.shape[0], 1, 1]), text_features, text_features, need_weights=False)[0]
+
+        fused_features = torch.concat((self.vision_head(img_features)+self.img_positional_embedding, self.text_head(text_features)+self.text_positional_embedding), 1)
+        att_fused_features = self.self_attetion.forward(fused_features, fused_features, fused_features, need_weights=False)[0]
+        fused_features = fused_features + att_fused_features
+        fused_features = self.ln1(fused_features)
+
+        features = self.cross_attention1.forward(img_features, fused_features, fused_features, need_weights=False)[0]
+        features = img_features + features
+        features = self.ln2(features)
+
+        features = self.dense1(features)
+        latent_features = self.relu1(features)
+
+        latent_features = latent_features.permute(0,2,1)
+        out = torch.reshape(latent_features, (features.shape[0], self.feature_dim, 7, 7))
+        out = self.decoder(out)
+
+        return out

Code/tokenizer_bert.py

@@ -0,0 +1,15 @@
+import torch
+from transformers import AutoTokenizer
+
+# tokenizer = AutoTokenizer.from_pretrained("roberta-base")
+tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
+print('bert-base-uncased tokenizer loaded')
+
+def padding_fn(data):
+    img, q, fix, hm, name = zip(*data)
+
+    input_ids = tokenizer(q, return_tensors="pt", padding=True)
+
+    return torch.stack(img), input_ids, torch.stack(fix), torch.stack(hm), name
+
+

README.md

@@ -12,13 +12,38 @@ $Root Directory
 │
 │─ README.md —— this file
 │
-|─ VisSalFormer —— Source code of the network to predict question-driven saliency 
+|─ Code —— Source code of the VisSalFormer model to predict question-driven saliency
 │  │
-│  │─ coming soon
+│  |─ environment.yml —— conda environment
-│   
+│  │
-└─ SalChartQA —— The dataset
+│  |─ env.py —— python envorinment $TORCH_HOME and $TRANSFORMERS_CACHE 
+│  │
+│  │─ dataset_new.py —— dataloader for SalChartQA
+│  │
+│  │─ evaluation.py —— evaluation script to load VisSalFormer weights and make predictions
+│  │
+│  │─ evaluation.sh —— bash script to run evaluation.py
+│  │
+│  │─ model_swin.py —— definition of the VisSalFormer model
+│  │
+│  │─ tokenizer_bert.py —— tokenizer of Bert
+│  │
+│  └─ VisSalFormer_weights.tar —— weights of VisSalFormer
+│
+└─ SalChartQA.zip —— The SalChartQA dataset
    │
-   │─ coming soon
+   │─ fixationByVis —— BubbleView data (mouse clicks) of AMT workers
+   │
+   │─ image_questions.json —— visualisation-question pairs
+   │
+   │─ raw_img —— original visualisations from the ChartQA dataset
+   │
+   │─ saliency_all —— saliency maps from all AMT workers
+   │
+   │─ saliency_ans —— saliency maps aggretated by all AMT workers who either answered a question correctly or wrongly
+   │
+   └─ unified_approved.csv —— responses from AMT workers
+
 ```
 
 If you think our work is useful to you, please consider citing our paper as:
@@ -28,9 +53,9 @@ If you think our work is useful to you, please consider citing our paper as:
   title = {SalChartQA: Question-driven Saliency on Information Visualisations},
   author = {Wang, Yao and Wang, Weitian and Abdelhafez, Abdullah and Elfares, Mayar and Hu, Zhiming and B{\^a}ce, Mihai and Bulling, Andreas},
   year = {2024},
-  pages = {1--20},
+  pages = {1--14},
   booktitle = {Proc. ACM SIGCHI Conference on Human Factors in Computing Systems (CHI)},
-  doi = {}
+  doi = {10.1145/3613904.3642942}
 }
 
 ```