files

BERTmodel.py

@@ -0,0 +1,257 @@
+from transformers import BertTokenizer, BertModel
+import torch
+import matplotlib.pyplot as plt
+import torch.nn as nn
+from transformers import BertTokenizer, BertModel, DataCollatorForLanguageModeling
+from datasets import load_dataset
+from torch.utils.data import DataLoader
+import torch.nn.functional as F
+
+from transformers import (
+    BertTokenizer, BertModel,
+    DataCollatorForLanguageModeling
+)
+import torch.optim as optim
+
+import os
+from transformers.models.bert.modeling_bert import BertOnlyMLMHead
+from models import TransformerVisualizer
+
+from transformers import (
+    BertTokenizer,
+    BertForMaskedLM,
+    BertForSequenceClassification,
+    BertForQuestionAnswering,
+)
+import torch
+import torch.nn.functional as F
+from models import TransformerVisualizer
+
+ 
+    
+class BERTVisualizer(TransformerVisualizer):
+    def __init__(self,task):
+        super().__init__()  
+        self.task = task
+        self.tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+        print('finding model', self.task)
+        if self.task == 'mlm':
+            self.model = BertForMaskedLM.from_pretrained(
+                "bert-base-uncased",
+                attn_implementation="eager"  # fallback to standard attention
+            ).to(self.device)
+        elif self.task == 'sst':
+            self.model = BertForSequenceClassification.from_pretrained("textattack/bert-base-uncased-SST-2",device_map=None)
+        elif self.task == 'mnli':
+            self.model = BertForSequenceClassification.from_pretrained("textattack/bert-base-uncased-MNLI", device_map=None)
+        else:
+            raise ValueError(f"Unsupported task: {self.task}")
+        print('model found')
+        #self.model.to(self.device)
+        print('self device junk')
+        self.model.eval()
+        print('self model eval')
+        self.num_attention_layers = len(self.model.bert.encoder.layer)
+        print('init finished')
+        
+    def tokenize(self, text, hypothesis = ''):
+        print('TTTokenize',text,'H:', hypothesis)
+        if len(hypothesis) == 0:
+            encoded = self.tokenizer(text, return_tensors='pt', return_attention_mask=True)
+        else:
+            encoded = self.tokenizer(text, hypothesis, return_tensors='pt', return_attention_mask=True)
+        input_ids = encoded['input_ids'].to(self.device)
+        attention_mask = encoded['attention_mask'].to(self.device)
+        tokens = self.tokenizer.convert_ids_to_tokens(input_ids[0])
+        return {
+            'input_ids': input_ids,
+            'attention_mask': attention_mask,
+            'tokens': tokens
+        }
+ 
+
+    def predict(self, task, text, hypothesis='', maskID = None):
+        
+        print(task,text,hypothesis)
+    
+
+
+        if task == 'mlm':
+
+            # Tokenize and find [MASK] position
+            print('Tokenize and find [MASK] position')  
+            inputs = self.tokenizer(text, return_tensors='pt', padding=False, truncation=True)
+            if maskID is not None and 0 <= maskID < inputs['input_ids'].size(1):
+                inputs['input_ids'][0][maskID] = self.tokenizer.mask_token_id
+                mask_index = maskID
+            else:
+                raise ValueError(f"Invalid maskID {maskID} for input length {inputs['input_ids'].size(1)}")
+
+
+
+            # Move to device
+            inputs = {k: v.to(self.device) for k, v in inputs.items()}
+
+            # Get embeddings 
+            embedding_layer = self.model.bert.embeddings.word_embeddings
+            inputs_embeds = embedding_layer(inputs['input_ids'])
+
+            # Forward through BERT encoder
+            
+            hidden_states = self.model.bert(inputs_embeds=inputs_embeds,
+                                    attention_mask=inputs['attention_mask']).last_hidden_state
+
+            # Predict logits via MLM head
+            logits = self.model.cls(hidden_states)
+            mask_logits = logits[0, mask_index]
+
+            top_probs, top_indices = torch.topk(mask_logits, k=10, dim=-1)
+            top_probs = F.softmax(top_probs, dim=-1)
+            decoded = self.tokenizer.convert_ids_to_tokens(top_indices.tolist())
+            
+            return decoded, top_probs
+    
+        elif task == 'sst':
+            print('input')
+            inputs = self.tokenizer(text, return_tensors='pt', padding=False, truncation=True).to(self.device)
+            print('output')
+            with torch.no_grad():
+                outputs = self.model(**inputs)
+                logits = outputs.logits  # shape: [1, 2]
+                probs = F.softmax(logits, dim=1).squeeze()
+
+            labels = ["negative", "positive"]
+            print('ready to return')
+            return labels, probs
+        
+        elif task == 'mnli':
+            inputs = self.tokenizer(text, hypothesis, return_tensors='pt', padding=True, truncation=True).to(self.device)
+
+            with torch.no_grad():
+                outputs = self.model(**inputs)
+                logits = outputs.logits
+                probs = F.softmax(logits, dim=1).squeeze()
+
+            labels = ["entailment", "neutral", "contradiction"]
+            return labels, probs
+        
+
+    def get_all_grad_attn_matrix(self, task, sentence, hypothesis='', maskID = 0):
+
+        print('GET GRAD:', task,'sentence',sentence, 'hypothesis', hypothesis)
+         
+        
+        
+        print('Tokenize')
+        if task == 'mnli':
+            inputs = self.tokenizer(sentence, hypothesis, return_tensors='pt', padding=False, truncation=True)
+        elif task == 'mlm':
+            inputs = self.tokenizer(sentence,  return_tensors='pt', padding=False, truncation=True)
+            if maskID is not None and 0 <= maskID < inputs['input_ids'].size(1):
+                inputs['input_ids'][0][maskID] = self.tokenizer.mask_token_id
+            else:
+                raise ValueError(f"Invalid maskID {maskID} for input length {inputs['input_ids'].size(1)}")
+        else:
+            inputs = self.tokenizer(sentence,  return_tensors='pt', padding=False, truncation=True)
+        tokens = self.tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])
+
+        inputs = {k: v.to(self.device) for k, v in inputs.items()}
+        print(inputs['input_ids'].shape)
+        print(tokens,len(tokens))
+        print('Input embeddings with grad')
+        embedding_layer = self.model.bert.embeddings.word_embeddings
+        inputs_embeds = embedding_layer(inputs["input_ids"])
+        inputs_embeds.requires_grad_()
+
+        print('Forward pass')
+        outputs = self.model.bert(
+            inputs_embeds=inputs_embeds,
+            attention_mask=inputs["attention_mask"],
+            output_attentions=True
+        )
+        attentions = outputs.attentions  # list of [1, heads, seq, seq]
+
+        print('Optional: store average attentions per layer')
+        mean_attns = [a.squeeze(0).mean(dim=0).detach().cpu() for a in attentions]
+
+        attn_matrices_all = []
+        grad_matrices_all = []
+        for target_layer in range(len(attentions)):
+            grad_matrix, attn_matrix = self.get_grad_attn_matrix(inputs_embeds, attentions, mean_attns, target_layer)
+            grad_matrices_all.append(grad_matrix.tolist())
+            attn_matrices_all.append(attn_matrix.tolist())
+        return grad_matrices_all, attn_matrices_all
+
+    def get_grad_attn_matrix(self,inputs_embeds, attentions, mean_attns, target_layer):
+         
+        
+        attn_matrix = mean_attns[target_layer] 
+        seq_len = attn_matrix.shape[0]
+        attn_layer = attentions[target_layer].squeeze(0).mean(dim=0)  # [seq, seq]
+
+
+        print('computing gradnorms now')
+        
+
+        grad_norms_list = []
+
+        for k in range(seq_len):
+            scalar = attn_layer[:, k].sum()  # ✅ total attention received by token k
+            
+            # Compute gradient: d scalar / d inputs_embeds
+            
+            grad = torch.autograd.grad(scalar, inputs_embeds, retain_graph=True)[0].squeeze(0)  # shape: [seq, hidden]
+            
+            grad_norms = grad.norm(dim=1)  # shape: [seq]
+            
+            grad_norms_list.append(grad_norms.unsqueeze(1))  # shape: [seq, 1]
+        
+        
+        grad_matrix = torch.cat(grad_norms_list, dim=1)  # shape: [seq, seq]
+        print('ready to send!')
+     
+        grad_matrix = grad_matrix[:seq_len, :seq_len]
+        attn_matrix = attn_matrix[:seq_len, :seq_len]
+
+        #tokens = self.tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])
+        
+        return grad_matrix, attn_matrix
+ 
+
+if __name__ == "__main__":
+    import sys
+
+    MODEL_CLASSES = {
+        "bert": BERTVisualizer,
+        "roberta": RoBERTaVisualizer,
+        "distilbert": DistilBERTVisualizer,
+        "bart": BARTVisualizer,
+    }
+
+    # Parse command-line args or fallback to default
+    model_name = sys.argv[1] if len(sys.argv) > 1 else "bert"
+    text = " ".join(sys.argv[2:]) if len(sys.argv) > 2 else "The quick brown fox jumps over the lazy dog."
+
+    if model_name.lower() not in MODEL_CLASSES:
+        print(f"Supported models: {list(MODEL_CLASSES.keys())}")
+        sys.exit(1)
+
+    # Instantiate the visualizer
+    visualizer_class = MODEL_CLASSES[model_name.lower()]
+    visualizer = visualizer_class()
+
+    # Tokenize
+    token_info = visualizer.tokenize(text)
+
+    # Report
+    print(f"\nModel: {model_name}")
+    print(f"Num attention layers: {visualizer.num_attention_layers}")
+    print(f"Tokens: {token_info['tokens']}")
+    print(f"Input IDs: {token_info['input_ids'].tolist()}")
+    print(f"Attention mask: {token_info['attention_mask'].tolist()}")
+
+
+"""
+usage for debug:
+python your_file.py bert "The rain in Spain falls mainly on the plain."
+"""

DISTILLBERTmodel.py

@@ -0,0 +1,210 @@
+import torch
+import matplotlib.pyplot as plt
+import torch.nn as nn
+from datasets import load_dataset
+from torch.utils.data import DataLoader
+import torch.nn.functional as F
+
+ 
+
+import os
+from transformers import DistilBertModel, DistilBertTokenizer
+from models import TransformerVisualizer
+
+from transformers import (
+     DistilBertTokenizer,
+    DistilBertForMaskedLM, DistilBertForSequenceClassification
+)    
+
+class DistilBERTVisualizer(TransformerVisualizer):
+    def __init__(self, task):
+        super().__init__()
+        self.task = task
+        self.tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
+        if self.task == 'mlm':
+            self.model = DistilBertForMaskedLM.from_pretrained('distilbert-base-uncased')
+        elif self.task == 'sst':
+            self.model = DistilBertForSequenceClassification.from_pretrained('distilbert-base-uncased-finetuned-sst-2-english')
+        elif self.task == 'mnli':
+            self.model = DistilBertForSequenceClassification.from_pretrained("textattack/distilbert-base-uncased-MNLI")
+
+
+        else:
+            raise NotImplementedError("Task not supported for DistilBERT")
+        
+
+        self.model.eval()
+        self.num_attention_layers = len(self.model.distilbert.transformer.layer)
+
+        self.model.to(self.device)
+
+
+        
+    def tokenize(self, text, hypothesis = ''):
+
+         
+
+        if len(hypothesis) == 0:
+            encoded = self.tokenizer(text, return_tensors='pt', return_attention_mask=True,padding=False, truncation=True)
+        else:
+            encoded = self.tokenizer(text, hypothesis, return_tensors='pt', return_attention_mask=True,padding=False, truncation=True)
+
+
+        input_ids = encoded['input_ids'].to(self.device)
+        attention_mask = encoded['attention_mask'].to(self.device)
+        tokens = self.tokenizer.convert_ids_to_tokens(input_ids[0])
+        return {
+            'input_ids': input_ids,
+            'attention_mask': attention_mask,
+            'tokens': tokens
+        }
+     
+    def predict(self, task, text, hypothesis='', maskID = 0):
+        
+        if task  == 'mlm':
+            inputs = self.tokenizer(text, return_tensors='pt', padding=False, truncation=True)
+            if maskID is not None and 0 <= maskID < inputs['input_ids'].size(1):
+                inputs['input_ids'][0][maskID] = self.tokenizer.mask_token_id
+                mask_index = maskID
+            else:
+                raise ValueError(f"Invalid maskID {maskID} for input of length {inputs['input_ids'].size(1)}")
+            inputs = {k: v.to(self.device) for k, v in inputs.items()}
+
+            with torch.no_grad():
+                outputs = self.model(**inputs)
+                logits = outputs.logits
+
+            mask_logits = logits[0, mask_index]
+            top_probs, top_indices = torch.topk(F.softmax(mask_logits, dim=-1), 10)
+            decoded = self.tokenizer.convert_ids_to_tokens(top_indices.tolist())
+            return decoded, top_probs
+
+        elif task == 'sst':
+            inputs = self.tokenizer(text, return_tensors='pt', padding=False, truncation=True).to(self.device)
+
+            with torch.no_grad():
+                outputs = self.model(**inputs)
+                logits = outputs.logits
+                probs = F.softmax(logits, dim=1).squeeze()
+
+            labels = ["negative", "positive"]
+            return labels, probs
+        elif task == 'mnli': 
+            inputs = self.tokenizer(text, hypothesis, return_tensors='pt', padding=True, truncation=True).to(self.device)
+
+            with torch.no_grad():
+                outputs = self.model(**inputs)
+                logits = outputs.logits
+                probs = F.softmax(logits, dim=1).squeeze()
+
+            labels = ["entailment", "neutral", "contradiction"]
+            return labels, probs
+
+        else:
+            raise NotImplementedError(f"Task '{task}' not supported for DistilBERT")
+
+    def get_all_grad_attn_matrix(self, task, sentence, hypothesis='', maskID = 0):
+        print(task, sentence,hypothesis) 
+
+        print('Tokenize')
+        if task == 'mnli':
+            inputs = self.tokenizer(sentence, hypothesis, return_tensors='pt', padding=False, truncation=True)
+        elif task == 'mlm':
+            inputs = self.tokenizer(sentence,  return_tensors='pt', padding=False, truncation=True)
+            if maskID is not None and 0 <= maskID < inputs['input_ids'].size(1):
+                inputs['input_ids'][0][maskID] = self.tokenizer.mask_token_id
+            else:
+                print(f"Invalid maskID {maskID} for input of length {inputs['input_ids'].size(1)}")
+                raise ValueError(f"Invalid maskID {maskID} for input of length {inputs['input_ids'].size(1)}")
+            inputs = {k: v.to(self.device) for k, v in inputs.items()}
+        else:
+            inputs = self.tokenizer(sentence,  return_tensors='pt', padding=False, truncation=True)
+        tokens = self.tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])
+        print(tokens)
+        inputs = {k: v.to(self.device) for k, v in inputs.items()}
+
+        print('Input embeddings with grad')
+        embedding_layer = self.model.distilbert.embeddings.word_embeddings
+        inputs_embeds = embedding_layer(inputs["input_ids"])
+        inputs_embeds.requires_grad_()
+
+        print('Forward pass')
+        outputs = self.model.distilbert(
+            inputs_embeds=inputs_embeds,
+            attention_mask=inputs["attention_mask"],
+            output_attentions=True,
+        )
+        attentions = outputs.attentions  # list of [1, heads, seq, seq]
+
+        print('Mean attentions per layer')
+        mean_attns = [a.squeeze(0).mean(dim=0).detach().cpu() for a in attentions]
+        
+
+        
+        attn_matrices_all = []
+        grad_matrices_all = []
+        for target_layer in range(len(attentions)):
+            grad_matrix, attn_matrix = self.get_grad_attn_matrix(inputs_embeds, attentions, mean_attns, target_layer)
+            grad_matrices_all.append(grad_matrix.tolist())
+            attn_matrices_all.append(attn_matrix.tolist())
+        return grad_matrices_all, attn_matrices_all
+    
+
+    def get_grad_attn_matrix(self,inputs_embeds, attentions, mean_attns, target_layer):
+        attn_matrix = mean_attns[target_layer]
+        seq_len = attn_matrix.shape[0]
+        attn_layer = attentions[target_layer].squeeze(0).mean(dim=0)
+
+        print('Computing grad norms')
+        grad_norms_list = []
+        for k in range(seq_len):
+            scalar = attn_layer[:, k].sum()
+            grad = torch.autograd.grad(scalar, inputs_embeds, retain_graph=True)[0].squeeze(0)
+            grad_norms = grad.norm(dim=1)
+            grad_norms_list.append(grad_norms.unsqueeze(1))
+
+        grad_matrix = torch.cat(grad_norms_list, dim=1)
+        grad_matrix = grad_matrix[:seq_len, :seq_len]
+        attn_matrix = attn_matrix[:seq_len, :seq_len]
+
+        return grad_matrix, attn_matrix
+
+
+
+if __name__ == "__main__":
+    import sys
+
+    MODEL_CLASSES = {
+        "bert": BERTVisualizer,
+        "roberta": RoBERTaVisualizer,
+        "distilbert": DistilBERTVisualizer,
+        "bart": BARTVisualizer,
+    }
+
+    # Parse command-line args or fallback to default
+    model_name = sys.argv[1] if len(sys.argv) > 1 else "bert"
+    text = " ".join(sys.argv[2:]) if len(sys.argv) > 2 else "The quick brown fox jumps over the lazy dog."
+
+    if model_name.lower() not in MODEL_CLASSES:
+        print(f"Supported models: {list(MODEL_CLASSES.keys())}")
+        sys.exit(1)
+
+    # Instantiate the visualizer
+    visualizer_class = MODEL_CLASSES[model_name.lower()]
+    visualizer = visualizer_class()
+
+    # Tokenize
+    token_info = visualizer.tokenize(text)
+
+    # Report
+    print(f"\nModel: {model_name}")
+    print(f"Num attention layers: {visualizer.num_attention_layers}")
+    print(f"Tokens: {token_info['tokens']}")
+    print(f"Input IDs: {token_info['input_ids'].tolist()}")
+    print(f"Attention mask: {token_info['attention_mask'].tolist()}")
+
+
+"""
+usage for debug:
+python your_file.py bert "The rain in Spain falls mainly on the plain."
+"""

ROBERTAmodel.py

@@ -0,0 +1,155 @@
+from transformers import RobertaTokenizer, RobertaForMaskedLM
+import torch
+import torch.nn.functional as F
+from models import TransformerVisualizer
+from transformers import (
+    RobertaForMaskedLM, RobertaForSequenceClassification, RobertaForQuestionAnswering,
+)
+
+class RoBERTaVisualizer(TransformerVisualizer):
+    def __init__(self, task):
+        super().__init__()
+        self.task = task
+        self.tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
+        if self.task == 'mlm':
+            self.model = RobertaForMaskedLM.from_pretrained("roberta-base")
+        elif self.task == 'sst':
+            self.model = RobertaForSequenceClassification.from_pretrained('textattack/roberta-base-SST-2')
+        elif self.task == 'mnli':
+            self.model = RobertaForSequenceClassification.from_pretrained("roberta-large-mnli")
+
+
+        self.model.to(self.device)
+        self.model.eval()
+        self.num_attention_layers = self.model.config.num_hidden_layers
+
+
+    def tokenize(self, text, hypothesis = ''):
+
+         
+
+        if len(hypothesis) == 0:
+            encoded = self.tokenizer(text, return_tensors='pt', return_attention_mask=True,padding=False, truncation=True)
+        else:
+            encoded = self.tokenizer(text, hypothesis, return_tensors='pt', return_attention_mask=True,padding=False, truncation=True)
+
+        input_ids = encoded['input_ids'].to(self.device)
+        attention_mask = encoded['attention_mask'].to(self.device)
+        tokens = self.tokenizer.convert_ids_to_tokens(input_ids[0])
+        print('First time tokenizing:', tokens, len(tokens))
+
+        response = {
+            'input_ids': input_ids,
+            'attention_mask': attention_mask,
+            'tokens': tokens
+        }
+        print(response)
+        return response
+
+    def predict(self, task, text, hypothesis='', maskID = None):
+        
+        
+
+        if task == 'mlm':
+            inputs = self.tokenizer(text, return_tensors='pt', padding=False, truncation=True)
+            if maskID is not None and 0 <= maskID < inputs['input_ids'].size(1):
+                inputs['input_ids'][0][maskID] = self.tokenizer.mask_token_id
+                mask_index = maskID
+            else:
+                raise ValueError(f"Invalid maskID {maskID} for input of length {inputs['input_ids'].size(1)}")
+            inputs = {k: v.to(self.device) for k, v in inputs.items()}
+
+            with torch.no_grad():
+                outputs = self.model(**inputs)
+                logits = outputs.logits
+
+            mask_logits = logits[0, mask_index]
+            top_probs, top_indices = torch.topk(F.softmax(mask_logits, dim=-1), 10)
+            decoded = self.tokenizer.convert_ids_to_tokens(top_indices.tolist())
+            return decoded, top_probs
+
+        elif task == 'sst':
+            inputs = self.tokenizer(text, return_tensors='pt', padding=False, truncation=True).to(self.device)
+
+            with torch.no_grad():
+                outputs = self.model(**inputs)
+                logits = outputs.logits
+                probs = F.softmax(logits, dim=1).squeeze()
+
+            labels = ["negative", "positive"]
+            return labels, probs
+        
+        elif task == 'mnli': 
+            inputs = self.tokenizer(text, hypothesis, return_tensors='pt', padding=True, truncation=True).to(self.device)
+
+            with torch.no_grad():
+                outputs = self.model(**inputs)
+                logits = outputs.logits
+                probs = F.softmax(logits, dim=1).squeeze()
+
+            labels = ["entailment", "neutral", "contradiction"]
+            return labels, probs
+
+        else:
+            raise NotImplementedError(f"Task '{task}' not supported for RoBERTa")
+
+
+    def get_all_grad_attn_matrix(self, task, sentence, hypothesis='', maskID = None):
+        print(task, sentence,  hypothesis)
+        print('Tokenize')
+        if task == 'mnli':
+            inputs = self.tokenizer(sentence, hypothesis, return_tensors='pt', padding=False, truncation=True)
+        elif task == 'mlm':
+            inputs = self.tokenizer(sentence,  return_tensors='pt', padding=False, truncation=True)
+            if maskID is not None and 0 <= maskID < inputs['input_ids'].size(1):
+                inputs['input_ids'][0][maskID] = self.tokenizer.mask_token_id
+        else:
+            inputs = self.tokenizer(sentence,  return_tensors='pt', padding=False, truncation=True)
+        tokens = self.tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])
+        print(tokens)
+        inputs = {k: v.to(self.device) for k, v in inputs.items()}
+
+        print('Input embeddings with grad')
+        embedding_layer = self.model.roberta.embeddings.word_embeddings
+        inputs_embeds = embedding_layer(inputs["input_ids"])
+        inputs_embeds.requires_grad_()
+
+        print('Forward pass')
+        outputs = self.model.roberta(
+            inputs_embeds=inputs_embeds,
+            attention_mask=inputs["attention_mask"],
+            output_attentions=True
+        )
+        attentions = outputs.attentions  # list of [1, heads, seq, seq]
+
+        print('Average attentions per layer')
+        mean_attns = [a.squeeze(0).mean(dim=0).detach().cpu() for a in attentions]
+
+        attn_matrices_all = []
+        grad_matrices_all = []
+        for target_layer in range(len(attentions)):
+            grad_matrix, attn_matrix = self.get_grad_attn_matrix(inputs_embeds, attentions, mean_attns, target_layer)
+            grad_matrices_all.append(grad_matrix.tolist())
+            attn_matrices_all.append(attn_matrix.tolist())
+        return grad_matrices_all, attn_matrices_all
+    
+    def get_grad_attn_matrix(self,inputs_embeds, attentions, mean_attns, target_layer):
+        
+        attn_matrix = mean_attns[target_layer]
+        seq_len = attn_matrix.shape[0]
+        attn_layer = attentions[target_layer].squeeze(0).mean(dim=0)  # [seq, seq]
+
+        print('Computing grad norms')
+        grad_norms_list = []
+        for k in range(seq_len):
+            scalar = attn_layer[:, k].sum()
+            grad = torch.autograd.grad(scalar, inputs_embeds, retain_graph=True)[0].squeeze(0)
+            grad_norms = grad.norm(dim=1)
+            grad_norms_list.append(grad_norms.unsqueeze(1))
+
+        grad_matrix = torch.cat(grad_norms_list, dim=1)
+        grad_matrix = grad_matrix[:seq_len, :seq_len]
+        attn_matrix = attn_matrix[:seq_len, :seq_len]
+
+        
+        return grad_matrix, attn_matrix

models.py

@@ -0,0 +1,34 @@
+from transformers import BertTokenizer, BertModel
+import torch
+import matplotlib.pyplot as plt
+import torch.nn as nn
+from transformers import BertTokenizer, BertModel, DataCollatorForLanguageModeling
+from datasets import load_dataset
+from torch.utils.data import DataLoader
+import torch.nn.functional as F
+
+from transformers import (
+    BertTokenizer, BertModel,
+    DataCollatorForLanguageModeling
+)
+import torch.optim as optim
+
+import os
+from transformers.models.bert.modeling_bert import BertOnlyMLMHead
+from transformers import RobertaModel, RobertaTokenizer
+from transformers import DistilBertModel, DistilBertTokenizer
+from transformers import BartModel, BartTokenizer
+
+
+
+class TransformerVisualizer():
+    def __init__(self):
+        self.device = torch.device('cpu')
+        
+    def predict(self, task, text):
+        return task, text,1
+    
+
+    def get_attention_gradient_matrix(self, task, text, target_layer):
+        return task, text,target_layer,1
+     

requirements.txt

@@ -0,0 +1,4 @@
+fastapi
+uvicorn
+transformers
+torch

server.py

@@ -0,0 +1,215 @@
+from fastapi import FastAPI, Request
+from pydantic import BaseModel
+
+import torch
+
+from fastapi.middleware.cors import CORSMiddleware
+
+from ROBERTAmodel import *
+from BERTmodel import *
+from DISTILLBERTmodel import *
+
+VISUALIZER_CLASSES = {
+    "BERT": BERTVisualizer,
+    "RoBERTa": RoBERTaVisualizer,
+    "DistilBERT": DistilBERTVisualizer,
+}
+
+VISUALIZER_CACHE = {}
+app = FastAPI()
+
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=["*"],  # or restrict to ["http://localhost:3000"]
+    allow_credentials=True,
+    allow_methods=["*"],
+    allow_headers=["*"],
+)
+
+MODEL_MAP = {
+    "BERT": "bert-base-uncased",
+    "RoBERTa": "roberta-base",
+    "DistilBERT": "distilbert-base-uncased",
+}
+
+class LoadModelRequest(BaseModel):
+    model: str
+    sentence: str
+    task:str
+    hypothesis:str
+
+class GradAttnModelRequest(BaseModel):
+    model: str
+    task: str
+    sentence: str
+    hypothesis:str
+    maskID: int | None = None
+
+class PredModelRequest(BaseModel):
+    model: str
+    sentence: str
+    task:str
+    hypothesis:str
+    maskID: int | None = None
+
+
+@app.post("/load_model")
+def load_model(req: LoadModelRequest):
+    print(f"\n--- /load_model request received ---")
+    print(f"Model: {req.model}")
+    print(f"Sentence: {req.sentence}")
+    print(f"Task: {req.task}")
+    print(f"hypothesis: {req.hypothesis}")
+
+
+    if req.model in VISUALIZER_CACHE:
+        del VISUALIZER_CACHE[req.model]
+    torch.cuda.empty_cache()
+
+    vis_class = VISUALIZER_CLASSES.get(req.model)
+    if vis_class is None:
+        return {"error": f"Unknown model: {req.model}"}
+
+    print("instantiating visualizer")
+    try:
+        vis = vis_class(task=req.task.lower())
+        print(vis)
+        VISUALIZER_CACHE[req.model] = vis
+        print("Visualizer instantiated")
+    except Exception as e:
+        print("Visualizer init failed:", e)
+        return {"error": f"Instantiation failed: {str(e)}"}
+
+    print('tokenizing')
+    try:
+        if req.task.lower() == 'mnli':
+            token_output = vis.tokenize(req.sentence, hypothesis=req.hypothesis)
+        else:
+            token_output = vis.tokenize(req.sentence)
+        print("0 Tokenization successful:", token_output["tokens"])
+    except Exception as e:
+        print("Tokenization failed:", e)
+        return {"error": f"Tokenization failed: {str(e)}"}
+
+    print('response ready')    
+    response = {
+        "model": req.model,
+        "tokens": token_output['tokens'],
+        "num_layers": vis.num_attention_layers,
+    }
+    print("load model successful")
+    print(response)
+    return response
+
+
+
+
+
+@app.post("/predict_model")
+def predict_model(req: PredModelRequest):
+    
+    print(f"\n--- /predict_model request received ---")
+    print(f"predict: Model: {req.model}")
+    print(f"predict: Task: {req.task}")
+    print(f"predict: sentence: {req.sentence}")
+    print(f"predict: hypothesis: {req.hypothesis}")
+    print(f"predict: maskID: {req.maskID}")
+
+
+
+    print('predict: instantiating')
+    try:
+        vis_class = VISUALIZER_CLASSES.get(req.model)
+        if vis_class is None:
+            return {"error": f"Unknown model: {req.model}"}
+        #if any(p.device.type == 'meta' for p in vis.model.parameters()):
+        #    vis.model = torch.nn.Module.to_empty(vis.model, device=torch.device("cpu"))
+    
+        vis = vis_class(task=req.task.lower())
+        VISUALIZER_CACHE[req.model] = vis
+        print("Model reloaded and cached.")
+    except Exception as e:
+        return {"error": f"Failed to reload model: {str(e)}"}
+    
+    print('predict: meta stuff')
+    
+
+    
+    print('predict: Run prediction')
+    try:
+        if req.task.lower() == 'mnli':
+            decoded, top_probs = vis.predict(req.task.lower(), req.sentence, hypothesis=req.hypothesis)
+        elif req.task.lower() == 'mlm':
+            decoded, top_probs = vis.predict(req.task.lower(), req.sentence, maskID=req.maskID)
+        
+        else:
+            decoded, top_probs = vis.predict(req.task.lower(), req.sentence)
+    except Exception as e:
+        decoded, top_probs = "error", e
+        print(e)
+
+    print('predict: response ready')
+    response = {
+        "decoded": decoded,
+        "top_probs": top_probs.tolist(),
+    }
+    print("predict: predict model successful")
+    if len(decoded) > 5:
+        print([(k,v[:5]) for k,v in response.items()])
+    else:
+        print(response)
+    return response
+
+
+
+@app.post("/get_grad_attn_matrix")
+def get_grad_attn_matrix(req: GradAttnModelRequest):
+    
+    try:
+        print(f"\n--- /get_grad_matrix request received ---")
+        print(f"grad:Model: {req.model}")
+        print(f"grad:Task: {req.task}")
+        print(f"grad:sentence: {req.sentence}")
+        print(f"grad: hypothesis: {req.hypothesis}")
+        print(f"predict: maskID: {req.maskID}")
+        
+        
+        
+        try:
+            vis_class = VISUALIZER_CLASSES.get(req.model)
+            if vis_class is None:
+                return {"error": f"Unknown model: {req.model}"}
+            #if any(p.device.type == 'meta' for p in vis.model.parameters()):
+            #    vis.model = torch.nn.Module.to_empty(vis.model, device=torch.device("cpu"))
+            vis = vis_class(task=req.task.lower())
+            VISUALIZER_CACHE[req.model] = vis
+            print("Model reloaded and cached.")
+        except Exception as e:
+            return {"error": f"Failed to reload model: {str(e)}"} 
+
+
+        
+        print("run function")
+        try:
+            if req.task.lower()=='mnli':
+                grad_matrix, attn_matrix = vis.get_all_grad_attn_matrix(req.task.lower(), req.sentence,hypothesis=req.hypothesis)
+            elif req.task.lower()=='mlm':
+                grad_matrix, attn_matrix = vis.get_all_grad_attn_matrix(req.task.lower(), req.sentence,maskID=req.maskID)
+            else:
+                grad_matrix, attn_matrix = vis.get_all_grad_attn_matrix(req.task.lower(), req.sentence)
+        except Exception as e:
+            print("Exception during grad/attn computation:", e)
+            grad_matrix, attn_matrix = e,e
+
+    
+        response = {
+            "grad_matrix": grad_matrix,
+            "attn_matrix": attn_matrix,
+        }
+        print('grad attn successful')
+        return response
+    except Exception as e:
+        print("SERVER EXCEPTION:", e)
+        return {"error": str(e)}
+    
+