Spaces:
Running
on
T4
Running
on
T4
| from transformers import RobertaTokenizer, RobertaForMaskedLM | |
| import torch | |
| import torch.nn.functional as F | |
| from models import TransformerVisualizer | |
| from transformers import ( | |
| RobertaForMaskedLM, RobertaForSequenceClassification | |
| ) | |
| import os,time | |
| import torch.autograd.functional as Fgrad | |
| CACHE_DIR = "/data/hf_cache" | |
| class RoBERTaVisualizer(TransformerVisualizer): | |
| def __init__(self, task): | |
| super().__init__() | |
| self.task = task | |
| TOKENIZER = 'roberta-base' | |
| LOCAL_PATH = os.path.join(CACHE_DIR, "tokenizers",TOKENIZER) | |
| self.tokenizer = RobertaTokenizer.from_pretrained(LOCAL_PATH, local_files_only=True) | |
| """ | |
| try: | |
| self.tokenizer = RobertaTokenizer.from_pretrained(LOCAL_PATH, local_files_only=True) | |
| except Exception as e: | |
| self.tokenizer = RobertaTokenizer.from_pretrained(TOKENIZER) | |
| self.tokenizer.save_pretrained(LOCAL_PATH) | |
| """ | |
| if self.task == 'mlm': | |
| MODEL = "roberta-base" | |
| LOCAL_PATH = os.path.join(CACHE_DIR, "models",MODEL) | |
| self.model = RobertaForMaskedLM.from_pretrained( LOCAL_PATH, local_files_only=True ).to(self.device) | |
| """ | |
| try: | |
| self.model = RobertaForMaskedLM.from_pretrained( LOCAL_PATH, local_files_only=True ) | |
| except Exception as e: | |
| self.model = RobertaForMaskedLM.from_pretrained( MODEL ) | |
| self.model.save_pretrained(LOCAL_PATH) | |
| """ | |
| elif self.task == 'sst': | |
| MODEL = 'textattack_roberta-base-SST-2' | |
| LOCAL_PATH = os.path.join(CACHE_DIR, "models",MODEL) | |
| self.model = RobertaForSequenceClassification.from_pretrained( LOCAL_PATH, local_files_only=True ).to(self.device) | |
| """ | |
| try: | |
| self.model = RobertaForSequenceClassification.from_pretrained( LOCAL_PATH, local_files_only=True ) | |
| except Exception as e: | |
| self.model = RobertaForSequenceClassification.from_pretrained( MODEL ) | |
| self.model.save_pretrained(LOCAL_PATH) | |
| """ | |
| elif self.task == 'mnli': | |
| MODEL = "roberta-large-mnli" | |
| LOCAL_PATH = os.path.join(CACHE_DIR, "models",MODEL) | |
| self.model = RobertaForSequenceClassification.from_pretrained( LOCAL_PATH, local_files_only=True).to(self.device) | |
| """ | |
| try: | |
| self.model = RobertaForSequenceClassification.from_pretrained( LOCAL_PATH, local_files_only=True) | |
| except Exception as e: | |
| self.model = RobertaForSequenceClassification.from_pretrained( MODEL) | |
| self.model.save_pretrained(LOCAL_PATH) | |
| """ | |
| self.model.to(self.device) | |
| # Force materialization of all layers (avoids meta device errors) | |
| with torch.no_grad(): | |
| dummy_ids = torch.tensor([[0, 1]], device=self.device) | |
| dummy_mask = torch.tensor([[1, 1]], device=self.device) | |
| _ = self.model(input_ids=dummy_ids, attention_mask=dummy_mask) | |
| 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] | |
| def scalar_outputs(inputs_embeds): | |
| outputs = self.model.roberta( | |
| inputs_embeds=inputs_embeds, | |
| attention_mask=inputs["attention_mask"], | |
| output_attentions=True | |
| ) | |
| attentions = outputs.attentions | |
| attentions_condensed = [a.mean(dim=0).mean(dim=0).sum(dim=0) for a in attentions] | |
| attentions_condensed= torch.vstack(attentions_condensed) | |
| return attentions_condensed | |
| start = time.time() | |
| jac = torch.autograd.functional.jacobian(scalar_outputs, inputs_embeds).to(torch.float16) | |
| print(jac.shape) | |
| jac = jac.norm(dim=-1).squeeze(dim=2) | |
| print(jac.shape) | |
| seq_len = jac.shape[0] | |
| print(seq_len) | |
| grad_matrices_all = [jac[ii,:,:].tolist() for ii in range(seq_len)] | |
| print(31,time.time()-start) | |
| attn_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) | |
| attn_matrix = mean_attns[target_layer] | |
| seq_len = attn_matrix.shape[0] | |
| attn_matrix = attn_matrix[:seq_len, :seq_len] | |
| print(4,attn_matrix.shape) | |
| attn_matrices_all.append(attn_matrix.tolist()) | |
| print(3,time.time()-start) | |
| """ | |
| 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) | |
| attn_matrix = mean_attns[target_layer] | |
| seq_len = attn_matrix.shape[0] | |
| attn_matrix = attn_matrix[:seq_len, :seq_len] | |
| attn_matrices_all.append(attn_matrix.tolist()) | |
| start = time.time() | |
| def scalar_outputs(inputs_embeds): | |
| outputs = self.model.roberta( | |
| inputs_embeds=inputs_embeds, | |
| attention_mask=inputs["attention_mask"], | |
| output_attentions=True | |
| ) | |
| attentions = outputs.attentions | |
| return attentions[target_layer].mean(dim=0).mean(dim=0).sum(dim=0) | |
| jac = torch.autograd.functional.jacobian(scalar_outputs, inputs_embeds).norm(dim=-1).squeeze(1) | |
| grad_matrices_all.append(jac.tolist()) | |
| print(1,time.time()-start) | |
| start = time.time() | |
| grad_norms_list = [] | |
| scalar_layer = attentions[target_layer].mean(dim=0).mean(dim=0) | |
| for k in range(seq_len): | |
| scalar = scalar_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)) | |
| print(2,time.time()-start) | |
| """ | |
| #print(grad_matrices_all) | |
| return grad_matrices_all, attn_matrices_all | |