import torch import torch.nn as nn from torch.nn import functional as F from torch.optim import Adam, SGD from torch.utils.data import DataLoader import pickle from ..bert import BERT from ..seq_model import BERTSM from ..classifier_model import BERTForClassification from ..optim_schedule import ScheduledOptim import tqdm import sys import time import numpy as np # import visualization from sklearn.metrics import precision_score, recall_score, f1_score import matplotlib.pyplot as plt import seaborn as sns import pandas as pd from collections import defaultdict import os class ECE(nn.Module): def __init__(self, n_bins=15): """ n_bins (int): number of confidence interval bins """ super(ECE, self).__init__() bin_boundaries = torch.linspace(0, 1, n_bins + 1) self.bin_lowers = bin_boundaries[:-1] self.bin_uppers = bin_boundaries[1:] def forward(self, logits, labels): softmaxes = F.softmax(logits, dim=1) confidences, predictions = torch.max(softmaxes, 1) labels = torch.argmax(labels,1) accuracies = predictions.eq(labels) ece = torch.zeros(1, device=logits.device) for bin_lower, bin_upper in zip(self.bin_lowers, self.bin_uppers): # Calculated |confidence - accuracy| in each bin in_bin = confidences.gt(bin_lower.item()) * confidences.le(bin_upper.item()) prop_in_bin = in_bin.float().mean() if prop_in_bin.item() > 0: accuracy_in_bin = accuracies[in_bin].float().mean() avg_confidence_in_bin = confidences[in_bin].mean() ece += torch.abs(avg_confidence_in_bin - accuracy_in_bin) * prop_in_bin return ece def accurate_nb(preds, labels): pred_flat = np.argmax(preds, axis=1).flatten() labels_flat = np.argmax(labels, axis=1).flatten() labels_flat = labels.flatten() return np.sum(pred_flat == labels_flat) class BERTTrainer: """ BERTTrainer pretrains BERT model on input sequence of strategies. BERTTrainer make the pretrained BERT model with one training method objective. 1. Masked Strategy Modelling : 3.3.1 Task #1: Masked SM """ def __init__(self, bert: BERT, vocab_size: int, train_dataloader: DataLoader, val_dataloader: DataLoader = None, test_dataloader: DataLoader = None, lr: float = 1e-4, betas=(0.9, 0.999), weight_decay: float = 0.01, warmup_steps=5000, with_cuda: bool = True, cuda_devices=None, log_freq: int = 10, same_student_prediction = False, workspace_name=None, code=None): """ :param bert: BERT model which you want to train :param vocab_size: total word vocab size :param train_dataloader: train dataset data loader :param test_dataloader: test dataset data loader [can be None] :param lr: learning rate of optimizer :param betas: Adam optimizer betas :param weight_decay: Adam optimizer weight decay param :param with_cuda: traning with cuda :param log_freq: logging frequency of the batch iteration """ cuda_condition = torch.cuda.is_available() and with_cuda self.device = torch.device("cuda:0" if cuda_condition else "cpu") print(cuda_condition, " Device used = ", self.device) available_gpus = list(range(torch.cuda.device_count())) # This BERT model will be saved every epoch self.bert = bert.to(self.device) # Initialize the BERT Language Model, with BERT model self.model = BERTSM(bert, vocab_size).to(self.device) # Distributed GPU training if CUDA can detect more than 1 GPU if with_cuda and torch.cuda.device_count() > 1: print("Using %d GPUS for BERT" % torch.cuda.device_count()) self.model = nn.DataParallel(self.model, device_ids=available_gpus) # Setting the train and test data loader self.train_data = train_dataloader self.val_data = val_dataloader self.test_data = test_dataloader # Setting the Adam optimizer with hyper-param self.optim = Adam(self.model.parameters(), lr=lr, betas=betas, weight_decay=weight_decay) self.optim_schedule = ScheduledOptim(self.optim, self.bert.hidden, n_warmup_steps=warmup_steps) # Using Negative Log Likelihood Loss function for predicting the masked_token self.criterion = nn.NLLLoss(ignore_index=0) self.log_freq = log_freq self.same_student_prediction = same_student_prediction self.workspace_name = workspace_name self.save_model = False self.code = code self.avg_loss = 10000 self.start_time = time.time() print("Total Parameters:", sum([p.nelement() for p in self.model.parameters()])) def train(self, epoch): self.iteration(epoch, self.train_data) def val(self, epoch): self.iteration(epoch, self.val_data, phase="val") def test(self, epoch): self.iteration(epoch, self.test_data, phase="test") def iteration(self, epoch, data_loader, phase="train"): """ loop over the data_loader for training or testing if on train status, backward operation is activated and also auto save the model every peoch :param epoch: current epoch index :param data_loader: torch.utils.data.DataLoader for iteration :param train: boolean value of is train or test :return: None """ # str_code = "train" if train else "test" # code = "masked_prediction" if self.same_student_prediction else "masked" self.log_file = f"{self.workspace_name}/logs/{self.code}/log_{phase}_pretrained.txt" # bert_hidden_representations = [] if epoch == 0: f = open(self.log_file, 'w') f.close() if phase == "val": self.avg_loss = 10000 # Setting the tqdm progress bar data_iter = tqdm.tqdm(enumerate(data_loader), desc="EP_%s:%d" % (phase, epoch), total=len(data_loader), bar_format="{l_bar}{r_bar}") avg_loss_mask = 0.0 total_correct_mask = 0 total_element_mask = 0 avg_loss_pred = 0.0 total_correct_pred = 0 total_element_pred = 0 avg_loss = 0.0 if phase == "train": self.model.train() else: self.model.eval() with open(self.log_file, 'a') as f: sys.stdout = f for i, data in data_iter: # 0. batch_data will be sent into the device(GPU or cpu) data = {key: value.to(self.device) for key, value in data.items()} # if i == 0: # print(f"data : {data[0]}") # 1. forward the next_sentence_prediction and masked_lm model # next_sent_output, mask_lm_output = self.model.forward(data["bert_input"], data["segment_label"]) if self.same_student_prediction: bert_hidden_rep, mask_lm_output, same_student_output = self.model.forward(data["bert_input"], data["segment_label"], self.same_student_prediction) else: bert_hidden_rep, mask_lm_output = self.model.forward(data["bert_input"], data["segment_label"], self.same_student_prediction) # embeddings = [h for h in bert_hidden_rep.cpu().detach().numpy()] # bert_hidden_representations.extend(embeddings) # 2-2. NLLLoss of predicting masked token word mask_loss = self.criterion(mask_lm_output.transpose(1, 2), data["bert_label"]) # 2-3. Adding next_loss and mask_loss : 3.4 Pre-training Procedure if self.same_student_prediction: # 2-1. NLL(negative log likelihood) loss of is_next classification result same_student_loss = self.criterion(same_student_output, data["is_same_student"]) loss = same_student_loss + mask_loss else: loss = mask_loss # 3. backward and optimization only in train if phase == "train": self.optim_schedule.zero_grad() loss.backward() self.optim_schedule.step_and_update_lr() # print(f"mask_lm_output : {mask_lm_output}") # non_zero_mask = (data["bert_label"] != 0).float() # print(f"bert_label : {data['bert_label']}") non_zero_mask = (data["bert_label"] != 0).float() predictions = torch.argmax(mask_lm_output, dim=-1) # print(f"predictions : {predictions}") predicted_masked = predictions*non_zero_mask # print(f"predicted_masked : {predicted_masked}") mask_correct = ((data["bert_label"] == predicted_masked)*non_zero_mask).sum().item() # print(f"mask_correct : {mask_correct}") # print(f"non_zero_mask.sum().item() : {non_zero_mask.sum().item()}") avg_loss_mask += loss.item() total_correct_mask += mask_correct total_element_mask += non_zero_mask.sum().item() # total_element_mask += data["bert_label"].sum().item() torch.cuda.empty_cache() post_fix = { "epoch": epoch, "iter": i, "avg_loss": avg_loss_mask / (i + 1), "avg_acc_mask": (total_correct_mask / total_element_mask * 100) if total_element_mask != 0 else 0, "loss": loss.item() } # next sentence prediction accuracy if self.same_student_prediction: correct = same_student_output.argmax(dim=-1).eq(data["is_same_student"]).sum().item() avg_loss_pred += loss.item() total_correct_pred += correct total_element_pred += data["is_same_student"].nelement() # correct = next_sent_output.argmax(dim=-1).eq(data["is_next"]).sum().item() post_fix["avg_loss"] = avg_loss_pred / (i + 1) post_fix["avg_acc_pred"] = total_correct_pred / total_element_pred * 100 post_fix["loss"] = loss.item() avg_loss +=loss.item() if i % self.log_freq == 0: data_iter.write(str(post_fix)) # if not train and epoch > 20 : # pickle.dump(mask_lm_output.cpu().detach().numpy(), open(f"logs/mask/mask_out_e{epoch}_{i}.pkl","wb")) # pickle.dump(data["bert_label"].cpu().detach().numpy(), open(f"logs/mask/label_e{epoch}_{i}.pkl","wb")) end_time = time.time() final_msg = { "epoch": f"EP{epoch}_{phase}", "avg_loss": avg_loss / len(data_iter), "total_masked_acc": total_correct_mask * 100.0 / total_element_mask if total_element_mask != 0 else 0, "time_taken_from_start": end_time - self.start_time } if self.same_student_prediction: final_msg["total_prediction_acc"] = total_correct_pred * 100.0 / total_element_pred print(final_msg) f.close() sys.stdout = sys.__stdout__ if phase == "val": self.save_model = False if self.avg_loss > (avg_loss / len(data_iter)): self.save_model = True self.avg_loss = (avg_loss / len(data_iter)) # pickle.dump(bert_hidden_representations, open(f"embeddings/{code}/{str_code}_embeddings_{epoch}.pkl","wb")) def save(self, epoch, file_path="output/bert_trained.model"): """ Saving the current BERT model on file_path :param epoch: current epoch number :param file_path: model output path which gonna be file_path+"ep%d" % epoch :return: final_output_path """ # if self.code: # fpath = file_path.split("/") # # output_path = fpath[0]+ "/"+ fpath[1]+f"/{self.code}/" + fpath[2] + ".ep%d" % epoch # output_path = "/",join(fpath[0]+ "/"+ fpath[1]+f"/{self.code}/" + fpath[-1] + ".ep%d" % epoch # else: output_path = file_path + ".ep%d" % epoch torch.save(self.bert.cpu(), output_path) self.bert.to(self.device) print("EP:%d Model Saved on:" % epoch, output_path) return output_path class BERTFineTuneTrainer: def __init__(self, bert: BERT, vocab_size: int, train_dataloader: DataLoader, test_dataloader: DataLoader = None, lr: float = 1e-4, betas=(0.9, 0.999), weight_decay: float = 0.01, warmup_steps=10000, with_cuda: bool = True, cuda_devices=None, log_freq: int = 10, workspace_name=None, num_labels=2, finetune_task=""): """ :param bert: BERT model which you want to train :param vocab_size: total word vocab size :param train_dataloader: train dataset data loader :param test_dataloader: test dataset data loader [can be None] :param lr: learning rate of optimizer :param betas: Adam optimizer betas :param weight_decay: Adam optimizer weight decay param :param with_cuda: traning with cuda :param log_freq: logging frequency of the batch iteration """ # Setup cuda device for BERT training, argument -c, --cuda should be true cuda_condition = torch.cuda.is_available() and with_cuda self.device = torch.device("cuda:0" if cuda_condition else "cpu") print(with_cuda, cuda_condition, " Device used = ", self.device) # This BERT model will be saved every epoch self.bert = bert for param in self.bert.parameters(): param.requires_grad = False # Initialize the BERT Language Model, with BERT model self.model = BERTForClassification(self.bert, vocab_size, num_labels).to(self.device) # Distributed GPU training if CUDA can detect more than 1 GPU if with_cuda and torch.cuda.device_count() > 1: print("Using %d GPUS for BERT" % torch.cuda.device_count()) self.model = nn.DataParallel(self.model, device_ids=cuda_devices) # Setting the train and test data loader self.train_data = train_dataloader self.test_data = test_dataloader self.optim = Adam(self.model.parameters(), lr=lr, weight_decay=weight_decay) #, eps=1e-9 # self.scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', patience=2, factor=0.1) if num_labels == 1: self.criterion = nn.MSELoss() elif num_labels == 2: self.criterion = nn.BCEWithLogitsLoss() # self.criterion = nn.CrossEntropyLoss() elif num_labels > 2: self.criterion = nn.CrossEntropyLoss() # self.criterion = nn.BCEWithLogitsLoss() # self.ece_criterion = ECE().to(self.device) self.log_freq = log_freq self.workspace_name = workspace_name self.finetune_task = finetune_task self.save_model = False self.avg_loss = 10000 self.start_time = time.time() self.probability_list = [] print("Total Parameters:", sum([p.nelement() for p in self.model.parameters()])) def train(self, epoch): self.iteration(epoch, self.train_data) def test(self, epoch): self.iteration(epoch, self.test_data, train=False) def iteration(self, epoch, data_loader, train=True): """ loop over the data_loader for training or testing if on train status, backward operation is activated and also auto save the model every peoch :param epoch: current epoch index :param data_loader: torch.utils.data.DataLoader for iteration :param train: boolean value of is train or test :return: None """ str_code = "train" if train else "test" self.log_file = f"{self.workspace_name}/logs/{self.finetune_task}/log_{str_code}_finetuned.txt" if epoch == 0: f = open(self.log_file, 'w') f.close() if not train: self.avg_loss = 10000 # Setting the tqdm progress bar data_iter = tqdm.tqdm(enumerate(data_loader), desc="EP_%s:%d" % (str_code, epoch), total=len(data_loader), bar_format="{l_bar}{r_bar}") avg_loss = 0.0 total_correct = 0 total_element = 0 plabels = [] tlabels = [] eval_accurate_nb = 0 nb_eval_examples = 0 logits_list = [] labels_list = [] if train: self.model.train() else: self.model.eval() self.probability_list = [] with open(self.log_file, 'a') as f: sys.stdout = f for i, data in data_iter: # 0. batch_data will be sent into the device(GPU or cpu) data = {key: value.to(self.device) for key, value in data.items()} if train: h_rep, logits = self.model.forward(data["bert_input"], data["segment_label"]) else: with torch.no_grad(): h_rep, logits = self.model.forward(data["bert_input"], data["segment_label"]) # print(logits, logits.shape) logits_list.append(logits.cpu()) labels_list.append(data["progress_status"].cpu()) # print(">>>>>>>>>>>>", progress_output) # print(f"{epoch}---nelement--- {data['progress_status'].nelement()}") # print(data["progress_status"].shape, logits.shape) progress_loss = self.criterion(logits, data["progress_status"]) loss = progress_loss if torch.cuda.device_count() > 1: loss = loss.mean() # 3. backward and optimization only in train if train: self.optim.zero_grad() loss.backward() # torch.nn.utils.clip_grad_norm_(self.model.parameters(), 1.0) self.optim.step() # progress prediction accuracy # correct = progress_output.argmax(dim=-1).eq(data["progress_status"]).sum().item() probs = nn.LogSoftmax(dim=-1)(logits) self.probability_list.append(probs) predicted_labels = torch.argmax(probs, dim=-1) true_labels = torch.argmax(data["progress_status"], dim=-1) plabels.extend(predicted_labels.cpu().numpy()) tlabels.extend(true_labels.cpu().numpy()) # Compare predicted labels to true labels and calculate accuracy correct = (predicted_labels == true_labels).sum().item() avg_loss += loss.item() total_correct += correct # total_element += true_labels.nelement() total_element += data["progress_status"].nelement() # print(">>>>>>>>>>>>>>", predicted_labels, true_labels, correct, total_correct, total_element) # if train: post_fix = { "epoch": epoch, "iter": i, "avg_loss": avg_loss / (i + 1), "avg_acc": total_correct / total_element * 100, "loss": loss.item() } # else: # logits = logits.detach().cpu().numpy() # label_ids = data["progress_status"].to('cpu').numpy() # tmp_eval_nb = accurate_nb(logits, label_ids) # eval_accurate_nb += tmp_eval_nb # nb_eval_examples += label_ids.shape[0] # # total_element += data["progress_status"].nelement() # # avg_loss += loss.item() # post_fix = { # "epoch": epoch, # "iter": i, # "avg_loss": avg_loss / (i + 1), # "avg_acc": tmp_eval_nb / total_element * 100, # "loss": loss.item() # } if i % self.log_freq == 0: data_iter.write(str(post_fix)) # precisions = precision_score(plabels, tlabels, average="weighted") # recalls = recall_score(plabels, tlabels, average="weighted") f1_scores = f1_score(plabels, tlabels, average="weighted") # if train: end_time = time.time() final_msg = { "epoch": f"EP{epoch}_{str_code}", "avg_loss": avg_loss / len(data_iter), "total_acc": total_correct * 100.0 / total_element, # "precisions": precisions, # "recalls": recalls, "f1_scores": f1_scores, "time_taken_from_start": end_time - self.start_time } # else: # eval_accuracy = eval_accurate_nb/nb_eval_examples # logits_ece = torch.cat(logits_list) # labels_ece = torch.cat(labels_list) # ece = self.ece_criterion(logits_ece, labels_ece).item() # end_time = time.time() # final_msg = { # "epoch": f"EP{epoch}_{str_code}", # "eval_accuracy": eval_accuracy, # "ece": ece, # "avg_loss": avg_loss / len(data_iter), # "precisions": precisions, # "recalls": recalls, # "f1_scores": f1_scores, # "time_taken_from_start": end_time - self.start_time # } # if self.save_model: # conf_hist = visualization.ConfidenceHistogram() # plt_test = conf_hist.plot(np.array(logits_ece), np.array(labels_ece), title= f"Confidence Histogram {epoch}") # plt_test.savefig(f"{self.workspace_name}/plots/confidence_histogram/{self.finetune_task}/conf_histogram_test_{epoch}.png",bbox_inches='tight') # plt_test.close() # rel_diagram = visualization.ReliabilityDiagram() # plt_test_2 = rel_diagram.plot(np.array(logits_ece), np.array(labels_ece),title=f"Reliability Diagram {epoch}") # plt_test_2.savefig(f"{self.workspace_name}/plots/confidence_histogram/{self.finetune_task}/rel_diagram_test_{epoch}.png",bbox_inches='tight') # plt_test_2.close() print(final_msg) # print("EP%d_%s, avg_loss=" % (epoch, str_code), avg_loss / len(data_iter), "total_acc=", total_correct * 100.0 / total_element) f.close() sys.stdout = sys.__stdout__ self.save_model = False if self.avg_loss > (avg_loss / len(data_iter)): self.save_model = True self.avg_loss = (avg_loss / len(data_iter)) def iteration_1(self, epoch_idx, data): try: data = {key: value.to(self.device) for key, value in data.items()} logits = self.model(data['input_ids'], data['segment_label']) # Ensure logits is a tensor, not a tuple loss_fct = nn.CrossEntropyLoss() loss = loss_fct(logits, data['labels']) # Backpropagation and optimization self.optim.zero_grad() loss.backward() self.optim.step() if self.log_freq > 0 and epoch_idx % self.log_freq == 0: print(f"Epoch {epoch_idx}: Loss = {loss.item()}") return loss except Exception as e: print(f"Error during iteration: {e}") raise # plt_test.show() # print("EP%d_%s, " % (epoch, str_code)) def save(self, epoch, file_path="output/bert_fine_tuned_trained.model"): """ Saving the current BERT model on file_path :param epoch: current epoch number :param file_path: model output path which gonna be file_path+"ep%d" % epoch :return: final_output_path """ if self.finetune_task: fpath = file_path.split("/") output_path = fpath[0]+ "/"+ fpath[1]+f"/{self.finetune_task}/" + fpath[2] + ".ep%d" % epoch else: output_path = file_path + ".ep%d" % epoch torch.save(self.model.cpu(), output_path) self.model.to(self.device) print("EP:%d Model Saved on:" % epoch, output_path) return output_path class BERTAttention: def __init__(self, bert: BERT, vocab_obj, train_dataloader: DataLoader, workspace_name=None, code=None, finetune_task=None, with_cuda=True): # available_gpus = list(range(torch.cuda.device_count())) cuda_condition = torch.cuda.is_available() and with_cuda self.device = torch.device("cuda:0" if cuda_condition else "cpu") print(with_cuda, cuda_condition, " Device used = ", self.device) self.bert = bert.to(self.device) # if with_cuda and torch.cuda.device_count() > 1: # print("Using %d GPUS for BERT" % torch.cuda.device_count()) # self.bert = nn.DataParallel(self.bert, device_ids=available_gpus) self.train_dataloader = train_dataloader self.workspace_name = workspace_name self.code = code self.finetune_task = finetune_task self.vocab_obj = vocab_obj def getAttention(self): # self.log_file = f"{self.workspace_name}/logs/{self.code}/log_attention.txt" labels = ['PercentChange', 'NumeratorQuantity2', 'NumeratorQuantity1', 'DenominatorQuantity1', 'OptionalTask_1', 'EquationAnswer', 'NumeratorFactor', 'DenominatorFactor', 'OptionalTask_2', 'FirstRow1:1', 'FirstRow1:2', 'FirstRow2:1', 'FirstRow2:2', 'SecondRow', 'ThirdRow', 'FinalAnswer','FinalAnswerDirection'] df_all = pd.DataFrame(0.0, index=labels, columns=labels) # Setting the tqdm progress bar data_iter = tqdm.tqdm(enumerate(self.train_dataloader), desc="attention", total=len(self.train_dataloader), bar_format="{l_bar}{r_bar}") count = 0 for i, data in data_iter: data = {key: value.to(self.device) for key, value in data.items()} a = self.bert.forward(data["bert_input"], data["segment_label"]) non_zero = np.sum(data["segment_label"].cpu().detach().numpy()) # Last Transformer Layer last_layer = self.bert.attention_values[-1].transpose(1,0,2,3) # print(last_layer.shape) head, d_model, s, s = last_layer.shape for d in range(d_model): seq_labels = self.vocab_obj.to_sentence(data["bert_input"].cpu().detach().numpy().tolist()[d])[1:non_zero-1] # df_all = pd.DataFrame(0.0, index=seq_labels, columns=seq_labels) indices_to_choose = defaultdict(int) for k,s in enumerate(seq_labels): if s in labels: indices_to_choose[s] = k indices_chosen = list(indices_to_choose.values()) selected_seq_labels = [s for l,s in enumerate(seq_labels) if l in indices_chosen] # print(len(seq_labels), len(selected_seq_labels)) for h in range(head): # fig, ax = plt.subplots(figsize=(12, 12)) # seq_labels = self.vocab_obj.to_sentence(data["bert_input"].cpu().detach().numpy().tolist()[d])#[1:non_zero-1] # seq_labels = self.vocab_obj.to_sentence(data["bert_input"].cpu().detach().numpy().tolist()[d])[1:non_zero-1] # indices_to_choose = defaultdict(int) # for k,s in enumerate(seq_labels): # if s in labels: # indices_to_choose[s] = k # indices_chosen = list(indices_to_choose.values()) # selected_seq_labels = [s for l,s in enumerate(seq_labels) if l in indices_chosen] # print(f"Chosen index: {seq_labels, indices_to_choose, indices_chosen, selected_seq_labels}") df_cm = pd.DataFrame(last_layer[h][d][indices_chosen,:][:,indices_chosen], index = selected_seq_labels, columns = selected_seq_labels) df_all = df_all.add(df_cm, fill_value=0) count += 1 # df_cm = pd.DataFrame(last_layer[h][d][1:non_zero-1,:][:,1:non_zero-1], index=seq_labels, columns=seq_labels) # df_all = df_all.add(df_cm, fill_value=0) # df_all = df_all.reindex(index=seq_labels, columns=seq_labels) # sns.heatmap(df_all, annot=False) # plt.title("Attentions") #Probabilities # plt.xlabel("Steps") # plt.ylabel("Steps") # plt.grid(True) # plt.tick_params(axis='x', bottom=False, top=True, labelbottom=False, labeltop=True, labelrotation=90) # plt.savefig(f"{self.workspace_name}/plots/{self.code}/{self.finetune_task}_attention_scores_over_[{h}]_head_n_data[{d}].png", bbox_inches='tight') # plt.show() # plt.close() print(f"Count of total : {count, head * self.train_dataloader.dataset.len}") df_all = df_all.div(count) # head * self.train_dataloader.dataset.len df_all = df_all.reindex(index=labels, columns=labels) sns.heatmap(df_all, annot=False) plt.title("Attentions") #Probabilities plt.xlabel("Steps") plt.ylabel("Steps") plt.grid(True) plt.tick_params(axis='x', bottom=False, top=True, labelbottom=False, labeltop=True, labelrotation=90) plt.savefig(f"{self.workspace_name}/plots/{self.code}/{self.finetune_task}_attention_scores.png", bbox_inches='tight') plt.show() plt.close()