import torch import torch.nn.functional as F import numpy as np import os import copy import logging import pandas as pd from sklearn.cluster import KMeans from utils.metrics import clustering_score from sklearn.metrics import accuracy_score, confusion_matrix from tqdm import trange, tqdm from torch.utils.data import (DataLoader, SequentialSampler, RandomSampler, TensorDataset) from losses import loss_map from utils.functions import save_model, restore_model, set_seed from utils.faster_mix_k_means_pytorch import K_Means as SemiSupKMeans from scipy.optimize import minimize_scalar from functools import partial from sklearn.metrics.cluster import normalized_mutual_info_score as nmi_score from sklearn.metrics import adjusted_rand_score as ari_score from scipy.optimize import linear_sum_assignment as linear_assignment class GCDManager: def __init__(self, args, data, model, logger_name = 'Discovery'): self.logger = logging.getLogger(logger_name) set_seed(args.seed) loader = data.dataloader self.loader = data.dataloader self.train_dataloader, self.eval_dataloader, self.test_dataloader = \ loader.train_outputs['loader'], loader.eval_outputs['loader'], loader.test_outputs['loader'] self.train_input_ids, self.train_input_mask, self.train_segment_ids, self.train_label_ids= \ loader.train_outputs['input_ids'], loader.train_outputs['input_mask'], loader.train_outputs['segment_ids'], loader.train_outputs['label_ids'] self.aug_train_dataloader = self.get_augment_dataloader(args, self.train_label_ids, data_aug = True) self.set_model_optimizer(args, data, model) self.num_labels = data.num_labels self.temperature=0.07 self.sup_con_weight = 0.5 self.loss_fct = loss_map[args.loss_fct] if not args.train: self.model = restore_model(self.model, args.model_output_dir) def set_model_optimizer(self, args, data, model): self.model = model.set_model(args, data, 'bert', args.freeze_train_bert_parameters) self.optimizer , self.scheduler = model.set_optimizer(self.model, len(data.dataloader.train_examples), args.train_batch_size, \ args.num_train_epochs, args.lr, args.warmup_proportion) self.device = model.device def batch_chunk(self, x): x1, x2 = torch.chunk(input=x, chunks=2, dim=1) x1, x2 = x1.squeeze(1), x2.squeeze(1) return x1, x2 def semisupvised_kmeans(self, args): # Semi-Kmeans feats, all_labels = self.get_outputs(args, mode = 'train') l_index = [k for k,i in enumerate(all_labels) if i !=-1] u_index = [k for k,i in enumerate(all_labels) if i ==-1] print('Fitting Semi-Supervised K-Means...') kmeans = SemiSupKMeans(k=self.num_labels, tolerance=1e-4, max_iterations=200, init='k-means++', n_init=100, random_state=args.seed, n_jobs=None, pairwise_batch_size=1024, mode=None) u_feats = feats[u_index] l_feats = feats[l_index] l_targets = all_labels[l_index] u_targets = all_labels[u_index] l_feats, u_feats, l_targets, u_targets = (torch.from_numpy(x).to(self.device) for x in (l_feats, u_feats, l_targets, u_targets)) kmeans.fit_mix(u_feats, l_feats, l_targets) self.semisupvised_kmeans_cluster = kmeans.cluster_centers_ def train(self, args, data): wait = 0 best_model = None best_eval_score = 0 criterion = loss_map['SupConLoss'] for epoch in trange(int(args.num_train_epochs), desc="Epoch"): self.model.train() tr_loss = 0 nb_tr_examples, nb_tr_steps = 0, 0 train_acc = 0 for step, batch in enumerate(tqdm(self.aug_train_dataloader, desc="Iteration")): batch = tuple(t.to(self.device) for t in batch) input_ids, input_mask, segment_ids, label_ids = batch with torch.set_grad_enabled(True): input_ids_a, input_ids_b = self.batch_chunk(input_ids) input_mask_a, input_mask_b = self.batch_chunk(input_mask) segment_ids_a, segment_ids_b = self.batch_chunk(segment_ids) label_ids = torch.chunk(input=label_ids, chunks=2, dim=1)[0][:, 0] x_a = self.model(input_ids_a, segment_ids_a, input_mask_a, mode = 'train') x_b = self.model(input_ids_b, segment_ids_b, input_mask_b, mode = 'train') aug_mlp_outputs_a = self.model.mlp_head(x_a) aug_mlp_outputs_b = self.model.mlp_head(x_b) norm_logits = F.normalize(aug_mlp_outputs_a) norm_aug_logits = F.normalize(aug_mlp_outputs_b) contrastive_feats = torch.cat((norm_logits, norm_aug_logits)) contrastive_logits, contrastive_labels = self.info_nce_logits(features=contrastive_feats) contrastive_loss = self.loss_fct(contrastive_logits, contrastive_labels) mask_lab = torch.from_numpy(np.array([0 if i ==-1 else 1 for i in label_ids])).bool() f1, f2 = [f[mask_lab] for f in contrastive_feats.chunk(2)] sup_con_feats = torch.cat([f1.unsqueeze(1), f2.unsqueeze(1)], dim=1) sup_con_labels = label_ids[mask_lab] sup_loss = criterion(features = sup_con_feats, labels = sup_con_labels, device = self.device) loss = self.sup_con_weight * sup_loss + (1 - self.sup_con_weight) * contrastive_loss self.optimizer.zero_grad() loss.backward() tr_loss += loss.item() nb_tr_examples += input_ids.size(0) nb_tr_steps += 1 self.optimizer.step() self.scheduler.step() train_loss = tr_loss / nb_tr_steps features, y_true = self.get_outputs(args, mode = 'eval') km = KMeans(n_clusters = int(data.n_known_cls), random_state=args.seed).fit(features) y_pred = km.labels_ eval_score = clustering_score(y_true, y_pred) eval_results = { 'train_loss': train_loss, 'eval_score': eval_score, 'best_score':best_eval_score, } self.logger.info("***** Epoch: %s: Eval results *****", str(epoch + 1)) for key in sorted(eval_results.keys()): self.logger.info(" %s = %s", key, str(eval_results[key])) if eval_score['ACC'] > best_eval_score: best_model = copy.deepcopy(self.model) wait = 0 best_eval_score = eval_score['ACC'] else: wait += 1 if wait >= args.wait_patient: break self.logger.info('GCD training finished...') self.model = best_model if args.save_model: save_model(self.model, args.model_output_dir) self.semisupvised_kmeans(args) def get_outputs(self, args, mode): if mode == 'train': dataloader = self.train_dataloader elif mode == 'eval': dataloader = self.eval_dataloader elif mode == 'test': dataloader = self.test_dataloader self.model.eval() total_labels = torch.empty(0,dtype=torch.long).to(self.device) total_features = torch.empty((0,args.feat_dim)).to(self.device) for batch in tqdm(dataloader, desc="Iteration"): batch = tuple(t.to(self.device) for t in batch) input_ids, input_mask, segment_ids, label_ids = batch with torch.set_grad_enabled(False): pooled_output = self.model(input_ids, segment_ids, input_mask) total_labels = torch.cat((total_labels,label_ids)) total_features = torch.cat((total_features, pooled_output)) feats = total_features.cpu().numpy() y_true = total_labels.cpu().numpy() return feats, y_true def info_nce_logits(self, features): b_ = 0.5 * int(features.size(0)) labels = torch.cat([torch.arange(b_) for i in range(2)], dim=0) labels = (labels.unsqueeze(0) == labels.unsqueeze(1)).float() labels = labels.to(self.device) features = F.normalize(features, dim=1) similarity_matrix = torch.matmul(features, features.T) mask = torch.eye(labels.shape[0], dtype=torch.bool).to(self.device) labels = labels[~mask].view(labels.shape[0], -1) similarity_matrix = similarity_matrix[~mask].view(similarity_matrix.shape[0], -1) positives = similarity_matrix[labels.bool()].view(labels.shape[0], -1) negatives = similarity_matrix[~labels.bool()].view(similarity_matrix.shape[0], -1) logits = torch.cat([positives, negatives], dim=1) labels = torch.zeros(logits.shape[0], dtype=torch.long).to(self.device) logits = logits / self.temperature return logits, labels def get_augment_dataloader(self, args, pseudo_labels, data_aug = False): train_input_ids = self.train_input_ids.unsqueeze(1) train_input_mask = self.train_input_mask.unsqueeze(1) train_segment_ids = self.train_segment_ids.unsqueeze(1) train_label_ids = torch.tensor(pseudo_labels).unsqueeze(1) train_input_ids = torch.cat(([train_input_ids, train_input_ids]), dim = 1) train_input_mask = torch.cat(([train_input_mask, train_input_mask]), dim = 1) train_segment_ids = torch.cat(([train_segment_ids, train_segment_ids]), dim = 1) train_label_ids = torch.cat(([train_label_ids, train_label_ids]), dim = 1) train_data = TensorDataset(train_input_ids, train_input_mask, train_segment_ids, train_label_ids) label_len = len(self.loader.train_labeled_examples) unlabelled_len = len(self.loader.train_unlabeled_examples) sample_weights = [1 if i < label_len else label_len / unlabelled_len for i in range(len(self.loader.train_examples))] sample_weights = torch.DoubleTensor(sample_weights) sampler = torch.utils.data.WeightedRandomSampler(sample_weights, num_samples=len(self.loader.train_examples)) train_dataloader = DataLoader(train_data, sampler = sampler, batch_size = args.train_batch_size) return train_dataloader def test(self, args, data): feats, y_true = self.get_outputs(args, mode = 'test') centers = self.semisupvised_kmeans_cluster print("self.semisupvised_kmeans_cluster", self.semisupvised_kmeans_cluster) dis = (torch.from_numpy(feats).to(self.device).unsqueeze(dim=1)-centers.unsqueeze(dim=0))**2 dis = dis.sum(dim = -1) u_mindist, y_pred = torch.min(dis, dim=1) y_pred = y_pred.cpu().numpy() test_results = clustering_score(y_true, y_pred) cm = confusion_matrix(y_true, y_pred) self.logger.info self.logger.info("***** Test: Confusion Matrix *****") self.logger.info("%s", str(cm)) self.logger.info("***** Test results *****") for key in sorted(test_results.keys()): self.logger.info(" %s = %s", key, str(test_results[key])) test_results['y_true'] = y_true test_results['y_pred'] = y_pred return test_results