File size: 10,610 Bytes
2d06dcc | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 | import logging
import copy
import os
import random
import torch
import torch.nn.functional as F
import numpy as np
import math
import pandas as pd
from .pretrain import PretrainDTCManager
from sklearn.cluster import KMeans
from sklearn.metrics import confusion_matrix
from tqdm import trange, tqdm
from utils.metrics import clustering_score, clustering_accuracy_score
from utils.functions import save_model, restore_model, set_seed
from utils.faster_mix_k_means_pytorch import K_Means
from sklearn.metrics import silhouette_score
from scipy.optimize import linear_sum_assignment
from collections import Counter
class DTCManager:
def __init__(self, args, data, model, logger_name = 'Discovery'):
pretrain_manager = PretrainDTCManager(args, data, model)
set_seed(args.seed)
self.logger = logging.getLogger(logger_name)
loader = data.dataloader
self.train_dataloader, self.eval_dataloader, self.test_dataloader = \
loader.train_unlabeled_outputs['loader'], loader.eval_outputs['loader'], loader.test_outputs['loader']
if args.pretrain:
self.pretrained_model = pretrain_manager.model
self.set_model_optimizer(args, data, model, pretrain_manager)
self.load_pretrained_model(self.pretrained_model)
else:
self.pretrained_model = restore_model(pretrain_manager.model, os.path.join(args.method_output_dir, 'pretrain'))
self.set_model_optimizer(args, data, model, pretrain_manager)
if args.train:
self.load_pretrained_model(self.pretrained_model)
else:
self.model = restore_model(self.model, args.model_output_dir)
def set_model_optimizer(self, args, data, model, pretrain_manager):
if args.cluster_num_factor > 1:
args.num_labels = self.num_labels = pretrain_manager.num_labels
else:
args.num_labels = self.num_labels = data.num_labels
num_train_examples = len(data.dataloader.train_unlabeled_examples)
self.model = model.set_model(args, data, 'bert', args.freeze_bert_parameters)
self.warmup_optimizer, self.warmup_scheduler = model.set_optimizer(self.model, num_train_examples, args.train_batch_size, \
args.num_warmup_train_epochs, args.lr, args.warmup_proportion)
self.optimizer, self.scheduler = model.set_optimizer(self.model, num_train_examples, args.train_batch_size, \
args.num_train_epochs, args.lr, args.warmup_proportion)
self.device = model.device
self.model.to(self.device)
def initialize_centroids(self, args):
self.logger.info("Initialize centroids...")
feats = self.get_outputs(args, mode = 'train', get_feats = True)
km = KMeans(n_clusters=args.num_labels, n_jobs=-1, random_state=args.seed)
km.fit(feats)
self.logger.info("Initialization finished...")
self.model.cluster_layer.data = torch.tensor(km.cluster_centers_).to(self.device)
def warmup_train(self, args):
probs = self.get_outputs(args, mode = 'train', get_probs = True)
p_target = target_distribution(probs)
for epoch in trange(int(args.num_warmup_train_epochs), desc="Warmup_Epoch"):
tr_loss, nb_tr_examples, nb_tr_steps = 0, 0, 0
self.model.train()
for step, batch in enumerate(tqdm(self.train_dataloader, desc="Warmup_Training")):
batch = tuple(t.to(self.device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
logits, q = self.model(input_ids, segment_ids, input_mask)
loss = F.kl_div(q.log(),torch.Tensor(p_target[step * args.train_batch_size: (step+1) * args.train_batch_size]).to(self.device))
loss.backward()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
self.warmup_optimizer.step()
self.warmup_scheduler.step()
self.warmup_optimizer.zero_grad()
eval_true, eval_pred = self.get_outputs(args, mode = 'eval')
eval_score = clustering_score(eval_true, eval_pred)['NMI']
eval_results = {
'loss': tr_loss,
'eval_score': round(eval_score, 2)
}
self.logger.info("***** Epoch: %s: Eval results *****", str(epoch))
for key in sorted(eval_results.keys()):
self.logger.info(" %s = %s", key, str(eval_results[key]))
return p_target
def get_outputs(self, args, mode = 'eval', get_feats = False, get_probs = False):
if mode == 'eval':
dataloader = self.eval_dataloader
elif mode == 'test':
dataloader = self.test_dataloader
elif mode == 'train':
dataloader = self.train_dataloader
self.model.eval()
total_labels = torch.empty(0,dtype=torch.long).to(self.device)
total_preds = torch.empty(0,dtype=torch.long).to(self.device)
total_features = torch.empty((0, args.num_labels)).to(self.device)
total_probs = torch.empty((0, args.num_labels)).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):
logits, probs = self.model(input_ids, segment_ids, input_mask)
total_labels = torch.cat((total_labels, label_ids))
total_features = torch.cat((total_features, logits))
total_probs = torch.cat((total_probs, probs))
if get_feats:
feats = total_features.cpu().numpy()
return feats
elif get_probs:
return total_probs.cpu().numpy()
else:
total_preds = total_probs.argmax(1)
y_pred = total_preds.cpu().numpy()
y_true = total_labels.cpu().numpy()
return y_true, y_pred
def train(self, args, data):
self.initialize_centroids(args)
self.logger.info('WarmUp Training start...')
self.p_target = self.warmup_train(args)
self.logger.info('WarmUp Training finished...')
ntrain = len(data.dataloader.train_unlabeled_examples)
Z = torch.zeros(ntrain, args.num_labels).float().to(self.device)
z_ema = torch.zeros(ntrain, args.num_labels).float().to(self.device)
z_epoch = torch.zeros(ntrain, args.num_labels).float().to(self.device)
best_model = None
best_eval_score = 0
for epoch in trange(int(args.num_train_epochs), desc="Epoch"):
# Fine-tuning with auxiliary distribution
tr_loss, nb_tr_examples, nb_tr_steps = 0, 0, 0
self.model.train()
for step, batch in enumerate(self.train_dataloader):
batch = tuple(t.to(self.device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
logits, q = self.model(input_ids, segment_ids, input_mask)
z_epoch[step * args.train_batch_size: (step+1) * args.train_batch_size, :] = q
kl_loss = F.kl_div(q.log(), torch.tensor(self.p_target[step * args.train_batch_size: (step+1) * args.train_batch_size]).to(self.device))
kl_loss.backward()
tr_loss += kl_loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
self.optimizer.step()
self.scheduler.step()
self.optimizer.zero_grad()
z_epoch = torch.tensor(self.get_outputs(args, mode = 'train', get_probs = True)).to(self.device)
Z = args.alpha * Z + (1. - args.alpha) * z_epoch
z_ema = Z * (1. / (1. - args.alpha ** (epoch + 1)))
if epoch % args.update_interval == 0:
self.logger.info('updating target ...')
self.p_target = target_distribution(z_ema).float().to(self.device)
self.logger.info('updating finished ...')
eval_true, eval_pred = self.get_outputs(args, mode = 'eval')
eval_score = clustering_score(eval_true, eval_pred)['NMI']
train_loss = tr_loss / nb_tr_steps
eval_results = {
'train_loss': train_loss,
'best_eval_score': best_eval_score,
'eval_score': round(eval_score, 2),
}
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 > best_eval_score:
best_model = copy.deepcopy(self.model)
wait = 0
best_eval_score = eval_score
elif eval_score > 0:
wait += 1
if wait >= args.wait_patient:
break
self.model = best_model
if args.save_model:
save_model(self.model, args.model_output_dir)
def test(self, args, data):
y_true, y_pred = self.get_outputs(args,mode = 'test')
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
if args.cluster_num_factor > 1:
test_results['estimate_k'] = args.num_labels
return test_results
def load_pretrained_model(self, pretrained_model):
pretrained_dict = pretrained_model.state_dict()
classifier_params = ['cluster_layer', 'classifier.weight','classifier.bias']
pretrained_dict = {k: v for k, v in pretrained_dict.items() if k not in classifier_params}
self.model.load_state_dict(pretrained_dict, strict=False)
def target_distribution(q):
weight = q ** 2 / q.sum(0)
return (weight.T / weight.sum(1)).T
|