File size: 5,195 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 | import torch
import logging
import copy
import torch.nn.functional as F
from tqdm import trange, tqdm
from sklearn.metrics import confusion_matrix
from losses import loss_map
from utils.metrics import clustering_score
from utils.functions import restore_model, save_model
class MCLManager:
def __init__(self, args, data, model, logger_name = 'Discovery'):
self.logger = logging.getLogger(logger_name)
self.num_labels = data.num_labels
loader = data.dataloader
self.train_dataloader, self.eval_dataloader, self.test_dataloader = \
loader.train_outputs['loader'], loader.eval_outputs['loader'], loader.test_outputs['loader']
backbone = args.backbone
args.backbone = backbone
self.model = model.set_model(args, data, 'bert')
self.optimizer , self.scheduler = model.set_optimizer(self.model, data.dataloader.num_train_examples, args.train_batch_size, \
args.num_train_epochs, args.lr, args.warmup_proportion)
self.device = model.device
self.loss_fct = loss_map[args.loss_fct]
if not args.train:
self.model = restore_model(self.model, args.model_output_dir)
def train(self, args, data):
best_model = None
wait = 0
best_eval_score = 0
for epoch in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
self.model.train()
for batch in tqdm(self.train_dataloader, desc="Training(All)"):
batch = tuple(t.to(self.device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
loss = self.model(input_ids, segment_ids, input_mask, label_ids, mode = 'train', loss_fct = self.loss_fct)
loss.backward()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
self.optimizer.step()
self.scheduler.step()
self.optimizer.zero_grad()
tr_loss = tr_loss / nb_tr_steps
y_true, y_pred = self.get_outputs(args, mode = 'eval')
eval_score = clustering_score(y_true, y_pred)['NMI']
eval_results = {
'train_loss': tr_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 > 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 get_outputs(self, args, mode = 'eval', get_feats = False):
if 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_logits = torch.empty((0, args.num_labels)).to(self.device)
total_features = torch.empty((0, args.feat_dim)).to(self.device)
total_preds = torch.empty(0, dtype=torch.long).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):
features, logits = self.model(input_ids, segment_ids, input_mask)
total_labels = torch.cat((total_labels, label_ids))
total_logits = torch.cat((total_logits, logits))
total_features = torch.cat((total_features, features))
if get_feats:
feats = total_features.cpu().numpy()
return feats
else:
total_probs = F.softmax(total_logits.detach(), dim = 1)
total_maxprobs, total_preds = total_probs.max(dim = 1)
y_true = total_labels.cpu().numpy()
y_pred = total_preds.cpu().numpy()
return y_true, y_pred
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
return test_results
|