VHM/train.py at main · XL2248/VHM · GitHub

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
# -*- coding: utf-8 -*-
import logging

import torch
import torch.cuda
import os
from beaver.data import build_dataset
from beaver.infer import beam_search
from beaver.loss import WarmAdam, LabelSmoothingLoss
from beaver.model import NMTModel
from beaver.utils import Saver
from beaver.utils import calculate_bleu
from beaver.utils import parseopt, get_device, printing_opt
from beaver.utils.metric import calculate_rouge

logging.basicConfig(format="%(asctime)s - %(message)s", level=logging.INFO)
opt = parseopt.parse_train_args()

device = get_device()

logging.info("\n" + printing_opt(opt))

saver = Saver(opt)

def get_parameter_number(model):
    total_num = sum(p.numel() for p in model.parameters())
    trainable_num = sum(p.numel() for p in model.parameters() if p.requires_grad)
    logging.info('Trainable', str(trainable_num), str(total_num))
    return {'Total': total_num, 'Trainable': trainable_num}

def valid(model, criterion_task1, criterion_task2, criterion_task3, valid_dataset, step, typ):
    model.eval()
    total_n = 0
    total_task1_loss = total_task2_loss =  total_task3_loss = 0.0
    task1_hypothesis, task1_references = [], []
    task2_hypothesis, task2_references = [], []
    task3_hypothesis, task3_references = [], []

    for i, (batch, flag) in enumerate(valid_dataset):

        scores, kl_loss = model(batch.src, batch.tgt, flag)
        _, predictions = scores.topk(k=1, dim=-1)
        if flag == 1:
            loss = criterion_task1(scores, batch.tgt)
        elif flag == 2:
            loss = criterion_task2(scores, batch.tgt)
        else: #if flag == 3:
            loss = criterion_task3(scores, batch.tgt)
#       _, predictions = scores.topk(k=1, dim=-1)

        if flag == 1:  # task1
            total_task1_loss += loss.data
            task1_hypothesis += [valid_dataset.fields["task1_tgt"].decode(p) for p in predictions]
            task1_references += [valid_dataset.fields["task1_tgt"].decode(t) for t in batch.tgt]
        elif flag == 2:
            total_task2_loss += loss.data
            task2_hypothesis += [valid_dataset.fields["task2_tgt"].decode(p) for p in predictions]
            task2_references += [valid_dataset.fields["task2_tgt"].decode(t) for t in batch.tgt]
        else:
            total_task3_loss += loss.data
            task3_hypothesis += [valid_dataset.fields["task3_tgt"].decode(p) for p in predictions]
            task3_references += [valid_dataset.fields["task3_tgt"].decode(t) for t in batch.tgt]

        total_n += 1

    bleu_task1 = calculate_bleu(task1_hypothesis, task1_references)
    bleu_task2 = calculate_bleu(task2_hypothesis, task2_references)
    bleu_task3 = calculate_bleu(task3_hypothesis, task3_references)
    rouge1_task1, rouge2_task1 = calculate_rouge(task1_hypothesis, task1_references)
    rouge1_task2, rouge2_task2 = calculate_rouge(task2_hypothesis, task2_references)
    rouge1_task3, rouge2_task3 = calculate_rouge(task3_hypothesis, task3_references)
    mean_task1_loss = total_task1_loss / total_n
    mean_task2_loss = total_task2_loss / total_n
    mean_task3_loss = total_task3_loss / total_n
    if typ == "test":
        with open(os.path.join(opt.model_path + "_gpu" + str(opt.gpu) + "_warmup" + str(opt.warm_up) + "_latent" + str(opt.latent_dim) + "_kl" + str(opt.kl_annealing_steps) + "_split" + str(opt.split), "prediction."+str(step)), "w", encoding="UTF-8") as out_file:
            out_file.write("\n".join(task3_hypothesis))
            out_file.write("\n")

    logging.info("type: %s \t loss-mt: %.2f \t loss-mls %.2f \t loss-cls %.2f \t bleu-mt: %3.2f\t bleu-mls: %3.2f \t bleu-cls: %3.2f \t rouge1-mt: %3.2f \t rouge1-mls: %3.2f \t rouge1-cls: %3.2f \t rouge2-mt: %3.2f \t rouge2-mls: %3.2f \t rouge2-cls: %3.2f"
                 % (typ, mean_task1_loss, mean_task2_loss, mean_task3_loss, bleu_task1, bleu_task2, bleu_task3, rouge1_task1, rouge1_task2, rouge1_task3, rouge2_task1, rouge2_task2, rouge2_task3))
    checkpoint = {"model": model.state_dict(), "opt": opt}
    saver.save(checkpoint, step, mean_task1_loss, mean_task2_loss, mean_task3_loss, bleu_task1, bleu_task2, bleu_task3, rouge1_task1, rouge1_task2, rouge1_task3, rouge2_task1, rouge2_task2, rouge2_task3, typ)


def train(model, criterion_task1, criterion_task2, criterion_task3, optimizer, train_dataset, valid_dataset, test_dataset):
    total_task1_loss = total_task2_loss = total_task3_loss = kl_mt_losses = kl_cls_losses = kl_mls_losses = 0.0
    model.zero_grad()

    for i, (batch, flag) in enumerate(train_dataset):
        kl_weights = min(optimizer.n_step * 1.0 / opt.kl_annealing_steps, 1.0)
        scores, kl_loss = model(batch.src, batch.tgt, flag, True)
#        logging.info(flag, batch.src.size(), batch.tgt.size(), scores.size())
        if flag == 1:
            loss = criterion_task1(scores, batch.tgt)
            kl_mt_loss = kl_loss
        elif flag == 2:
            loss = criterion_task2(scores, batch.tgt)
            kl_mls_loss = kl_loss
        else: # flag == 3:
            loss = criterion_task3(scores, batch.tgt)
            kl_cls_loss = kl_loss
        loss += kl_weights * kl_loss
        loss.backward()

        if flag == 1:  # task1
            total_task1_loss += loss.data
            kl_mt_losses += kl_mt_loss.data
        elif flag == 2:
            total_task2_loss += loss.data
            kl_mls_losses += kl_mls_loss.data
        else:
            total_task3_loss += loss.data
            kl_cls_losses += kl_cls_loss.data

        if (i + 1) % opt.grad_accum == 0:
            optimizer.step()
            model.zero_grad()

            if optimizer.n_step % opt.report_every == 0:
                mean_task1_loss = total_task1_loss / opt.report_every / opt.grad_accum * 2
                mean_task2_loss = total_task2_loss / opt.report_every / opt.grad_accum * 2
                mean_task3_loss = total_task3_loss / opt.report_every / opt.grad_accum * 2
                mean_kl_mt = kl_mt_losses / opt.report_every / opt.grad_accum * 2
                mean_kl_cls = kl_cls_losses / opt.report_every / opt.grad_accum * 2
                mean_kl_mls = kl_mls_losses / opt.report_every / opt.grad_accum * 2
                logging.info("step: %7d\t loss-mt: %.4f \t loss-mls: %.4f \t loss-cls: %.4f \t kl_weights: %.8f \t kl-mt: %.4f \t kl-mls: %.4f \t kl-cls: %.4f"
                             % (optimizer.n_step, mean_task1_loss, mean_task2_loss, mean_task3_loss, kl_weights, mean_kl_mt, mean_kl_mls, mean_kl_cls))
                total_task1_loss = total_task2_loss = total_task3_loss = kl_mt_losses = kl_cls_losses = kl_mls_losses = 0.0

            if optimizer.n_step % opt.save_every == 0: # and optimizer.n_step > 100000:
                with torch.set_grad_enabled(False):
                    valid(model, criterion_task1, criterion_task2, criterion_task3, valid_dataset, optimizer.n_step, "valid")
                    if optimizer.n_step > 400000 and optimizer.n_step % (2 * opt.save_every) == 0:
                        valid(model, criterion_task1, criterion_task2, criterion_task3, test_dataset, optimizer.n_step, "test")
                model.train()
        if optimizer.n_step % 850000 == 0:
            logging.info("Training DONE all steps  %7d" % optimizer.n_step)
            break;
        del loss


def main():
    logging.info("Build dataset...")
    train_dataset = build_dataset(opt, opt.train, opt.vocab, device, train=True)
    valid_dataset = build_dataset(opt, opt.valid, opt.vocab, device, train=False)
    test_dataset = build_dataset(opt, opt.test, opt.vocab, device, train=False)
    fields = valid_dataset.fields = train_dataset.fields = test_dataset.fields
    logging.info("Build model...")

    pad_ids = {"src": fields["src"].pad_id,
               "task1_tgt": fields["task1_tgt"].pad_id,
               "task2_tgt": fields["task2_tgt"].pad_id,
               "task3_tgt": fields["task3_tgt"].pad_id}
    vocab_sizes = {"src": len(fields["src"].vocab),
                   "task1_tgt": len(fields["task1_tgt"].vocab),
                   "task2_tgt": len(fields["task1_tgt"].vocab),
                   "task3_tgt": len(fields["task1_tgt"].vocab)}

#    model = NMTModel.load_model(opt, pad_ids, vocab_sizes).to(device)
    criterion_task1 = LabelSmoothingLoss(opt.label_smoothing, vocab_sizes["task1_tgt"], pad_ids["task1_tgt"]).to(device)
    criterion_task2 = LabelSmoothingLoss(opt.label_smoothing, vocab_sizes["task2_tgt"], pad_ids["task2_tgt"]).to(device)
    criterion_task3 = LabelSmoothingLoss(opt.label_smoothing, vocab_sizes["task3_tgt"], pad_ids["task3_tgt"]).to(device)
    checkpoint_num = []
    if os.path.exists(opt.model_path + "_gpu" + str(opt.gpu) + "_warmup" + str(opt.warm_up) + "_latent" + str(opt.latent_dim) + "_kl" + str(opt.kl_annealing_steps) + "_split" + str(opt.split)):
        files= os.listdir(opt.model_path + "_gpu" + str(opt.gpu) + "_warmup" + str(opt.warm_up) + "_latent" + str(opt.latent_dim) + "_kl" + str(opt.kl_annealing_steps) + "_split" + str(opt.split))
        for fil in files:
            if not os.path.isdir(fil) and len(fil) > 20:
                checkpoint_num.append(int(fil.split("-")[-1]))
        if len(checkpoint_num) > 0:
            opt.train_from = opt.model_path + "_gpu" + str(opt.gpu) + "_warmup" + str(opt.warm_up) + "_latent" + str(opt.latent_dim) + "_kl" + str(opt.kl_annealing_steps) + "_split" + str(opt.split) + "/checkpoint-step-%06d" % max(checkpoint_num)
    model = NMTModel.load_model(opt, pad_ids, vocab_sizes).to(device)
    logging.info("model parameters:", get_parameter_number(model))
    for name, parameters in model.named_parameters():
        print(name, ':', str(parameters.size()))
    n_step = int(opt.train_from.split("-")[-1]) if opt.train_from else 1
    optimizer = WarmAdam(model.parameters(), opt.lr, opt.hidden_size, opt.warm_up, n_step)

    logging.info("start training...")
    train(model, criterion_task1, criterion_task2, criterion_task3, optimizer, train_dataset, valid_dataset, test_dataset)


if __name__ == '__main__':
    main()