IWHVI/flowvae_run.py at master · artsobolev/IWHVI · 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
import argparse
import os

from tqdm import tqdm
import numpy as np
import pandas as pd

import tensorflow as tf

import flow_vae
import utils
import datasets


def main():
    datasets_available = [f[4:] for f in dir(datasets) if f.startswith('get_') and callable(getattr(datasets, f))]

    argparser = argparse.ArgumentParser(allow_abbrev=False)
    argparser.add_argument('--dataset', choices=datasets_available, default='dynamic_mnist')
    argparser.add_argument('--datasets_dir', default='./datasets/')
    argparser.add_argument('--no_extra_steps', type=int, default=0)
    argparser.add_argument('--q_extra_steps', type=int, default=0)

    argparser.add_argument('--learning_rate', type=float, default=1e-3)
    argparser.add_argument('--annealing_factor', type=float, default=1.0)
    argparser.add_argument('--annealing_speed', type=int, default=100)
    argparser.add_argument('--kl_warm_up_over', type=int, default=1)

    argparser.add_argument('--epochs', type=int, default=3000)
    # StepSchedules are in format [(step_to_begin_at, value_from_now_on_until_next_step)]
    argparser.add_argument('--train_learning_rate_step_schedule', nargs='*',
                           type=utils.make_typed_tuple(int, float), default=[])
    argparser.add_argument('--train_iwae_samples_step_schedule', nargs='*',
                           type=utils.make_typed_tuple(int, int), default=[])
    argparser.add_argument('--train_batch_size', type=int, default=256)
    argparser.add_argument('--train_attempts', type=int, default=10)

    argparser.add_argument('--val_iwae_samples', type=int, default=1000)
    argparser.add_argument('--val_batch_size', type=int, default=50 * 1000)

    argparser.add_argument('--evaluate_every', type=int, default=25)
    argparser.add_argument('--save_every', type=int, default=50)
    argparser.add_argument('--save_path', default=None)

    flow_vae.utils.add_model_args(argparser)
    args = argparser.parse_args()

    dataset = getattr(datasets, 'get_%s' % args.dataset)(args.datasets_dir)
    train_data = dataset.train
    val_data = dataset.validation

    dat_train = []
    dat_val = []

    print('Aguments:')
    for param_name, param_value in sorted(vars(args).items()):
        print('--{:30}: {}'.format(param_name, param_value))
    print('\n')

    print('{} dataset details:'.format(args.dataset))
    for split_name in ['train', 'test', 'validation']:
        print('{}: {}'.format(split_name, getattr(dataset, split_name).num_examples))
    print('\n')

    sess = tf.InteractiveSession()
    vae = flow_vae.utils.get_model(args)

    q_gradients = vae.build_iwhvi_gradients(scope='encoder/')
    all_gradients = vae.build_iwhvi_gradients()

    lr = tf.placeholder(tf.float32, shape=(), name='learning_rate')
    main_optimizer = tf.train.AdamOptimizer(learning_rate=lr)
    train_op = main_optimizer.apply_gradients(all_gradients)
    q_finetune_op = main_optimizer.apply_gradients(q_gradients)

    train_iwae_schedule = utils.StepSchedule(args.train_iwae_samples_step_schedule, default=1)
    train_lr_schedule = utils.StepSchedule(args.train_learning_rate_step_schedule, default=args.learning_rate)

    saver = tf.train.Saver(max_to_keep=100)
    best_val_saver = tf.train.Saver()
    if args.save_path is not None:
        train_writer = tf.summary.FileWriter(args.save_path + '/logs')
        train_writer.add_graph(sess.graph)
        # TODO: add summaries

    for attempt in range(args.train_attempts):
        utils.print_over('Attempt #{}'.format(attempt + 1))

        sess.run(tf.global_variables_initializer())
        best_val_score = -np.inf
        try:
            tqdm_t = tqdm(range(args.epochs), unit='epoch', desc='Training')
            for epoch in tqdm_t:
                np_lr = train_lr_schedule.at(epoch) * args.annealing_factor ** (epoch / args.annealing_speed)
                np_kl_coef = np.min([epoch / args.kl_warm_up_over, 1])
                iwae_samples = train_iwae_schedule.at(epoch)

                for train_batch in utils.batched_dataset(train_data, args.train_batch_size):
                    binarized_train_batch_x, binarized_train_batch_y = utils.binarize_batch(train_batch)
                    sess.run(train_op, {
                        vae.input_x: binarized_train_batch_x,
                        vae.output_y: binarized_train_batch_y,
                        vae.m_iwae_samples: iwae_samples,
                        vae.kl_coef: np_kl_coef,
                        lr: np_lr
                    })

                    if epoch >= args.no_extra_steps:
                        for _ in range(args.q_extra_steps):
                            sess.run(q_finetune_op, {
                                vae.input_x: binarized_train_batch_x,
                                vae.output_y: binarized_train_batch_y,
                                vae.m_iwae_samples: iwae_samples,
                                vae.kl_coef: np_kl_coef,
                                lr: np_lr
                            })

                if epoch % (5 if epoch < 50 else args.evaluate_every) == 0:
                    avg_evidence_train, = flow_vae.utils.calculate_evidence(
                        sess, train_data, vae, iwae_samples, args.train_batch_size, n_repeats=1)
                    avg_evidence_val, = flow_vae.utils.calculate_evidence(
                        sess, val_data, vae, iwae_samples, args.val_batch_size, n_repeats=1)

                    kl_q, = flow_vae.utils.calculate_kl_q(
                        sess, val_data, vae, args.val_iwae_samples, batch_size=100, n_repeats=1)

                    dat_train.append([epoch, avg_evidence_train])
                    dat_val.append([epoch, avg_evidence_val])

                    utils.print_over("Epoch: {:4}, train_evidence: {:.5f}, "
                                     "val_evidence: {:.5f}, KL(q(z)||p(z)): {:.5f}".format(
                        epoch, avg_evidence_train, avg_evidence_val, kl_q))

                    if avg_evidence_val > best_val_score:
                        best_val_score = avg_evidence_val
                        utils.save_weights(sess, args.save_path, suffix='best_val', saver=best_val_saver)

                    if any(np.isnan([avg_evidence_train, avg_evidence_val])):
                        if epoch < 300:
                            tqdm_t.close()
                            utils.print_over('nans detected, restarting')
                            break
                        else:
                            raise RuntimeError('nans detected')

                if epoch % args.save_every == 0 and epoch:
                    utils.save_weights(sess, args.save_path, suffix='ep%d' % epoch, saver=saver)
            else:
                break
        except KeyboardInterrupt:
            print('Manual stop')
            break

    utils.save_weights(sess, args.save_path, suffix='final', saver=saver)
    avg_evi_val, = flow_vae.utils.calculate_evidence(sess, val_data, vae, args.val_iwae_samples, batch_size_x=1,
                                                     n_repeats=1, tqdm_desc='Calculating val log-evidence')

    print('*' * 30)
    print('* The final model log-evidence on validation is', avg_evi_val)
    print('*' * 30)

    if args.save_path is not None:
        saver.restore(sess, os.path.join(args.save_path, 'model-weights-best_val'))
        avg_evi_val, = flow_vae.utils.calculate_evidence(sess, val_data, vae, args.val_iwae_samples, batch_size_x=1,
                                                         n_repeats=1, tqdm_desc='Calculating val log-evidence')

        print('* The best_val model evidence on validation is', avg_evi_val)
        print('*' * 30)

    if args.save_path is not None:
        dat0 = np.array(dat_train)
        dat1 = np.array(dat_val)

        df0 = pd.DataFrame({'epoch': dat0[:, 0], 'train': dat0[:, 1]})
        df1 = pd.DataFrame({'epoch': dat1[:, 0], 'val': dat1[:, 1]})

        df = pd.concat([df0, df1], ignore_index=True, axis=1)
        df.to_csv(os.path.join(args.save_path, 'data.csv'), index=False)


if __name__ == "__main__":
    main()