-
Notifications
You must be signed in to change notification settings - Fork 2
/
train_jtvae.py
401 lines (340 loc) · 21.1 KB
/
train_jtvae.py
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
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
import os
import sys
import math
import torch
import datetime
import argparse
import numpy as np
import torch.nn as nn
import torch.optim as optim
import torch.optim.lr_scheduler as lr_scheduler
import shutil
import inspect
from multiprocessing import Pool
from jtvae_models.VAE import JunctionTreeVAE
from lib.data_utils import JunctionTreeFolder
import jtvae_models.GraphEncoder
import jtvae_models.TreeEncoder
import jtvae_models.ParallelAltDecoder
import jtvae_models.ParallelAltDecoderV1
import lib.plot_utils, lib.logger
import jtvae_models.jtvae_utils
from jtvae_models.jtvae_utils import *
parser = argparse.ArgumentParser()
parser.add_argument('--save_dir', required=True)
parser.add_argument('--hidden_size', type=int, default=450)
parser.add_argument('--batch_size', type=int, default=32)
parser.add_argument('--latent_size', type=int, default=56)
parser.add_argument('--depthT', type=int, default=40)
parser.add_argument('--depthG', type=int, default=10)
parser.add_argument('--lr', type=float, default=1e-3)
parser.add_argument('--clip_norm', type=float, default=50.0)
parser.add_argument('--beta', type=float, default=0.0)
parser.add_argument('--step_beta', type=float, default=0.002)
parser.add_argument('--max_beta', type=float, default=1.0)
parser.add_argument('--epoch', type=int, default=10)
# parser.add_argument('--anneal_rate', type=float, default=0.9)
parser.add_argument('--print_iter', type=int, default=1000)
parser.add_argument('--tree_encoder_arch', type=str, default='baseline')
parser.add_argument('--decoder_version', type=str, default='default')
parser.add_argument('--warmup_epoch', type=int, default=1)
parser.add_argument('--use_flow_prior', type=eval, default=True, choices=[True, False])
parser.add_argument('--limit_data', type=int, default=None)
parser.add_argument('--resume', type=eval, default=False, choices=[True, False])
if __name__ == "__main__":
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
args = parser.parse_args()
print(args)
model = JunctionTreeVAE(args.hidden_size, args.latent_size, args.depthT, args.depthG,
decode_nuc_with_lstm=True, device=device, tree_encoder_arch=args.tree_encoder_arch,
use_flow_prior=args.use_flow_prior, decoder_version=args.decoder_version).to(device)
print(model)
for param in model.parameters():
if param.dim() == 1:
nn.init.constant_(param, 0)
elif param.dim() >= 2:
nn.init.xavier_normal_(param)
print("Model #Params: %dK" % (sum([x.nelement() for x in model.parameters()]) / 1000,))
optimizer = optim.Adam(model.parameters(), lr=args.lr, amsgrad=True)
# scheduler = lr_scheduler.ExponentialLR(optimizer, args.anneal_rate)
param_norm = lambda m: math.sqrt(sum([p.norm().item() ** 2 for p in m.parameters()]))
grad_norm = lambda m: math.sqrt(sum([p.grad.norm().item() ** 2 for p in m.parameters() if p.grad is not None]))
total_step = 0
beta = args.beta
meters = np.zeros(8)
cur_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
save_dir = '/'.join(args.save_dir.split('/')[:-1] + [cur_time + '-' + args.save_dir.split('/')[-1]])
if not os.path.exists(save_dir):
os.makedirs(save_dir)
backup_dir = os.path.join(save_dir, 'backup')
if not os.path.exists(backup_dir):
os.makedirs(backup_dir)
shutil.copy(__file__, backup_dir)
shutil.copy(inspect.getfile(JunctionTreeVAE), backup_dir)
shutil.copy(inspect.getfile(JunctionTreeFolder), backup_dir)
shutil.copy(inspect.getfile(jtvae_models.jtvae_utils), backup_dir)
shutil.copy(inspect.getfile(jtvae_models.GraphEncoder), backup_dir)
shutil.copy(inspect.getfile(jtvae_models.TreeEncoder), backup_dir)
shutil.copy(inspect.getfile(jtvae_models.ParallelAltDecoder), backup_dir)
shutil.copy(inspect.getfile(jtvae_models.ParallelAltDecoderV1), backup_dir)
lib.plot_utils.set_output_dir(save_dir)
lib.plot_utils.suppress_stdout()
logger = lib.logger.CSVLogger(
'run.csv', save_dir,
['Epoch', 'Beta', 'Validation_Entropy', 'Validation_Neg_Log_Prior', 'Validation_KL',
'Validation_Node_Acc', 'Validation_Nuc_Stop_Acc', 'Validation_Nuc_Ord_Acc',
'Validation_Nuc_Acc', 'Validation_Topo_Acc', 'Validation_recon_acc_with_reg',
'Validation_post_valid_with_reg', 'Validation_post_fe_deviation_with_reg',
'Validation_post_fe_deviation_len_normed_with_reg',
'Validation_recon_acc_no_reg', 'Validation_post_valid_no_reg',
'Validation_post_fe_deviation_no_reg',
'Validation_post_fe_deviation_len_normed_no_reg', 'Prior_valid_with_reg',
'Prior_fe_deviation_with_reg', 'Prior_fe_deviation_len_normed_with_reg',
'Prior_valid_no_reg', 'Prior_fe_deviation_no_reg',
'Prior_fe_deviation_len_normed_no_reg',
'Prior_valid_no_reg_greedy', 'Prior_fe_deviation_no_reg_greedy',
'Prior_fe_deviation_len_normed_no_reg_greedy',
'Prior_uniqueness_no_reg_greedy', 'Validation_mutual_information', 'Validation_NLL_IW_100',
'Validation_active_units'])
mp_pool = Pool(8)
jtvae_models.jtvae_utils.model = model
if args.resume:
'''load warm-up results'''
if args.use_flow_prior:
if args.limit_data is not None:
weight_path = '/home/zichao/scratch/JTRNA/output/20200514-160741-512-64-5-10-amsgrad-stability-mb-2e-3-sb-1e-4-limit-data/model.epoch-17'
epoch_to_start = 18
beta = 0.0012
else:
if args.tree_encoder_arch == 'baseline':
weight_path = '/home/zichao/scratch/JTRNA/output/corrected-treeenc-jtvae/20200606-234107-corrected-treeencoder-mb-2e-3-sb-5e-4/model.epoch-4'
epoch_to_start = 5
beta = 0.0000
else:
weight_path = '/home/zichao/scratch/JTRNA/output/branched-treeenc-corrected/20200606-200347-branched-treedec-corrected-mb-2e-3-sb-5e-4/model.epoch-6'
epoch_to_start = 7
beta = 0.0005
else:
weight_path = '/home/zichao/scratch/JTRNA/output/20200514-163629-512-64-5-10-amsgrad-stability-mb-2e-3-sb-5e-4-no-flow-prior/model.epoch-8'
epoch_to_start = 9
beta = 0.0015
all_weights = torch.load(weight_path)
model.load_state_dict(all_weights['model_weights'])
optimizer.load_state_dict(all_weights['opt_weights'])
print('Loaded weights:', weight_path)
print('Loaded beta:', beta)
else:
epoch_to_start = 1
lib.plot_utils.set_first_tick(epoch_to_start)
for epoch in range(epoch_to_start, args.epoch + 1):
if epoch > args.warmup_epoch:
beta = min(args.max_beta, beta + args.step_beta)
loader = JunctionTreeFolder('data/rna_jt_32-512/train-split', args.batch_size,
num_workers=8, tree_encoder_arch=args.tree_encoder_arch,
limit_data=args.limit_data)
for batch in loader:
total_step += 1
model.zero_grad()
ret_dict = model(batch)
loss = ret_dict['sum_hpn_pred_loss'] / ret_dict['nb_hpn_targets'] + \
ret_dict['sum_nuc_pred_loss'] / ret_dict['nb_nuc_targets'] + \
ret_dict['sum_stop_pred_loss'] / ret_dict['nb_stop_targets'] + \
beta * (ret_dict['entropy_loss'] + ret_dict['prior_loss'])
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), args.clip_norm)
optimizer.step()
neg_entropy = float(ret_dict['entropy_loss'])
neg_log_prior = float(ret_dict['prior_loss'])
kl_div = neg_entropy + neg_log_prior
hpn_pred_acc, stop_translation_nuc_acc, ord_nuc_acc, all_nuc_pred_acc, stop_acc = \
ret_dict['nb_hpn_pred_correct'] / ret_dict['nb_hpn_targets'], \
ret_dict['nb_stop_trans_pred_correct'] / ret_dict['nb_stop_trans_targets'], \
ret_dict['nb_ord_nuc_pred_correct'] / ret_dict['nb_ord_nuc_targets'], \
ret_dict['nb_nuc_pred_correct'] / ret_dict['nb_nuc_targets'], \
ret_dict['nb_stop_pred_correct'] / ret_dict['nb_stop_targets'],
meters = meters + np.array(
[neg_entropy, neg_log_prior, kl_div, hpn_pred_acc * 100, stop_translation_nuc_acc * 100,
ord_nuc_acc * 100, all_nuc_pred_acc * 100, stop_acc * 100])
if total_step % args.print_iter == 0:
meters /= args.print_iter
print(
"[%d] Beta: %.4f, Entropy: %.2f, Neg_log_prior: %.2f, KL: %.2f, Node: %.2f, Nucleotide stop: %.2f, Nucleotide ord: %.2f, Nucleotide: %.2f, Topo: %.2f, PNorm: %.2f, GNorm: %.2f" % (
total_step, beta, -meters[0], meters[1], meters[2], meters[3], meters[4], meters[5], meters[6],
meters[7],
param_norm(model), grad_norm(model)))
lib.plot_utils.plot('Train_Entropy', -meters[0], index=0)
lib.plot_utils.plot('Train_Neg_Log_Prior', meters[1], index=0)
lib.plot_utils.plot('Train_KL', meters[2], index=0)
lib.plot_utils.plot('Train_Node_Acc', meters[3], index=0)
lib.plot_utils.plot('Train_Nucleotide_Stop', meters[4], index=0)
lib.plot_utils.plot('Train_Nucleotide_Ord', meters[5], index=0)
lib.plot_utils.plot('Train_Nucleotide_All', meters[6], index=0)
lib.plot_utils.plot('Train_Topo_Acc', meters[7], index=0)
lib.plot_utils.flush()
sys.stdout.flush()
meters *= 0
lib.plot_utils.tick(index=0)
del loss, kl_div
# scheduler.step(epoch)
# print("learning rate: %.6f" % scheduler.get_lr()[0])
# save model at the end of each epoch
torch.save(
{'model_weights': model.state_dict(),
'opt_weights': optimizer.state_dict()},
os.path.join(save_dir, "model.epoch-" + str(epoch)))
# validation step
print('End of epoch %d,' % (epoch), 'starting validation')
valid_batch_size = 128
loader = JunctionTreeFolder('data/rna_jt_32-512/validation-split', valid_batch_size,
num_workers=8, tree_encoder_arch=args.tree_encoder_arch)
# turns out there is a very large graph in the validation set, therefore we have to use a smaller batch size
nb_iters = 20000 // valid_batch_size # 20000 is the size of the validation set
post_max_iters = min(10, nb_iters) # for efficiency
total = 0
# from tqdm import trange
# bar = trange(nb_iters, desc='', leave=True)
# loader = loader.__iter__()
nb_encode, nb_decode = 4, 4
recon_acc, post_valid, post_fe_deviation, post_fe_deviation_len_normed = 0, 0, 0., 0.
recon_acc_noreg, post_valid_noreg, post_fe_deviation_noreg, post_fe_deviation_noreg_len_normed = 0, 0, 0., 0.
valid_kl, valid_node_acc, valid_nuc_stop_acc, valid_nuc_ord_acc, \
valid_nuc_acc, valid_topo_acc = 0., 0., 0., 0., 0., 0.
valid_entropy, valid_neg_log_prior = 0., 0.
all_means = []
total_mi = 0.
nll_iw = 0.
with torch.no_grad():
for i, batch in enumerate(loader):
# for i in bar:
tree_batch, graph_encoder_input, tree_encoder_input = batch
# tree_batch, graph_encoder_input, tree_encoder_input = next(loader)
graph_vectors, tree_vectors = model.encode(graph_encoder_input, tree_encoder_input)
if i < post_max_iters:
all_seq = [''.join(tree.rna_seq) for tree in tree_batch]
all_struct = [''.join(tree.rna_struct) for tree in tree_batch]
ret = evaluate_posterior(all_seq, all_struct, graph_vectors, tree_vectors,
mp_pool, nb_encode=nb_encode, nb_decode=nb_decode,
enforce_rna_prior=True)
total += nb_encode * nb_decode * valid_batch_size
recon_acc += np.sum(ret['recon_acc'])
post_valid += np.sum(ret['posterior_valid'])
post_fe_deviation += np.sum(ret['posterior_fe_deviation'])
post_fe_deviation_len_normed += np.sum(ret['posterior_fe_deviation_len_normed'])
ret = evaluate_posterior(all_seq, all_struct, graph_vectors, tree_vectors,
mp_pool, nb_encode=nb_encode, nb_decode=nb_decode,
enforce_rna_prior=False)
recon_acc_noreg += np.sum(ret['recon_acc'])
post_valid_noreg += np.sum(ret['posterior_valid'])
post_fe_deviation_noreg += np.sum(ret['posterior_fe_deviation'])
post_fe_deviation_noreg_len_normed += np.sum(ret['posterior_fe_deviation_len_normed'])
# bar.set_description(
# 'streaming recon acc: %.2f, streaming post valid: %.2f, streaming post free energy deviation: %.2f'
# % (recon_acc / total * 100, post_valid / total * 100, post_fe_deviation / post_valid))
#
# bar.refresh()
total_mi += model.calc_mi((graph_encoder_input, tree_encoder_input),
graph_latent_vec=graph_vectors, tree_latent_vec=tree_vectors)
all_mean = torch.cat([model.g_mean(graph_vectors), model.t_mean(tree_vectors)], dim=-1)
all_means.append(all_mean.cpu().detach().numpy())
# trite accuracy measures
(z_vecs, graph_z_vecs, tree_z_vecs), (entropy, log_pz) = model.rsample(graph_vectors, tree_vectors)
graph_z_vecs, tree_z_vecs = graph_z_vecs[:, 0, :], tree_z_vecs[:, 0, :]
ret_dict = model.decoder(tree_batch, tree_z_vecs, graph_z_vecs)
valid_entropy += float(entropy.mean())
valid_neg_log_prior += -float(log_pz.mean())
valid_kl += float(- entropy.mean() - log_pz.mean())
valid_node_acc += ret_dict['nb_hpn_pred_correct'] / ret_dict['nb_hpn_targets']
valid_nuc_stop_acc += ret_dict['nb_stop_trans_pred_correct'] / ret_dict['nb_stop_trans_targets']
valid_nuc_ord_acc += ret_dict['nb_ord_nuc_pred_correct'] / ret_dict['nb_ord_nuc_targets']
valid_nuc_acc += ret_dict['nb_nuc_pred_correct'] / ret_dict['nb_nuc_targets']
valid_topo_acc += ret_dict['nb_stop_pred_correct'] / ret_dict['nb_stop_targets']
lib.plot_utils.plot('Validation_Entropy', valid_entropy / nb_iters, index=1)
lib.plot_utils.plot('Validation_Neg_Log_Prior', valid_neg_log_prior / nb_iters, index=1)
lib.plot_utils.plot('Validation_KL', valid_kl / nb_iters, index=1)
lib.plot_utils.plot('Validation_Node_Acc', valid_node_acc / nb_iters * 100, index=1)
lib.plot_utils.plot('Validation_Nuc_Stop_Acc', valid_nuc_stop_acc / nb_iters * 100, index=1)
lib.plot_utils.plot('Validation_Nuc_Ord_Acc', valid_nuc_ord_acc / nb_iters * 100, index=1)
lib.plot_utils.plot('Validation_Nuc_Acc', valid_nuc_acc / nb_iters * 100, index=1)
lib.plot_utils.plot('Validation_Topo_Acc', valid_topo_acc / nb_iters * 100, index=1)
# posterior decoding with enforced RNA regularity
lib.plot_utils.plot('Validation_recon_acc_with_reg', recon_acc / total * 100, index=1)
lib.plot_utils.plot('Validation_post_valid_with_reg', post_valid / total * 100, index=1)
lib.plot_utils.plot('Validation_post_fe_deviation_with_reg',
post_fe_deviation / post_valid, index=1)
lib.plot_utils.plot('Validation_post_fe_deviation_len_normed_with_reg',
post_fe_deviation_noreg_len_normed / post_valid, index=1)
# posterior decoding without RNA regularity
lib.plot_utils.plot('Validation_recon_acc_no_reg', recon_acc_noreg / total * 100, index=1)
lib.plot_utils.plot('Validation_post_valid_no_reg', post_valid_noreg / total * 100, index=1)
lib.plot_utils.plot('Validation_post_fe_deviation_no_reg',
post_fe_deviation_noreg / post_valid_noreg, index=1)
lib.plot_utils.plot('Validation_post_fe_deviation_len_normed_no_reg',
post_fe_deviation_noreg_len_normed / post_valid_noreg, index=1)
######################## sampling from the prior ########################
sampled_g_z = torch.as_tensor(np.random.randn(1000, args.latent_size).
astype(np.float32)).to(device)
sampled_t_z = torch.as_tensor(np.random.randn(1000, args.latent_size).
astype(np.float32)).to(device)
sampled_z = torch.cat([sampled_g_z, sampled_t_z], dim=-1)
if args.use_flow_prior:
sampled_z = model.latent_cnf(sampled_z, None, reverse=True).view(
*sampled_z.size())
sampled_g_z = sampled_z[:, :args.latent_size]
sampled_t_z = sampled_z[:, args.latent_size:]
######################## evaluate prior with regularity constraints ########################
ret = evaluate_prior(sampled_g_z, sampled_t_z, 1000, 1, mp_pool, enforce_rna_prior=True)
lib.plot_utils.plot('Prior_valid_with_reg', np.sum(ret['prior_valid']) / 10,
index=1) # /1000 * 100 = /10
lib.plot_utils.plot('Prior_fe_deviation_with_reg',
np.sum(ret['prior_fe_deviation']) / np.sum(ret['prior_valid']), index=1)
lib.plot_utils.plot('Prior_fe_deviation_len_normed_with_reg',
np.sum(ret['prior_fe_deviation_len_normed']) / np.sum(ret['prior_valid']), index=1)
######################## evaluate prior without regularity constraints ########################
ret = evaluate_prior(sampled_g_z, sampled_t_z, 1000, 1, mp_pool, enforce_rna_prior=False)
lib.plot_utils.plot('Prior_valid_no_reg', np.sum(ret['prior_valid']) / 10, index=1) # /1000 * 100 = /10
lib.plot_utils.plot('Prior_fe_deviation_no_reg',
np.sum(ret['prior_fe_deviation']) / np.sum(ret['prior_valid']), index=1)
lib.plot_utils.plot('Prior_fe_deviation_len_normed_no_reg',
np.sum(ret['prior_fe_deviation_len_normed']) / np.sum(ret['prior_valid']), index=1)
######################## evaluate prior without regularity constraints and greedy ########################
ret = evaluate_prior(sampled_g_z, sampled_t_z, 1000, 1, mp_pool,
enforce_rna_prior=False, prob_decode=False)
decoded_seq = [''.join(tree.rna_seq) for tree in ret['all_parsed_trees'][:1000] if
type(tree) is RNAJunctionTree and tree.is_valid]
lib.plot_utils.plot('Prior_valid_no_reg_greedy', np.sum(ret['prior_valid']) / 10,
index=1) # /1000 * 100 = /10
lib.plot_utils.plot('Prior_fe_deviation_no_reg_greedy',
np.sum(ret['prior_fe_deviation']) / np.sum(ret['prior_valid']), index=1)
lib.plot_utils.plot('Prior_fe_deviation_len_normed_no_reg_greedy',
np.sum(ret['prior_fe_deviation_len_normed']) / np.sum(ret['prior_valid']), index=1)
if len(decoded_seq) == 0:
lib.plot_utils.plot('Prior_uniqueness_no_reg_greedy', 0.,
index=1)
else:
lib.plot_utils.plot('Prior_uniqueness_no_reg_greedy', len(set(decoded_seq)) / len(decoded_seq) * 100,
index=1)
######################## mutual information ########################
cur_mi = total_mi / nb_iters
lib.plot_utils.plot('Validation_mutual_information', cur_mi, index=1)
######################## active units ########################
all_means = np.concatenate(all_means, axis=0)
au_mean = np.mean(all_means, axis=0, keepdims=True)
au_var = all_means - au_mean
ns = au_var.shape[0]
au_var = (au_var ** 2).sum(axis=0) / (ns - 1)
delta = 0.01
au = (au_var >= delta).sum().item()
lib.plot_utils.plot('Validation_active_units', au, index=1)
lib.plot_utils.plot('Beta', beta, index=1)
tocsv = {'Epoch': epoch}
for name, val in lib.plot_utils._since_last_flush.items():
if lib.plot_utils._ticker_registry[name] == 1:
tocsv[name] = list(val.values())[0]
logger.update_with_dict(tocsv)
lib.plot_utils.set_xlabel_for_tick(index=1, label='epoch')
lib.plot_utils.flush()
lib.plot_utils.tick(index=1)
if mp_pool is not None:
mp_pool.close()
mp_pool.join()
logger.close()