-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathmain_e2e.py
229 lines (177 loc) · 9 KB
/
main_e2e.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
import argparse
from tools.args import *
from tools.utils import *
import torch
import random
import time
from modules.emb2emb import E2E
from collections import OrderedDict
from evaluation import Evaluator
import json
import gc
from tools.lazy_reader import *
parser = argparse.ArgumentParser(description="embedding to embedding")
add_e2e_args(parser)
args = parser.parse_args()
print('\n'.join('%s: %s' % (k, str(v)) for k, v in sorted(dict(vars(args)).items())))
exp_path = get_exp_path("../saved_exps/" + args.model_name + "/")
log_path = exp_path + "exp.log"
assert args.src_emb_path is not None and args.tgt_emb_path is not None
args.src_lang = args.src_emb_path.split("/")[-1].split(".")[1]
args.tgt_lang = args.tgt_emb_path.split("/")[-1].split(".")[1]
if args.valid_option == "unsup":
# unsupervised validation metric
VALIDATION_METRIC_s2t = args.src_lang + "-" + args.tgt_lang + '-mean_cosine-csls_knn_10-S2T-10000'
VALIDATION_METRIC_t2s = args.tgt_lang + "-" + args.src_lang + '-mean_cosine-csls_knn_10-S2T-10000'
elif args.valid_option == "train":
# validation metric on the sampled training set
VALIDATION_METRIC_s2t = "valid-" + args.src_lang + "-" + args.tgt_lang + '-precision_at_1-csls_knn_10'
VALIDATION_METRIC_t2s = "valid-" + args.tgt_lang + "-" + args.src_lang + '-precision_at_1-csls_knn_10'
else:
raise NotImplementedError
# results on the test set with csls
VALIDATION_METRIC_SUP_s2t = args.src_lang + "-" + args.tgt_lang + '-precision_at_1-csls_knn_10'
VALIDATION_METRIC_SUP_t2s = args.tgt_lang + "-" + args.src_lang + '-precision_at_1-csls_knn_10'
# results on the test set with csls-d
DENSITY_METRIC_SUP_s2t = args.src_lang + "-" + args.tgt_lang + '-precision_at_1-density'
DENSITY_METRIC_SUP_t2s = args.tgt_lang + "-" + args.src_lang + '-precision_at_1-density'
best_valid_s2t_metric = 1e-12
best_valid_t2s_metric = 1e-12
# best csls results selected by validation
best_valid_density_s2t_metric = 1e-12
best_valid_density_t2s_metric = 1e-12
# best csls-d results selected by validation
best_valid_density_s2t_train_metric = 1e-12
best_valid_density_t2s_train_metric = 1e-12
# true best on test
best_csls_s2t = 1e-12
best_csls_t2s = 1e-12
best_density_s2t = 1e-12
best_density_t2s = 1e-12
args.cuda = torch.cuda.is_available()
random.seed(args.random_seed)
np.random.seed(args.random_seed)
torch.manual_seed(args.random_seed)
if args.cuda:
torch.cuda.manual_seed(args.random_seed)
device = torch.device('cuda') if args.cuda else torch.device("cpu")
assert args.tgt_emb_path.endswith("bin")
src_dict, np_src_emb, np_src_freqs = load_bin_embeddings(args, True)
tgt_dict, np_tgt_emb, np_tgt_freqs = load_bin_embeddings(args, False)
gb_size = 1073741824
print("Size of the src and tgt embedding in Gigabytes: %f, %f" %
((np_src_emb.size * np_src_emb.itemsize / gb_size, np_tgt_emb.size * np_tgt_emb.itemsize / gb_size)))
# prepare embeddings
src_emb = torch.from_numpy(np_src_emb).float().to(device)
tgt_emb = torch.from_numpy(np_tgt_emb).float().to(device)
normalize_embeddings(src_emb, args.normalize_embeddings)
normalize_embeddings(tgt_emb, args.normalize_embeddings)
# prepare model and evaluator
src_emb_for_mog = src_emb[:args.src_train_most_frequent]
tgt_emb_for_mog = tgt_emb[:args.tgt_train_most_frequent]
if args.t2s_s_var == 0:
args.t2s_s_var = args.s_var
args.s2t_t_var = args.t_var
if args.init_var:
src_var = load_txt_var(args, True, src_dict.word2id)
tgt_var = load_txt_var(args, False, tgt_dict.word2id)
src_var = torch.from_numpy(src_var).float().to(device)
tgt_var = torch.from_numpy(tgt_var).float().to(device)
src_dict.var = src_var
tgt_dict.var = tgt_var
model = E2E(args, src_dict, tgt_dict, src_emb_for_mog, tgt_emb_for_mog, device).to(device)
evaluator = Evaluator(model, src_emb, tgt_emb)
# prepare supervised batches
batch_size = args.batch_size
if args.supervise_id:
src_in_dict = src_emb[model.dict[:, 0], :].to(device)
tgt_in_dict = tgt_emb[model.dict[:, 1], :].to(device)
tot_sup_batches = math.ceil(src_in_dict.size(0) / args.batch_size)
sup_batch_inds = np.random.permutation(range(src_in_dict.size(0)))
def get_sup_batches(batch_inds, batch_id):
if batch_size > src_in_dict.size(0):
return src_in_dict, tgt_in_dict
src_batch = src_in_dict[batch_inds[batch_id * batch_size: min((batch_id + 1) * batch_size, src_in_dict.size(0))]]
tgt_batch = tgt_in_dict[batch_inds[batch_id * batch_size: min((batch_id + 1) * batch_size, tgt_in_dict.size(0))]]
return src_batch, tgt_batch
def check_dict(dict):
for i in range(10):
print(src_dict.id2word[dict[i][0].item()], tgt_dict.id2word[dict[i][1].item()])
train_step = 1
n_words_proc = 0
tic = time.time()
training_stats = {"disc_loss": [], "S2T_nll": [], "T2S_nll": [], "adv_loss": [], "sup_loss": [],
"Sup_S2T": [], "Sup_T2S": [],
"Cos_S": [], "Cos_T": [], "flow_tot_loss": [], "Diag_S": [], "Diag_T": [],
"BT_S2T": [], "BT_T2S": [], "step": train_step}
src_sampler = words_sampler_iterator(probs=np_src_freqs, buffer_size=args.src_base_batch_size,
batch_size=args.batch_size, uniform_sample=args.uniform_sample)
tgt_sampler = words_sampler_iterator(probs=np_tgt_freqs, buffer_size=args.tgt_base_batch_size,
batch_size=args.batch_size, uniform_sample=args.uniform_sample)
base_src_idx = src_sampler.retrieve_cache()
base_tgt_idx = tgt_sampler.retrieve_cache()
for src_idx, tgt_idx in zip(src_sampler, tgt_sampler):
if train_step > 0 and train_step % args.display_steps == 0:
ss = 'Step=%i, %i samples/s' % (train_step, int(n_words_proc / (time.time() - tic)))
flow_lr = model.flow_scheduler.get_lr()[0]
ss += ", flow lr=%.6f" % flow_lr
for k, v in training_stats.items():
if type(v) != list or len(v) == 0:
continue
ss += ", %s=%.4f" % (k, np.mean(v))
print(ss)
for k, v in training_stats.items():
if type(training_stats[k]) == list:
del training_stats[k][:]
if train_step % 1000 == 0:
n_words_proc = 0
tic = time.time()
if args.supervise_id:
sup_src_batch, sup_tgt_batch = src_in_dict, tgt_in_dict
else:
sup_src_batch = sup_tgt_batch = None
model.flow_step(base_src_idx, base_tgt_idx, src_idx, tgt_idx, training_stats, sup_src_batch, sup_tgt_batch)
n_words_proc += len(src_idx) * 2
if train_step > 0 and train_step % args.valid_steps == 0:
gc.collect()
to_log = OrderedDict({'train_iters': train_step, 'exp_path': exp_path})
evaluator.all_eval(to_log, train=True, unsup_eval=args.valid_option=="unsup")
if to_log[VALIDATION_METRIC_s2t] > best_valid_s2t_metric:
model.set_save_s2t_path(exp_path + "best_s2t_params.bin")
model.save_best_s2t()
best_valid_s2t_metric = to_log[VALIDATION_METRIC_s2t]
best_valid_csls_s2t_metric = to_log[VALIDATION_METRIC_SUP_s2t]
best_valid_density_s2t_metric = to_log[DENSITY_METRIC_SUP_s2t]
if to_log[VALIDATION_METRIC_t2s] > best_valid_t2s_metric:
model.set_save_t2s_path(exp_path + "best_t2s_params.bin")
model.save_best_t2s()
best_valid_t2s_metric = to_log[VALIDATION_METRIC_t2s]
best_valid_csls_t2s_metric = to_log[VALIDATION_METRIC_SUP_t2s]
best_valid_density_t2s_metric = to_log[DENSITY_METRIC_SUP_t2s]
if to_log[VALIDATION_METRIC_SUP_s2t] > best_csls_s2t:
best_csls_s2t = to_log[VALIDATION_METRIC_SUP_s2t]
if to_log[VALIDATION_METRIC_SUP_t2s] > best_csls_t2s:
best_csls_t2s = to_log[VALIDATION_METRIC_SUP_t2s]
if to_log[DENSITY_METRIC_SUP_s2t] > best_density_s2t:
best_density_s2t = to_log[DENSITY_METRIC_SUP_s2t]
if to_log[DENSITY_METRIC_SUP_t2s] > best_density_t2s:
best_density_t2s = to_log[DENSITY_METRIC_SUP_t2s]
print("Selected via metric with %s!" % args.valid_option)
print(f"----------------- best valid csls s2t = {best_valid_csls_s2t_metric}, "
f"best valid csls t2s = {best_valid_csls_t2s_metric} --------------")
print(f"----------------- best valid density s2t = {best_valid_density_s2t_metric}, "
f"best valid density t2s = {best_valid_density_t2s_metric} --------------")
print("Evaluation on the test set!")
print(f"----------------- best test csls s2t = {best_csls_s2t}, best test csls t2s = {best_csls_t2s} --------------")
print(f"----------------- best test density s2t = {best_density_s2t}, best test density t2s = {best_density_t2s} --------------")
print(json.dumps(to_log))
if model.flow_scheduler.get_lr()[0] < args.min_lr:
print('Learning rate < 1e-6. BREAK.')
break
train_step += 1
if train_step > args.n_steps:
print("Reach maximum training step. BREAK.")
break
training_stats["step"] = train_step
if args.export_emb:
model.export_embeddings(src_emb, tgt_emb, exp_path)