-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathseqskip_train_rnb1.py
338 lines (272 loc) · 15.2 KB
/
seqskip_train_rnb1.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
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Tue Dec 11 00:45:08 2018
@author: mimbres
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
from torch.optim.lr_scheduler import StepLR
from torch.backends import cudnn
import numpy as np
import glob, os
import argparse
from tqdm import trange, tqdm
from spotify_data_loader import SpotifyDataloader
cudnn.benchmark = True
parser = argparse.ArgumentParser(description="Sequence Skip Prediction")
parser.add_argument("-c","--config",type = str, default = "./config_init_dataset.json")
parser.add_argument("-s","--save_path",type = str, default = "./save/exp_rnb1/")
parser.add_argument("-l","--load_continue_latest",type = str, default = None)
parser.add_argument("-f","--feature_dim",type = int, default = 64)
parser.add_argument("-r","--relation_dim",type = int, default = 8)
parser.add_argument("-w","--class_num",type = int, default = 2)
parser.add_argument("-e","--epochs",type = int, default= 1000)
parser.add_argument("-t","--test_episode", type = int, default = 1000)
parser.add_argument("-lr","--learning_rate", type = float, default = 0.001)
parser.add_argument("-b","--train_batch_size", type = int, default = 1024)
parser.add_argument("-g","--gpu",type=int, default=0)
#parser.add_argument("-e","--embed_hidden_unit",type=int, default=2)
args = parser.parse_args()
# Hyper Parameters
FEATURE_DIM = args.feature_dim
RELATION_DIM = args.relation_dim
CLASS_NUM = args.class_num
EPOCHS = args.epochs
TEST_EPISODE = args.test_episode
LEARNING_RATE = args.learning_rate
TR_BATCH_SZ = args.train_batch_size
GPU = args.gpu
# Model-save directory
MODEL_SAVE_PATH = args.save_path
os.makedirs(os.path.dirname(MODEL_SAVE_PATH), exist_ok=True)
# Trainset stats: 2072002577 items from 124950714 sessions
print('Initializing dataloader...')
mtrain_loader = SpotifyDataloader(config_fpath=args.config,
mtrain_mode=True,
data_sel=(0, 99965071), # 80% 트레인
batch_size=TR_BATCH_SZ,
shuffle=True) # shuffle은 True로 해야됨 나중에...
mval_loader = SpotifyDataloader(config_fpath=args.config,
mtrain_mode=True, # True, because we use part of trainset as testset
data_sel=(99965071, 124950714),#(99965071, 124950714), # 20%를 테스트
batch_size=2048,
shuffle=True)
#Feature encoder:
class MLP(nn.Module):
def __init__(self, input_sz, hidden_sz, output_sz):
super(MLP, self).__init__()
self.fc1 = nn.Linear(input_sz, hidden_sz)
self.relu = nn.ReLU()
self.fc2 = nn.Linear(hidden_sz, output_sz)
def forward(self, x):
out = self.fc1(x)
out = self.relu(out)
out = self.fc2(out)
return out
# Model: Relation Nets
class RelationNetwork(nn.Module):
"""docstring for RelationNetwork"""
def __init__(self,
num_sup_max=10, num_que_max=10,
in_feat_sup_sz=64, in_feat_que_sz=64,
in_log_sup_sz=41, in_log_que_sz=0,
in_label_sup_sz=3): # 64,64,41,0,3
super(RelationNetwork, self).__init__()
self.num_sup_max = num_sup_max
self.num_que_max = num_que_max
self.in_feat_sup_sz = in_feat_sup_sz
self.in_feat_que_sz = in_feat_que_sz
self.in_log_sup_sz = in_log_sup_sz
self.in_log_que_sz = in_log_que_sz
self.in_label_sup_sz = in_label_sup_sz
self.layer1_input_sz = in_feat_sup_sz + in_feat_que_sz + in_log_sup_sz + in_log_que_sz + in_label_sup_sz
self.layer1 = nn.Sequential(
nn.Linear(self.layer1_input_sz, 512), # bx8x7x1*172 (bx8x7x1*213) -> bx8x7x1*512
nn.LayerNorm(512),
nn.ReLU())
self.layer2 = nn.Sequential(
nn.Linear(512, 256),
nn.LayerNorm(256),
nn.ReLU())
self.fc1 = nn.Linear(256,64)
self.fc2 = nn.Linear(64,1)
# Option: embedding for log data
# Option: classifier
self.classifier = nn.Linear(10,1)
def forward(self, x_sup, x_que, x_log_sup, y_log_que, label_sup):
relation_pairs = self.pack_relation_pairs(x_sup, x_que, x_log_sup, y_log_que, label_sup) # bx8x7x1x172 (bx8x7x1x213)
out = self.layer1(relation_pairs) #bx7x8x1*512
out = self.layer2(out) #bx7x8x1*256
out = F.relu(self.fc1(out)) # bx7x8x1*64
out = torch.sigmoid(self.fc2(out)) # bx7x8*1
out = out.view(-1,10,10,1)
return out
def pack_relation_pairs(self, x_feat_sup, x_feat_que, x_log_sup, x_log_que, label_sup):
# x_feat_sup: bx7x1*64, x_feat_que: bx8x1*64
# _extras: concat support logs(d=41) and labels(d=3) to feat_support. QUERY SHOULD NOT INCLUDE THESE...
_extras_sup = torch.cat((x_log_sup, label_sup), 2).unsqueeze(2) # bx7x1*44
x_feat_sup = torch.cat((x_feat_sup, _extras_sup), 3) # bx7x1*108
x_feat_sup_ext = x_feat_sup.unsqueeze(1).repeat(1,10,1,1,1) # bx8x7x1*108
if self.in_log_que_sz is not 0: # As default, we don't use x_log_que
_extras_que = x_log_que.unsqueeze(2) # (bx8x1*41)
x_feat_que = torch.cat((x_feat_que, _extras_que), 3) # (bx8x1*108)
x_feat_que_ext = x_feat_que.unsqueeze(2).repeat(1,1,10,1,1)
#x_feat_que_ext = x_feat_que.unsqueeze(1).repeat(1,10,1,1,1) # bx7x8x1*64 (bx7x8x1*105)
#x_feat_que_ext = torch.transpose(x_feat_que_ext,1,2) # bx8x7x1*64 (bx8x7x1*105)
x_relation_pairs = torch.cat((x_feat_sup_ext, x_feat_que_ext), 4) # bx8x7x1*172 (bx8x7x1*213)
return x_relation_pairs
# Init neural net
#FeatEnc = MLP(input_sz=29, hidden_sz=512, output_sz=64).apply(weights_init).cuda(GPU)
FeatEnc = MLP(input_sz=29, hidden_sz=512, output_sz=64).cuda(GPU)
RN = RelationNetwork().cuda(GPU)
FeatEnc_optim = torch.optim.Adam(FeatEnc.parameters(), lr=LEARNING_RATE)
RN_optim = torch.optim.Adam(RN.parameters(), lr=LEARNING_RATE)
FeatEnc_scheduler = StepLR(FeatEnc_optim, step_size=100000, gamma=0.2)
RN_scheduler = StepLR(RN_optim, step_size=100000, gamma=0.2)
#relation_net_optim#
#%%
hist_trloss = list()
hist_tracc = list()
hist_vloss = list()
hist_vacc = list()
np.set_printoptions(precision=3)
def validate():
tqdm.write("Validation...")
total_vloss = 0
total_vcorrects = 0
total_vquery = 0
val_sessions_iter = iter(mval_loader)
for val_session in trange(len(val_sessions_iter), desc='val-sessions', position=2, ascii=True):
FeatEnc.eval(); RN.eval();
x_sup, x_que, x_log_sup, x_log_que, label_sup, label_que, num_items, index = val_sessions_iter.next() # FIXED 13.Dec. SEPARATE LOGS. QUERY SHOULT NOT INCLUDE LOGS
x_sup, x_que = Variable(x_sup).cuda(GPU), Variable(x_que).cuda(GPU)
x_log_sup, x_log_que = Variable(x_log_sup).cuda(GPU), Variable(x_log_que).cuda(GPU)
label_sup = Variable(label_sup).cuda(GPU)
num_support = num_items[:,0].detach().numpy().flatten() # If num_items was odd number, query has one more item.
num_query = num_items[:,1].detach().numpy().flatten()
batch_sz = num_items.shape[0]
x_sup = x_sup.unsqueeze(2) # 1x7*29 --> 1x7x1*29
x_que = x_que.unsqueeze(2) # 1x8*29 --> 1x8x1*29
x_feat_sup = FeatEnc(x_sup) # 1x7x1*64
x_feat_que = FeatEnc(x_que) # 1x8x1*64
y_hat_relation = RN(x_feat_sup, x_feat_que, x_log_sup, x_log_que, label_sup) # bx8x7*1
y_sup_ext = label_sup[:,:,1].detach().cpu().view(-1,1,10,1).repeat(1,10,1,1) # bx8x7*1
y_que_ext = label_que[:,:,1].view(-1,10,1,1).repeat(1,1,10,1) # bx7x8*1
y_relation = (y_sup_ext==y_que_ext).float().view(-1,10,10,1) # bx8x7*1
y_mask = np.zeros((batch_sz,10,10,1), dtype=np.float32)
for b in np.arange(batch_sz):
y_mask[b,:num_query[b],:num_support[b],0] = 1
y_mask = torch.FloatTensor(y_mask).cuda(GPU)
loss = F.mse_loss(input=y_hat_relation*y_mask, target=y_relation.cuda(GPU)*y_mask)
total_vloss += loss.item()
decision = torch.FloatTensor(np.zeros((batch_sz, 10, 10 ,2))).detach().cpu() # bx8x7*2 (b x que x sup x class)
decision[:,:,:,0] = (y_hat_relation.detach().cpu()*(y_sup_ext==0).float() + (1-y_hat_relation.detach().cpu())*(y_sup_ext==1).float()).view(-1,10,10)
decision[:,:,:,1] = (y_hat_relation.detach().cpu()*(y_sup_ext==1).float() + (1-y_hat_relation.detach().cpu())*(y_sup_ext==0).float()).view(-1,10,10)
decision = decision * y_mask.detach().cpu().repeat(1,1,1,2)
y_pred = torch.argmax(decision.sum(2),2).numpy()
sim_score = (y_hat_relation.detach()*y_mask).cpu().numpy()
sample_sup = label_sup[0,:num_support[0],1].detach().long().cpu().numpy().flatten()
sample_que = label_que[0,:num_query[0],1].detach().long().numpy().flatten()
sample_pred = y_pred[0,:num_query[0]].flatten()
total_vcorrects += np.sum((y_pred == label_que[:,:,1].long().numpy()) * y_mask[:,:,0,0].cpu().numpy())
total_vquery += np.sum(num_query)
if (val_session+1)%4000 == 0:
tqdm.write(np.array2string(sim_score[0,:,:,0]))
tqdm.write("S:" + np.array2string(sample_sup) +'\n'+
"Q:" + np.array2string(sample_que) + '\n' +
"P:" + np.array2string(sample_pred) )
tqdm.write("val_session:{0:} vloss:{1:.6f} vacc:{2:.4f}".format(val_session,loss.item(), total_vcorrects/total_vquery))
hist_vloss.append(total_vloss/val_session)
hist_vacc.append(total_vcorrects/total_vquery)
# Main
if args.load_continue_latest is None:
START_EPOCH = 0
else:
latest_fpath = max(glob.iglob(MODEL_SAVE_PATH + "check*.pth"),key=os.path.getctime)
checkpoint = torch.load(latest_fpath)
tqdm.write("Loading saved model from '{0:}'... loss: {1:.6f}".format(latest_fpath,checkpoint['loss']))
FeatEnc.load_state_dict(checkpoint['FE_state'])
RN.load_state_dict(checkpoint['RN_state'])
FeatEnc_optim.load_state_dict(checkpoint['FE_opt_state'])
RN_optim.load_state_dict(checkpoint['RN_opt_state'])
FeatEnc_scheduler.load_state_dict(checkpoint['FE_sch_state'])
RN_scheduler.load_state_dict(checkpoint['RN_sch_state'])
START_EPOCH = checkpoint['ep']
for epoch in trange(START_EPOCH, EPOCHS, desc='epochs', position=0, ascii=True):
tqdm.write('Train...')
tr_sessions_iter = iter(mtrain_loader)
total_corrects = 0
total_query = 0
total_trloss = 0
for session in trange(len(tr_sessions_iter), desc='sessions', position=1, ascii=True):
FeatEnc.train(); RN.train();
x_sup, x_que, x_log_sup, x_log_que, label_sup, label_que, num_items, index = tr_sessions_iter.next() # FIXED 13.Dec. SEPARATE LOGS. QUERY SHOULT NOT INCLUDE LOGS
x_sup, x_que = Variable(x_sup).cuda(GPU), Variable(x_que).cuda(GPU)
x_log_sup, x_log_que = Variable(x_log_sup).cuda(GPU), Variable(x_log_que).cuda(GPU)
label_sup = Variable(label_sup).cuda(GPU)
# Sample data for 'support' and 'query': ex) 15 items = 7 sup, 8 queries...
num_support = num_items[:,0].detach().numpy().flatten() # If num_items was odd number, query has one more item.
num_query = num_items[:,1].detach().numpy().flatten()
batch_sz = num_items.shape[0]
x_sup = x_sup.unsqueeze(2) # 1x7*29 --> 1x7x1*29
x_que = x_que.unsqueeze(2) # 1x8*29 --> 1x8x1*29
# - feature encoder
x_feat_sup = FeatEnc(x_sup) # 1x7x1*64
x_feat_que = FeatEnc(x_que) # 1x8x1*64
# - relation network
y_hat_relation = RN(x_feat_sup, x_feat_que, x_log_sup, x_log_que, label_sup) # bx8x7*1
# Prepare ground-truth simlarity score and mask
y_sup_ext = label_sup[:,:,1].detach().cpu().view(-1,1,10,1).repeat(1,10,1,1) # bx8x7*1
y_que_ext = label_que[:,:,1].view(-1,10,1,1).repeat(1,1,10,1) # bx7x8*1
y_relation = (y_sup_ext==y_que_ext).float().view(-1,10,10,1) # bx8x7*1
y_mask = np.zeros((batch_sz,10,10,1), dtype=np.float32)
for b in np.arange(batch_sz):
y_mask[b,:num_query[b],:num_support[b],0] = 1
y_mask = torch.FloatTensor(y_mask).cuda(GPU)
# Calcultate MSE loss
loss = F.mse_loss(input=y_hat_relation*y_mask, target=y_relation.cuda(GPU)*y_mask)
total_trloss += loss.item()
# Update Nets
FeatEnc.zero_grad()
RN.zero_grad()
loss.backward()
#torch.nn.utils.clip_grad_norm_(FeatEnc.parameters(), 0.5)
#torch.nn.utils.clip_grad_norm_(RN.parameters(), 0.5)
FeatEnc_optim.step()
RN_optim.step()
# Decision
decision = torch.FloatTensor(np.zeros((batch_sz, 10, 10 ,2))).detach().cpu() # bx8x7*2 (b x que x sup x class)
decision[:,:,:,0] = (y_hat_relation.detach().cpu()*(y_sup_ext==0).float() + (1-y_hat_relation.detach().cpu())*(y_sup_ext==1).float()).view(-1,10,10)
decision[:,:,:,1] = (y_hat_relation.detach().cpu()*(y_sup_ext==1).float() + (1-y_hat_relation.detach().cpu())*(y_sup_ext==0).float()).view(-1,10,10)
decision = decision * y_mask.detach().cpu().repeat(1,1,1,2)
y_pred = torch.argmax(decision.sum(2),2).numpy()
sim_score = (y_hat_relation.detach()*y_mask).cpu().numpy()
# Prepare display
sample_sup = label_sup[0,:num_support[0],1].detach().long().cpu().numpy().flatten()
sample_que = label_que[0,:num_query[0],1].detach().long().numpy().flatten()
sample_pred = y_pred[0,:num_query[0]].flatten()
# Acc
total_corrects += np.sum((y_pred == label_que[:,:,1].long().numpy()) * y_mask[:,:,0,0].cpu().numpy())
total_query += np.sum(num_query)
if (session+1)%5000 == 0:
hist_trloss.append(total_trloss/5000)
hist_tracc.append(total_corrects/total_query)
tqdm.write(np.array2string(sim_score[0,:,:,0]))
tqdm.write("S:" + np.array2string(sample_sup) +'\n'+
"Q:" + np.array2string(sample_que) + '\n' +
"P:" + np.array2string(sample_pred) )
tqdm.write("tr_session:{0:} tr_loss:{1:.6f} tr_acc:{2:.4f}".format(session, hist_trloss[-1], hist_tracc[-1]))
total_corrects = 0
total_query = 0
total_trloss = 0
if (session+1)%40000 == 0:
# Validation
validate()
# Save
torch.save({'ep': epoch, 'sess':session, 'FE_state': FeatEnc.state_dict(), 'RN_state': RN.state_dict(), 'loss': loss, 'hist_vacc': hist_vacc,
'hist_vloss': hist_vloss, 'hist_trloss': hist_trloss, 'FE_opt_state': FeatEnc_optim.state_dict(), 'RN_opt_state': RN_optim.state_dict(),
'FE_sch_state': FeatEnc_scheduler.state_dict(), 'RN_sch_state': RN_scheduler.state_dict()}, MODEL_SAVE_PATH + "check_{0:}_{1:}.pth".format(epoch, session))
del loss, x_feat_sup, x_feat_que, y_hat_relation