-
Notifications
You must be signed in to change notification settings - Fork 58
/
rnn_vae.py
393 lines (325 loc) · 16.8 KB
/
rnn_vae.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
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Variational Animal Motion Embedding 0.1 Toolbox
© K. Luxem & P. Bauer, Department of Cellular Neuroscience
Leibniz Institute for Neurobiology, Magdeburg, Germany
https://github.com/LINCellularNeuroscience/VAME
Licensed under GNU General Public License v3.0
"""
import torch
from torch import nn
import torch.utils.data as Data
from torch.autograd import Variable
from torch.optim.lr_scheduler import StepLR
import os
import numpy as np
from pathlib import Path
from vame.util.auxiliary import read_config
from vame.model.dataloader import SEQUENCE_DATASET
from vame.model.rnn_model import RNN_VAE, RNN_VAE_LEGACY
# make sure torch uses cuda for GPU computing
use_gpu = torch.cuda.is_available()
if use_gpu:
print("Using CUDA")
print('GPU active:',torch.cuda.is_available())
print('GPU used:',torch.cuda.get_device_name(0))
else:
torch.device("cpu")
def reconstruction_loss(x, x_tilde, reduction):
mse_loss = nn.MSELoss(reduction=reduction)
rec_loss = mse_loss(x_tilde,x)
return rec_loss
def future_reconstruction_loss(x, x_tilde, reduction):
mse_loss = nn.MSELoss(reduction=reduction)
rec_loss = mse_loss(x_tilde,x)
return rec_loss
def cluster_loss(H, kloss, lmbda, batch_size):
gram_matrix = (H.T @ H) / batch_size
_ ,sv_2, _ = torch.svd(gram_matrix)
sv = torch.sqrt(sv_2[:kloss])
loss = torch.sum(sv)
return lmbda*loss
def kullback_leibler_loss(mu, logvar):
# see Appendix B from VAE paper:
# Kingma and Welling. Auto-Encoding Variational Bayes. ICLR, 2014
# https://arxiv.org/abs/1312.6114
# 0.5 * sum(1 + log(sigma^2) - mu^2 - sigma^2)
# I'm using torch.mean() here as the sum() version depends on the size of the latent vector
KLD = -0.5 * torch.mean(1 + logvar - mu.pow(2) - logvar.exp())
return KLD
def kl_annealing(epoch, kl_start, annealtime, function):
"""
Annealing of Kullback-Leibler loss to let the model learn first
the reconstruction of the data before the KL loss term gets introduced.
"""
if epoch > kl_start:
if function == 'linear':
new_weight = min(1, (epoch-kl_start)/(annealtime))
elif function == 'sigmoid':
new_weight = float(1/(1+np.exp(-0.9*(epoch-annealtime))))
else:
raise NotImplementedError('currently only "linear" and "sigmoid" are implemented')
return new_weight
else:
new_weight = 0
return new_weight
def gaussian(ins, is_training, seq_len, std_n=0.8):
if is_training:
emp_std = ins.std(1)*std_n
emp_std = emp_std.unsqueeze(2).repeat(1, 1, seq_len)
emp_std = emp_std.permute(0,2,1)
noise = Variable(ins.data.new(ins.size()).normal_(0, 1))
return ins + (noise*emp_std)
return ins
def train(train_loader, epoch, model, optimizer, anneal_function, BETA, kl_start,
annealtime, seq_len, future_decoder, future_steps, scheduler, mse_red,
mse_pred, kloss, klmbda, bsize, noise):
model.train() # toggle model to train mode
train_loss = 0.0
mse_loss = 0.0
kullback_loss = 0.0
kmeans_losses = 0.0
fut_loss = 0.0
loss = 0.0
seq_len_half = int(seq_len / 2)
for idx, data_item in enumerate(train_loader):
data_item = Variable(data_item)
data_item = data_item.permute(0,2,1)
if use_gpu:
data = data_item[:,:seq_len_half,:].type('torch.FloatTensor').cuda()
fut = data_item[:,seq_len_half:seq_len_half+future_steps,:].type('torch.FloatTensor').cuda()
else:
data = data_item[:,:seq_len_half,:].type('torch.FloatTensor').to()
fut = data_item[:,seq_len_half:seq_len_half+future_steps,:].type('torch.FloatTensor').to()
if noise == True:
data_gaussian = gaussian(data,True,seq_len_half)
else:
data_gaussian = data
if future_decoder:
data_tilde, future, latent, mu, logvar = model(data_gaussian)
rec_loss = reconstruction_loss(data, data_tilde, mse_red)
fut_rec_loss = future_reconstruction_loss(fut, future, mse_pred)
kmeans_loss = cluster_loss(latent.T, kloss, klmbda, bsize)
kl_loss = kullback_leibler_loss(mu, logvar)
kl_weight = kl_annealing(epoch, kl_start, annealtime, anneal_function)
loss = rec_loss + fut_rec_loss + BETA*kl_weight*kl_loss + kl_weight*kmeans_loss
fut_loss += fut_rec_loss.item()
else:
data_tilde, latent, mu, logvar = model(data_gaussian)
rec_loss = reconstruction_loss(data, data_tilde, mse_red)
kl_loss = kullback_leibler_loss(mu, logvar)
kmeans_loss = cluster_loss(latent.T, kloss, klmbda, bsize)
kl_weight = kl_annealing(epoch, kl_start, annealtime, anneal_function)
loss = rec_loss + BETA*kl_weight*kl_loss + kl_weight*kmeans_loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
# torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm = 5)
train_loss += loss.item()
mse_loss += rec_loss.item()
kullback_loss += kl_loss.item()
kmeans_losses += kmeans_loss.item()
if idx % 1000 == 0:
print('Epoch: %d. loss: %.4f' %(epoch, loss.item()))
scheduler.step() #be sure scheduler is called before optimizer in >1.1 pytorch
if future_decoder:
print('Train loss: {:.3f}, MSE-Loss: {:.3f}, MSE-Future-Loss {:.3f}, KL-Loss: {:.3f}, Kmeans-Loss: {:.3f}, weight: {:.2f}'.format(train_loss / idx,
mse_loss /idx, fut_loss/idx, BETA*kl_weight*kullback_loss/idx, kl_weight*kmeans_losses/idx, kl_weight))
else:
print('Train loss: {:.3f}, MSE-Loss: {:.3f}, KL-Loss: {:.3f}, Kmeans-Loss: {:.3f}, weight: {:.2f}'.format(train_loss / idx,
mse_loss /idx, BETA*kl_weight*kullback_loss/idx, kl_weight*kmeans_losses/idx, kl_weight))
return kl_weight, train_loss/idx, kl_weight*kmeans_losses/idx, kullback_loss/idx, mse_loss/idx, fut_loss/idx
def test(test_loader, epoch, model, optimizer, BETA, kl_weight, seq_len, mse_red, kloss, klmbda, future_decoder, bsize):
model.eval() # toggle model to inference mode
test_loss = 0.0
mse_loss = 0.0
kullback_loss = 0.0
kmeans_losses = 0.0
loss = 0.0
seq_len_half = int(seq_len / 2)
with torch.no_grad():
for idx, data_item in enumerate(test_loader):
# we're only going to infer, so no autograd at all required
data_item = Variable(data_item)
data_item = data_item.permute(0,2,1)
if use_gpu:
data = data_item[:,:seq_len_half,:].type('torch.FloatTensor').cuda()
else:
data = data_item[:,:seq_len_half,:].type('torch.FloatTensor').to()
if future_decoder:
recon_images, _, latent, mu, logvar = model(data)
rec_loss = reconstruction_loss(data, recon_images, mse_red)
kl_loss = kullback_leibler_loss(mu, logvar)
kmeans_loss = cluster_loss(latent.T, kloss, klmbda, bsize)
loss = rec_loss + BETA*kl_weight*kl_loss+ kl_weight*kmeans_loss
else:
recon_images, latent, mu, logvar = model(data)
rec_loss = reconstruction_loss(data, recon_images, mse_red)
kl_loss = kullback_leibler_loss(mu, logvar)
kmeans_loss = cluster_loss(latent.T, kloss, klmbda, bsize)
loss = rec_loss + BETA*kl_weight*kl_loss + kl_weight*kmeans_loss
# torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm = 5)
test_loss += loss.item()
mse_loss += rec_loss.item()
kullback_loss += kl_loss.item()
kmeans_losses += kmeans_loss
print('Test loss: {:.3f}, MSE-Loss: {:.3f}, KL-Loss: {:.3f}, Kmeans-Loss: {:.3f}\n'.format(test_loss / idx,
mse_loss /idx, BETA*kl_weight*kullback_loss/idx, kl_weight*kmeans_losses/idx))
return mse_loss /idx, test_loss/idx, kl_weight*kmeans_losses
def train_model(config):
config_file = Path(config).resolve()
cfg = read_config(config_file)
legacy = cfg['legacy']
model_name = cfg['model_name']
pretrained_weights = cfg['pretrained_weights']
pretrained_model = cfg['pretrained_model']
print("Train Variational Autoencoder - Model name: %s \n" %model_name)
if not os.path.exists(os.path.join(cfg['project_path'],'model','best_model',"")):
os.mkdir(os.path.join(cfg['project_path'],'model','best_model',""))
os.mkdir(os.path.join(cfg['project_path'],'model','best_model','snapshots',""))
os.mkdir(os.path.join(cfg['project_path'],'model','model_losses',""))
# make sure torch uses cuda for GPU computing
use_gpu = torch.cuda.is_available()
if use_gpu:
print("Using CUDA")
print('GPU active:',torch.cuda.is_available())
print('GPU used: ',torch.cuda.get_device_name(0))
else:
torch.device("cpu")
print("warning, a GPU was not found... proceeding with CPU (slow!) \n")
#raise NotImplementedError('GPU Computing is required!')
""" HYPERPARAMTERS """
# General
CUDA = use_gpu
SEED = 19
TRAIN_BATCH_SIZE = cfg['batch_size']
TEST_BATCH_SIZE = int(cfg['batch_size']/4)
EPOCHS = cfg['max_epochs']
ZDIMS = cfg['zdims']
BETA = cfg['beta']
SNAPSHOT = cfg['model_snapshot']
LEARNING_RATE = cfg['learning_rate']
NUM_FEATURES = cfg['num_features']
if legacy == False:
NUM_FEATURES = NUM_FEATURES - 2
TEMPORAL_WINDOW = cfg['time_window']*2
FUTURE_DECODER = cfg['prediction_decoder']
FUTURE_STEPS = cfg['prediction_steps']
# RNN
hidden_size_layer_1 = cfg['hidden_size_layer_1']
hidden_size_layer_2 = cfg['hidden_size_layer_2']
hidden_size_rec = cfg['hidden_size_rec']
hidden_size_pred = cfg['hidden_size_pred']
dropout_encoder = cfg['dropout_encoder']
dropout_rec = cfg['dropout_rec']
dropout_pred = cfg['dropout_pred']
noise = cfg['noise']
scheduler_step_size = cfg['scheduler_step_size']
# Loss
MSE_REC_REDUCTION = cfg['mse_reconstruction_reduction']
MSE_PRED_REDUCTION = cfg['mse_prediction_reduction']
KMEANS_LOSS = cfg['kmeans_loss']
KMEANS_LAMBDA = cfg['kmeans_lambda']
KL_START = cfg['kl_start']
ANNEALTIME = cfg['annealtime']
anneal_function = cfg['anneal_function']
optimizer_scheduler = cfg['scheduler']
BEST_LOSS = 999999
convergence = 0
print('Latent Dimensions: %d, Time window: %d, Beta: %d, lr: %.4f\n' %(ZDIMS, cfg['time_window'], BETA, LEARNING_RATE))
# simple logging of diverse losses
train_losses = []
test_losses = []
kmeans_losses = []
kl_losses = []
weight_values = []
mse_losses = []
fut_losses = []
torch.manual_seed(SEED)
if legacy == False:
RNN = RNN_VAE
else:
RNN = RNN_VAE_LEGACY
if CUDA:
torch.cuda.manual_seed(SEED)
model = RNN(TEMPORAL_WINDOW,ZDIMS,NUM_FEATURES,FUTURE_DECODER,FUTURE_STEPS, hidden_size_layer_1,
hidden_size_layer_2, hidden_size_rec, hidden_size_pred, dropout_encoder,
dropout_rec, dropout_pred).cuda()
else: #cpu support ...
torch.cuda.manual_seed(SEED)
model = RNN(TEMPORAL_WINDOW,ZDIMS,NUM_FEATURES,FUTURE_DECODER,FUTURE_STEPS, hidden_size_layer_1,
hidden_size_layer_2, hidden_size_rec, hidden_size_pred, dropout_encoder,
dropout_rec, dropout_pred).to()
if pretrained_weights:
if os.path.exists(os.path.join(cfg['project_path'],'model','best_model',pretrained_model+'_'+cfg['Project']+'.pkl')): #TODO, fix this path seeking....
print("Loading pretrained Model: %s\n" %pretrained_model)
model.load_state_dict(torch.load(os.path.join(cfg['project_path'],'model','best_model',pretrained_model+'_'+cfg['Project']+'.pkl'), strict=False))
""" DATASET """
trainset = SEQUENCE_DATASET(os.path.join(cfg['project_path'],"data", "train",""), data='train_seq.npy', train=True, temporal_window=TEMPORAL_WINDOW)
testset = SEQUENCE_DATASET(os.path.join(cfg['project_path'],"data", "train",""), data='test_seq.npy', train=False, temporal_window=TEMPORAL_WINDOW)
train_loader = Data.DataLoader(trainset, batch_size=TRAIN_BATCH_SIZE, shuffle=True, drop_last=True)
test_loader = Data.DataLoader(testset, batch_size=TEST_BATCH_SIZE, shuffle=True, drop_last=True)
optimizer = torch.optim.Adam(model.parameters(), lr=LEARNING_RATE, amsgrad=True)
if optimizer_scheduler:
print('Scheduler step size: %d, Scheduler gamma: %.2f\n' %(scheduler_step_size, cfg['scheduler_gamma']))
scheduler = StepLR(optimizer, step_size=scheduler_step_size, gamma=cfg['scheduler_gamma'], last_epoch=-1)
else:
scheduler = StepLR(optimizer, step_size=scheduler_step_size, gamma=1, last_epoch=-1)
for epoch in range(1,EPOCHS):
weight, train_loss, km_loss, kl_loss, mse_loss, fut_loss = train(train_loader, epoch, model, optimizer,
anneal_function, BETA, KL_START,
ANNEALTIME, TEMPORAL_WINDOW, FUTURE_DECODER,
FUTURE_STEPS, scheduler, MSE_REC_REDUCTION,
MSE_PRED_REDUCTION, KMEANS_LOSS, KMEANS_LAMBDA,
TRAIN_BATCH_SIZE, noise)
current_loss, test_loss, test_list = test(test_loader, epoch, model, optimizer,
BETA, weight, TEMPORAL_WINDOW, MSE_REC_REDUCTION,
KMEANS_LOSS, KMEANS_LAMBDA, FUTURE_DECODER, TEST_BATCH_SIZE)
for param_group in optimizer.param_groups:
print('lr: {}'.format(param_group['lr']))
# logging losses
train_losses.append(train_loss)
test_losses.append(test_loss)
kmeans_losses.append(km_loss)
kl_losses.append(kl_loss)
weight_values.append(weight)
mse_losses.append(mse_loss)
fut_losses.append(fut_loss)
# save best model
if weight > 0.99 and current_loss <= BEST_LOSS:
BEST_LOSS = current_loss
print("Saving model!\n")
if use_gpu:
torch.save(model.state_dict(), os.path.join(cfg['project_path'],"model", "best_model",model_name+'_'+cfg['Project']+'.pkl'))
else:
torch.save(model.state_dict(), os.path.join(cfg['project_path'],"model", "best_model",model_name+'_'+cfg['Project']+'.pkl'))
convergence = 0
else:
convergence += 1
if epoch % SNAPSHOT == 0:
print("Saving model snapshot!\n")
torch.save(model.state_dict(), os.path.join(cfg['project_path'],'model','best_model','snapshots',model_name+'_'+cfg['Project']+'_epoch_'+str(epoch)+'.pkl'))
if convergence > cfg['model_convergence']:
print('Model converged. Please check your model with vame.evaluate_model(). \n'
'You can also re-run vame.rnn_model() to further improve your model. \n'
'Hint: Set "model_convergence" in your config.yaml to a higher value. \n'
'\n'
'Next: \n'
'Use vame.behavior_segmentation() to identify behavioral motifs in your dataset!')
#return
break
# save logged losses
np.save(os.path.join(cfg['project_path'],'model','model_losses','train_losses_'+model_name), train_losses)
np.save(os.path.join(cfg['project_path'],'model','model_losses','test_losses_'+model_name), test_losses)
np.save(os.path.join(cfg['project_path'],'model','model_losses','kmeans_losses_'+model_name), kmeans_losses)
np.save(os.path.join(cfg['project_path'],'model','model_losses','kl_losses_'+model_name), kl_losses)
np.save(os.path.join(cfg['project_path'],'model','model_losses','weight_values_'+model_name), weight_values)
np.save(os.path.join(cfg['project_path'],'model','model_losses','mse_train_losses_'+model_name), mse_losses)
np.save(os.path.join(cfg['project_path'],'model','model_losses','mse_test_losses_'+model_name), current_loss)
np.save(os.path.join(cfg['project_path'],'model','model_losses','fut_losses_'+model_name), fut_losses)
if convergence < cfg['model_convergence']:
print('Model seemed to have not reached convergence. You may want to check your model \n'
'with vame.evaluate_model(). If your satisfied you can continue with \n'
'Use vame.behavior_segmentation() to identify behavioral motifs!\n\n'
'OPTIONAL: You can re-run vame.rnn_model() to improve performance.')