-
Notifications
You must be signed in to change notification settings - Fork 50
/
Copy pathmain.py
executable file
·165 lines (131 loc) · 7.43 KB
/
main.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
import torch
import torchvision
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
from torch.optim.lr_scheduler import ReduceLROnPlateau
import numpy as np
import random
import time
from models.models import ConvLSTM,PhyCell, EncoderRNN
from data.moving_mnist import MovingMNIST
from constrain_moments import K2M
from skimage.measure import compare_ssim as ssim
import argparse
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
parser = argparse.ArgumentParser()
parser.add_argument('--root', type=str, default='data/')
parser.add_argument('--batch_size', type=int, default=16, help='batch_size')
parser.add_argument('--nepochs', type=int, default=2001, help='nb of epochs')
parser.add_argument('--print_every', type=int, default=1, help='')
parser.add_argument('--eval_every', type=int, default=10, help='')
parser.add_argument('--save_name', type=str, default='phydnet', help='')
args = parser.parse_args()
mm = MovingMNIST(root=args.root, is_train=True, n_frames_input=10, n_frames_output=10, num_objects=[2])
train_loader = torch.utils.data.DataLoader(dataset=mm, batch_size=args.batch_size, shuffle=True, num_workers=0)
mm = MovingMNIST(root=args.root, is_train=False, n_frames_input=10, n_frames_output=10, num_objects=[2])
test_loader = torch.utils.data.DataLoader(dataset=mm, batch_size=args.batch_size, shuffle=False, num_workers=0)
constraints = torch.zeros((49,7,7)).to(device)
ind = 0
for i in range(0,7):
for j in range(0,7):
constraints[ind,i,j] = 1
ind +=1
def train_on_batch(input_tensor, target_tensor, encoder, encoder_optimizer, criterion,teacher_forcing_ratio):
encoder_optimizer.zero_grad()
# input_tensor : torch.Size([batch_size, input_length, channels, cols, rows])
input_length = input_tensor.size(1)
target_length = target_tensor.size(1)
loss = 0
for ei in range(input_length-1):
encoder_output, encoder_hidden, output_image,_,_ = encoder(input_tensor[:,ei,:,:,:], (ei==0) )
loss += criterion(output_image,input_tensor[:,ei+1,:,:,:])
decoder_input = input_tensor[:,-1,:,:,:] # first decoder input = last image of input sequence
use_teacher_forcing = True if random.random() < teacher_forcing_ratio else False
for di in range(target_length):
decoder_output, decoder_hidden, output_image,_,_ = encoder(decoder_input)
target = target_tensor[:,di,:,:,:]
loss += criterion(output_image,target)
if use_teacher_forcing:
decoder_input = target # Teacher forcing
else:
decoder_input = output_image
# Moment regularization # encoder.phycell.cell_list[0].F.conv1.weight # size (nb_filters,in_channels,7,7)
k2m = K2M([7,7]).to(device)
for b in range(0,encoder.phycell.cell_list[0].input_dim):
filters = encoder.phycell.cell_list[0].F.conv1.weight[:,b,:,:] # (nb_filters,7,7)
m = k2m(filters.double())
m = m.float()
loss += criterion(m, constraints) # constrains is a precomputed matrix
loss.backward()
encoder_optimizer.step()
return loss.item() / target_length
def trainIters(encoder, nepochs, print_every=10,eval_every=10,name=''):
train_losses = []
best_mse = float('inf')
encoder_optimizer = torch.optim.Adam(encoder.parameters(),lr=0.001)
scheduler_enc = ReduceLROnPlateau(encoder_optimizer, mode='min', patience=2,factor=0.1,verbose=True)
criterion = nn.MSELoss()
for epoch in range(0, nepochs):
t0 = time.time()
loss_epoch = 0
teacher_forcing_ratio = np.maximum(0 , 1 - epoch * 0.003)
for i, out in enumerate(train_loader, 0):
input_tensor = out[1].to(device)
target_tensor = out[2].to(device)
loss = train_on_batch(input_tensor, target_tensor, encoder, encoder_optimizer, criterion, teacher_forcing_ratio)
loss_epoch += loss
train_losses.append(loss_epoch)
if (epoch+1) % print_every == 0:
print('epoch ',epoch, ' loss ',loss_epoch, ' time epoch ',time.time()-t0)
if (epoch+1) % eval_every == 0:
mse, mae,ssim = evaluate(encoder,test_loader)
scheduler_enc.step(mse)
torch.save(encoder.state_dict(),'save/encoder_{}.pth'.format(name))
return train_losses
def evaluate(encoder,loader):
total_mse, total_mae,total_ssim,total_bce = 0,0,0,0
t0 = time.time()
with torch.no_grad():
for i, out in enumerate(loader, 0):
input_tensor = out[1].to(device)
target_tensor = out[2].to(device)
input_length = input_tensor.size()[1]
target_length = target_tensor.size()[1]
for ei in range(input_length-1):
encoder_output, encoder_hidden, _,_,_ = encoder(input_tensor[:,ei,:,:,:], (ei==0))
decoder_input = input_tensor[:,-1,:,:,:] # first decoder input= last image of input sequence
predictions = []
for di in range(target_length):
decoder_output, decoder_hidden, output_image,_,_ = encoder(decoder_input, False, False)
decoder_input = output_image
predictions.append(output_image.cpu())
input = input_tensor.cpu().numpy()
target = target_tensor.cpu().numpy()
predictions = np.stack(predictions) # (10, batch_size, 1, 64, 64)
predictions = predictions.swapaxes(0,1) # (batch_size,10, 1, 64, 64)
mse_batch = np.mean((predictions-target)**2 , axis=(0,1,2)).sum()
mae_batch = np.mean(np.abs(predictions-target) , axis=(0,1,2)).sum()
total_mse += mse_batch
total_mae += mae_batch
for a in range(0,target.shape[0]):
for b in range(0,target.shape[1]):
total_ssim += ssim(target[a,b,0,], predictions[a,b,0,]) / (target.shape[0]*target.shape[1])
cross_entropy = -target*np.log(predictions) - (1-target) * np.log(1-predictions)
cross_entropy = cross_entropy.sum()
cross_entropy = cross_entropy / (args.batch_size*target_length)
total_bce += cross_entropy
print('eval mse ', total_mse/len(loader), ' eval mae ', total_mae/len(loader),' eval ssim ',total_ssim/len(loader), ' time= ', time.time()-t0)
return total_mse/len(loader), total_mae/len(loader), total_ssim/len(loader)
phycell = PhyCell(input_shape=(16,16), input_dim=64, F_hidden_dims=[49], n_layers=1, kernel_size=(7,7), device=device)
convcell = ConvLSTM(input_shape=(16,16), input_dim=64, hidden_dims=[128,128,64], n_layers=3, kernel_size=(3,3), device=device)
encoder = EncoderRNN(phycell, convcell, device)
def count_parameters(model):
return sum(p.numel() for p in model.parameters() if p.requires_grad)
print('phycell ' , count_parameters(phycell) )
print('convcell ' , count_parameters(convcell) )
print('encoder ' , count_parameters(encoder) )
trainIters(encoder,args.nepochs,print_every=args.print_every,eval_every=args.eval_every,name=args.save_name)
#encoder.load_state_dict(torch.load('save/encoder_phydnet.pth'))
#encoder.eval()
#mse, mae,ssim = evaluate(encoder,test_loader)