loss_value2 = []
# loss_img_value = []
scheduler.step()
for i_batch, sample in enumerate(dataloader):
data = sample['frames'].squeeze(0).cuda(gpu_id)
flows = sample['flows'].squeeze(0)
expectedOut = Variable(data)
inputData = Variable(data[:, 0:FRA, :])
optimizer.zero_grad()
y = data[:, 0:FRA + PRE, :].data
dct = model.forward(inputData)
dct1 = model1.forward(inputData)
dct2 = model2.forward(inputData)
'''Plot DCT coefficients'''
coef = dct.data
coef1 = dct1.data
coef2 = dct2.data
coefP = torch.t(coef[0, :, :]).data.cpu().detach().numpy()
coefP1 = torch.t(coef1[0, :, :]).data.cpu().detach().numpy()
coefP2 = torch.t(coef2[0, :, :]).data.cpu().detach().numpy()
coefR = torch.t(y[0, :, :])
maxcp, idxcp = coefR[:, 0].max(0)
plt.figure()
t = np.arange(0, FRA + PRE, 1)
gt, = plt.plot(t, coefR[idxcp].data.cpu().detach().numpy(), label="GT")
for epoch in range(start_epoch, EPOCH + 1):
loss_value = []
loss_valueR = []
schedulerrgb.step()
for i_batch, sample in enumerate(dataloader):
data = sample['alframes'].squeeze(0).cuda(gpu_id)
dataori = sample['frames'].squeeze(0).cuda(gpu_id)
of = sample['of'].squeeze(0).cuda(
gpu_id) #could be done in different GPU's
ofori = sample['ofori'].squeeze(0) #.cuda(gpu_id)
expectedOut = Variable(data)
inputData = Variable(data[:, 0:N_FRAME, :])
optimizerrgb.zero_grad()
output = modelrgb.forward(inputData)
with torch.no_grad():
ofout = ofmodel.forward(Variable(of[:, 0:FRA, :]))
out = warp(output[:, FRA, :].view(3, x_fra, y_fra).unsqueeze(0),
ofout[:, FRA, :].view(2, x_fra, y_fra).unsqueeze(0))
showframes(
out.view(3, x_fra, y_fra).unsqueeze(0).unsqueeze(2),
expectedOut[:, N_FRAME, :].view(3, x_fra,
y_fra).unsqueeze(0).unsqueeze(2),
1, '')
# lossR = loss_mse(output[:, FRA, :], expectedOut[:, FRA, :])
loss = loss_mse(output[:, FRA, :], expectedOut[:, FRA, :]) \
+ loss_mse(output[:, 0:FRA, :], expectedOut[:, 0:FRA, :])/(2*FRA) # probar amb expectedout com a GT
torch.cat(c_list_VI, out=c_VI)
C[3, :, :] = c_VI.permute(1, 2, 0, 3).reshape(-1, x_fra * y_fra * CLen).unsqueeze(0)
C = C.permute(1,0,2).reshape(-1, x_fra * y_fra * CLen * 4).unsqueeze(0)
C_t.append(C)
print(len(C_t))
print('Time for spacial encoding in time T: ', time.time() - t0_sp)
## REGULAR DYAN
Ctens_t = torch.Tensor(FRA, 2, CLen*x_fra*y_fra*4).cuda(gpu_id)
torch.cat(C_t, out=Ctens_t)
print(Ctens_t.shape, type(Ctens_t))
Cpred = model_ti.forward(Ctens_t.permute(1,0,2))
print('Time per forward: ', time.time() - t0_epoch)
## GPU LACKS MEMORY - TOO MUCH TIME --> ENCODERS COULD BE PARALLELIZED
sys.exit()
loss_val = np.mean(np.array(loss_value))
## Checkpoint + prints
if epoch % saveEvery ==0 :
save_checkpoint({ 'epoch': epoch + 1,
'state_dict': model.state_dict(),