Exemplo n.º 1
0
    idxStart = bi * batch_size
    inputData_np = test_X_std[idxStart:(idxStart + batch_size), :, :]
    inputData_np_ori = test_X[idxStart:(idxStart + batch_size), :, :]
    # outputData_np = test_Y_std[idxStart:(idxStart+batch_size),:,:]
    # outputData_np_ori = test_Y[idxStart:(idxStart+batch_size),:,:]

    inputData = Variable(
        torch.from_numpy(inputData_np)).cuda()  #(batch, 73, frameNum)
    # outputGT = Variable(torch.from_numpy(outputData_np)).cuda()  #(batch, 73, frameNum)
    #outputGT = Variable(torch.from_numpy(inputData_np)).cuda()  #(batch, 73, frameNum)

    # ===================forward=====================
    if "vae" in model.__class__.__name__:
        output, mu, logvar = model(inputData)
        loss, recon_loss, kld_loss = modelZoo.vae_loss_function(
            output, inputData, mu, logvar, criterion, args.weight_kld)
    else:
        output = model(inputData)
        #loss = criterion(output, outputGT)
        loss = criterion(output, inputData)

    print('loss: {}'.format(loss))
    #De-standardaize
    output_np = output.data.cpu().numpy()  #(batch, featureDim, frames)
    output_np = output_np * Xstd + Xmean

    output_np = np.swapaxes(output_np, 1, 2)  #(batch, frames, featureDim)
    output_np = np.reshape(output_np, (-1, featureDim))
    output_np = np.swapaxes(output_np, 0, 1)  #(featureDim, frames)

    inputData_np_ori = np.swapaxes(inputData_np_ori, 1,
Exemplo n.º 2
0
        inputData = Variable(
            torch.from_numpy(inputData_np)).cuda()  #(batch, 73, frameNum)

        inputData_speech_cuda = Variable(
            torch.from_numpy(inputdata_speech[:, :, 0])).cuda()
        #outputGT = Variable(torch.from_numpy(outputData_np)).cuda()  #(batch, 73, frameNum)
        #outputGT = Variable(torch.from_numpy(inputData_np)).cuda()  #(batch, 73, frameNum)

        #################### VAE Only ####################
        # ===================forward=====================
        output, mu, logvar = model(inputData, inputData_speech_cuda)
        #loss = criterion(output, inputData)
        #loss = modelZoo.vae_loss_function(output, inputData, mu, logvar,criterion)
        #loss, recon_loss, kld_loss = modelZoo.vae_loss_function(output, inputData, mu, logvar,criterion,args.weight_kld)
        loss, recon_loss, kld_loss = modelZoo.vae_loss_function(
            output, inputData[:, :-1, :], mu, logvar, criterion,
            args.weight_kld)  #ignore label in the inputData

        # ===================backward====================
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

        # ===================log========================
        # print('model: {}, epoch [{}/{}], loss:{:.4f} (recon: {:.4f}, kld {:.4f})'
        #             .format(checkpointFolder_base, epoch +pretrain_epoch, num_epochs, loss.item(), recon_loss.item(), kld_loss.item()))
        avgLoss += loss.item() * batch_size
        avgReconLoss += recon_loss.item() * batch_size
        avgKLDLoss += kld_loss.item() * batch_size

        if tensorboard_bLog: