batch_size = 100
train_epoch = 30
input_size = 196
data_set = DataLoader(
datasets.MNIST(
"..\\data\\",
train = True,
download = False,
transform = transforms.Compose(
[transforms.Resize((56, 56)), transforms.ToTensor(), transforms.Normalize([0.5], [0.5])]
),
),
batch_size = batch_size,
shuffle = True,
)
batch_number = data_set.__len__()
G = Generator(input_size).cuda()
D = Discriminator().cuda()
gopt = optim.Adam(G.parameters(), lr = 4e-3)
dopt = optim.Adam(D.parameters(), lr = 5e-4)
lossFunc = nn.BCELoss()
real_labels = Var(torch.ones((batch_size, 1))).cuda()
fake_labels = Var(torch.zeros((batch_size, 1))).cuda()
for epoch in range(train_epoch):
for k, (bx, _) in enumerate(data_set):
dopt.zero_grad()
bx = bx.cuda()
# =============== 判别器训练 =================
real_out = D(bx)
d_loss_real = lossFunc(real_out, real_labels)
# print(np.shape(batch_data))
# # image = image.data.numpy()
# # image = np.sum(image[0], axis=0)
# # plt.imshow(image)
# # plt.show()
# # #print(np.shape(image))
# # #print(np.shape(heat_map))
#predict = model_AE.forward(Variable(image.type(torch.FloatTensor)).to(device))
#predict = model_AE.forward(Variable(image.type(torch.FloatTensor).to(device)))
optim.zero_grad()
predict = model_AE.forward(batch_data)
# # #print(predict)
output = criterion(predict, batch_label)
# optim.zero_grad()
output.backward()
optim.step()
loss_mean += output / (data_loader.__len__())
loss_mean_list.append(loss_mean)
print(loss_mean)
torch.save(model_AE, saved_model_path + "body_tracking_model.pth")
# train
print(cnt)
# print(loss_mean_list)
# f = open(saved_model_path + "loss_mean_list.txt", "w+")
# f.write(str(loss_mean_list))
# f.close()
