DNN2_net.load_state_dict(checkpoint_load['DNN2_model_par'])
optimizer_CNN = optim.RMSprop(CNN_net.parameters(), lr=lr,alpha=0.95, eps=1e-8)
optimizer_DNN1 = optim.RMSprop(DNN1_net.parameters(), lr=lr,alpha=0.95, eps=1e-8)
optimizer_DNN2 = optim.RMSprop(DNN2_net.parameters(), lr=lr,alpha=0.95, eps=1e-8)
# print("----------------------")
# print(DNN2_net(DNN1_net(CNN_net(3200))))
#for epoch in range(N_epochs):
test_flag=0
CNN_net.train()
DNN1_net.train()
DNN2_net.train()
loss_sum=0
err_sum=0
# for i in range(N_batches):
# [inp,lab]=create_batches_rnd(batch_size,data_folder,wav_lst_tr,snt_tr,wlen,lab_dict,0.2)
# pout=DNN2_net(DNN1_net(CNN_net(inp)))
# pred=torch.max(pout,dim=1)[1]
# loss = cost(pout, lab.long())
# err = torch.mean((pred!=lab.long()).float())
eps=1e-8)
optimizer_DNN2 = optim.RMSprop(DNN2_net.parameters(),
lr=lr,
alpha=0.95,
eps=1e-8)
# 开始进行训练，1500次
# 开始进行训练
for epoch in range(N_epochs):
test_flag = 0
CNN_net.train(
) # 这里因为里面包含了的dropout和batchnorm，因此需要指明是在训练还是在验证，在训练时则要.train,在验证时要.eval
AttentionModule.train()
DNN1_net.train() # 注意要在自动forward之前
DNN2_net.train()
loss_sum = 0
err_sum = 0
for i in range(N_batches): # 处理一批 for循环一共有800批
# 获取数据 batch_size = 128
[inp, lab] = create_batches_rnd(batch_size, data_folder, wav_lst_tr,
snt_tr, wlen, lab_dict, 0.2)
# 进行训练
output = CNN_net(inp)
# under deprecate
# print(output.shape) [128, 6420] 128条片段，每个是6420的
# o1, o3 = output.split([1, 1], dim=1) # 切割列