# optimizer_DNN2.step()
# loss_sum=loss_sum+loss.detach()
# err_sum=err_sum+err.detach()
# loss_tot=loss_sum/N_batches
# err_tot=err_sum/N_batches
# Full Validation new
# if epoch%N_eval_epoch==0:
CNN_net.eval()
DNN1_net.eval()
DNN2_net.eval()
test_flag=1
# loss_sum=0
# err_sum=0
# err_sum_snt=0
correct = 0
with torch.no_grad():
for i in range(snt_te):
#[fs,signal]=scipy.io.wavfile.read(data_folder+wav_lst_te[i])
#signal=signal.astype(float)/32768
# print('error total:' + str(err_tot.item()))
# 最开始是
# loss_totall:5.845455169677734
# error total:0.9918749928474426
# second epoch
# loss_totall: 5.143251895904541
# error total: 0.9721777439117432
# the 4th time
# loss_totall: 4.186425685882568
# error total: 0.8866991996765137
# epoch3, loss_tr = 4.186426 err_tr = 0.886699
# Full Validation new
if epoch % N_eval_epoch == 0: # 每8个写入到文件一下
# os.system("pause")
CNN_net.eval() # 说明是在验证
AttentionModule.eval()
DNN1_net.eval()
DNN2_net.eval()
test_flag = 1
loss_sum = 0
err_sum = 0
err_sum_snt = 0
with torch.no_grad(): # 由于是三个不同的模型拼装所以no_grad上下文不再自动进行梯度下降
for i in range(snt_te): # 列表长度
# [fs,signal]=scipy.io.wavfile.read(data_folder+wav_lst_te[i])
# signal=signal.astype(float)/32768
[signal, fs] = sf.read(data_folder + wav_lst_te[i])