Pedestrian_Alignment icon indicating copy to clipboard operation
Pedestrian_Alignment copied to clipboard

Comment in resnet52_market.m

Open ahangchen opened this issue 8 years ago • 2 comments

In resnet52_market, fc ans softmax layers are commented.

When running train_id_net_res_market_new, missing 'label' layer will cause error, so it's needed to uncomment this lines when train base branch.

When running train_id_net_res_market_align, this code block should be commented, otherwise another error will be thrown out saying layer 'local_fc751's der is empty. But commenting this block also causes an error saying that 'label_local' layer is missed.

Hope for a better instruction for this code.

ahangchen avatar Aug 08 '17 10:08 ahangchen

Hi, You're right. I add more illustration in the README. Do the problem still keep appear? When training the basic network, we need the loss and dropout layers. When training the whole network, we need to remove the loss in the resnet52_market.m. We just want the net structure and params from ImageNet.

layumi avatar Aug 09 '17 01:08 layumi

In net2, we need top1 loss for label_local, in net3, we don't need the losses. If that's right, there exists a conflict when training the align net containing net2 and net3. My solution is commenting out the loss in resnet52_market, and adding the loss layer additionally only for net2 in train_id_net_res_market_align, then the training run successfully, but this is different from the origin code so I am still confused about the correctness of the solution.

ahangchen avatar Aug 09 '17 03:08 ahangchen