search

JinghuiZhou / awesome_face_antispoofing

Single Shot Face Anti-spoofing.
286 stars 84 forks source link

Bug report:About train_acc and val_acc show err in visdom #12

Open zygmcc opened 5 years ago

zygmcc commented 5 years ago

About main.py:line 229 In this part: `vis.plot_many_stack({'train_loss': train_loss.value()[0], \ 'val_loss': val_loss.value()[0]}, win_name="Loss")

vis.plot_many_stack({'train_acc': train_acc.value()[0], \ 'val_acc': val_acc.value()[0]}, win_name='Acc')`

When using GPU training，var train_acc and val_acc should use .cpu() to load in memory，So I change source code like this: if opt.use_gpu: val_acc.add(v_accuracy.cpu()) else: val_acc.add(v_accuracy) And if opt.use_gpu: train_acc.add(epoch_acc.cpu()) else: train_acc.add(epoch_acc)

zygmcc commented 5 years ago

image image On resnet18 I fell sad

JaryHuang commented 5 years ago

image image On resnet18 I fell sad Hello, I think it is not err in visdom. First,I repeat training the resnet18 with batch_size=16 three times.I found: screenshot from 2019-01-12 15-45-19 it is unstable. So,I guee the reason may be as follows: 1.Due to the train data is small. And the train data shuffle.when batch=16,is it small 2.the resnet18 is not robust enough. I just change the batch_size=64 screenshot from 2019-01-12 15-53-16 it is relatiuvely stable. so you bug just is the training error.

justein commented 5 years ago

@zygmcc would you mind share the way to solve the 'tensor(acc=0.985,dtype=...) is not JSON serializable' problem ?