is_train=True when testing?

LucyLu-LX commented 6 years ago

https://github.com/liuziwei7/voxel-flow/blob/e796773c4a93ab5b8bf9671f0dd11b6ecfc12007/voxel_flow_train.py#L189

I am wondering why is_train=True when testing model. I compared the results between when is_train=True and when is_train=False. Oddly, is_train=True produces the correct result when testing.

liuziwei7 commented 6 years ago

It is used to accommodate for the legacy batch norm bug (i.e. running mean and variance) in the early TensorFlow version. But the network parameters are not updated during the process.

Dee-Ma commented 5 years ago

voxel-flow/voxel_flow_train.py

Line 189 in e796773

model = Voxel_flow_model(is_train=True) I am wondering why is_train=True when testing model. I compared the results between when is_train=True and when is_train=False. Oddly, is_train=True produces the correct result when testing.

Hi,

I am also using the Pytorch code to reproduce the DVF. I just got some questions and wish your kind help.

How do you get the 'convert_bn' and 'SyncBarchNorm2d' that are used in 'voxel_flow.py'? Since I could not find them in 'core/ops/sync_bn' folder.
What is your running environment for this code? As I just got 'torch.cuda.is_available()=False' even the appropriate CUDA and Pytorch were installed.....

Any help would be appreciated.

Dee-Ma commented 5 years ago

@LucyLu-LX

Hi,

I am also using the Pytorch code to reproduce the DVF. I just got some questions and wish your kind help.

How do you get the 'convert_bn' and 'SyncBarchNorm2d' that are used in 'voxel_flow.py'? Since I could not find them in 'core/ops/sync_bn' folder.

What is your running environment for this code? As I just got 'torch.cuda.is_available()=False' even the appropriate CUDA and Pytorch were installed.....

Any help would be appreciated.

liuziwei7 / voxel-flow

is_train=True when testing? #7