Pretrained Resnet3d giving reversed model

1

Hi,
I am trying to use a pretrained model of resnet3d on my gpu. But somehow it is getting reversed. I am not sure how this is possible. I am using

self.resnet3d = torch.hub.load('facebookresearch/pytorchvideo', 'slow_r50', pretrained=True)

And I am getting,

Net(
(blocks): ModuleList(
(0): ResNetBasicStem(
(pool): MaxPool3d(kernel_size=(1, 3, 3), stride=(1, 2, 2), padding=[0, 1, 1], dilation=1, ceil_mode=False)
(activation): ReLU()
(norm): BatchNorm3d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv): Conv3d(3, 64, kernel_size=(1, 7, 7), stride=(1, 2, 2), padding=[0, 3, 3], bias=False)
)
(1): ResStage(
(res_blocks): ModuleList(
(0): ResBlock(
(activation): ReLU()
(branch2): BottleneckBlock(
(norm_c): BatchNorm3d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_c): Conv3d(64, 256, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
(act_b): ReLU()
(norm_b): BatchNorm3d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_b): Conv3d(64, 64, kernel_size=(1, 3, 3), stride=(1, 1, 1), padding=(0, 1, 1), bias=False)
(act_a): ReLU()
(norm_a): BatchNorm3d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_a): Conv3d(64, 64, kernel_size=(1, 1, 1), stride=(1, 1, 1), padding=[0, 0, 0], bias=False)
)
(branch1_norm): BatchNorm3d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(branch1_conv): Conv3d(64, 256, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
)
(1): ResBlock(
(activation): ReLU()
(branch2): BottleneckBlock(
(norm_c): BatchNorm3d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_c): Conv3d(64, 256, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
(act_b): ReLU()
(norm_b): BatchNorm3d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_b): Conv3d(64, 64, kernel_size=(1, 3, 3), stride=(1, 1, 1), padding=(0, 1, 1), bias=False)
(act_a): ReLU()
(norm_a): BatchNorm3d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_a): Conv3d(256, 64, kernel_size=(1, 1, 1), stride=(1, 1, 1), padding=[0, 0, 0], bias=False)
)
)
(2): ResBlock(
(activation): ReLU()
(branch2): BottleneckBlock(
(norm_c): BatchNorm3d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_c): Conv3d(64, 256, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
(act_b): ReLU()
(norm_b): BatchNorm3d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_b): Conv3d(64, 64, kernel_size=(1, 3, 3), stride=(1, 1, 1), padding=(0, 1, 1), bias=False)
(act_a): ReLU()
(norm_a): BatchNorm3d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_a): Conv3d(256, 64, kernel_size=(1, 1, 1), stride=(1, 1, 1), padding=[0, 0, 0], bias=False)
)
)
)
)
(2): ResStage(
(res_blocks): ModuleList(
(0): ResBlock(
(activation): ReLU()
(branch2): BottleneckBlock(
(norm_c): BatchNorm3d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_c): Conv3d(128, 512, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
(act_b): ReLU()
(norm_b): BatchNorm3d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_b): Conv3d(128, 128, kernel_size=(1, 3, 3), stride=(1, 2, 2), padding=(0, 1, 1), bias=False)
(act_a): ReLU()
(norm_a): BatchNorm3d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_a): Conv3d(256, 128, kernel_size=(1, 1, 1), stride=(1, 1, 1), padding=[0, 0, 0], bias=False)
)
(branch1_norm): BatchNorm3d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(branch1_conv): Conv3d(256, 512, kernel_size=(1, 1, 1), stride=(1, 2, 2), bias=False)
)
(1): ResBlock(
(activation): ReLU()
(branch2): BottleneckBlock(
(norm_c): BatchNorm3d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_c): Conv3d(128, 512, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
(act_b): ReLU()
(norm_b): BatchNorm3d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_b): Conv3d(128, 128, kernel_size=(1, 3, 3), stride=(1, 1, 1), padding=(0, 1, 1), bias=False)
(act_a): ReLU()
(norm_a): BatchNorm3d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_a): Conv3d(512, 128, kernel_size=(1, 1, 1), stride=(1, 1, 1), padding=[0, 0, 0], bias=False)
)
)
(2): ResBlock(
(activation): ReLU()
(branch2): BottleneckBlock(
(norm_c): BatchNorm3d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_c): Conv3d(128, 512, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
(act_b): ReLU()
(norm_b): BatchNorm3d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_b): Conv3d(128, 128, kernel_size=(1, 3, 3), stride=(1, 1, 1), padding=(0, 1, 1), bias=False)
(act_a): ReLU()
(norm_a): BatchNorm3d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_a): Conv3d(512, 128, kernel_size=(1, 1, 1), stride=(1, 1, 1), padding=[0, 0, 0], bias=False)
)
)
(3): ResBlock(
(activation): ReLU()
(branch2): BottleneckBlock(
(norm_c): BatchNorm3d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_c): Conv3d(128, 512, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
(act_b): ReLU()
(norm_b): BatchNorm3d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_b): Conv3d(128, 128, kernel_size=(1, 3, 3), stride=(1, 1, 1), padding=(0, 1, 1), bias=False)
(act_a): ReLU()
(norm_a): BatchNorm3d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_a): Conv3d(512, 128, kernel_size=(1, 1, 1), stride=(1, 1, 1), padding=[0, 0, 0], bias=False)
)
)
)
)
(3): ResStage(
(res_blocks): ModuleList(
(0): ResBlock(
(activation): ReLU()
(branch2): BottleneckBlock(
(norm_c): BatchNorm3d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_c): Conv3d(256, 1024, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
(act_b): ReLU()
(norm_b): BatchNorm3d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_b): Conv3d(256, 256, kernel_size=(1, 3, 3), stride=(1, 2, 2), padding=(0, 1, 1), bias=False)
(act_a): ReLU()
(norm_a): BatchNorm3d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_a): Conv3d(512, 256, kernel_size=(3, 1, 1), stride=(1, 1, 1), padding=[1, 0, 0], bias=False)
)
(branch1_norm): BatchNorm3d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(branch1_conv): Conv3d(512, 1024, kernel_size=(1, 1, 1), stride=(1, 2, 2), bias=False)
)
(1): ResBlock(
(activation): ReLU()
(branch2): BottleneckBlock(
(norm_c): BatchNorm3d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_c): Conv3d(256, 1024, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
(act_b): ReLU()
(norm_b): BatchNorm3d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_b): Conv3d(256, 256, kernel_size=(1, 3, 3), stride=(1, 1, 1), padding=(0, 1, 1), bias=False)
(act_a): ReLU()
(norm_a): BatchNorm3d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_a): Conv3d(1024, 256, kernel_size=(3, 1, 1), stride=(1, 1, 1), padding=[1, 0, 0], bias=False)
)
)
(2): ResBlock(
(activation): ReLU()
(branch2): BottleneckBlock(
(norm_c): BatchNorm3d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_c): Conv3d(256, 1024, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
(act_b): ReLU()
(norm_b): BatchNorm3d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_b): Conv3d(256, 256, kernel_size=(1, 3, 3), stride=(1, 1, 1), padding=(0, 1, 1), bias=False)
(act_a): ReLU()
(norm_a): BatchNorm3d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_a): Conv3d(1024, 256, kernel_size=(3, 1, 1), stride=(1, 1, 1), padding=[1, 0, 0], bias=False)
)
)
(3): ResBlock(
(activation): ReLU()
(branch2): BottleneckBlock(
(norm_c): BatchNorm3d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_c): Conv3d(256, 1024, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
(act_b): ReLU()
(norm_b): BatchNorm3d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_b): Conv3d(256, 256, kernel_size=(1, 3, 3), stride=(1, 1, 1), padding=(0, 1, 1), bias=False)
(act_a): ReLU()
(norm_a): BatchNorm3d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_a): Conv3d(1024, 256, kernel_size=(3, 1, 1), stride=(1, 1, 1), padding=[1, 0, 0], bias=False)
)
)
(4): ResBlock(
(activation): ReLU()
(branch2): BottleneckBlock(
(norm_c): BatchNorm3d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_c): Conv3d(256, 1024, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
(act_b): ReLU()
(norm_b): BatchNorm3d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_b): Conv3d(256, 256, kernel_size=(1, 3, 3), stride=(1, 1, 1), padding=(0, 1, 1), bias=False)
(act_a): ReLU()
(norm_a): BatchNorm3d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_a): Conv3d(1024, 256, kernel_size=(3, 1, 1), stride=(1, 1, 1), padding=[1, 0, 0], bias=False)
)
)
(5): ResBlock(
(activation): ReLU()
(branch2): BottleneckBlock(
(norm_c): BatchNorm3d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_c): Conv3d(256, 1024, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
(act_b): ReLU()
(norm_b): BatchNorm3d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_b): Conv3d(256, 256, kernel_size=(1, 3, 3), stride=(1, 1, 1), padding=(0, 1, 1), bias=False)
(act_a): ReLU()
(norm_a): BatchNorm3d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_a): Conv3d(1024, 256, kernel_size=(3, 1, 1), stride=(1, 1, 1), padding=[1, 0, 0], bias=False)
)
)
)
)
(4): ResStage(
(res_blocks): ModuleList(
(0): ResBlock(
(activation): ReLU()
(branch2): BottleneckBlock(
(norm_c): BatchNorm3d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_c): Conv3d(512, 2048, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
(act_b): ReLU()
(norm_b): BatchNorm3d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_b): Conv3d(512, 512, kernel_size=(1, 3, 3), stride=(1, 2, 2), padding=(0, 1, 1), bias=False)
(act_a): ReLU()
(norm_a): BatchNorm3d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_a): Conv3d(1024, 512, kernel_size=(3, 1, 1), stride=(1, 1, 1), padding=[1, 0, 0], bias=False)
)
(branch1_norm): BatchNorm3d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(branch1_conv): Conv3d(1024, 2048, kernel_size=(1, 1, 1), stride=(1, 2, 2), bias=False)
)
(1): ResBlock(
(activation): ReLU()
(branch2): BottleneckBlock(
(norm_c): BatchNorm3d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_c): Conv3d(512, 2048, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
(act_b): ReLU()
(norm_b): BatchNorm3d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_b): Conv3d(512, 512, kernel_size=(1, 3, 3), stride=(1, 1, 1), padding=(0, 1, 1), bias=False)
(act_a): ReLU()
(norm_a): BatchNorm3d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_a): Conv3d(2048, 512, kernel_size=(3, 1, 1), stride=(1, 1, 1), padding=[1, 0, 0], bias=False)
)
)
(2): ResBlock(
(activation): ReLU()
(branch2): BottleneckBlock(
(norm_c): BatchNorm3d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_c): Conv3d(512, 2048, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
(act_b): ReLU()
(norm_b): BatchNorm3d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_b): Conv3d(512, 512, kernel_size=(1, 3, 3), stride=(1, 1, 1), padding=(0, 1, 1), bias=False)
(act_a): ReLU()
(norm_a): BatchNorm3d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv_a): Conv3d(2048, 512, kernel_size=(3, 1, 1), stride=(1, 1, 1), padding=[1, 0, 0], bias=False)
)
)
)
)
(5): ResNetBasicHead(
(output_pool): AdaptiveAvgPool3d(output_size=1)
(proj): Linear(in_features=2048, out_features=400, bias=True)
(dropout): Dropout(p=0.5, inplace=False)
(pool): AvgPool3d(kernel_size=(8, 7, 7), stride=(1, 1, 1), padding=(0, 0, 0))
)
)
)

As you can see the Convolutions have been reversed ie(conv_c, conv_b, conv_a) along with the batchnorm and while in google collab it is coming straight forward.
Can someone please help me out.
Thanks

2

The order of execution is defined in the forward method of each module while printing the model should print the modules in the order of their initialization (at least that’s what I’ve seen so far).
Are you seeing the forward pass “reversed” or just the printing of the submodules?

3

Thanks @ptrblck for your reply. Yes the forward is working correctly. But I don’t understand that in the google collab it is getting initialised the right way while in my gpu it is getting reversed. How can that be possible as I am using importing the model the same way.

4

I don’t know as I’m not familiar enough with the order of initialization of Python objects. I would assume they are created in the same order as used in the __init__ method and that’s what I’ve seen in all my setups so far. However, based on your description it seems as if that’s not the case for your setups so check the Python version etc. to narrow down if some internal init orders were changed.