Sharing model weights for Activity Net

[arnavc1712]

Hi,
Would it be possible to share model weights for the activity net training?

[kylemin]

Hi again.
Please refer to the previous issue: link.
Thank you.

[arnavc1712]

Hi Kyle,
Thanks for the revert. Yes but that would require training the model again. Would it be possible to share model weights after training on activity net?

[kylemin]

It was already uploaded in the same Google drive link that was mentioned in the issue!
Thank you.

[arnavc1712]

Hi Kyle,
Thank you for the quick reply. I tried loading the weights into the model and also added the 'tfilter' layer as described in the paper. However the numbers are extremely low. Where could I be going wrong?
0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95
val result at 0: 0.01 || 0.87 | 0.06, 0.01, 0.01, 0.01, 0.01, 0.01, 0.01, 0.01, 0.01, 0.00

[kylemin]

I think your tfilter layer is not defined properly. How did you add the layer? Please provide me the corresponding line of your code.

[arnavc1712]

Hi Kyle,
I used this in the init() function - self.tfilter = nn.Conv1d(num_class, num_class, kernel_size=13, stride=1, padding=12, dilation=2, bias=False, groups=num_class)
And this in the forward() function tcam = F.relu(self.tfilter(tcam.permute([0,2,1]))).permute([0,2,1])

[kylemin]

Please try without relu activation, as we just applied tfilter without it.

[memoryjing]

Hi Kyle,
Thanks very much for sharing the code. Could you please help me with where to add the filter?

[kylemin]

Hi @memoryjing,
You can add it at the end of the forward function.

...
tcam = (cls_x_r+cls_x_ratself.omega) * self.mul_r + (cls_x_f+cls_x_fatself.omega) * self.mul_f
tcam = self.tfilter(tcam.permute(0,2,1)).permute(0,2,1)
return ...