Question about input to ASU

[liamkl]

Hi team,

I'm trying to understand your code. In this line below it looks like you are feeding a 4D tensor of shape [batch, feature, stocks, length] to a Conv1D layer:

DeepTrader/src/model/ASU.py

Line 159 in 4b00b2a

x = self.bn_start(self.start_conv(x))

Am I misunderstanding something? Otherwise, wouldn't this raise an error? as Conv1D is expecting a 3D tensor of shape [batch, num_channels, length]. I appreciate any insight you can give. Thanks!

[YorkLouis]

I have encountered a similar issue.

[choasLC]

not just that
I also confuse that why they claimed the proposed market score unit can not be used in CSI 100, since it still can provides a risk signal to the agent; even though, Chinese stocks do not allow "short"

[SteamerLee]

That would be reasonable and make sense if changing all torch.nn.Conv1d() to torch.nn.Conv2d() since Conv1d cannot be applied to 4D tensor here.

[fxvvv]

That would be reasonable and make sense if changing all torch.nn.Conv1d() to torch.nn.Conv2d() since Conv1d cannot be applied to 4D tensor here.

Can changing all torch.nn.Conv1d() to torch.nn.Conv2d() successfully produce results? Have you tried this approach? Thank you

[SteamerLee]

That would be reasonable and make sense if changing all torch.nn.Conv1d() to torch.nn.Conv2d() since Conv1d cannot be applied to 4D tensor here.

Can changing all torch.nn.Conv1d() to torch.nn.Conv2d() successfully produce results? Have you tried this approach? Thank you

Yes, I tried it, and it works.

[fxvvv]

That would be reasonable and make sense if changing all torch.nn.Conv1d() to torch.nn.Conv2d() since Conv1d cannot be applied to 4D tensor here.

Can changing all torch.nn.Conv1d() to torch.nn.Conv2d() successfully produce results? Have you tried this approach? Thank you

Yes, I tried it, and it works.

Oh, thank you, i tried this method and it works. But then i meet the issue " loss has nan ", have you ever encountered the same problem？

[SteamerLee]

That would be reasonable and make sense if changing all torch.nn.Conv1d() to torch.nn.Conv2d() since Conv1d cannot be applied to 4D tensor here.

Can changing all torch.nn.Conv1d() to torch.nn.Conv2d() successfully produce results? Have you tried this approach? Thank you

Yes, I tried it, and it works.

Oh, thank you, i tried this method and it works. But then i meet the issue " loss has nan ", have you ever encountered the same problem？

I got the normal losses before. But expect for changing the above Conv2d(), I may change other places for my tasks. So there may need other customizations when using in different tasks. Thank you!