implement groups for conv1d and conv2d
#637
Conversation
jimexist
changed the title
groups for conv1d and conv2d
|
Thanks for putting this together! I'm just dropping a comment here to give you a heads up so that you can be on the look out for any updates and PRs to conv in the meantime |
| @@ -34,6 +37,8 @@ def __init__( | |||
| kernel_size: int, | |||
| stride: int = 1, | |||
| padding: int = 0, | |||
| dilation: int = 1, | |||
| groups: int = 1, | |||
| bias: bool = True, | |||
| ): | |||
| super().__init__() | |||
There was a problem hiding this comment.
Why the shape of the self.weight doesn't change?
I think the shape should be (out_channel, kernel_size, in_channels/groups) and (in_channel%groups ==0).
| lambda x: (x, x) if isinstance(x, int) else x, | ||
| (kernel_size, stride, padding), | ||
| (kernel_size, stride, padding, dilation), | ||
| ) | ||
| scale = math.sqrt(1 / (in_channels * kernel_size[0] * kernel_size[1])) | ||
| self.weight = mx.random.uniform( |
| @@ -2767,6 +2768,22 @@ inline void run_conv_checks(const array& in, const array& wt, int n_dim) { | |||
| << " input: " << in.shape() << " and weight: " << wt.shape(); | |||
There was a problem hiding this comment.
If the group>1, then the last dimension of weight * group == the last dimension of input shape
|
After I fixed the bugs, I had a try on depthwise conv1d, very slow. |
|
@jimexist any updates on this? |
|
What's the status of this PR? I think we could move it forward as no-one else is working on groups right now? Wdyt? Otherwise let's close it so someone else can pick this up. |
in that case let me see if i can move on and implement this, also rebasing after: |
|
Is that PR still active? I would like to work on it. |
@Stealeristaken feel free to take over since i am bit busy and engaged in March |
Proposed changes
Checklist
Put an
xin the boxes that apply.pre-commit run --all-filesto format my code / installed pre-commit prior to committing changes