Weight initialization in custom Conv3d seems different from nn.Conv3d

[qy-777]

Hi, I noticed that the outputs from my custom Conv3d and PyTorch’s nn.Conv3d were very different,
so I added some simple print() statements in both constructors (added at level1 question 60)
to check their weights.

Here’s what I got:

[RefModel] in=3, out=64, kernel=(3, 5, 7), stride=1, padding=0, bias=False
[RefModel Conv3d] weight shape: (64, 3, 3, 5, 7)
mean=-0.000240, std=0.032601, min=-0.056343, max=0.056341

[ModelNew] in=3, out=64, kernel=(3, 5, 7), stride=1, padding=(0, 0, 0), bias=False
[ModelNew Conv3d] weight shape: (64, 3, 3, 5, 7)
mean=0.009656, std=1.001391, min=-4.590487, max=4.220885

The shapes match, but the distributions are very different — nn.Conv3d has much smaller weights (std≈0.03),
while the custom one uses torch.randn() (std≈1.0).
This likely causes large output differences even if the kernels are correct.

[yuxuan-z19]

@qy-777 maybe try resetting the random seed before creating both RefModel and ModelNew to ensure they start from the same initialization state.

[qy-777]

@qy-777 maybe try resetting the random seed before creating both RefModel and ModelNew to ensure they start from the same initialization state.

This isn’t a situation that can be fixed just by changing the seed. You’ll understand once you try to reproduce the issue yourself.

[jason-yoo-108]

I also find this to be a problem with run_and_check. Specifically, for problems that use torch modules with parameter initializations such as nn.Linear and nn.Conv3D, there is no mechanism that ensures Model and ModelNew's corresponding parameters are set identically. Ideally, both models should receive parameters as their inputs. This seems like a serious problem.

Example Problem with Issue (only one of the matmul matrix to nn.Linear is explicitly specified): https://github.com/ScalingIntelligence/KernelBench/blob/main/KernelBench/level2/9_Matmul_Subtract_Multiply_ReLU.py

Example Problem without Issue (both matmul matrices are explicitly specified): https://github.com/ScalingIntelligence/KernelBench/blob/main/KernelBench/level1/1_Square_matrix_multiplication_.py