feat: support aten.resize_ converter #2874
Conversation
github-actions
bot
added
component: tests
|
@apbose @zewenli98 The converter for However, while investigating the cause of the CI/CD failure, I found that tensors in uninitialized locations sometimes contain extremely large values, whereas the description always showed values very close to zero (in the range of
|
|
Use run_tests custom compare and check the original elements and the shape to verify correctness |
|
@chohk88 I was talking about the function |
chohk88
force-pushed
the
aten_resize_converter
branch
from
eaecfb2 to
8f6805a
Compare
| ): | ||
| comp_func = comparator[0] | ||
| args = comparator[1] | ||
| self.assertTrue(comp_func(output_trt, output_cpu, *args)) | ||
|
|
There was a problem hiding this comment.
Is there a specific case where the len(cuda_inputs) == 1 is required? In general I assume the len(cuda_inputs) would be 1 in most cases. And since it conditioned on res_trt could you highlight the cases where it would be required?
There was a problem hiding this comment.
Thank you for your comment!
As you mentioned, in most cases, the length of cuda_inputs is 1. Below are the outputs when I set a breakpoint in the original code for two cases where torch.ops.aten.resize_.default(x, target_shape) returns tensors with shapes (3,) and (10, 15, 10).
In both cases, the length of cuda_inputs is also 1. When cuda_inputs has a length of 1, res_trt and res_cpu are not lists of length 1 but are torch.Tensors with shapes (3,) and (10, 15, 10) for each case. Therefore, when we use zip in for output_trt, output_cpu, comparator in zip(res_trt, res_cpu, comparators), and the comparators list has a length of 1, it results in comparing fewer elements because one dimension is lost from res_trt and res_cpu.
chohk88
force-pushed
the
aten_resize_converter
branch
from
4c80772 to
6bcadc5
Compare
Description
Support converter for
aten.resize_operation: https://pytorch.org/docs/stable/generated/torch.Tensor.resize_.html#torch.Tensor.resize_One critical aspect of this operation is handling cases where the target size (output size) is larger than the input tensor size. When
the target sizeis larger thanthe input tensor size, the values of the additional elements are unpredictable. In PyTorch, these additional elements are not initialized, which can result in values that are close tozerobut are not guaranteed to bezero.In the converter developed for this PR, the additional elements are initialized to
zerousingnumpy.zeros. While this approach ensures a predictable output, it does not exactly replicate the behavior of PyTorch, where the values of the additional elements are not initialized and can be unpredictable.Fixes # (issue)
Type of change
Checklist: