-
Notifications
You must be signed in to change notification settings - Fork 7.2k
Closed
Description
Restoring the tolerances to
vision/test/prototype_transforms_kernel_infos.py
Lines 64 to 67 in f32600b
| DEFAULT_PIL_REFERENCE_CLOSENESS_KWARGS = { | |
| (("TestKernels", "test_against_reference"), torch.float32, "cpu"): dict(atol=1e-5, rtol=0, agg_method="mean"), | |
| (("TestKernels", "test_against_reference"), torch.uint8, "cpu"): dict(atol=1e-5, rtol=0, agg_method="mean"), | |
| } |
that will be cranked up by #6934 surfaces failures on the following kernels in the reference test:
-
adjust_brightness_image_tensor -
adjust_contrast_image_tensor -
adjust_gamma_image_tensor -
adjust_hue_image_tensor -
adjust_saturation_image_tensor -
adjust_sharpness_image_tensor -
affine_image_tensor -
affine_mask -
autocontrast_image_tensor -
center_crop_image_tensor -
crop_image_tensor -
elastic_image_tensor -
elastic_mask -
equalize_image_tensor -
five_crop_image_tensor -
horizontal_flip_image_tensor -
invert_image_tensor -
pad_image_tensor -
perspective_image_tensor -
perspective_mask -
posterize_image_tensor -
resize_image_tensor -
resize_mask -
resized_crop_image_tensor -
resized_crop_mask -
rotate_image_tensor -
rotate_mask -
solarize_image_tensor -
ten_crop_image_tensor -
vertical_flip_image_tensor
We need to go through each of them to see if there is a systematic problem here. However, this is not reason to "panic" since
- these kernels are used in the training pipeline and we are seeing no accuracy drop compared to v1
- these kernels are also tested for consistency against v1 and there everything is fine.