function get_dimensions is only available in the nightly version

[hukkai]

🐛 Describe the bug

In the reference codes for training ImageNet classification, function torchvision.transforms.functional.get_dimensions is used here:

vision/references/classification/transforms.py

Line 152 in 6274080

_, H, W = F.get_dimensions(batch)

However, as I update my torch vision version to the latest one 0.12.0, I cannot find this function. This function is available at version 0.13.0.dev20220502.

Besides, is it possible to provide training logs for the method (ResNet50 80.858 top1 accuracy on ImageNet) mentioned in this blog (https://pytorch.org/blog/how-to-train-state-of-the-art-models-using-torchvision-latest-primitives/)? Thanks!

Versions

Collecting environment information...
PyTorch version: 1.11.0
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A

OS: macOS 11.6.4 (x86_64)
GCC version: Could not collect
Clang version: 13.0.0 (clang-1300.0.29.30)
CMake version: Could not collect
Libc version: N/A

Python version: 3.8.1 (default, Jan 8 2020, 16:15:59) [Clang 4.0.1 (tags/RELEASE_401/final)] (64-bit runtime)
Python platform: macOS-10.16-x86_64-i386-64bit
Is CUDA available: False
CUDA runtime version: No CUDA
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

Versions of relevant libraries:
[pip3] numpy==1.22.3
[pip3] pytorchvideo==0.1.5
[pip3] torch==1.11.0
[pip3] torchvision==0.12.0
[conda] blas 1.0 mkl
[conda] mkl 2021.4.0 hecd8cb5_637
[conda] mkl-service 2.4.0 py38h9ed2024_0
[conda] mkl_fft 1.3.1 py38h4ab4a9b_0
[conda] mkl_random 1.2.2 py38hb2f4e1b_0
[conda] numpy 1.19.5 pypi_0 pypi
[conda] pytorchvideo 0.1.5 pypi_0 pypi
[conda] torch 1.11.0 pypi_0 pypi
[conda] torchvision 0.12.0 pypi_0 pypi

cc @datumbox

[pmeier]

The references on the main branch are in general only compatible with the library on the main branch. If you want to run the references with our stable binaries, you should also look at the reference scripts for the same version. For 0.12.0 the corresponding call looks like

vision/references/classification/transforms.py

Line 144 in 9b5a3fe

W, H = F.get_image_size(batch)

You can access it locally with git checkout v0.12.0.

[pmeier]

Besides, is it possible to provide training logs for the method (ResNet50 80.858 top1 accuracy on ImageNet) mentioned in this blog (https://pytorch.org/blog/how-to-train-state-of-the-art-models-using-torchvision-latest-primitives/)

cc @datumbox