Update code for pytorch 1.9.0

[wanchichen]

The latest version of pytorch removed some internal code that this repo imports, such as:

• container_abcs from torch._six
• int_classes from torch._six

Causing anyone running the current TimeSformer code on new versions of pytorch to receive compile-time errors.
The fix is similar to NVIDIA/apex#1049, checking the user's pytorch version before importing the necessary components.

[smivv]

Hello, sorry for bothering, but why this PR still wasn't merged?

[wanchichen]

The original author of the paper/repo (@gberta) no longer works at Facebook, so he doesn’t have permissions to merge it. So unless someone there with the necessary permissions takes notice it likely won’t be merged for the foreseeable future.

[smivv]

@wanchichen, thank you for your answer!

By the way, does your branch works fine? 😄
I consider your PR as the most recent update, so probably will use it.

[wanchichen]

Yeah the branch should work without issues.

[ShoufaChen]

Hi, @wanchichen
Thanks for your sharing.

One small fix:
you should add

import torch

at the top of timesformer/datasets/multigrid_helper.py.

[wanchichen]

Thank you @ShoufaChen , it has been fixed.

[MeSveta]

hi, I'm trying to set the environment but without any success. How I should update the pull request? manually? since it is not merged.

[wanchichen]

You can just clone from my fork: git clone https://github.com/wanchichen/TimeSformer.git

[MeSveta]

Hello William, Many thanks!! Can I please ask you another question? I'm more interested in working with the HowTo100 dataset , I'm trying to understand where to get the translation from a label that is a number to the actual class of the video ( such as 'preparing breakfast'). Any ideas? Thanks:)

…

On Thu, Jul 13, 2023 at 10:00 AM William Chen ***@***.***> wrote: You can just clone from my fork: git clone https://github.com/wanchichen/TimeSformer.git — Reply to this email directly, view it on GitHub <#47 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A6ORCENXNUFLJVMIURDPUG3XP6MJ7ANCNFSM47AGWPXQ> . You are receiving this because you commented.Message ID: ***@***.***>

[jromerooo2]

@MeSveta From my understanding you should create your own function that maps those features to an integer and/or viceversa. There should be a file inside of the dataset that contains the video path or video as an array and labels (sometimes the labels are in the actual name of the video, this depends on the structure of the dataset).

An example for the described above could be:
class_labels = sorted({str(path).split("/")[2] for path in all_video_file_paths})
label2id = {label: i for i, label in enumerate(class_labels)}
id2label = {i: label for label, i in label2id.items()}

This splits the path of the file in the current index i of the iteration and assigns it to a set. The set can't contain any duplicates so that is already covered for us.

Note that the code heavily relies on the assumption that your dataset is structured as the following:
'subset/train/Class/class_1.mp4'

So please modify if your dataset looks different.