for classification traing

[AlexNmSED]

Hello, When I was doing feature extraction, I encountered the above problem, do you have a solution? Are you facing the same problem.

[AlexNmSED]

I use the Pytorch 1.8

[clemsgrs]

hey, could you tell me which python version you're using?

[AlexNmSED]

Yes ， Python 3.8.16 . Should I be using a higher base python?

[clemsgrs]

try upgrading to python 3.9

[AlexNmSED]

In order to reduce other conflicts. Can you share the conda environment configuration file? This will reduce the hassle. Thank you for your help.

XinYu Hao, Master Student JYU

---Original--- From: "clement @.> Date: Tue, May 23, 2023 11:59 AM To: @.>; Cc: @.**@.>; Subject: Re: [clemsgrs/hipt] for distributed training (Issue #11)

try upgrading to python 3.9

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

[clemsgrs]

I don't use conda, I use Docker instead. But you should be good to go in a conda environment with python 3.9.10 and torch 1.10.2. Then just pip install -r requirements.txt

[clemsgrs]

If you still run into some distributed-related issues, you can try upgrading to torch 1.12

[AlexNmSED]

Thank you.

I will do it.

XinYu Hao, Master Student The Alpha Lab (http://thealphalab.org) School of Software, Dalian University of Technology, Development Zone, Dalian 116620, China @.***

---Original--- From: "clement @.> Date: Tue, May 23, 2023 12:12 PM To: @.>; Cc: @.**@.>; Subject: Re: [clemsgrs/hipt] for distributed training (Issue #11)

If you still run into some distributed-related issues, you can try upgrading to torch 1.12

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

[AlexNmSED]

Thank you for your help. The feature extraction is going fine.

[AlexNmSED]

Hello , when I try agg_method: 'self_att', I encountered the following problems:

slide_pos_embed: use: False type: '1d' AssertionError: was expecting embedding dimension of 192, but got 1

slide_pos_embed: use: True type: '1d' AssertionError: query should be unbatched 2D or batched 3D tensor but received 4-D query tensor

slide_pos_embed: use: True type: '2d' TypeError: forward() missing 1 required positional argument: 'coords'

How should I use self-attention, can you give me some advice? Thank you for your help.

[clemsgrs]

hi, agg_method: 'self_att' is not fully supported yet my idea was to add a 4th level of self-attention on top of the slide-level Transformer to manage cases where the label is assigned to a group of slides (rather than a single slide), but i haven't had time to thoroughly test it.

I'd suggest using agg_method: 'concat', which will simply concatenate regions extracted from multiple slides into a single -- longer -- sequence.