TwT

Some questions.

1

It's been a very interesting job and it has inspired me a lot, thank you for sharing this paper! After reading, I have some doubts. Can you help me answer...

Hello, I have a small question: why to test with a feature after the BN layer and not before it?

There are some questions related to the cluster center in the code

1

I found that when calculating the loss in the dynamic memory bank, the cluster center feature is not used, or the instance-level feature is used. May I be the one...

Why log loss[-sum(log f (yis; xsi ))] is used in multi classification problem instead of cross entropy？ For example, I have three categories. When the prediction y of the network...

I download the dataset and after unzipping the folder the format is: GTAV └ images └00001.png └00002.png ...... └ labels └00001.png ..... Noting your readme file, I chang it to...

The benchmark test encountered a problem

1

### Issue dirname: 2024-08-05-18-58-25--test2 test_cases: 27 model: ourModel edit_format: whole commit_hash: c45688c-dirty pass_rate_1: 25.9 pass_rate_2: 33.3 percent_cases_well_formed: 100.0 **error_outputs: 10** num_malformed_responses: 0 num_with_malformed_responses: 0 **user_asks: 9** lazy_comments: 0 syntax_errors: 0...

How to inference without docker?

2

### Describe the issue How to run swe-bench without docker? ### Suggest an improvement to documentation _No response_

How to draw the figures in hacl？

1

 We have recently been exploring methods to reduce the modality gap in multimodal large models, but have been struggling to determine the most effective way to validate the efficacy...

Want a docker image tar file

3

### Describe the issue I always encounter bugs when building Docker, specifically with this code (Line 121 in swebench/harness/docker_build.py): response = client.api.build( path=str(build_dir), tag=image_name, rm=True, forcerm=True, decode=True, platform=platform, nocache=nocache, )...