Feedback and Assistance Request for MARS_developer Tool

[www1987wl]

Description

Dear Developer,
Hello! I am a graduate student specializing in biomedical sensors. First, I would like to thank your team for the open-source MARS_developer tool, which has greatly assisted my research. However, I have encountered several issues during its usage and would appreciate your guidance or help.

Issues

1. Download Issue

   • Problem: When setting download_MARS_checkpoints=True, I am unable to automatically download the checkpoint files. This might be due to my location in China, necessitating manual download and decompression.
   • Attempted Fixes: Manual download and decompression.

2. Path Issue

   • Problem: After decompression, the path for the pre-trained model's checkpoint files uses \ instead of \\, causing Python parsing errors.
   • Attempted Fixes: Modified the relevant path settings in evaluate_detection.py.

3. Code Errors

   • Location: training_input.py at line 2435
   • Original Code: batched_background_heatmaps = tf.contrib.image.rotate(batched_heatmaps, angles)
   • Modified Code: batched_background_heatmaps = tf.contrib.image.rotate(batched_background_heatmaps, angles)

4. Exception Handling

   • Problem: The original code did not catch the NotImplementedError for certain Tensor types.
   • Original Code:

     if np.prod(shape) < 1000:
         return constant(value, shape=shape, dtype=dtype, name=name)
     except TypeError:

   • Modified Code:

     except (TypeError, NotImplementedError):
         # Happens when shape is a Tensor, list with Tensor elements, etc.
         pass
     return None

Additional Observations

• Functionality Issue: Errors occur when there are multiple best checkpoints in the plot_training_progress function in evaluate_pose.py.
• Data Representation Issue: Discrepancies in the graphical representation of the pre-trained model between the ground truth and the predicted results.

Questions

1. Is this tool still being maintained?
2. May I update this tool on my own to make it more user-friendly and to further my learning and practice?
3. Could I receive some advice and materials from you, such as whether it is necessary to refactor based on PyTorch, or are there more convenient alternatives? Also, could you provide descriptions of the overall design framework?

I am curious about these issues and would like to know more about the maintenance status of this tool. I look forward to your reply and thank you for taking the time to read this.

Best regards,
Wang Lu
[email protected]