Segment_WSI migration from TF1 to TF2

[ishaq-ansari]

Summary

This PR addresses issue #12 by adding a TensorFlow 2 compatibility layer to all files associated with WSI segmentation, enabling inference to run in a TF2 environment while maintaining compatibility with existing TF1.15 legacy code.

Implementation Details

• Added TF2-compatible replacements for deprecated TF1 APIs used in WSI segmentation.
• Preserved variable naming and checkpoint loading behavior between TF1 and TF2.

nas_network.py

• Introduced a lightweight configuration class to:
  • Store and access hparams in a TF2-compatible way
  • Support dynamic updates to configuration values when required
• Replaces deprecated TF1 hparams APIs while preserving existing behavior.

conv2d_ws.py

Implemented TF2 equivalents for deprecated internal TF1 layer utilities:

• _variable_getter_for_compat
  • Replicates layers._build_variable_getter
• _add_variable_to_collections
  • Replicates layers._add_variable_to_collections
  • Preserves expected variable collections behavior
    nas_network.py and conv2d_ws.py implementations are based on the TF1.15 internal source code:
• TF1 hparams:
  https://raw.githubusercontent.com/tensorflow/tensorflow/v1.15.0/tensorflow/contrib/training/python/training/hparam.py
• TF1 layers utilities:
  https://raw.githubusercontent.com/tensorflow/tensorflow/v1.15.0/tensorflow/contrib/layers/python/layers/layers.py

How to Run / Test

1. Clone the repository

2. Map TF1 checkpoint to TF2 format

   • Update the TF1 checkpoint path in map_checkpoint_to_deeplabv3plus.py (line 318), for example:

     /home/iansari/model_test/model.ckpt-5000
     

   • Run:

     python map_checkpoint_to_deeplabv3plus.py

   • This generates TF2-compatible checkpoints in the same directory with prefix:

     model_mapped_tf2.ckpt
     

3. Run segmentation

   • Update paths in run_glom_detection_gpu.sh (lines 21–31)
   • Set checkpoint path using the mapped TF2 checkpoint prefix:

     /home/iansari/model/model_mapped_tf2.ckpt
     

   • Submit the job:

     sbatch run_glom_detection_gpu.sh

4. Output

   • Segmentation results are saved as .json files in the root-level output/ directory.

Notes

• This PR focuses on inference compatibility and does not introduce training changes.
• Output behavior is consistent with TF1 inference when using the same checkpoint.