GSOC’25 Project: Porting Chemception Model to PyTorch

#1

Hi everyone!
I’m Bibhusundar Mohapatra , a 3rd year undergraduate student in Biotechnology and Medical Engineering at NIT Rourkela, Odisha, India .

I’m thrilled to share that I will be working with DeepChem on the project “Porting Chemception Model to PyTorch” as part of Google Summer of Code 2025 (GSoC’25) .

Over the summer, I’ll be focusing on enhancing the interoperability and performance of the Chemception deep learning model by porting it from TensorFlow to PyTorch. I’ll be posting regular updates on my progress and would love to engage with the community throughout this journey.

Looking forward to learning, contributing, and having insightful discussions. Stay tuned for more updates!

Regards,
Bibhusundar Mohapatra

#2

Week 1 updates (june 2 - june 9)

  • Built intuition on Chemception architecture.
  • Studied the original TensorFlow implementation of ChemCeption and understood its core components.
  • Reviewed PyTorch best practices for model implementation and testing.
  • Implemented the Stem layer and got it merged into the codebase. Link to PR (https://github.com/deepchem/deepchem/pull/4286)
  • Implemented InceptionA and InceptionB layers and raised a pull request (PR).Link to PR(https://github.com/deepchem/deepchem/pull/4453)
  • Added unit tests for each implemented layer, including forward pass verification.

Link to slides

#3

Week 2 updates (Jun 9 - Jun 16)

  • Built more intuition on the ChemCeption architecture and its modular design.
  • Worked on implementing the InceptionResNetC and ReductionA layers.
  • Resolved a few issues raised in PR #4453.
  • Added corresponding unit tests for both layers, including forward pass checks.

Link to slides

#4

Week 3 updates (Jun 16 - Jun 22)

  • Built deeper intuition on ChemCeption’s Reduction layers
  • Implemented the ReductionA, ReductionB, and ReductionC layers locally with complete functionality.
  • Almost ready to raise the PR for these implementations after final review.
  • Resolved issues in PR #4453 and added unit tests for all layers, including forward pass checks.

Link to slides

#5

Week 3 updates (Jun 23 - Jun 29)

  • Built deeper intuition on ChemCeption’s Reduction layers, implemented ReductionA/B with full functionality locally, and am almost ready to raise the PR after final review.
  • Gone through the forum post by Vighnesh Somnath on his GSoC project with DeepChem, where he built a comprehensive transfer learning framework for molecular property prediction, implementing SmilesToImage/SmilesToSeq featurizers and ChemCeption/Smiles2Vec models based on the ChemNet protocol.
  • Gone through the ChemNet paper (https://arxiv.org/abs/1712.02734) which proposed a weakly supervised transfer learning protocol for chemistry using rule-based descriptors, and evaluated its model-agnostic performance across various datasets.

Link to slides

#6

Week 5 (June 30 - June 6)

Link to slides

#7

Week 6 (July 7 - july 13)

  • Got the InceptionC, ReductionA and ReductionB layer pull request (PR) merged into DeepChem’s codebase.Link to PR(https://github.com/deepchem/deepchem/pull/4471)
  • Explored the TorchModel class in depth and reviewed its implementation in other models.
  • Looking forward to building the base model class and the DeepChem wrapper class along with Tests.

Link to slides