Google Summer of Code 2023: Porting SeqToSeq and DTNN Model

#1

About Me

I’m Rakshit Kumar Singh currently a first year Student at NIT Rourkela, India.

I am excited to be a part of the DeepChem community. I have been passionate about machine learning for scientific applications, and I am eager to contribute to the development of this project. I believe that machine learning has the potential to make a real impact on the world.

Contact Details

About the Project

PyTorch is becoming increasingly popular as the preferred backend for developing algorithms, thanks to its sustainability. DeepChem has decided to port its TensorFlow and Jax Models to PyTorch. This project aims to successfully port two of DeepChem’s models. SeqToSeq and DTNN Model are being ported under this project. SeqToSeq is a neural network that maps variable-length input sequences to variable-length output sequences. DTNN stands for Deep Tensor Neural Network, which is a deep learning architecture that uses tensors to represent molecular structures for predicting chemical properties and activities. The implementation will include documentation changes, usage examples, and a tutorial if required.

#2

Progress of week 1 (29 May - 4 June):

#3

Progress of Week 2 (5 June - 12 June):

#4

Progress of Week 3 (13 June - 19 June):

  • Created prototype of DTNNStep Layer.
  • DTNNEmbedding Layer Merged.

Related PR:

#5

Progress of Week 4 (20 June - 26 June):

  • Completed DTNNStep Layer.
  • Improved Documentation.
  • Wrote tests for DTNNStep Layer.
  • Created prototype of DTNNGather Layer.

Related PR:

#6

Progress of Week 5 (27 June - 3 July):

  • Completed DTNNGather Layer.
  • Merged DTNNStep Layer.
  • Worked on DTNNModel Class.

Related PR:

#7

Progress of Week 6 (4 July - 10 July):

  • Completed Coding part of DTNNModel.
  • Fixed Parameter Error in DTNNEmbedding and DTNNStep. (Merged)

Related PR:

#8

Progress of Week 7 (11 July - 17 July):

  • Merged DTNNGather Layer.
  • Completed DTNNModel Layer.
  • Worked on VariationalRandomizer Layer.
  • Worked on SeqToSeq Model Prototype.

Related PR:

#9

Progress of Week 8 (18 July - 24 July):

  • Made the ProtoType of SeqToSeq Model.
  • Made Encoder of SeqToSeq Model.
  • Made Decoder of SeqToSeq Model.

Related PR:

#10

Progress of Week 9 (25 July - 31 July):

  • Improved documentation of DTNN Model.
  • Did Literature Review of SeqToSeq.
#11

Progress of Week 10 (1 August - 7 August):

  • Looked into many different implementations of SeqToSeq Model.
  • Separated the compute_features_on_batch into a unique file for all batch utilities.
#12

Progress of Week 11 (8 August - 14 August):

  • Did more study on SeqToSeq Model.
  • Improved documentation of batch_coulomb_matrix_features.
#13

Progress of Week 12 (15 August - 21 August):

  • Made the EncoderRNN PR.
  • Made the DecoderRNN PR.

Related PRs:

GSoC ‘23 | Porting DTNN and SeqToSeq Model to PyTorch | Report
#14

Progress of Week 13 (22 August - 28 August):

  • EncoderRNN PR Merged.
  • DTNNModel PR Merged.
  • Worked on DecoderRNN.

Related PRs:

#15

Progress of Week 13 (29 August - 4 September):

  • Completed DecoderRNN.
  • Completed Variational Randomizer.

Related PRs: