Deepchem Open Source Fellow (funded by Deep Forest Sciences)

#1

Hi everyone, I am Anshuman Mishra.I’ll be contributing for Deepchem under Deepchem Open Source Fellowship funded by Deep Forest Sciences.

About Me

I am in Prefinal year of my Bachelor of Technology, in Electronics & Communication Engineering at National Insitute of Technology, Warangal, India. I’ve always been interested in Machine Learning. On a side note, Chemistry has been my favorite subject since school. Thus, in Deepchem, I found perfect blend of these two. As a result, I have kept myself involved with Deepchem Community since February, 2022.

Project

According to this post on the Forum, “Growing support for Neural ODEs and CFD is on top priority list for Deepchem ‘’. The Convolutional ODE model to be contributed through this project in the original paper was used to model turbulent fluid flow by using Neural ODE. Thus it is an ideal choice to start with, for Neural ODE models for deepchem.

A Neural ODE which can learn the ‘equation’ or system dynamics to be more precise, by being fed data can be extremely powerful in not just medicine, but biology, chemistry, physics and engineering too. The conjoining of dynamical systems and deep learning has become a topic of great interest.
The project proposal can be found here .

Contact

Email : anshuman.mishra.h@gmail.com

Linkedin

2 Likes
#2

Week 0 Updates ( June, 12 - 18 )

PR : #2911

Have been actively working on the Encoder layer for the Conv OPE

Had 2 Code Reviews with @arunppsg and @ScienceStanley

Primarily worked with improving code quality and making the model more flexible (especially user being able to provide layer size, kernel size, stride size, dropout for each layer.

Discussed in Office hours to rather porting CNN tf implementation to PyTorch and use that as encoder in the ODE model would be more meaningful contribution to Deepchem.

Next Step

  1. Tweaks in current code to convert it into PyTorch ported model of tf cnn of Deepchem
  2. Resolving type issues, [CI failing]
1 Like
#3

Working on PR :
#2951 #2952 #2953

Have been working parallely on porting Deepchem’s Keras implementation of CNN to PyTorch
Started working on ODELayer, Dynamics Layers
Got valuable review from Bharath on focusing on CNN and get it merged in

Experimenting with CNN to reuse the layer by introducing an extra parameter ‘transpose’ used as a switch between ConvNd and ConvTransposeNd, for Encoder, Decoder as well as ODELayer : Colab,

Up Next :
CNN Layer tests : overfitting, save-reload etc,
Torchmodel wrapper for CNN

1 Like
#4

PR : #2963 #2964

This week I have been primarily working on porting CNN to PyTorch. Have been facing test failures in CI Testing most of time went there in fixing. Torch Model wrapper is still giving issues while calling model.fit()

On a side note, I figured out a bug in existing CNN implementation (tensorflow version) and opened a Pull Request fixing the same.

The priority now is to get the CNN merged, Next , I’m diving deep to fix errors in TorchModel wrapper and fix a couple of bugs in the docs.

1 Like
#5

PR : #2963 #2964 #2988

This week I majorly focused on the TorchModel wrapper and resolved the fit error that I’ve been facing will forward porpagating. The model runs fine, however the model gives poor score which will be item of action for the next week.

Fixed a bug #2964 and fixed a failing test in the CI this week #2988

Currently I am doing background research and planning on the design of Neural ODE API for deepchem, during which came across deepxde.

#6

Hi, all this week I worked on wrapping up the CNN PR to get merged.
The main things that I spent most of my time around was

  • Fixing the poor score of the torchmodel wrapper by using the LazyLinear layer. See this

  • Added some more unit tests for the model

1. test_cnn_regression_uncertainty()
2. test_residual_cnn_classification()
3. test_1d_cnn_reload()
#7

PR : #2963

This week I was mainly working on minor tasks like restoring the original unit tests for tf-cnn and enable model to be backwards compatible, resolving the recurring CI failures.
I observed that Model performance suddenly drops sometimes so I spent 2 days in fixing the same.

Next Step :

I’m waiting for the Code review, which will take time. I reached out Peter, Tony, Stanley and Abhishek to review the PR as Bharath said it will be good to have multiple people review it.

#8

PR : #2963 #3013

Hi Folks !

This week I mainly finished up the CNN PR’s work. The PR is ready to merge and is in the state of final review.
I’ve been working on the design document of Neural ODE API, wrote some pseudo code for the same. In last office hours, Bharath gave me this and this resource to cite on, as I haven’t made any design document before . I’ve been working towards the same.
As a side project for the fellowship, I took up GAN and WGAN model, to port from Tensorflow to PyTorch, the PR number for the same is #3013 .

I’ll be working with GAN and the nODE in the upcoming with.

#9

Hi,
This week I’ve been primarily working on the design document of Neural ODE API.
Nothing much in the WGAN part.

#10

Hi,
In the last two week, I worked on multiple versions of design document, getting them reviewed and working on feedback. The design is pretty much finalized now, all we need to brainstorm now is on the test and evaluation schema apart from some examples.

#11

Hi folks !
This week I mainly worked on finalizing the design document. I’m currently working on the implementation of the API.

1 Like
#12

Hi,
This week I’ve been mainly working with the implementation of Neural ODE API. I’m parallaly working on a tutorial as well to demonstrate the usage.

PR : #3050
Tutorial : Link

The PR is currently is in early stage of it’s development and there is huge scope of improvement, thus I’ll continue to work on the same for the next week.

#13

Hi Folks,
After getting the code reviewed last week, it got the feedback to make the code more deepchemic. I’m working on the same. Also what features to expose in the api and the interface is something that still requires brainstorming which I’ve been onto lately.