Bharath Ramsundar comments

Results 191 comments of


                                            Bharath Ramsundar

Introduce yourself here first!

Thanks for starting this effort! **Your name:** Bharath Ramsundar **Your job other than this review:** Entrepreneur. Lead developer of the DeepChem project. **What you hope to get out of participating:**...

Support for COLL Dataset

Sounds like a useful dataset!

Difference in prediction results between just trained model and reloaded model

In general models autosave when `fit()` is called (the `save()` method is a vestigial remnant of older implementations that we should clean up). Try `restore()` instead of `reload()` as well

nan value from weave on HIV dataset

I think the issue is that Weave is unstable for some hyperparameters and sometimes produces NaNs. I'd recommend playing with the learning rate to see if you can stabilize trainiing....

tighten test_lcnn_regression bound

I believe that this looks reasonable as a change. @austereantelope is this ready to merge in?

torch model.restore() error on CUDA mismatch

Thank you for the bug report! I'll mark this as a bug and open for contribution

Roadmap to port Keras layers to Torch

Yes that would be a great idea. The way we've done this before is ensuring the existing TF tests pass out of box for the PyTorch reimplementation

Roadmap to port Keras layers to Torch

Also #2863 is related to this list. Converting the remaining TF models to PyTorch may necessitate adding some new layers in as well

Error in prediction for multitask datasets

Here's a workaround if you're running into this issue: Make dummy labels and add them to your dataset for prediction. Here's a snippet that fixes the code sample above: ```...

Some of Dataset object methods does not work with ConvMol featurizer

We are generally migrating away from TensorFlow to PyTorch. I anticipate that over the next few releases a lot of this old TF code will be migrated forward. We are...