Revised to work with Tensorflow 2.4

ur-whitelab / hoomd-tf

A plugin that allows the use of Tensorflow in Hoomd-Blue for GPU-accelerated ML+MD

https://hoomd-tf.readthedocs.io

MIT License

30 stars 8 forks source link

Revised to work with Tensorflow 2.4 #285

Closed whitead closed 3 years ago

whitead commented 3 years ago

I removed batch fraction (sample weight) - this was not strictly necessary but I found it to be cluttered and we weren't using it. Had to write a custom training loop - we need to check that metrics are working correctly.

whitead commented 3 years ago

@hgandhi2411 moviepy that was an accident! I didn't mean to have those changes. Will revise.

whitead commented 3 years ago

train_on_batch is still not working due to #261. Trying to decide if we should update example/docs to use work around, write our own wrapper or wait on resolving https://github.com/tensorflow/tensorflow/issues/47116

whitead commented 3 years ago

Just a note for myself: we can get around this by moving the box to be a weight or variable.

whitead commented 3 years ago

@mehradans92 and @hgandhi2411 this is ready for review. Had to do a bit of a hack and make the box be an N x 3 x 3 sparse tensor. Should be invisible to users though.