Now, I understood that I need to install some virtual environment to work with USPTO database, for Google Colab. There are no hints about this in "https://deepchem.readthedocs.io/"
Does anyone here have any tips?
Does anyone here have any tips?
Hey, that does not seem to be necessary. Here is a colab notebook https://colab.research.google.com/drive/14DEzysC0rLYysC-9bAh4l-W-zCwg6tcw?usp=sharing where I demonstrate how to use it starting from scratch.
If you still have issues with using it, you could share your notebook with me.
now it works!
I'm waiting for someone to let me in. Time - Thursday 7.30 PM PST. Friday 9.00 AM IST
OH is at 7:00 pm PST, the developer call is at 7:30 pm PST
I think you may have gotten the calls mixed up, sorry!
There will be OH tomorrow at 9:00am PST / 10:30pm IST
Can you make it then?
Ok. I'll mail you to add me to OH hours meeting. Otherwise I'll still be confused.
Are there anyone who encountered the same?
I figured it out by following this - https://scikit-learn.org/stable/developers/advanced_installation.html#building-from-source
Hi all, I'm new to deepchem so I have a noob question. I wanted to train a deep learning model on the QM9 dataset so I started with a DTNN model with the same hyperparamters as in the qm9_DTNN.py example. This got me very poor pearson r2 scores, but I saw in issue #2475 on github that this might be a bug related to that type of model. However, after trying to train a GraphConv model on it this only gave me slightly better results (~0.33 pearson_r2 after 100 epochs). I used the default hyperparameters for the GraphConv model. Does anyone know if this can be resolved by hyperparameter tuning and if so, can you give me some advice on how to do this?
Since I want to create a new dataset that is outside of deepchem :)
Hi, is there a way to have Keras model that takes in 2 inputs and runs it through 2 parallel set of layers that have graph conv and graph pool layers in the set of layers with the outputs of the 2 sets being combined with layers.Add layer then a dense layer after that then output. All the examples of custom graph convolution models seem to be made for only one input set
1 reply
@rbharath I am trying to do that but every example of using the graph layers with deepchem for a custom model used one input variable and I tried making the call function input to take in multiple inputs like in the way keras.model would be set up to take multiple inputs however function didn't accept that. Then tried to build the keras.model by calling the inputs and output layers using the keras.model and set 2 inputs and an output for the model however it seems that the output of the graphconv layer isn't accepted by either the ker.batch normilisation layer (unknown shape error) or dc.graph pool layer (didn't accept data type from the graph conv I think)
@rbharath Are deepchem featurization methods parallelizable? I found this issue that was closed in 2016 deepchem/deepchem#267. Does the user still bear responsibility to parallelize featurization using multiple dataloaders? Or is there an argument/feature for user to specify if featurization should be parallelized?