Pytorch NN is not working?

sayred1 / tf.keras_GCN

Translation of AGCN implemented with tf to an implementation of tf.keras.

1 stars 0 forks source link

Pytorch NN is not working? #2

Closed qzhu2017 closed 5 years ago

qzhu2017 commented 5 years ago

@yanxon It seems that there is a strange behavior in pytorch NN.linear method. I built a 3-layer linear neuron model from SKlearn, the results are good (pytorch->SK-ANN.py). But when I tried to do the same thing for pytorch, I just failed to obtain the same results. (Torch_net.py)

I think Howard met a similar issue before. Can you help me to resolve this issue?

sayred1 commented 5 years ago

Still currently looking into it! With the skip connection and gg readout, I get better results, but the model never goes below an MAE of one. I'm putting together a training process similar to the AGCN repo where they train/optimize over batches of molecules as opposed to training/optimizing every specified epoch to see how the loss is fluctuating.

qzhu2017 commented 5 years ago

NN is fixed. The problem is due to incorrect assignment of output shape.