keras-io

/

drug-molecule-generation-with-VAE

like 7

Thanks for your great work @camocazi

Currently what's included in the repo is only the decoder part of the network, which is not very useful on its own.

from huggingface_hub import from_pretrained_keras
model = from_pretrained_keras("keras-io/drug-molecule-generation-with-VAE")
model.summary()

Model: "decoder"
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 input_6 (InputLayer)           [(None, 435)]        0           []                               
                                                                                                  
 dense_8 (Dense)                (None, 128)          55808       ['input_6[0][0]']                
                                                                                                  
 dropout_5 (Dropout)            (None, 128)          0           ['dense_8[0][0]']                
                                                                                                  
 dense_9 (Dense)                (None, 256)          33024       ['dropout_5[0][0]']              
                                                                                                  
 dropout_6 (Dropout)            (None, 256)          0           ['dense_9[0][0]']                
                                                                                                  
 dense_10 (Dense)               (None, 512)          131584      ['dropout_6[0][0]']              
                                                                                                  
 dropout_7 (Dropout)            (None, 512)          0           ['dense_10[0][0]']               
                                                                                                  
 dense_11 (Dense)               (None, 72000)        36936000    ['dropout_7[0][0]']              
                                                                                                  
 reshape_2 (Reshape)            (None, 5, 120, 120)  0           ['dense_11[0][0]']               
                                                                                                  
 tf.compat.v1.transpose_1 (TFOp  (None, 5, 120, 120)  0          ['reshape_2[0][0]']              
 Lambda)                                                                                          
                                                                                                  
 tf.__operators__.add_1 (TFOpLa  (None, 5, 120, 120)  0          ['reshape_2[0][0]',              
 mbda)                                                            'tf.compat.v1.transpose_1[0][0]'
                                                                 ]                                
                                                                                                  
 dense_12 (Dense)               (None, 1320)         677160      ['dropout_7[0][0]']              
                                                                                                  
 tf.math.truediv_1 (TFOpLambda)  (None, 5, 120, 120)  0          ['tf.__operators__.add_1[0][0]'] 
                                                                                                  
 reshape_3 (Reshape)            (None, 120, 11)      0           ['dense_12[0][0]']               
                                                                                                  
 softmax_2 (Softmax)            (None, 5, 120, 120)  0           ['tf.math.truediv_1[0][0]']      
                                                                                                  
 softmax_3 (Softmax)            (None, 120, 11)      0           ['reshape_3[0][0]']              
                                                                                                  
==================================================================================================
Total params: 37,833,576
Trainable params: 37,833,576
Non-trainable params: 0

This PR includes the encoder, decoder, and sampling layer in the model object (and includes their weights after training with the published code).

model.summary()

Model: "molecule_generator_1"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 encoder (Functional)        [(None, 435),             451925    
                              (None, 435)]                       
                                                                 
 decoder (Functional)        [(None, 5, 120, 120),     37833576  
                              (None, 120, 11)]                   
                                                                 
 dense_13 (Dense)            multiple                  436       
                                                                 
=================================================================
Total params: 38,285,941
Trainable params: 38,285,937
Non-trainable params: 4

Also updates the diagrams of the encoder and decoder graphs.

Super! Thank you for clearly describing