Deformable Shape Completion with Graph Convolutional Autoencoders-public

it is based on Deformable Shape Completion with Graph Convolutional Autoencoders paper's implementation.

this is incomplete yet.

Model Arch

model_exp2.py is actual model file.
modelwrapper2.py is current model file.

called by

main.py -> modelwrapper.py -> model_exp2.py (it include model and loss )

main2.py -> modelwrapper2.py -> feastnet.py

ISSUE

adam optimizer gradient variables is too big. need to divide at more GPU.
( it seem to allow custom training loop. not in keras Sequencial Model. )

Arch Detail

Input = (batch_size = ?, vertice_size = 5023, Feature_input = 3)
    class model
        - encoder
            * LinearLayer(output=[ ?, 5023, 16 ])
            * batchNorm [ ](output=[ ?, 5023, 16 ])
            * activation[relu | tanh | leakyrelu](output=[ ?, 5023, 16 ])
    
            * GCONV1(output=[ ?, 5023, 32 ])
            * batchNorm [ ](output=[ ?, 5023, 32 ])
            * activation[relu | tanh | leakyrelu](output=[ ?, 5023, 32 ])
    
            * GCONV2(output=[ ?, 5023, 64 ])
            * batchNorm [ ](output=[ ?, 5023, 64 ])
            * activation[relu | tanh | leakyrelu](output=[ ?, 5023, 64 ])
    
            * GCONV3(output=[ ?, 5023, 96 ])
            * batchNorm [ ](output=[ ?, 5023, 96 ])
            * activation[relu | tanh | leakyrelu](output=[ ?, 5023, 96 ]) 
    
            * GCONV3(output=[ ?, 5023, 128])
            * activation[relu | tanh | leakyrelu](output=[ ?, 5023, 128 ])
            * Pooling[](output=[?, 1, 128])   

            * LinearLayer(output=[ ?, 1 , 128 ]) -> For z_log_var
            * LinearLayer(output=[ ?, 1 , 128 ]) -> For z_log_var   
        
        - decoder
            * LinearLayer(output=[ ?, 5023*128 ]) 
            * batchNorm [ ](output=[ ?, 5023, 96 ])
            * activation[relu | tanh | leakyrelu](output=[ ?, 5023, 128 ])
            * Reshape(output=[ ? 5023, 128 ])
          
            * GCONV1(output=[ ?, 5023, 128])
            * batchNorm [ ](output=[ ?, 5023, 128 ])
            * activation[relu | tanh | leakyrelu](output=[ ?, 5023, 96 ])            
          
            * GCONV2(output=[ ?, 5023, 96 ])
            * batchNorm [ ](output=[ ?, 5023, 96 ])
            * activation[relu | tanh | leakyrelu](output=[ ?, 5023, 96 ])
          
            * GCONV3(output=[ ?, 5023, 64 ])
            * batchNorm [ ](output=[ ?, 5023, 64 ])
            * activation[relu | tanh | leakyrelu](output=[ ?, 5023, 64 ])                        
          
            * GCONV4(output=[ ?, 5023, 32 ])
            * batchNorm [ ](output=[ ?, 5023, 32 ])
            * activation[ relu | tanh | leakyrelu ](output=[ ?, 5023, 32 ]) 
          
            * GCONV5(output=[ ?, 5023, 16 ])
            * batchNorm [ ](output=[ ?, 5023, 16 ])
            * activation[ relu | tanh | leakyrelu ](output=[ ?, 5023, 16 ])
          
            * LinearLayer(output=[ ?, 5023, 3 ])

Configurations

numpy
tensorflow-gpu 2.0 or 2.1
libigl for python.
mesh lib(optional) : https://github.com/MPI-IS/mesh

just using requirements.txt

data preprocessing.

1. first download CoMA dataset files.(it's need login.)

2. extract files. and remove readme in CoMA dataset.

3. make one more folder(I actually named ./unrpocessed_dataset), and put renamed CoMa folder( rename plain ) on.

4. run fileconverter.py. before using fileconverter, and write your own data root path(for me. ./unprocessed_dataset) _target_input_dir variable in fileconverter.

run model

python main2.py [ train | test | summary ]