search

MrGiovanni / UNetPlusPlus

[IEEE TMI] Official Implementation for UNet++
Other
2.26k stars 538 forks source link

when training model resnet34+xnet configuration, the default decoder_filter parameters seems not work #38

Open brb-chen opened 4 years ago

brb-chen commented 4 years ago

main train script configuration to: model = Xnet(backbone_name=config.backbone, input_shape=(config.input_deps, config.input_rows, config.input_cols), n_upsample_blocks=4, decoder_filters=(64,64,128,256,512), encoder_weights=config.weights, decoder_block_type=config.decoder_block_type, classes=config.nb_class, activation=config.activation)

and builder.py in xnet model to: ` if downterm[i+1] is not None:

interm[(n_upsample_blocks+1)*i+j+1] = up_block(decoder_filters[n_upsample_blocks-i-2],

                interm[(n_upsample_blocks+1)*i+j+1] = up_block(decoder_filters[i],
                                  i+1, j+1, upsample_rate=upsample_rate,
                                  skip=interm[(n_upsample_blocks+1)*i+j],
                                  use_batchnorm=use_batchnorm)(downterm[i+1])
            else:
                interm[(n_upsample_blocks+1)*i+j+1] = None
            # print("\n{} = {} + {}\n".format(interm[(n_upsample_blocks+1)*i+j+1],
            #                             interm[(n_upsample_blocks+1)*i+j], 
            #                             downterm[i+1]))
        else:
            #interm[(n_upsample_blocks+1)*i+j+1] = up_block(decoder_filters[n_upsample_blocks-i-2], 
            interm[(n_upsample_blocks+1)*i+j+1] = up_block(decoder_filters[i],
                              i+1, j+1, upsample_rate=upsample_rate,
                              skip=interm[(n_upsample_blocks+1)*i : (n_upsample_blocks+1)*i+j+1],
                              use_batchnorm=use_batchnorm)(interm[(n_upsample_blocks+1)*(i+1)+j])
            # print("\n{} = {} + {}\n".format(interm[(n_upsample_blocks+1)*i+j+1],
            #                             interm[(n_upsample_blocks+1)*i : (n_upsample_blocks+1)*i+j+1], 
            #                             interm[(n_upsample_blocks+1)*(i+1)+j]))

`

when i adapt resnet34 as backbone when try to training xnet model with my own data(5125123), after detailed checked the downsampling layer and skip-connection layers, and the up-block(transpose currently), it's seems the default decoder filter parameters didn't work, as the concatenate operation the up-block require the input as same dimension, so after checked the details network and skip-connection configuration, i changed the decoder filters, did anyone meet the same situation? just to confirm that, thanks!