thauptmann / 3d-resnet-builder-for-keras Goto Github PK

At the moment only standard 3D convolutions are used. Different types of spatial-temporal convolutions should be implemented and a choice should be added.
E.g:

• (2+1)D
• FAST

Run demo.py with different configurations and add them to the table on the main page.

Hi!
Thank you for sharing your code, which has so many options.
However, I met some problems.

When I directly download the dataset and run the demo. I got the following errors.
ValueError: Dimensions must be equal, but are 8 and 16 for '{{node three_d_convolution_res_net/sequential_36/sequential_2/residual_conv_block_1/Add}} = Add[T=DT_HALF](three_d_convolution_res_net/sequential_36/sequential_2/residual_conv_block_1/sequential_13/sequential_12/batch_normalization_8/Cast_3, three_d_convolution_res_net/sequential_36/sequential_2/residual_conv_block_1/sequential_15/batch_normalization_10/Cast_3)' with input shapes: [?,?,8,8,128], [?,?,16,16,128].

I thought maybe it is the problem with the input size.
So I used my own code with your ResNet architecture part. My input is size is [16 (frame), 112, 112, 3] (3 channels not 1 in your demo). with batchszie = 8.
However, this time it returned the following error:

tensorflow.python.framework.errors_impl.InvalidArgumentError: required broadcastable shapes at loc(unknown) [Op:Add]

Then I get into the network part, I found in 'residual_conv_block_1', after calculating

intermediate_output = self.resnet_conv_block(inputs, training=training)
shortcut = self.shortcut_conv(inputs, training=training)

The shape of inputs is [8, 3, 27, 27, 64], the shape of intermediate_output is [8, 1, 7, 7, 128], the shape of shortcut is [8, 2, 14, 14, 128], so they can not be added output_sum = tf.add(intermediate_output, shortcut) here.

Do you have any ideas about these problems?
I am sorry for my long issue.
Thank you very much.

Hello, and thank you for your work on this module.

I'm interested in training a 3D ResNet to predict a 3D image from a 2-channel 3D image, where input shape is (240, 240, 240, 2) and output shape is (240, 240, 240, 1). I'll probably need to add some dilated convolution layers to the model to preserve those dimensions, but I got stuck before that point.

I tried this to start with:

model = three_d_resnet_builder.build_three_d_resnet_50(input_shape= (240, 240, 240, 2), 
                                                       output_shape= (240, 240, 240, 1), 
                                                       output_activation='sigmoid', 
                                                       regularizer='l2',
                                                       squeeze_and_excitation=False, 
                                                       kernel_name='3D')

but get the exception:
TypeError: int() argument must be a string, a bytes-like object or a number, not 'tuple'

I believe it is complaining about the output_shape not being an integer when it uses it to create a Dense layer in the resnet_tail section.

Is it possible to leave the output layers without a tail?