Siamese Network using Rstudio Keras

I'm trying to implement a siamese network using Rstudio Keras package. The network I'm trying to implement is the same network that you can see in this post.

So, basically, I'm porting the code to R and using Rstudio Keras implementation. So far my code looks like this:

    library(keras)

    inputShape <- c(105, 105, 1)
    leftInput <- layer_input(inputShape)
    rightInput <- layer_input(inputShape)

    model<- keras_model_sequential()

    model %>%
      layer_conv_2d(filter=64,
                    kernel_size=c(10,10),
                    activation = "relu",
                    input_shape=inputShape,
                    kernel_initializer = initializer_random_normal(0, 1e-2),
                    kernel_regularizer = regularizer_l2(2e-4)) %>%
      layer_max_pooling_2d() %>%

      layer_conv_2d(filter=128,
                    kernel_size=c(7,7),
                    activation = "relu",
                    kernel_initializer = initializer_random_normal(0, 1e-2),
                    kernel_regularizer = regularizer_l2(2e-4),
                    bias_initializer = initializer_random_normal(0.5, 1e-2)) %>%
      layer_max_pooling_2d() %>%

      layer_conv_2d(filter=128,
                    kernel_size=c(4,4),
                    activation = "relu",
                    kernel_initializer = initializer_random_normal(0, 1e-2),
                    kernel_regularizer = regularizer_l2(2e-4),
                    bias_initializer = initializer_random_normal(0.5, 1e-2)) %>%
      layer_max_pooling_2d() %>%

      layer_conv_2d(filter=256,
                    kernel_size=c(4,4),
                    activation = "relu",
                    kernel_initializer = initializer_random_normal(0, 1e-2),
                    kernel_regularizer = regularizer_l2(2e-4),
                    bias_initializer = initializer_random_normal(0.5, 1e-2)) %>%

      layer_flatten() %>%
      layer_dense(4096, 
                  activation = "sigmoid",
                  kernel_initializer = initializer_random_normal(0, 1e-2),
                  kernel_regularizer = regularizer_l2(1e-3),
                  bias_initializer = initializer_random_normal(0.5, 1e-2)) 

    encoded_left <- leftInput %>% model
    encoded_right <- rightInput %>% model

However, when running the last two lines, I get the following error:

Error in py_call_impl(callable, dots$args, dots$keywords) : 
  AttributeError: 'Model' object has no attribute '_losses'

Detailed traceback: 
  File "/home/rstudio/.virtualenvs/r-tensorflow/lib/python2.7/site-packages/tensorflow/contrib/keras/python/keras/engine/topology.py", line 432, in __call__
    output = super(Layer, self).__call__(inputs, **kwargs)
  File "/home/rstudio/.virtualenvs/r-tensorflow/lib/python2.7/site-packages/tensorflow/python/layers/base.py", line 441, in __call__
    outputs = self.call(inputs, *args, **kwargs)
  File "/home/rstudio/.virtualenvs/r-tensorflow/lib/python2.7/site-packages/tensorflow/contrib/keras/python/keras/models.py", line 560, in call
    return self.model.call(inputs, mask)
  File "/home/rstudio/.virtualenvs/r-tensorflow/lib/python2.7/site-packages/tensorflow/contrib/keras/python/keras/engine/topology.py", line 1743, in call
    output_tensors, _, _ = self.run_internal_graph(inputs, masks)
  File "/home/rstudio/.virtualenvs/r-tensorflow/lib/python2.7/site-packages/tensorflow/contrib/keras/python

I have been looking at similar implementations and questions all over StackOverflow, but I could not find a solution. I think I might be missing something really obvious.

Any ideas how to solve this?

Solution 1:

As pointed out by Daniel Falbel in his comment, the solution was updating R-keras package and then updating the tensorflow installation.

However, tensorflow package in R was not installing the latest 1.3 tensorflow version (it was reinstalling the 1.2 version).

To fix this problem, the URL to the correct version can be supplied to the install_tensorflow function. The URLs for the different implementations can be found here. I was using Linux in this case. Running this command should solve the problem for anyone who comes across the same issue:

library(tensorflow)
install_tensorflow(package_url = "https://pypi.python.org/packages/b8/d6/af3d52dd52150ec4a6ceb7788bfeb2f62ecb6aa2d1172211c4db39b349a2/tensorflow-1.3.0rc0-cp27-cp27mu-manylinux1_x86_64.whl#md5=1cf77a2360ae2e38dd3578618eacc03b")

Solution 2:

I tried GAN and also got this error. When I use the same code on the CPU version of tensorflow is was ok, but on the GPU version wasn't.

I found that this problem was caused by using the kernel_regularizer parameter on the GPU version. You can remove the parameter and try it again. I don't know why this solved the issue. I guess it may be a bug when processing reused models.