Write a LayerΒΆ

The first thing to note: you never have to write a layer. Tensorpack layers are nothing but wrappers of symbolic functions. You can use any symbolic functions you have written or seen elsewhere with or without tensorpack layers.

If you would like, you can make a symbolic function become a "layer" by following some simple rules, and then gain benefits from the framework.

Take a look at the Convolutional Layer implementation for an example of how to define a layer:

def Conv2D(x, out_channel, kernel_shape,
           padding='SAME', stride=1,
           W_init=None, b_init=None,
           nl=tf.nn.relu, split=1, use_bias=True):

Basically, a tensorpack layer is just a symbolic function, but with the following rules:

  • It is decorated by @layer_register.

  • The first argument is its "input". It must be a tensor or a list of tensors.

  • It returns either a tensor or a list of tensors as its "output".

By making a symbolic function a "layer", the following things will happen:

  • You will need to call the function with a scope name as the first argument, e.g. Conv2D('conv0', x, 32, 3). Everything happening in this function will be under the variable scope conv0. You can register the layer with use_scope=False to disable this feature.

  • Static shapes of input/output will be printed to screen (if you register with log_shape=True).

  • argscope will work for all its arguments except the input tensor(s).

  • It will work with LinearWrap: you can use it if the output of one layer matches the input of the next layer.

There are also some (non-layer) symbolic functions in the tfutils.symbolic_functions module. There is not a rule about what kind of symbolic functions should be made a layer -- they are quite similar anyway. However, in general, I define the following symbolic functions as layers:

  • Functions which contain variables. A variable scope is almost always helpful for such functions.

  • Functions which are commonly referred to as "layers", such as pooling. This makes a model definition more straightforward.