Utility Layers

class ConcatLayer

Concatenates multiple blobs into a single blob along the specified dimension. Except in the concatenation dimension, the shapes of the blobs being concatenated have to be the same.

dim

Default 3 (channel). The dimension to concatenate.

bottoms

Names of the blobs to be concatenated.

tops

Name of the concatenated output blob.

class MemoryOutputLayer

Takes some blobs in the network and collect their data during forward pass of the network as a list of julia Array objects. Useful when doing in-memory testing for collecting the output. After running the forward pass of the network, the outputs field of the corresponding layer state object will contain a vector of the same size as the bottoms attribute. Each element of the vector is a list of tensors (julia Array objects), each tensor corresponds to the output in a mini-batch.

bottoms

A list of names of the blobs in the network to store.

class HDF5OutputLayer

Takes some blobs in the network and writes them to a HDF5 file. Note that the target HDF5 file will be overwritten when the network is first constructed, but later iterations will append data for each mini-batch. This is useful for storing the final predictions or the intermediate representations (feature extraction) of a network.

filename

The path to the target HDF5 file.

force_overwrite

Default false. When the layer tries to create the target HDF5 file, if this attribute is enabled, it will overwrite any existing file (with a warning printed). Otherwise, it will raise an exception and refuse to overwrite the existing file.

bottoms

A list of names of the blobs in the network to store.

datasets

Default []. Should either be empty or a list of Symbol of the same length as bottoms. Each blob will be stored as an HDF5 dataset in the target HDF5 file. If this attribute is given, the corresponding symbol in this list is used as the dataset name instead of the original blob’s name.

class IdentityLayer

An Identity layer maps inputs to outputs without changing anything. This can be useful as a glue layer to rename some blobs. There is no data-copying for this layer.

tops
bottoms

Blob names for output and input. This layer can take multiple input blobs and produce the corresponding number of output blobs. The shapes of the input blobs do not need to be the same.

class Index2OnehotLayer

A utility layer that could convert category class into one-hot encoded vector. For example, for K classes, input j is converted into a vector of size K, with all zeros, but the (j-1)-th entry 1.

dim

The dimension to operate on. The input must have size 1 on this dimension, i.e. size(input, dim) == 1. And the value should be integers from 0 to (K-1).

n_class

Number of categories, i.e. K as described above.

tops
bottoms

Blob names for output and input. This layer can take multiple input blobs and produce the corresponding number of output blobs. The shapes of the input blobs do not need to be the same. But they will be operated on the same dimension, and the n_class for them are the same.

class ReshapeLayer

Reshapes a blob. Can be useful if, for example, you want to make the flat output from an InnerProductLayer meaningful by assigning each dimension spatial information.

Internally, no data is copied. The total number of elements in the blob tensor after reshaping has to be the same as the original blob tensor.

shape

Has to be an NTuple of Int specifying the new shape. Note that the new shape does not include the last (mini-batch) dimension of a data blob. So a reshape layer cannot change the mini-batch size of a data blob.

tops
bottoms

Blob names for output and input. This layer can take multiple input blobs and produce the corresponding number of output blobs. The shapes of the input blobs do not need to be the same. But the feature dimensions (i.e. the product of the first 3 dimensions) have to be the same.

class SplitLayer

A Split layer produces identical copies of the input. The number of copies is determined by the length of the tops property. During back propagation, derivatives from all the output copies are added together and propagated down.

This layer is typically used as a helper to implement some more complicated layers.

bottoms

Input blob names, only one input blob is allowed.

tops

Output blob names, should be more than one output blobs.

no_copy

Default false. When true, no data is copied in the forward pass. In this case, all the output blobs share data. When, for example, an in-place layer is used to modify one of the output blobs, all the other output blobs will also change.