- 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.
Default 3 (channel). The dimension to concatenate.
Names of the blobs to be concatenated.
Name of the concatenated output blob.
- 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.
The path to the target HDF5 file.
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.
A list of names of the blobs in the network to store.
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.
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 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.
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.
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.
Input blob names, only one input blob is allowed.
Output blob names, should be more than one output blobs.
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.