Description

Memory Wrap is an extension to image classification models that improves both data-efficiency and model interpretability, adopting a sparse content-attention mechanism between the input and some memories of past training samples.

Installation

This is a PyTorch implementation of Memory Wrap. To install Memory Wrap run the following command:

pip install memorywrap

The library contains two main classes:

• MemoryWrapLayer: it is the Memory Wrap variant described in the paper that uses both the input encoding and the memory encoding to compute the output;
• BaselineMemory: it is the baseline that uses only the memory encoding to compute the output.

Usage

Instantiate the layer

memorywrap = MemoryWrapLayer(encoder_output_dim, output_dim, classifier=None, distance='cosine')

or, for the baseline that uses only the memory to output the prediction:

memorywrap = BaselineMemory(encoder_output_dim, output_dim,  classifier=None, distance='cosine')

where

• encoder_output_dim (int) is the output dimension of the last layer of the encoder

• output_dim (int) is the desired output dimensione. In the case of the paper output_dim is equal to the number of classes;

• classifier (torch.nn.Module): Classifier on top of MemoryWrap. Inputs dimensions must be equal at encoder_output_dim*2 for MemoryWrapLayer and encoder_output_dim for BaselineMemory. By default is an MLP as described in the paper. An alternative is to use a linear layer. (e.g. torch.nn.Linear(encoder_output_dim*2, output_dim). Default: torch.nn.Sequential( torch.nn.Linear(encoder_output_dim*2, encoder_output_dim*4), torch.nn.ReLU(), torch.nn.Linear(encoder_output_dim*4, output_dim)

• distance (str): Distance to use to compute the similarity between input and memory set. Allowed values are: cosine, l2 and dot for respectively cosine similarity, l2 distance and dot product distance. Default=cosine

Forward call

Add the forward call to your forward function.

output_memorywrap = memorywrap(input_encoding, memory_encoding, return_weights=False)

where input_encoding and memory_encoding are the outputs of the the encoder of rispectively the current input and the memory set.
The last argument of the Memory Wrap's call function is a boolean flag controlling the number of outputs returned. If the flag is True, then the layer returns both the output and the sparse attention weight associated to each memory sample; if the flag is False, then the layer return only the output.

Additional information

Here you can find link to additional source of information about Memory Wrap:

• Paper
• GitHub repo
• Jupyter notebook