Metadata-Version: 2.1
Name: memorywrap
Version: 1.0.5
Summary: Memory Wrap: an extension for image classification models
Home-page: UNKNOWN
Author: La Rosa Biagio
Author-email: larosa@diag.uniroma1.it
License: UNKNOWN
Platform: UNKNOWN
Classifier: Programming Language :: Python :: 3
Classifier: License :: OSI Approved :: MIT License
Classifier: Operating System :: OS Independent
Description-Content-Type: text/markdown
Requires-Dist: entmax
Requires-Dist: torch (>1.5)

# 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
```python
memorywrap = MemoryWrapLayer(encoder_dim,output_dim,mlp_activation=torch.nn.ReLU())
```
or, for the baseline that uses only the memory to output the prediction:
```python
memorywrap = BaselineMemory(encoder_dim,output_dim,mlp_activation=torch.nn.ReLU())
```
where
- *encoder_dim* is the output dimension of the last layer of the encoder 
- *output_dim* is the desired output dimensione. In the case of the paper *output_dim* is equal to the **number of classes**;
- *mlp_activation* s the activation function that must be used in the hidden layer of multi-layer perceptron.

## Forward call
Add the forward call to your forward function.
```python
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. <br>
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:
- <a href="https://arxiv.org/abs/2106.01440">Paper</a>
- <a href="">GitHub repo</a>
- <a href="https://colab.research.google.com/drive/1OPjcpTH7X8EV1ev361iuhVzd2Jfp9kFA">Jupyter notebook</a>


