pytorch-dnc/README.md

491 lines
16 KiB
Markdown
Raw Normal View History

2017-12-18 15:08:45 +08:00
# Differentiable Neural Computers and family, for Pytorch
Includes:
1. Differentiable Neural Computers (DNC)
2. Sparse Access Memory (SAM)
3. Sparse Differentiable Neural Computers (SDNC)
2017-10-26 23:29:05 +08:00
2017-12-11 03:34:26 +08:00
<!-- START doctoc generated TOC please keep comment here to allow auto update -->
<!-- DON'T EDIT THIS SECTION, INSTEAD RE-RUN doctoc TO UPDATE -->
2017-12-18 15:10:13 +08:00
- [Install](#install)
- [From source](#from-source)
- [Architecure](#architecure)
- [Usage](#usage)
- [DNC](#dnc)
2017-12-18 15:11:32 +08:00
- [Example usage](#example-usage)
- [Debugging](#debugging)
2017-12-18 15:10:13 +08:00
- [SDNC](#sdnc)
2017-12-18 15:11:32 +08:00
- [Example usage](#example-usage-1)
- [Debugging](#debugging-1)
2017-12-18 15:10:13 +08:00
- [SAM](#sam)
2017-12-18 15:11:32 +08:00
- [Example usage](#example-usage-2)
- [Debugging](#debugging-2)
2017-12-19 04:08:03 +08:00
- [Tasks](#tasks)
2017-12-19 16:49:48 +08:00
- [Copy task (with curriculum and generalization)](#copy-task-with-curriculum-and-generalization)
2017-12-19 04:08:03 +08:00
- [Generalizing Addition task](#generalizing-addition-task)
2017-12-19 16:49:48 +08:00
- [Generalizing Argmax task](#generalizing-argmax-task)
2017-12-18 16:08:04 +08:00
- [Code Structure](#code-structure)
2017-12-18 15:10:13 +08:00
- [General noteworthy stuff](#general-noteworthy-stuff)
2017-12-11 03:34:26 +08:00
<!-- END doctoc generated TOC please keep comment here to allow auto update -->
2017-10-30 01:23:28 +08:00
2017-10-30 01:23:28 +08:00
[![Build Status](https://travis-ci.org/ixaxaar/pytorch-dnc.svg?branch=master)](https://travis-ci.org/ixaxaar/pytorch-dnc) [![PyPI version](https://badge.fury.io/py/dnc.svg)](https://badge.fury.io/py/dnc)
2017-10-27 22:43:01 +08:00
This is an implementation of [Differentiable Neural Computers](http://people.idsia.ch/~rupesh/rnnsymposium2016/slides/graves.pdf), described in the paper [Hybrid computing using a neural network with dynamic external memory, Graves et al.](https://www.nature.com/articles/nature20101)
2017-12-18 15:08:45 +08:00
and Sparse DNCs (SDNCs) and Sparse Access Memory (SAM) described in [Scaling Memory-Augmented Neural Networks with Sparse Reads and Writes](http://papers.nips.cc/paper/6298-scaling-memory-augmented-neural-networks-with-sparse-reads-and-writes.pdf).
2017-10-26 23:29:05 +08:00
## Install
```bash
pip install dnc
```
### From source
2017-11-27 14:32:41 +08:00
```
git clone https://github.com/ixaxaar/pytorch-dnc
cd pytorch-dnc
pip install -r ./requirements.txt
pip install -e .
```
2017-11-29 18:11:50 +08:00
For using fully GPU based SDNCs or SAMs, install FAISS:
```bash
conda install faiss-gpu -c pytorch
```
`pytest` is required to run the test
2017-11-29 18:11:50 +08:00
2017-11-01 15:26:24 +08:00
## Architecure
<img src="./docs/dnc.png" height="600" />
2017-10-26 23:29:05 +08:00
## Usage
2017-12-11 03:13:52 +08:00
### DNC
**Constructor Parameters**:
2017-10-26 23:29:05 +08:00
Following are the constructor parameters:
Following are the constructor parameters:
| Argument | Default | Description |
| --- | --- | --- |
| input_size | `None` | Size of the input vectors |
| hidden_size | `None` | Size of hidden units |
| rnn_type | `'lstm'` | Type of recurrent cells used in the controller |
| num_layers | `1` | Number of layers of recurrent units in the controller |
| num_hidden_layers | `2` | Number of hidden layers per layer of the controller |
| bias | `True` | Bias |
| batch_first | `True` | Whether data is fed batch first |
| dropout | `0` | Dropout between layers in the controller |
2017-11-13 03:04:59 +08:00
| bidirectional | `False` | If the controller is bidirectional (Not yet implemented |
| nr_cells | `5` | Number of memory cells |
| read_heads | `2` | Number of read heads |
| cell_size | `10` | Size of each memory cell |
| nonlinearity | `'tanh'` | If using 'rnn' as `rnn_type`, non-linearity of the RNNs |
| gpu_id | `-1` | ID of the GPU, -1 for CPU |
| independent_linears | `False` | Whether to use independent linear units to derive interface vector |
| share_memory | `True` | Whether to share memory between controller layers |
Following are the forward pass parameters:
2017-10-26 23:29:05 +08:00
| Argument | Default | Description |
| --- | --- | --- |
2017-11-11 00:02:14 +08:00
| input | - | The input vector `(B*T*X)` or `(T*B*X)` |
| hidden | `(None,None,None)` | Hidden states `(controller hidden, memory hidden, read vectors)` |
| reset_experience | `False` | Whether to reset memory |
| pass_through_memory | `True` | Whether to pass through memory |
2017-10-26 23:29:05 +08:00
2017-12-18 15:11:32 +08:00
#### Example usage
2017-10-26 23:29:05 +08:00
```python
from dnc import DNC
rnn = DNC(
input_size=64,
hidden_size=128,
rnn_type='lstm',
num_layers=4,
nr_cells=100,
cell_size=32,
read_heads=4,
batch_first=True,
gpu_id=0
)
(controller_hidden, memory, read_vectors) = (None, None, None)
output, (controller_hidden, memory, read_vectors) = \
2018-06-29 23:34:10 +08:00
rnn(torch.randn(10, 4, 64), (controller_hidden, memory, read_vectors), reset_experience=True)
2017-10-26 23:29:05 +08:00
```
2017-12-11 03:13:52 +08:00
2017-12-18 15:11:32 +08:00
#### Debugging
2017-11-13 03:04:59 +08:00
The `debug` option causes the network to return its memory hidden vectors (numpy `ndarray`s) for the first batch each forward step.
These vectors can be analyzed or visualized, using visdom for example.
```python
from dnc import DNC
rnn = DNC(
input_size=64,
hidden_size=128,
rnn_type='lstm',
num_layers=4,
nr_cells=100,
cell_size=32,
read_heads=4,
batch_first=True,
gpu_id=0,
debug=True
)
(controller_hidden, memory, read_vectors) = (None, None, None)
output, (controller_hidden, memory, read_vectors), debug_memory = \
2018-06-29 23:34:10 +08:00
rnn(torch.randn(10, 4, 64), (controller_hidden, memory, read_vectors), reset_experience=True)
2017-11-13 03:04:59 +08:00
```
Memory vectors returned by forward pass (`np.ndarray`):
| Key | Y axis (dimensions) | X axis (dimensions) |
| --- | --- | --- |
| `debug_memory['memory']` | layer * time | nr_cells * cell_size
| `debug_memory['link_matrix']` | layer * time | nr_cells * nr_cells
| `debug_memory['precedence']` | layer * time | nr_cells
| `debug_memory['read_weights']` | layer * time | read_heads * nr_cells
| `debug_memory['write_weights']` | layer * time | nr_cells
| `debug_memory['usage_vector']` | layer * time | nr_cells
2017-12-11 03:13:52 +08:00
### SDNC
**Constructor Parameters**:
Following are the constructor parameters:
| Argument | Default | Description |
| --- | --- | --- |
| input_size | `None` | Size of the input vectors |
| hidden_size | `None` | Size of hidden units |
| rnn_type | `'lstm'` | Type of recurrent cells used in the controller |
| num_layers | `1` | Number of layers of recurrent units in the controller |
| num_hidden_layers | `2` | Number of hidden layers per layer of the controller |
| bias | `True` | Bias |
| batch_first | `True` | Whether data is fed batch first |
| dropout | `0` | Dropout between layers in the controller |
| bidirectional | `False` | If the controller is bidirectional (Not yet implemented |
| nr_cells | `5000` | Number of memory cells |
| read_heads | `4` | Number of read heads |
2017-12-11 19:46:47 +08:00
| sparse_reads | `4` | Number of sparse memory reads per read head |
| temporal_reads | `4` | Number of temporal reads |
2017-12-11 03:13:52 +08:00
| cell_size | `10` | Size of each memory cell |
| nonlinearity | `'tanh'` | If using 'rnn' as `rnn_type`, non-linearity of the RNNs |
| gpu_id | `-1` | ID of the GPU, -1 for CPU |
| independent_linears | `False` | Whether to use independent linear units to derive interface vector |
| share_memory | `True` | Whether to share memory between controller layers |
Following are the forward pass parameters:
| Argument | Default | Description |
| --- | --- | --- |
| input | - | The input vector `(B*T*X)` or `(T*B*X)` |
| hidden | `(None,None,None)` | Hidden states `(controller hidden, memory hidden, read vectors)` |
| reset_experience | `False` | Whether to reset memory |
| pass_through_memory | `True` | Whether to pass through memory |
2017-12-18 15:11:32 +08:00
#### Example usage
2017-12-11 03:13:52 +08:00
```python
from dnc import SDNC
rnn = SDNC(
input_size=64,
hidden_size=128,
rnn_type='lstm',
num_layers=4,
nr_cells=100,
cell_size=32,
read_heads=4,
sparse_reads=4,
batch_first=True,
gpu_id=0
)
(controller_hidden, memory, read_vectors) = (None, None, None)
output, (controller_hidden, memory, read_vectors) = \
2018-06-29 23:34:10 +08:00
rnn(torch.randn(10, 4, 64), (controller_hidden, memory, read_vectors), reset_experience=True)
2017-12-11 03:13:52 +08:00
```
2017-12-18 15:11:32 +08:00
#### Debugging
2017-12-11 03:13:52 +08:00
The `debug` option causes the network to return its memory hidden vectors (numpy `ndarray`s) for the first batch each forward step.
These vectors can be analyzed or visualized, using visdom for example.
```python
from dnc import SDNC
rnn = SDNC(
input_size=64,
hidden_size=128,
rnn_type='lstm',
num_layers=4,
nr_cells=100,
cell_size=32,
read_heads=4,
batch_first=True,
sparse_reads=4,
2017-12-11 19:46:47 +08:00
temporal_reads=4,
2017-12-11 03:13:52 +08:00
gpu_id=0,
debug=True
)
(controller_hidden, memory, read_vectors) = (None, None, None)
output, (controller_hidden, memory, read_vectors), debug_memory = \
2018-06-29 23:34:10 +08:00
rnn(torch.randn(10, 4, 64), (controller_hidden, memory, read_vectors), reset_experience=True)
2017-12-11 03:13:52 +08:00
```
Memory vectors returned by forward pass (`np.ndarray`):
| Key | Y axis (dimensions) | X axis (dimensions) |
| --- | --- | --- |
| `debug_memory['memory']` | layer * time | nr_cells * cell_size
2017-12-11 19:46:47 +08:00
| `debug_memory['visible_memory']` | layer * time | sparse_reads+2*temporal_reads+1 * nr_cells
| `debug_memory['read_positions']` | layer * time | sparse_reads+2*temporal_reads+1
| `debug_memory['link_matrix']` | layer * time | sparse_reads+2*temporal_reads+1 * sparse_reads+2*temporal_reads+1
| `debug_memory['rev_link_matrix']` | layer * time | sparse_reads+2*temporal_reads+1 * sparse_reads+2*temporal_reads+1
| `debug_memory['precedence']` | layer * time | nr_cells
2017-12-11 03:13:52 +08:00
| `debug_memory['read_weights']` | layer * time | read_heads * nr_cells
| `debug_memory['write_weights']` | layer * time | nr_cells
| `debug_memory['usage']` | layer * time | nr_cells
2017-12-18 15:08:45 +08:00
### SAM
**Constructor Parameters**:
Following are the constructor parameters:
| Argument | Default | Description |
| --- | --- | --- |
| input_size | `None` | Size of the input vectors |
| hidden_size | `None` | Size of hidden units |
| rnn_type | `'lstm'` | Type of recurrent cells used in the controller |
| num_layers | `1` | Number of layers of recurrent units in the controller |
| num_hidden_layers | `2` | Number of hidden layers per layer of the controller |
| bias | `True` | Bias |
| batch_first | `True` | Whether data is fed batch first |
| dropout | `0` | Dropout between layers in the controller |
| bidirectional | `False` | If the controller is bidirectional (Not yet implemented |
| nr_cells | `5000` | Number of memory cells |
| read_heads | `4` | Number of read heads |
| sparse_reads | `4` | Number of sparse memory reads per read head |
| cell_size | `10` | Size of each memory cell |
| nonlinearity | `'tanh'` | If using 'rnn' as `rnn_type`, non-linearity of the RNNs |
| gpu_id | `-1` | ID of the GPU, -1 for CPU |
| independent_linears | `False` | Whether to use independent linear units to derive interface vector |
| share_memory | `True` | Whether to share memory between controller layers |
Following are the forward pass parameters:
| Argument | Default | Description |
| --- | --- | --- |
| input | - | The input vector `(B*T*X)` or `(T*B*X)` |
| hidden | `(None,None,None)` | Hidden states `(controller hidden, memory hidden, read vectors)` |
| reset_experience | `False` | Whether to reset memory |
| pass_through_memory | `True` | Whether to pass through memory |
2017-12-18 15:11:32 +08:00
#### Example usage
2017-12-18 15:08:45 +08:00
```python
from dnc import SAM
rnn = SAM(
input_size=64,
hidden_size=128,
rnn_type='lstm',
num_layers=4,
nr_cells=100,
cell_size=32,
read_heads=4,
sparse_reads=4,
batch_first=True,
gpu_id=0
)
(controller_hidden, memory, read_vectors) = (None, None, None)
output, (controller_hidden, memory, read_vectors) = \
2018-06-29 23:34:10 +08:00
rnn(torch.randn(10, 4, 64), (controller_hidden, memory, read_vectors), reset_experience=True)
2017-12-18 15:08:45 +08:00
```
2017-12-18 15:11:32 +08:00
#### Debugging
2017-12-18 15:08:45 +08:00
The `debug` option causes the network to return its memory hidden vectors (numpy `ndarray`s) for the first batch each forward step.
These vectors can be analyzed or visualized, using visdom for example.
```python
from dnc import SAM
rnn = SAM(
input_size=64,
hidden_size=128,
rnn_type='lstm',
num_layers=4,
nr_cells=100,
cell_size=32,
read_heads=4,
batch_first=True,
sparse_reads=4,
gpu_id=0,
debug=True
)
(controller_hidden, memory, read_vectors) = (None, None, None)
output, (controller_hidden, memory, read_vectors), debug_memory = \
2018-06-29 23:34:10 +08:00
rnn(torch.randn(10, 4, 64), (controller_hidden, memory, read_vectors), reset_experience=True)
2017-12-18 15:08:45 +08:00
```
Memory vectors returned by forward pass (`np.ndarray`):
| Key | Y axis (dimensions) | X axis (dimensions) |
| --- | --- | --- |
| `debug_memory['memory']` | layer * time | nr_cells * cell_size
| `debug_memory['visible_memory']` | layer * time | sparse_reads+2*temporal_reads+1 * nr_cells
| `debug_memory['read_positions']` | layer * time | sparse_reads+2*temporal_reads+1
| `debug_memory['read_weights']` | layer * time | read_heads * nr_cells
| `debug_memory['write_weights']` | layer * time | nr_cells
| `debug_memory['usage']` | layer * time | nr_cells
2017-12-19 03:59:12 +08:00
## Tasks
### Copy task (with curriculum and generalization)
2017-10-26 23:29:05 +08:00
The copy task, as descibed in the original paper, is included in the repo.
From the project root:
2017-10-28 02:44:12 +08:00
```bash
2017-11-14 23:46:49 +08:00
python ./tasks/copy_task.py -cuda 0 -optim rmsprop -batch_size 32 -mem_slot 64 # (like original implementation)
2017-11-12 18:26:33 +08:00
python ./tasks/copy_task.py -cuda 0 -lr 0.001 -rnn_type lstm -nlayer 1 -nhlayer 2 -dropout 0 -mem_slot 32 -batch_size 1000 -optim adam -sequence_max_length 8 # (faster convergence)
2017-12-11 03:13:52 +08:00
For SDNCs:
python ./tasks/copy_task.py -cuda 0 -lr 0.001 -rnn_type lstm -memory_type sdnc -nlayer 1 -nhlayer 2 -dropout 0 -mem_slot 100 -mem_size 10 -read_heads 1 -sparse_reads 10 -batch_size 20 -optim adam -sequence_max_length 10
2017-12-11 03:13:52 +08:00
and for curriculum learning for SDNCs:
python ./tasks/copy_task.py -cuda 0 -lr 0.001 -rnn_type lstm -memory_type sdnc -nlayer 1 -nhlayer 2 -dropout 0 -mem_slot 100 -mem_size 10 -read_heads 1 -sparse_reads 4 -temporal_reads 4 -batch_size 20 -optim adam -sequence_max_length 4 -curriculum_increment 2 -curriculum_freq 10000
2017-11-12 18:26:33 +08:00
```
For the full set of options, see:
```
python ./tasks/copy_task.py --help
2017-10-28 02:44:12 +08:00
```
The copy task can be used to debug memory using [Visdom](https://github.com/facebookresearch/visdom).
Additional step required:
```bash
pip install visdom
python -m visdom.server
2017-10-26 23:29:05 +08:00
```
2017-10-28 02:44:12 +08:00
Open http://localhost:8097/ on your browser, and execute the copy task:
```bash
python ./tasks/copy_task.py -cuda 0
2017-10-26 23:29:05 +08:00
```
2017-10-28 02:44:12 +08:00
The visdom dashboard shows memory as a heatmap for batch 0 every `-summarize_freq` iteration:
2017-11-01 17:49:39 +08:00
![Visdom dashboard](./docs/dnc-mem-debug.png)
2017-10-28 02:44:12 +08:00
2017-12-19 03:59:12 +08:00
### Generalizing Addition task
The adding task is as described in [this github pull request](https://github.com/Mostafa-Samir/DNC-tensorflow/pull/4#issue-199369192).
This task
- creates one-hot vectors of size `input_size`, each representing a number
- feeds a sentence of them to a network
- the output of which is added to get the sum of the decoded outputs
The task first trains the network for sentences of size ~100, and then tests if the network genetalizes for lengths ~1000.
```bash
python ./tasks/adding_task.py -cuda 0 -lr 0.0001 -rnn_type lstm -memory_type sam -nlayer 1 -nhlayer 1 -nhid 100 -dropout 0 -mem_slot 1000 -mem_size 32 -read_heads 1 -sparse_reads 4 -batch_size 20 -optim rmsprop -input_size 3 -sequence_max_length 100
2017-12-19 03:59:12 +08:00
```
### Generalizing Argmax task
2017-12-19 03:59:12 +08:00
The second adding task is similar to the first one, except that the network's output at the last time step is expected to be the argmax of the input.
2017-12-19 03:59:12 +08:00
```bash
python ./tasks/argmax_task.py -cuda 0 -lr 0.0001 -rnn_type lstm -memory_type dnc -nlayer 1 -nhlayer 1 -nhid 100 -dropout 0 -mem_slot 100 -mem_size 10 -read_heads 2 -batch_size 1 -optim rmsprop -sequence_max_length 15 -input_size 10 -iterations 10000
2017-12-19 03:59:12 +08:00
```
2017-12-18 16:08:04 +08:00
## Code Structure
1. DNCs:
- [dnc/dnc.py](dnc/dnc.py) - Controller code.
- [dnc/memory.py](dnc/memory.py) - Memory module.
2. SDNCs:
- [dnc/sdnc.py](dnc/sdnc.py) - Controller code, inherits [dnc.py](dnc/dnc.py).
- [dnc/sparse_temporal_memory.py](dnc/sparse_temporal_memory.py) - Memory module.
- [dnc/flann_index.py](dnc/flann_index.py) - Memory index using kNN.
3. SAMs:
- [dnc/sam.py](dnc/sam.py) - Controller code, inherits [dnc.py](dnc/dnc.py).
- [dnc/sparse_memory.py](dnc/sparse_memory.py) - Memory module.
- [dnc/flann_index.py](dnc/flann_index.py) - Memory index using kNN.
4. Tests:
- All tests are in [./tests](./tests) folder.
2017-10-28 02:44:12 +08:00
2017-10-26 23:29:05 +08:00
## General noteworthy stuff
1. SDNCs use the [FLANN approximate nearest neigbhour library](https://github.com/mariusmuja/flann), with its python binding [pyflann3](https://github.com/primetang/pyflann) and [FAISS](https://github.com/facebookresearch/faiss).
FLANN can be installed either from pip (automatically as a dependency), or from source (e.g. for multithreading via OpenMP):
```bash
# install openmp first: e.g. `sudo pacman -S openmp` for Arch.
git clone git://github.com/mariusmuja/flann.git
cd flann
mkdir build
cd build
cmake ..
make -j 4
sudo make install
```
FAISS can be installed using:
```bash
conda install faiss-gpu -c pytorch
```
FAISS is much faster, has a GPU implementation and is interoperable with pytorch tensors.
We try to use FAISS by default, in absence of which we fall back to FLANN.
2017-12-20 05:35:01 +08:00
2. `nan`s in the gradients are common, try with different batch sizes
2017-10-26 23:29:05 +08:00
2017-11-01 15:26:24 +08:00
Repos referred to for creation of this repo:
2017-11-01 15:38:32 +08:00
- [deepmind/dnc](https://github.com/deepmind/dnc)
- [ypxie/pytorch-NeuCom](https://github.com/ypxie/pytorch-NeuCom)
- [jingweiz/pytorch-dnc](https://github.com/jingweiz/pytorch-dnc)
2017-11-01 15:26:24 +08:00