Differentiable Neural Computers and family, for Pytorch

Includes:

1. Differentiable Neural Computers (DNC)
2. Sparse Access Memory (SAM)
3. Sparse Differentiable Neural Computers (SDNC)

• Install
  • From source
• Architecure
• Usage
  • DNC
    • Example usage
    • Debugging
  • SDNC
    • Example usage
    • Debugging
  • SAM
    • Example usage
    • Debugging
• Tasks
  • Copy task (with curriculum and generalization)
  • Generalizing Addition task
  • Generalizing Argmax task
• Code Structure
• General noteworthy stuff

This is an implementation of Differentiable Neural Computers, described in the paper Hybrid computing using a neural network with dynamic external memory, Graves et al. and Sparse DNCs (SDNCs) and Sparse Access Memory (SAM) described in Scaling Memory-Augmented Neural Networks with Sparse Reads and Writes.

Install

pip install dnc

From source

git clone https://github.com/ixaxaar/pytorch-dnc
cd pytorch-dnc
pip install -r ./requirements.txt
pip install -e .

For using fully GPU based SDNCs or SAMs, install FAISS:

conda install faiss-gpu -c pytorch

pytest is required to run the test

Architecure

Usage

DNC

Constructor Parameters:

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
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:

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

Example usage

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) = \
  rnn(torch.randn(10, 4, 64), (controller_hidden, memory, read_vectors, reset_experience=True))

Debugging

The debug option causes the network to return its memory hidden vectors (numpy ndarrays) for the first batch each forward step. These vectors can be analyzed or visualized, using visdom for example.

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 = \
  rnn(torch.randn(10, 4, 64), (controller_hidden, memory, read_vectors, reset_experience=True))

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

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
sparse_reads	4	Number of sparse memory reads per read head
temporal_reads	4	Number of temporal reads
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

Example usage

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) = \
  rnn(torch.randn(10, 4, 64), (controller_hidden, memory, read_vectors, reset_experience=True))

Debugging

The debug option causes the network to return its memory hidden vectors (numpy ndarrays) for the first batch each forward step. These vectors can be analyzed or visualized, using visdom for example.

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,
  temporal_reads=4,
  gpu_id=0,
  debug=True
)

(controller_hidden, memory, read_vectors) = (None, None, None)

output, (controller_hidden, memory, read_vectors), debug_memory = \
  rnn(torch.randn(10, 4, 64), (controller_hidden, memory, read_vectors, reset_experience=True))

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['link_matrix']	layer * time	sparse_reads+2temporal_reads+1 * sparse_reads+2temporal_reads+1
debug_memory['rev_link_matrix']	layer * time	sparse_reads+2temporal_reads+1 * sparse_reads+2temporal_reads+1
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']	layer * time	nr_cells

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

Example usage

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) = \
  rnn(torch.randn(10, 4, 64), (controller_hidden, memory, read_vectors, reset_experience=True))

Debugging

The debug option causes the network to return its memory hidden vectors (numpy ndarrays) for the first batch each forward step. These vectors can be analyzed or visualized, using visdom for example.

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 = \
  rnn(torch.randn(10, 4, 64), (controller_hidden, memory, read_vectors, reset_experience=True))

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

Tasks

Copy task (with curriculum and generalization)

The copy task, as descibed in the original paper, is included in the repo.

From the project root:

python ./tasks/copy_task.py -cuda 0 -optim rmsprop -batch_size 32 -mem_slot 64 # (like original implementation)

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)

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

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

For the full set of options, see:

python ./tasks/copy_task.py --help

The copy task can be used to debug memory using Visdom.

Additional step required:

pip install visdom
python -m visdom.server

Open http://localhost:8097/ on your browser, and execute the copy task:

python ./tasks/copy_task.py -cuda 0

The visdom dashboard shows memory as a heatmap for batch 0 every -summarize_freq iteration:

Generalizing Addition task

The adding task is as described in this github pull request. 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.

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

Generalizing Argmax task

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.

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

Code Structure

1. DNCs:

• dnc/dnc.py - Controller code.
• dnc/memory.py - Memory module.

2. SDNCs:

• dnc/sdnc.py - Controller code, inherits dnc.py.
• dnc/sparse_temporal_memory.py - Memory module.
• dnc/flann_index.py - Memory index using kNN.

3. SAMs:

• dnc/sam.py - Controller code, inherits dnc.py.
• dnc/sparse_memory.py - Memory module.
• dnc/flann_index.py - Memory index using kNN.

4. Tests:

• All tests are in ./tests folder.

General noteworthy stuff

1. SDNCs use the FLANN approximate nearest neigbhour library, with its python binding pyflann3 and FAISS.

FLANN can be installed either from pip (automatically as a dependency), or from source (e.g. for multithreading via OpenMP):

# 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:

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.

2. nans in the gradients are common, try with different batch sizes

Repos referred to for creation of this repo:

• deepmind/dnc
• ypxie/pytorch-NeuCom
• jingweiz/pytorch-dnc