Differentiable Neural Computers and Sparse Differentiable Neural Computers, for Pytorch

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 the Sparse version of the DNC (the SDNC) described in Scaling Memory-Augmented Neural Networks with Sparse Reads and Writes.

Install

pip install dnc

For using sparse DNCs, additional libraries are required:

FAISS

SDNCs require an additional library: facebookresearch/faiss. A compiled version of the library with intel SSE + CUDA 8 support ships with this library. If that does not work, one might need to manually compile faiss, as detailed below:

Installing FAISS

Needs libopenblas.so in /usr/lib/.

This has been tested on Arch Linux. Other distributions might have different libopenblas path or cuda root dir or numpy include files dir.

git clone https://github.com/facebookresearch/faiss.git
cd faiss
cp ./example_makefiles/makefile.inc.Linux ./makefile.inc
# change libopenblas path
sed -i "s/lib64\/libopenblas\.so\.0/lib\/libopenblas\.so/g" ./makefile.inc
# add option for nvcc to work properly with g++ > 5
sed -i "s/std c++11 \-lineinfo/std c++11 \-lineinfo \-Xcompiler \-D__CORRECT_ISO_CPP11_MATH_H_PROTO/g" ./makefile.inc
# change CUDA ROOT
sed -i "s/CUDAROOT=\/usr\/local\/cuda-8.0\//CUDAROOT=\/opt\/cuda\//g" ./makefile.inc
# change numpy include files (for v3.6)
sed -i "s/PYTHONCFLAGS=\-I\/usr\/include\/python2.7\/ \-I\/usr\/lib64\/python2.7\/site\-packages\/numpy\/core\/include\//PYTHONCFLAGS=\-I\/usr\/include\/python3.6m\/ \-I\/usr\/lib\/python3.6\/site\-packages\/numpy\/core\/include/g"

# build
make
cd gpu
make
cd ..
make py
cd gpu
make py
cd ..

mkdir /tmp/faiss
find -name "*.so" -exec cp {} /tmp/faiss \;
find -name "*.a" -exec cp {} /tmp/faiss \;
find -name "*.py" -exec cp {} /tmp/faiss \;
mv /tmp/faiss .
cd faiss

# convert to python3
2to3 -w ./*.py
rm -rf *.bak

# Fix relative imports
for i in *.py; do
  filename=`echo $i | cut -d "." -f 1`
  echo $filename
  find -name "*.py" -exec sed -i "s/import $filename/import \.$filename/g" {} \;
  find -name "*.py" -exec sed -i "s/from $filename import/from \.$filename import/g" {} \;
done

cd ..

git clone https://github.com/ixaxaar/pytorch-dnc
mv faiss pytorch-dnc
cd pytorch-dnc
sudo pip install -e .

Architecure

Usage

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 (This is a parameter for the forward pass
pass_through_memory	True	Whether to pass through memory (This is a parameter for the forward pass

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

Example copy task

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)

python3 ./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 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:

General noteworthy stuff

1. DNCs converge faster with Adam and RMSProp learning rules, SGD generally converges extremely slowly. The copy task, for example, takes 25k iterations on SGD with lr 1 compared to 3.5k for adam with lr 0.01.
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