Differentiable Neural Computer, 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.

Install

pip install dnc

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 `(BTX)` or `(TBX)`
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

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

Repos referred to for creation of this repo:

5.2 KiB Raw Blame History