Differentiable Neural Computers, Sparse Access Memory and Sparse Differentiable Neural Computers, for Pytorch

Related tags

Deep Learning pytorch rnn dnc differentiable-neural-computers
Overview

Differentiable Neural Computers and family, for Pytorch

Includes:

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

Build Status PyPI version

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:

Visdom dashboard

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:
  1. SDNCs:
  1. SAMs:
  1. Tests:

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.

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

Repos referred to for creation of this repo:

Comments
  • copy_task.py sample fails.

    copy_task.py sample fails.

    testing with command line:

    python 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 -iterations 100

    I get multiple errors when it finishes, first on the generate_data call which has undefined parameters:

    input_data, target_output, loss_weights = generate_data(random_length, input_size)

    NameError: name 'input_size' is not defined

    And then after fixing that I get: output = output[:, -1, :].sum().data.cpu().numpy()[0] IndexError: too many indices for array

    Looks like that bit of code hasn't been used. I have tried to fix it but I'm unclear of the solution for the second issue as I'm new to pytorch, thanks in advance for any fixes.

    ChrisP.

    opened by chrispugmire 3
  • Problem of the Softmax on Read Mode

    Problem of the Softmax on Read Mode

    https://github.com/ixaxaar/pytorch-dnc/blob/1db78511fe5622ade1c554d265a5d9d729c8801d/dnc/memory.py#L235

    Should the softmax be applied on the last dimension? (i.e. the dimension of the read mode)

    Currently, each read mode would always return 1 if the model has only one read head.

    opened by yat011 3
  • Question about the running speed of Pyflann and Faiss for the SAM model

    Question about the running speed of Pyflann and Faiss for the SAM model

    I can't normally install Faiss environment because of certain force majeure, so I wonder how using Faiss-gpu or Pyflann will influence the actual training speed of the SAM model. Let's say for example, in the copy task, how are the actual epoch times when using these two methods? Can you give me a rough reference?

    opened by zoharli 3
  • PySide dependency error

    PySide dependency error

    I followed your instructions to run and visualize copy_task.py in visdom, but am encountering some dependency errors. I am using Python 3.6.

    First error when running python ./tasks/copy_task.py -cuda 0:

    File "C:\Users\alexander.d.payne\AppData\Local\Programs\Python\Python36\lib\site-packages\pyflann\bindings\flann_ctypes.py", line 171, in <module> raise ImportError('Cannot load dynamic library. Did you compile FLANN?') ImportError: Cannot load dynamic library. Did you compile FLANN?

    pip installed pyflann, and ran again:

    File "C:\Users\alexander.d.payne\AppData\Local\Programs\Python\Python36\lib\site-packages\pyflann\__init__.py", line 27, in <module> from index import * ModuleNotFoundError: No module named 'index'

    pip installed index, and ran again:

    C:\Users\alexander.d.payne\Documents\pytorch-dnc-master>pip install index Collecting index Downloading ...files.pythonhosted.org/packages/7f/59/65da893e04f3eb49f73e6770e0999c57230669a484b14ca574154e9b75d3/index-0.2.tar.gz Collecting PySide (from index) Downloading ...files.pythonhosted.org/packages/36/ac/ca31db6f2225844d37a41b10615c3d371587677efd074db29855e7035de6/PySide-1.2.4.tar.gz (9.3MB) 100% |████████████████████████████████| 9.3MB 3.2MB/s Complete output from command python setup.py egg_info: only these python versions are supported: [(2, 6), (2, 7), (3, 2), (3, 3), (3, 4)] Command "python setup.py egg_info" failed with error code 1 in C:\Users\ALEXAN~1.PAY\AppData\Local\Temp\pip-install-aph21f59\PySide\

    Even if I switched to Python 3.4 I don't see a torch option for that version at their website https://pytorch.org/, is there anyway around this? Thank you.

    opened by apayne19 3
  • Issues when using pytorch 0.4

    Issues when using pytorch 0.4

    I get errors when trying to run both DNC and SDNC examples with pytorch 0.4.0. For DNC:

    (py36) [mammadli@server test]$ python test_dnc.py 
/amd/home/mammadli/tools/anaconda2/envs/py36/lib/python3.6/site-packages/h5py/__init__.py:36: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.
  from ._conv import register_converters as _register_converters
/amd/home/mammadli/tools/anaconda2/envs/py36/lib/python3.6/site-packages/dnc/dnc.py:118: UserWarning: nn.init.orthogonal is now deprecated in favor of nn.init.orthogonal_.
  orthogonal(self.output.weight)
/amd/home/mammadli/tools/anaconda2/envs/py36/lib/python3.6/site-packages/dnc/dnc.py:133: UserWarning: nn.init.xavier_uniform is now deprecated in favor of nn.init.xavier_uniform_.
  xavier_uniform(h)
/amd/home/mammadli/tools/anaconda2/envs/py36/lib/python3.6/site-packages/dnc/util.py:95: UserWarning: Implicit dimension choice for softmax has been deprecated. Change the call to include dim=X as an argument.
  soft_max_2d = F.softmax(input_2d)
Traceback (most recent call last):
  File "test_dnc.py", line 19, in <module>
    rnn(torch.randn(10, 4, 64).cuda(), (controller_hidden, memory, read_vectors), True)
  File "/amd/home/mammadli/tools/anaconda2/envs/py36/lib/python3.6/site-packages/torch/nn/modules/module.py", line 491, in __call__
    result = self.forward(*input, **kwargs)
  File "/amd/home/mammadli/tools/anaconda2/envs/py36/lib/python3.6/site-packages/dnc/dnc.py", line 265, in forward
    inputs = [self.output(i) for i in inputs]
  File "/amd/home/mammadli/tools/anaconda2/envs/py36/lib/python3.6/site-packages/dnc/dnc.py", line 265, in <listcomp>
    inputs = [self.output(i) for i in inputs]
  File "/amd/home/mammadli/tools/anaconda2/envs/py36/lib/python3.6/site-packages/torch/nn/modules/module.py", line 491, in __call__
    result = self.forward(*input, **kwargs)
  File "/amd/home/mammadli/tools/anaconda2/envs/py36/lib/python3.6/site-packages/torch/nn/modules/linear.py", line 55, in forward
    return F.linear(input, self.weight, self.bias)
  File "/amd/home/mammadli/tools/anaconda2/envs/py36/lib/python3.6/site-packages/torch/nn/functional.py", line 992, in linear
    return torch.addmm(bias, input, weight.t())
RuntimeError: Expected object of type torch.FloatTensor but found type torch.cuda.FloatTensor for argument #4 'mat1'

    For SDNC:

    (py36) [mammadli@server test]$ python test_dnc.py 
/amd/home/mammadli/tools/anaconda2/envs/py36/lib/python3.6/site-packages/h5py/__init__.py:36: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.
  from ._conv import register_converters as _register_converters
/amd/home/mammadli/tools/anaconda2/envs/py36/lib/python3.6/site-packages/dnc/dnc.py:118: UserWarning: nn.init.orthogonal is now deprecated in favor of nn.init.orthogonal_.
  orthogonal(self.output.weight)
/amd/home/mammadli/tools/anaconda2/envs/py36/lib/python3.6/site-packages/dnc/sparse_temporal_memory.py:65: UserWarning: nn.init.orthogonal is now deprecated in favor of nn.init.orthogonal_.
  T.nn.init.orthogonal(self.interface_weights.weight)
/amd/home/mammadli/tools/anaconda2/envs/py36/lib/python3.6/site-packages/dnc/dnc.py:133: UserWarning: nn.init.xavier_uniform is now deprecated in favor of nn.init.xavier_uniform_.
  xavier_uniform(h)
Traceback (most recent call last):
  File "test_dnc.py", line 20, in <module>
    rnn(torch.randn(10, 4, 64).cuda(), (controller_hidden, memory, read_vectors), True)
  File "/amd/home/mammadli/tools/anaconda2/envs/py36/lib/python3.6/site-packages/torch/nn/modules/module.py", line 491, in __call__
    result = self.forward(*input, **kwargs)
  File "/amd/home/mammadli/tools/anaconda2/envs/py36/lib/python3.6/site-packages/dnc/dnc.py", line 219, in forward
    controller_hidden, mem_hidden, last_read = self._init_hidden(hx, batch_size, reset_experience)
  File "/amd/home/mammadli/tools/anaconda2/envs/py36/lib/python3.6/site-packages/dnc/dnc.py", line 144, in _init_hidden
    mhx = self.memories[0].reset(batch_size, erase=reset_experience)
  File "/amd/home/mammadli/tools/anaconda2/envs/py36/lib/python3.6/site-packages/dnc/sparse_temporal_memory.py", line 126, in reset
    'read_positions': cuda(T.arange(0, c).expand(b, c), gpu_id=self.gpu_id).long()
  File "/amd/home/mammadli/tools/anaconda2/envs/py36/lib/python3.6/site-packages/dnc/util.py", line 30, in cuda
    return var(x.pin_memory(), requires_grad=grad).cuda(gpu_id, async=True)
RuntimeError: invalid argument 3: Source tensor must be contiguous at /opt/conda/conda-bld/pytorch_1524590031827/work/aten/src/THC/generic/THCTensorCopy.c:114
    opened by Rahim16 3
  • TypeError: cat received an invalid combination of arguments - got (list, int), but expected one of:

    TypeError: cat received an invalid combination of arguments - got (list, int), but expected one of:

    I'm trying to run your example for SAM, but I'm running into the following error:

    Traceback (most recent call last):
  File "dnc_test.py", line 20, in <module>
    rnn(torch.randn(10, 4, 64), (controller_hidden, memory, read_vectors), reset_experience=True)
  File "/home/xxx/miniconda3/envs/my_env/lib/python3.6/site-packages/torch/nn/modules/module.py", line 357, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/xxx/git_repos/pytorch-dnc/dnc/dnc.py", line 222, in forward
    inputs = [T.cat([input[:, x, :], last_read], 1) for x in range(max_length)]
  File "/home/xxx/git_repos/pytorch-dnc/dnc/dnc.py", line 222, in <listcomp>
    inputs = [T.cat([input[:, x, :], last_read], 1) for x in range(max_length)]
TypeError: cat received an invalid combination of arguments - got (list, int), but expected one of:
 * (sequence[torch.cuda.FloatTensor] seq)
 * (sequence[torch.cuda.FloatTensor] seq, int dim)
      didn't match because some of the arguments have invalid types: (list, int)

    I tried loading the input tensor onto the gpu with .cuda() and transforming last_read to a Tensor with .data, but that led to other issues.

    There's also a typo in your documentation:

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

    should be

    output, (controller_hidden, memory, read_vectors) = \
  rnn(torch.randn(10, 4, 64), (controller_hidden, memory, read_vectors), reset_experience=True)
    opened by kierkegaard13 3
  • bug in cosine distance?

    bug in cosine distance?

    I believe there's a bug in the function from dnc.util for computing cosine distance. First, I think you are trying to compute cosine similarity, not distance (sim = 1 - dist). Second, I think the current function implements neither cosine similarity nor distance. Here's a modified variant that returns the correct output for cosine similarity.

    def bcos(a, b, normBy=2):
    """Batchwise cosine similarity
    
    Arguments:
        a: 3D tensor of shape [b,m,w]
        b: 3D tensor of shape [b,r,w]
    Returns:
        cos: batchwise cosine similarity of shape [b,r,m]
    """
    dot = torch.bmm(a, b.transpose(1,2)) # [b,m,w] @ [b,w,r] -> [b,m,r]
    a_norm = torch.norm(a, normBy, dim=2).unsqueeze(2) # [b,m,1]
    b_norm = torch.norm(b, normBy, dim=2).unsqueeze(1) # [b,1,r]
    cos = dot / (a_norm * b_norm) # [b,m,r]

    return cos.transpose(1,2)  # [b,r,m]
    opened by rfeinman 2
  • reset_experience meaning

    reset_experience meaning

    Looking at the code it isn't clear to me what reset_experience does. Since the memory is an argument to the forward, is setting reset_experience equivalent to calling dnc(controller_state, None, read_vectors)?

    If this is the case, then when using dnc with inputs of shape [batch, time, feature] , reset_experience will clear the memory between batches. If we continuously call dnc with inputs of shape [batch, 1, feature], then we do not want to reset. Is this correct?

    opened by smorad 1
  • fix bug in function \theta for batchwise cosine similarity

    fix bug in function \theta for batchwise cosine similarity

    I wasn't able to keep the options "dimA" and "dimB" for non-default similarity dimensions, but I don't see those used anywhere in the repo.

    opened by rfeinman 1
  • Error when running copy_task.py

    Error when running copy_task.py

    I executed copy task as following the cmd line in README. python copy_task.py -cuda 0 -optim rmsprop -batch_size 32 -mem_slot 64 Having NameError at line 366, name 'input_size' is not defined. I can change the input_size to args.input_size, but I think there are additional problems other than that. Function generate_data at requires 3 arguments, but only 2 are given. Is generate_data in line 366 a different function defined in line 78?

    https://github.com/ixaxaar/pytorch-dnc/blob/016b541223bf801f3f3a617fa3942cc12ef71be9/tasks/copy_task.py#L78 https://github.com/ixaxaar/pytorch-dnc/blob/016b541223bf801f3f3a617fa3942cc12ef71be9/tasks/copy_task.py#L366

    Thank you,

    opened by jin8 1
  • When running adding task -- ModuleNotFoundError: No module named 'index'

    When running adding task -- ModuleNotFoundError: No module named 'index' 

      File "tasks/adding_task.py", line 25, in <module>
    from dnc.dnc import DNC
  File "/home/vanangamudi/projects/cloned/pytorch-dnc/dnc/__init__.py", line 5, in <module>
    from .sdnc import SDNC
  File "/home/vanangamudi/projects/cloned/pytorch-dnc/dnc/sdnc.py", line 15, in <module>
    from .sparse_temporal_memory import SparseTemporalMemory
  File "/home/vanangamudi/projects/cloned/pytorch-dnc/dnc/sparse_temporal_memory.py", line 11, in <module>
    from .flann_index import FLANNIndex
  File "/home/vanangamudi/projects/cloned/pytorch-dnc/dnc/flann_index.py", line 9, in <module>
    from pyflann import *
  File "/home/vanangamudi/env/torch/lib/python3.6/site-packages/pyflann/__init__.py", line 27, in <module>
    from index import *
ModuleNotFoundError: No module named 'index'
    opened by vanangamudi 1
  • pytorch LTS support (1.8.2) or stable (1.11.1)

    pytorch LTS support (1.8.2) or stable (1.11.1)

    Hello!

    I was wondering if someone can confirm that this package still runs under pytroch lts or current stable (1.11.1)?

    I'm getting a curious error. Note this is for CPU training. Maybe someone can confirm this is only broken under cpu training.

    Thank you!

    `03:44 $ python ./tasks/adding_task.py -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 Namespace(batch_size=20, check_freq=100, clip=50, cuda=-1, dropout=0.0, input_size=3, iterations=2000, lr=0.0001, mem_size=32, mem_slot=1000, memory_type='sam', nhid=100, nhlayer=1, nlayer=1, optim='rmsprop', read_heads=1, rnn_type='lstm', sequence_max_length=100, sparse_reads=4, summarize_freq=100, temporal_reads=2, visdom=False) Using CPU.

    SAM(3, 100, num_hidden_layers=1, nr_cells=1000, read_heads=1, cell_size=32) SAM( (lstm_layer_0): LSTM(35, 100, batch_first=True) (rnn_layer_memory_shared): SparseMemory( (interface_weights): Linear(in_features=100, out_features=70, bias=True) ) (output): Linear(in_features=132, out_features=3, bias=True) )

    Iteration 0/2000 Falling back to FLANN (CPU). For using faster, GPU based indexes, install FAISS: "conda install faiss-gpu -c pytorch" Traceback (most recent call last): File "./tasks/adding_task.py", line 222, in loss.backward() File "/home/eziegenbalg/.conda/envs/default/lib/python3.8/site-packages/torch/tensor.py", line 245, in backward torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs) File "/home/eziegenbalg/.conda/envs/default/lib/python3.8/site-packages/torch/autograd/init.py", line 145, in backward Variable._execution_engine.run_backward( RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.FloatTensor [1, 1000]], which is output 0 of AsStridedBackward, is at version 70; expected version 69 instead. Hint: enable anomaly detection to find the operation that failed to compute its gradient, with torch.autograd.set_detect_anomaly(True).

    ^C (default) ✘-INT ~/pytorch-dnc [master|✚ 2] 03:45 $ `

    opened by ziegenbalg 5
  • Question about allocation weighting

    Question about allocation weighting

    The paper describes the allocation weighting vector as:

    $a_t[\phi_t[j]] = (1 - u_t[\phi_t[j]]) \prod u_t[\psi_t[j]] $

    In your part of the code where you calculate the right-part product you do this:

    v = var(sorted_usage.data.new(batch_size, 1).fill_(1))
cat_sorted_usage = T.cat((v, sorted_usage), 1)
prod_sorted_usage = T.cumprod(cat_sorted_usage, 1)[:, :-1]

    Why do you create the var "v", which contains "ones", and concatenate it? This does not seem the same as in the paper.

    Thanks, Peter

    opened by PeterDeWachter1998 0
  • A question about memory initialization.

    A question about memory initialization.

    Hi,

    I am a bit confused about how we save memory states in DNC. To be more specific, at the starting point of training, we have to initialize the memory with no doubt (fill all 0s in code). Having finished the training process, I think the memory values should be saved for testing usages. But it turns out that you reset the memory hidden states to be 0s AGAIN! (just as the erase part of dnc/memory.py,Line 69-75).

    Could you please give me some explanations about this? Thank you in advance! Really need your help.

    opened by LiUzHiAn 1
  • batch_first argument doesn’t work

    batch_first argument doesn’t work

    Hi, I just want to let you know that with current implementation (file dnc.py, line 76:86), the batch_first will always be True. It is trivial but sometime troublesome. Have a nice day.

    opened by Trungmaster5 1
  • Unresolved reference 'output'

    Unresolved reference 'output'

    https://github.com/ixaxaar/pytorch-dnc/blob/016b541223bf801f3f3a617fa3942cc12ef71be9/dnc/dnc.py#L270

    The variable named 'output' (as can be seen above) is an unresolved reference. Kindly, please fix it.

    opened by denizetkar 1
Releases(1.0.1)
Owner
ixaxaar
ixaxaar
GitHub
PyTorch Code of "Memory In Memory: A Predictive Neural Network for Learning Higher-Order Non-Stationarity from Spatiotemporal Dynamics"

Memory In Memory Networks It is based on the paper Memory In Memory: A Predictive Neural Network for Learning Higher-Order Non-Stationarity from Spati

Yang Li 12 May 30, 2022
Segcache: a memory-efficient and scalable in-memory key-value cache for small objects

Segcache: a memory-efficient and scalable in-memory key-value cache for small objects This repo contains the code of Segcache described in the followi

TheSys Group @ CMU CS 78 Jan 7, 2023
Episodic-memory - Ego4D Episodic Memory Benchmark

Ego4D Episodic Memory Benchmark EGO4D is the world's largest egocentric (first p

null 3 Feb 18, 2022
Implementation of a memory efficient multi-head attention as proposed in the paper, "Self-attention Does Not Need O(n²) Memory"

Memory Efficient Attention Pytorch Implementation of a memory efficient multi-head attention as proposed in the paper, Self-attention Does Not Need O(

Phil Wang 180 Jan 5, 2023
TargetAllDomainObjects - A python wrapper to run a command on against all users/computers/DCs of a Windows Domain

TargetAllDomainObjects A python wrapper to run a command on against all users/co

Podalirius 19 Dec 13, 2022
Neural Turing Machine (NTM) & Differentiable Neural Computer (DNC) with pytorch & visdom

Neural Turing Machine (NTM) & Differentiable Neural Computer (DNC) with pytorch & visdom Sample on-line plotting while training(avg loss)/testing(writ

Jingwei Zhang 269 Nov 15, 2022
Pytorch Implementation of Google's Parallel Tacotron 2: A Non-Autoregressive Neural TTS Model with Differentiable Duration Modeling

Parallel Tacotron2 Pytorch Implementation of Google's Parallel Tacotron 2: A Non-Autoregressive Neural TTS Model with Differentiable Duration Modeling

Keon Lee 170 Dec 27, 2022
Pytorch implementation of DeepMind's differentiable neural computer paper.

DNC pytorch This is a Pytorch implementation of DeepMind's Differentiable Neural Computer (DNC) architecture introduced in their recent Nature paper:

Yuanpu Xie 91 Nov 21, 2022
Memory Efficient Attention (O(sqrt(n)) for Jax and PyTorch

Memory Efficient Attention This is unofficial implementation of Self-attention Does Not Need O(n^2) Memory for Jax and PyTorch. Implementation is almo

Amin Rezaei 126 Dec 27, 2022
Memory-efficient optimum einsum using opt_einsum planning and PyTorch kernels.

opt-einsum-torch There have been many implementations of Einstein's summation. numpy's numpy.einsum is the least efficient one as it only runs in sing

Haoyan Huo 9 Nov 18, 2022
Easily benchmark PyTorch model FLOPs, latency, throughput, max allocated memory and energy consumption

⏱ pytorch-benchmark Easily benchmark model inference FLOPs, latency, throughput, max allocated memory and energy consumption Install pip install pytor

Lukas Hedegaard 21 Dec 22, 2022
Official pytorch implementation of Rainbow Memory (CVPR 2021)

Rainbow Memory: Continual Learning with a Memory of Diverse Samples

Clova AI Research 91 Dec 17, 2022
This is a pytorch implementation of the NeurIPS paper GAN Memory with No Forgetting.

GAN Memory for Lifelong learning This is a pytorch implementation of the NeurIPS paper GAN Memory with No Forgetting. Please consider citing our paper

Miaoyun Zhao 43 Dec 27, 2022
Implementation of Hierarchical Transformer Memory (HTM) for Pytorch

Hierarchical Transformer Memory (HTM) - Pytorch Implementation of Hierarchical Transformer Memory (HTM) for Pytorch. This Deepmind paper proposes a si

Phil Wang 63 Dec 29, 2022
PyTorch implementation of Memory-based semantic segmentation for off-road unstructured natural environments.

MemSeg: Memory-based semantic segmentation for off-road unstructured natural environments Introduction This repository is a PyTorch implementation of

null 11 Nov 28, 2022
Space Time Recurrent Memory Network - Pytorch

Space Time Recurrent Memory Network - Pytorch (wip) Implementation of Space Time Recurrent Memory Network, recurrent network competitive with attentio

Phil Wang 50 Nov 7, 2021
This is the official PyTorch implementation for "Mesa: A Memory-saving Training Framework for Transformers".

Mesa: A Memory-saving Training Framework for Transformers This is the official PyTorch implementation for Mesa: A Memory-saving Training Framework for

Zhuang AI Group 105 Dec 6, 2022
A library for low-memory inferencing in PyTorch.

Pylomin Pylomin (PYtorch LOw-Memory INference) is a library for low-memory inferencing in PyTorch. Installation ... Usage For example, the following c

null 3 Oct 26, 2022
A human-readable PyTorch implementation of "Self-attention Does Not Need O(n^2) Memory"

memory_efficient_attention.pytorch A human-readable PyTorch implementation of "Self-attention Does Not Need O(n^2) Memory" (Rabe&Staats'21). def effic

Ryuichiro Hataya 7 Dec 26, 2022