theano.gpuarray.dnn – cuDNN

cuDNN is an NVIDIA library with functionality used by deep neural networks. It provides optimized versions of some operations like the convolution. cuDNN is not currently installed with CUDA. You must download and install it yourself.

To install it, decompress the downloaded file and make the *.h and *.so* files available to the compilation environment. There are at least three possible ways of doing so:

  • The easiest is to include them in your CUDA installation. Copy the *.h files to CUDA_ROOT/include and the *.so* files to CUDA_ROOT/lib64 (by default, CUDA_ROOT is /usr/local/cuda on Linux).

  • Alternatively, on Linux, you can set the environment variables LD_LIBRARY_PATH, LIBRARY_PATH and CPATH to the directory extracted from the download. If needed, separate multiple directories with : as in the PATH environment variable.

    example:

    export LD_LIBRARY_PATH=/home/user/path_to_CUDNN_folder/lib64:$LD_LIBRARY_PATH
export CPATH=/home/user/path_to_CUDNN_folder/include:$CPATH
export LIBRARY_PATH=/home/user/path_to_CUDNN_folder/lib64:$LD_LIBRARY_PATH

  • And as a third way, also on Linux, you can copy the *.h files to /usr/include and the *.so* files to /lib64.

By default, Theano will detect if it can use cuDNN. If so, it will use it. If not, Theano optimizations will not introduce cuDNN ops. So Theano will still work if the user did not introduce them manually.

To get an error if Theano can not use cuDNN, use this Theano flag: optimizer_including=cudnn.

Note

cuDNN v5.1 is supported in Theano master version. So it dropped cuDNN v3 support. Theano 0.8.0 and 0.8.1 support only cuDNN v3 and v4. Theano 0.8.2 will support only v4 and v5.

Note

Starting in cuDNN v3, multiple convolution implementations are offered and it is possible to use heuristics to automatically choose a convolution implementation well suited to the parameters of the convolution.

The Theano flag dnn.conv.algo_fwd allows to specify the cuDNN convolution implementation that Theano should use for forward convolutions. Possible values include :

  • small (default) : use a convolution implementation with small memory usage
  • none : use a slower implementation with minimal memory usage
  • large : use a sometimes faster implementation with large memory usage
  • fft : use the Fast Fourier Transform implementation of convolution (very high memory usage)
  • guess_once : the first time a convolution is executed, the implementation to use is chosen according to cuDNN’s heuristics and reused for every subsequent execution of the convolution.
  • guess_on_shape_change : like guess_once but a new convolution implementation selected every time the shapes of the inputs and kernels don’t match the shapes from the last execution.
  • time_once : the first time a convolution is executed, every convolution implementation offered by cuDNN is executed and timed. The fastest is reused for every subsequent execution of the convolution.
  • time_on_shape_change : like time_once but a new convolution implementation selected every time the shapes of the inputs and kernels don’t match the shapes from the last execution.

The Theano flag dnn.conv.algo_bwd allows to specify the cuDNN convolution implementation that Theano should use for gradient convolutions. Possible values include :

  • none (default) : use the default non-deterministic convolution implementation
  • deterministic : use a slower but deterministic implementation
  • fft : use the Fast Fourier Transform implementation of convolution (very high memory usage)
  • guess_once : the first time a convolution is executed, the implementation to use is chosen according to cuDNN’s heuristics and reused for every subsequent execution of the convolution.
  • guess_on_shape_change : like guess_once but a new convolution implementation selected every time the shapes of the inputs and kernels don’t match the shapes from the last execution.
  • time_once : the first time a convolution is executed, every convolution implementation offered by cuDNN is executed and timed. The fastest is reused for every subsequent execution of the convolution.
  • time_on_shape_change : like time_once but a new convolution implementation selected every time the shapes of the inputs and kernels don’t match the shapes from the last execution.

guess_* and time_* flag values take into account the amount of available memory when selecting an implementation. This means that slower implementations might be selected if not enough memory is available for the faster implementations.

Note

Normally you should not call GPU Ops directly, but the CPU interface currently does not allow all options supported by cuDNN ops. So it is possible that you will need to call them manually.

Note

The documentation of CUDNN tells that, for the 2 following operations, the reproducibility is not guaranteed with the default implementation: cudnnConvolutionBackwardFilter and cudnnConvolutionBackwardData. Those correspond to the gradient wrt the weights and the gradient wrt the input of the convolution. They are also used sometimes in the forward pass, when they give a speed up.

The Theano flag dnn.conv.algo_bwd can be use to force the use of a slower but deterministic convolution implementation.

Note

There is a problem we do not understand yet when cudnn paths are used with symbolic links. So avoid using that.

Note

cudnn.so* must be readable and executable by everybody. cudnn.h must be readable by everybody.

List of Implemented Operations

class theano.gpuarray.dnn.DnnBase(files=None, c_func=None)[source]

Creates a handle for cudnn and pulls in the cudnn libraries and headers.

class theano.gpuarray.dnn.GpuDnnBatchNorm(mode='per-activation', running_averages=False, inplace_running_mean=False, inplace_running_var=False, inplace_output=False)[source]

Base Op for cuDNN Batch Normalization.

Parameters:
  • mode ({'per-activation', 'spatial'}) – Whether to normalize per activation (in this mode, bias and scale tensor dimensions are 1xCxHxW) or share normalization factors across spatial dimensions (in this mode, bias and scale tensor dimensions are 1xCx1x1).
  • epsilon – Epsilon value used in the batch normalization formula. Minimum allowed value is 1e-5 (imposed by cuDNN).
  • running_average_factor (float) – Factor for updating the values or running_mean and running_var. If the factor is close to one, the running averages will update quickly, if the factor is close to zero it will update slowly.
  • running_mean (tensor or None) – Previous value of the running mean. If this is given, the new value running_mean * (1 - r_a_factor) + batch mean * r_a_factor will be returned as one of the outputs of this function. running_mean and running_var should either both be given or both be None.
  • running_var (tensor or None) – Previous value of the running variance. If this is given, the new value running_var * (1 - r_a_factor) + (m / (m - 1)) * batch var * r_a_factor will be returned as one of the outputs of this function, where m is the product of lengths of the averaged-over dimensions. running_mean and running_var should either both be given or both be None.
class theano.gpuarray.dnn.GpuDnnBatchNormInference(mode='per-activation', inplace=False)[source]

Base Op for cuDNN Batch Normalization.

Parameters:
  • mode ({'per-activation', 'spatial'}) – Whether to normalize per activation (in this mode, bias and scale tensor dimensions are 1xCxHxW) or share normalization factors across spatial dimensions (in this mode, bias and scale tensor dimensions are 1xCx1x1).
  • epsilon – Epsilon value used in the batch normalization formula. Minimum allowed value is 1e-5 (imposed by cuDNN).
class theano.gpuarray.dnn.GpuDnnConv(algo=None, inplace=False)[source]

The forward convolution.

Parameters:
  • image –
  • kernel –
  • descr – The convolution descriptor.
  • algo ({'small', 'none', 'large', 'fft', 'fft_tiling', 'winograd', 'guess_once',) – ‘guess_on_shape_change’, ‘time_once’, ‘time_on_shape_change’} Default is the value of config.dnn.conv.algo_fwd.
static get_out_shape(ishape, kshape, border_mode, subsample)[source]

This function computes the output shape for a convolution with the specified parameters. ishape and kshape can be symbolic or scalar.

class theano.gpuarray.dnn.GpuDnnConvDesc(border_mode, subsample=(1, 1), conv_mode='conv', precision='float32')[source]

This Op builds a convolution descriptor for use in the other convolution operations.

See the doc of dnn_conv() for a description of the parameters

class theano.gpuarray.dnn.GpuDnnConvGradI(inplace=False, algo=None)[source]

The convolution gradient with respect to the inputs.

Parameters:
  • image –
  • kernel –
  • descr – The convolution descriptor.
  • algo ({'none', 'deterministic', 'fft', 'fft_tiling', 'winograd', 'guess_once',) – ‘guess_on_shape_change’, ‘time_once’, ‘time_on_shape_change’} Default is the value of config.dnn.conv.algo_bwd_data.
class theano.gpuarray.dnn.GpuDnnConvGradW(inplace=False, algo=None)[source]

The convolution gradient with respect to the weights.

Parameters:
  • image –
  • kernel –
  • descr – The convolution descriptor.
  • algo ({'none', 'deterministic', 'fft', 'small', 'guess_once',) – ‘guess_on_shape_change’, ‘time_once’, ‘time_on_shape_change’} Default is the value of config.dnn.conv.algo_bwd_filter.
class theano.gpuarray.dnn.GpuDnnPool(mode='max')[source]
Parameters:
  • img (tensor) – The image 4d or 5d tensor.
  • ws (tensor) – Window size.
  • stride (tensor) – (dx, dy) or (dx, dy, dz).
  • mode ({'max', 'average_inc_pad', 'average_exc_pad'}) – The old deprecated name ‘average’ corresponds to ‘average_inc_pad’.
  • pad (tensor) – (padX, padY) or (padX, padY, padZ)
class theano.gpuarray.dnn.GpuDnnPoolDesc(ws=(1, 1), stride=(1, 1), mode='max', pad=(0, 0))[source]

This Op builds a pooling descriptor for use in the other pooling operations.

ws, stride and pad must have the same length.

Parameters:
  • ws (tuple) – Window size.
  • stride (tuple) – (dx, dy) or (dx, dy, dz).
  • mode ({'max', 'average_inc_pad', 'average_exc_pad'}) – The old deprecated name ‘average’ corresponds to ‘average_inc_pad’.
  • pad (tuple) – (padX, padY) or (padX, padY, padZ)

Note

Not used anymore. Only needed to reload old pickled files.

class theano.gpuarray.dnn.GpuDnnPoolGrad(mode='max')[source]

The pooling gradient.

Parameters:
  • inp – The input of the pooling.
  • out – The output of the pooling in the forward.
  • out_grad – Same size as out, but is the corresponding gradient information.
  • ws (tensor variable) – Window size.
  • stride (tensor variable) – (dx, dy) or (dx, dy, dz).
  • mode ({'max', 'average_inc_pad', 'average_exc_pad'}) – The old deprecated name ‘average’ corresponds to ‘average_inc_pad’.
  • pad (tensor) – (padX, padY) or (padX, padY, padZ)
class theano.gpuarray.dnn.GpuDnnSoftmax(algo, mode)[source]

Op for the cuDNN Softmax.

algo
: {‘fast’, ‘accurate’, ‘log’}
Indicating whether, respectively, computations should be optimized for speed, for accuracy, or if cuDNN should rather compute the log-softmax instead.
mode
: {‘instance’, ‘channel’}
Indicating whether the softmax should be computed per image across ‘c01’ or per spatial location ‘01’ per image across ‘c’.
class theano.gpuarray.dnn.GpuDnnSoftmaxBase(algo, mode)[source]

Op for the cuDNN Softmax.

Parameters:
  • algo ({'fast', 'accurate', 'log'}) – Indicating whether, respectively, computations should be optimized for speed, for accuracy, or if cuDNN should rather compute the log-softmax instead.
  • mode ({'instance', 'channel'}) – Indicating whether the softmax should be computed per image across ‘c01’ or per spatial location ‘01’ per image across ‘c’.
class theano.gpuarray.dnn.GpuDnnSoftmaxGrad(algo, mode)[source]

Op for the cuDNN SoftmaxGrad.

Parameters:
  • algo – ‘fast’, ‘accurate’ or ‘log’ indicating whether, respectively, computations should be optimized for speed, for accuracy, or if cuDNN should rather compute the gradient of the log-softmax instead.
  • mode – ‘instance’ or ‘channel’ indicating whether the softmax should be computed per image across ‘c01’ or per spatial location ‘01’ per image across ‘c’.
class theano.gpuarray.dnn.RNNBlock(dtype, hidden_size, num_layers, rnn_mode, input_mode='linear', direction_mode='unidirectional', context_name=None)[source]

An object that allow us to use CuDNN v5 RNN implementation. TODO: make an example how to use. You can check Theano tests test_dnn_rnn_gru() and test_dnn_rnn_lstm() in the file theano/gpuarray/tests/test_dnn.py for now.

Parameters:
  • dtype (data type of computation) –
  • hidden_size (int) –
  • num_layers (int) –
  • rnn_mode ({'rnn_relu', 'rnn_tanh', 'lstm', 'gru'}) – See cudnn documentation for cudnnRNNMode_t.
  • input_mode ({'linear', 'skip'}) – linear: input will be multiplied by a biased matrix skip: No operation is performed on the input. The size must match the hidden size.
  • direction_mode ({'unidirectional', 'bidirectional'}) –
    unidirectional: The network operates recurrently from the
    first input to the last.

    bidirectional: The network operates from first to last then from last to first and concatenates the results at each layer.
theano.gpuarray.dnn.dnn_batch_normalization_test(inputs, gamma, beta, mean, var, mode='per-activation', epsilon=0.0001)[source]

Performs batch normalization of the given inputs, using the given mean and variance.

Parameters:
  • mode ({'per-activation', 'spatial'}) – Whether to normalize per activation or share normalization factors across spatial dimensions (i.e., all dimensions past the second).
  • gamma (tensor) – Scale factors. Must match the dimensionality of inputs, but have sizes of 1 for all axes normalized over (i.e., in the first dimension for mode='per-activation'`, and additionally in all dimensions past the second for ``mode='spatial').
  • beta (tensor) – Biases. Must match the tensor layout of gamma.
  • mean (tensor) – Means. Usually these are running averages computed during training. Must match the tensor layout of gamma.
  • var (tensor) – Variances. Usually these are running averages computed during training. Must match the tensor layout of gamma.
  • epsilon (float) – Epsilon value used in the batch normalization formula. Minimum allowed value is 1e-5 (imposed by cuDNN).
Returns:

out – Batch-normalized inputs.
Return type:

tensor

Notes

Requires cuDNN 5 and Theano 0.9dev2 or more recent.

For 4d tensors, the returned value is equivalent to:

axes = (0,) if mode == 'per-activation' else (0, 2, 3)
gamma, beta, mean, var = (T.addbroadcast(t, *axes)
                          for t in (gamma, beta, mean, var))
out = (inputs - mean) * gamma / T.sqrt(var + epsilon) + beta

For 5d tensors, the axes would be (0, 2, 3, 4).

theano.gpuarray.dnn.dnn_batch_normalization_train(inputs, gamma, beta, mode='per-activation', epsilon=0.0001, running_average_factor=0.1, running_mean=None, running_var=None)[source]

Performs batch normalization of the given inputs, using the mean and variance of the inputs.

Parameters:
  • mode ({'per-activation', 'spatial'}) – Whether to normalize per activation or share normalization factors across spatial dimensions (i.e., all dimensions past the second).
  • gamma (tensor) – Learnable scale factors. Must match the dimensionality of inputs, but have sizes of 1 for all axes normalized over (i.e., in the first dimension for mode='per-activation'`, and additionally in all dimensions past the second for ``mode='spatial').
  • beta (tensor) – Learnable biases. Must match the tensor layout of gamma.
  • epsilon (float) – Epsilon value used in the batch normalization formula. Minimum allowed value is 1e-5 (imposed by cuDNN).
  • running_average_factor (float) – Factor for updating the values or running_mean and running_var. If the factor is close to one, the running averages will update quickly, if the factor is close to zero it will update slowly.
  • running_mean (tensor or None) – Previous value of the running mean. If this is given, the new value running_mean * (1 - r_a_factor) + batch mean * r_a_factor will be returned as one of the outputs of this function. running_mean and running_var should either both be given or both be None.
  • running_var (tensor or None) – Previous value of the running variance. If this is given, the new value running_var * (1 - r_a_factor) + (m / (m - 1)) * batch var * r_a_factor will be returned as one of the outputs of this function, where m is the product of lengths of the averaged-over dimensions. running_mean and running_var should either both be given or both be None.
Returns:

  • out (tensor) – Batch-normalized inputs.
  • mean (tensor) – Means of inputs across the normalization axes.
  • invstd (tensor) – Inverse standard deviations of inputs across the normalization axes.
  • new_running_mean (tensor) – New value of the running mean (only if both running_mean and running_var were given).
  • new_running_var (tensor) – New value of the running variance (only if both running_var and running_mean were given).

Notes

Requires cuDNN 5 and Theano 0.9dev2 or more recent.

For 4d tensors, returned values are equivalent to:

axes = 0 if mode == 'per-activation' else (0, 2, 3)
mean = inputs.mean(axes, keepdims=True)
var = inputs.var(axes, keepdims=True)
invstd = T.inv(T.sqrt(var + epsilon))
out = (inputs - mean) * gamma * invstd + beta

m = T.cast(T.prod(inputs.shape) / T.prod(mean.shape), 'float32')
running_mean = running_mean * (1 - running_average_factor) + \
               mean * running_average_factor
running_var = running_var * (1 - running_average_factor) + \
              (m / (m - 1)) * var * running_average_factor

For 5d tensors, the axes are (0, 2, 3, 4).

theano.gpuarray.dnn.dnn_conv(img, kerns, border_mode='valid', subsample=(1, 1), conv_mode='conv', direction_hint=None, workmem=None, algo=None, precision=None)[source]

GPU convolution using cuDNN from NVIDIA.

The memory layout to use is ‘bc01’, that is ‘batch’, ‘channel’, ‘first dim’, ‘second dim’ in that order.

Parameters:
  • img – Images to do the convolution over.
  • kerns – Convolution filters.
  • border_mode – One of ‘valid’, ‘full’, ‘half’; additionally, the padding size could be directly specified by an integer or a pair of integers.
  • subsample – Perform subsampling of the output (default: (1, 1)).
  • conv_mode – Perform convolution (kernels flipped) or cross-correlation. One of ‘conv’, ‘cross’ (default: ‘conv’).
  • direction_hint – Used by graph optimizers to change algorithm choice. By default, GpuDnnConv will be used to carry out the convolution. If border_mode is ‘valid’, subsample is (1, 1) and direction_hint is ‘bprop weights’, it will use GpuDnnConvGradW. If border_mode is ‘full’, subsample is (1, 1) and direction_hint is not ‘forward!’, it will use GpuDnnConvGradI. This parameter is used internally by graph optimizers and may be removed at any time without a deprecation period. You have been warned.
  • algo ({'none', 'small', 'large', 'fft', 'guess_once', 'guess_on_shape_change', 'time_once', 'time_on_shape_change'}) – Convolution implementation to use. Some of its values may require certain versions of cuDNN to be installed. Default is the value of config.dnn.conv.algo_fwd.
  • precision ({'as_input_f32', 'as_input', 'float16', 'float32', 'float64'}) – Description of the dtype in which the computation of the convolution should be done. Possible values are ‘as_input’, ‘float16’, ‘float32’ and ‘float64’. Default is the value of config.dnn.conv.precision.

Warning

The cuDNN library only works with GPUs that have a compute capability of 3.0 or higer. This means that older GPUs will not work with this Op.

theano.gpuarray.dnn.dnn_conv3d(img, kerns, border_mode='valid', subsample=(1, 1, 1), conv_mode='conv', direction_hint=None, algo='none', precision=None)[source]

GPU convolution using cuDNN from NVIDIA.

The memory layout to use is ‘bc012’, that is ‘batch’, ‘channel’, ‘first dim’, ‘second dim’, ‘third dim’ in that order.

Parameters:
  • img – Images to do the convolution over.
  • kerns – Convolution filters.
  • border_mode – One of ‘valid’, ‘full’, ‘half’; additionally, the padding size could be directly specified by an integer or a pair of integers.
  • subsample – Perform subsampling of the output (default: (1, 1)).
  • conv_mode – Perform convolution (kernels flipped) or cross-correlation. One of ‘conv’, ‘cross’ (default: ‘conv’).
  • direction_hint – Used by graph optimizers to change algorithm choice. By default, GpuDnnConv will be used to carry out the convolution. If border_mode is ‘valid’, subsample is (1, 1) and direction_hint is ‘bprop weights’, it will use GpuDnnConvGradW. If border_mode is ‘full’, subsample is (1, 1) and direction_hint is not ‘forward!’, it will use GpuDnnConvGradI. This parameter is used internally by graph optimizers and may be removed at any time without a deprecation period. You have been warned.
  • algo (convolution implementation to use. Only 'none' is implemented) – for the conv3d. Default is the value of config.dnn.conv.algo_fwd.
  • precision ({'as_input_f32', 'as_input', 'float16', 'float32', 'float64'}) – Description of the dtype in which the computation of the convolution should be done. Possible values are ‘as_input’, ‘float16’, ‘float32’ and ‘float64’. Default is the value of config.dnn.conv.precision.

Warning

The cuDNN library only works with GPUs that have a compute capability of 3.0 or higer. This means that older GPUs will not work with this Op.

theano.gpuarray.dnn.dnn_gradinput(kerns, topgrad, img_shp, border_mode='valid', subsample=(1, 1), conv_mode='conv', precision=None)[source]

TODO: document this

theano.gpuarray.dnn.dnn_gradinput3d(kerns, topgrad, img_shp, border_mode='valid', subsample=(1, 1, 1), conv_mode='conv', precision=None)[source]

3d version of dnn_gradinput.

theano.gpuarray.dnn.dnn_gradweight(img, topgrad, kerns_shp, border_mode='valid', subsample=(1, 1), conv_mode='conv', precision=None)[source]

TODO: document this

theano.gpuarray.dnn.dnn_gradweight3d(img, topgrad, kerns_shp, border_mode='valid', subsample=(1, 1, 1), conv_mode='conv', precision=None)[source]

3d version of dnn_gradweight

theano.gpuarray.dnn.dnn_pool(img, ws, stride=None, mode='max', pad=None)[source]

GPU pooling using cuDNN from NVIDIA.

The memory layout to use is ‘bc01’, that is ‘batch’, ‘channel’, ‘first dim’, ‘second dim’ in that order.

ws, stride and pad must have the same length.

Parameters:
  • img – Images to do the pooling over.
  • ws (tuple) – Subsampling window size. Should have 2 or 3 elements.
  • stride (tuple) – Subsampling stride (default: (1, 1) or (1, 1, 1)).
  • mode ({'max', 'average_inc_pad', 'average_exc_pad', 'sum'}) –
  • pad (tuple) – (padX, padY) or (padX, padY, padZ) default: (0, 0) or (0, 0, 0)

Warning

The cuDNN library only works with GPU that have a compute capability of 3.0 or higer. This means that older GPU will not work with this Op.

Notes

This Op implements the ignore_border=True of max_pool_2d.

theano.gpuarray.dnn.version(raises=True)[source]

Return the current cuDNN version we link with.

This also does a check that the header version matches the runtime version.

Raises:If True, raise an exception if cuDNN is not present or badly installed. Otherwise, return -1.