class GPU(Backend):
"""
Sets up a NervanaGPU based backend for matrix operations.
Note that some functions defined in the generic Backend class such as are
cross-map pooling and normalization and adaDelta are not implemented for
this backend.
"""
default_dtype = np.float32
def __init__(self, rng_seed, stochastic_round=False, device_id=0):
self.ng = NervanaGPU(stochastic_round=stochastic_round)
logger.info("Initialized NervanaGPU with stochastic_round=%s",
stochastic_round)
self.rng_seed = rng_seed
self.rng_init()
self.device_id = device_id if device_id is not None else 0
def __getstate__(self):
"""
Defines what and how we go about serializing an instance of this class.
Returns:
self.__dict__: The full contents of the backend class instance,
except for the mem_pool which is on device and
cannot be serialized.
"""
if hasattr(self, 'mem_pool') and self.mem_pool is not None:
self.mem_pool_pickle = {'shape': self.mem_pool.shape,
'dtype': np.float32}
self.mem_pool = None
return self.__dict__
def __setstate__(self, state):
"""
Defines how we go about deserializing into an instance of this class.
Arguments:
self.__dict__: The full contents of the backend class instance,
except for the mem_pool which is on device and
cannot be serialized.
"""
self.__dict__.update(state)
self.mem_pool = self.ng.empty(self.mem_pool_pickle['shape'],
dtype=self.mem_pool_pickle['dtype'])
def init_mempool(self, shape, dtype=default_dtype):
"""
Allocates a memory pool for temporary storage
"""
self.mem_pool = self.ng.empty(shape, dtype=dtype)
def alloc_host_mem(self, shape, dtype):
return drv.pagelocked_empty(shape, dtype, order="C", mem_flags=0)
def create_stream(self):
return drv.Stream()
def async_copy(self, dest, src, stream=None):
drv.memcpy_htod_async(dest.gpudata, src, stream)
def rng_init(self):
"""
Initialize and seed the pseudo random number genrator. Random numbers
are generated on the host using numpy, then transfered to device.
"""
seed = None
if 'rng_seed' in self.__dict__:
seed = self.rng_seed
logger.info("Seeding random number generator with: %s", str(seed))
np.random.seed(seed)
def flop_timing_init(self, decorate_fc, decorate_conv, decorate_ew):
"""
Initialize FLOP timing. Wraps the specified MOP calls via a decorator
to record elapsed time and number of operations.
Arguments:
decorate_fc (list): string giving the function names of fully
connected layer forward/backward/update calls
to time.
decorate_conv (list): string giving the function names of
convolutional layer forward/backward/update
calls to time.
decorate_ew (list): string giving the function names of element-wise
calls to time.
Notes:
Must be called prior to first flop_timing_start call
"""
self.start = drv.Event()
self.end = drv.Event()
self.flop_timer = FlopsDecorator(self)
self.flop_timer.decorate(decorate_fc=decorate_fc,
decorate_conv=decorate_conv,
decorate_ew=decorate_ew)
def flop_timinig_start(self):
"""
Start a new FLOP timer.
Returns:
None: dummy value (not used)
"""
return self.start.record()
def flop_timing_finish(self, start_time):
"""
Complete current FLOP timing.
Arguments:
start_time (unused): ignored.
Returns:
float: elapsed time in seconds since prior flop_timing_start call.
"""
self.end.record()
self.end.synchronize()
return self.end.time_since(self.start)
def uniform(self, low=0.0, high=1.0, shape=1, dtype=default_dtype,
persist_values=True, name=None, allocator=drv.mem_alloc):
"""
generate numpy random number and convert to a GPUTensor.
If called with dype=None it will probably explode
"""
ary = np.random.uniform(low, high, shape)
return GPUTensor(ary.shape, dtype, allocator=allocator, name=name,
rounding=self.ng.round_mode).set(ary)
def normal(self, loc=0.0, scale=1.0, size=1, dtype=default_dtype,
persist_values=True, name=None, allocator=drv.mem_alloc):
"""
Gaussian/Normal random number sample generation
"""
ary = np.random.normal(loc, scale, size)
return GPUTensor(ary.shape, dtype, allocator=allocator, name=name,
rounding=self.ng.round_mode).set(ary)
def fprop_fc(self, out, inputs, weights, layer=None):
"""
Forward propagate the inputs of a fully connected network layer to
produce output pre-activations (ready for transformation by an
activation function).
Arguments:
out (GPUTensor): Where to store the forward propagated results.
inputs (GPUTensor): Will be either the dataset input values (first
layer), or the outputs from the previous layer.
weights (GPUTensor): The weight coefficient values for this layer.
layer (Layer): The layer object.
"""
self.ng.dot(weights, inputs, out)
def bprop_fc(self, out, weights, deltas, layer=None):
"""
Backward propagate the error through a fully connected network layer.
Arguments:
out (GPUTensor): Where to store the backward propagated errors.
weights (GPUTensor): The weight coefficient values for this layer.
deltas (GPUTensor): The error values for this layer
layer (Layer): The layer object.
"""
self.ng.dot(weights.T, deltas, out)
def update_fc(self, out, inputs, deltas, layer=None):
"""
Compute the updated gradient for a fully connected network layer.
Arguments:
out (GPUTensor): Where to store the updated gradient value.
inputs (GPUTensor): Will be either the dataset input values (first
layer), or the outputs from the previous layer.
deltas (GPUTensor): The error values for this layer
layer (Layer): The layer object.
"""
self.ng.dot(deltas, inputs.T, out)
def fprop_conv(self, out, inputs, weights, ofmshape, ofmsize, ofmlocs,
ifmshape, links, nifm, padding, stride, ngroups, fpropbuf,
local=False):
"""
Forward propagate the inputs of a convolutional network layer to
produce output pre-activations (ready for transformation by an
activation function).
Arguments:
out (GPUTensor): Where to store the forward propagated results.
inputs (GPUTensor): Will be either the dataset input values (first
layer), or the outputs from the previous layer.
weights (GPUTensor): The weight coefficient values for this layer.
ofmshape (tuple): Dimensions of each output feature map (typically
number of height and width neurons).
ofmsize (int): Total size of each output feature map.
ofmlocs (GPUTensor): Indices giving the location of each element
in each output feature map stored in out.
ifmshape (tuple): Dimensions of each input feature map (typically
number of height and width neurons). For this
backend we expect these values to be square.
links (GPUTensor): Input receptive field indices.
nifm (int): Total number of input feature maps.
padding (int): Number of additional elements to include along each
dimension of each local receptive field during the
convolution operation.
stride (int): Number of neurons to shift the filter at each step.
ngroups (int): Number of groups.
fpropbuf (GPUTensor): Temporary storage buffer used to hold the
convolved outputs for a single receptive
field. Not used for this backend.
local (bool, optional): Whether to do local filtering (True) or
convolution (False, the default)
"""
'''
N: Number of images in mini-batch
C: Number of input feature maps
K: Number of output feature maps
D: Depth of input image
H: Height of input image
W: Width of input image
T: Depth of filter kernel
R: Height of filter kernel
S: Width of filter kernel
'''
self.ng.fprop_conv(layer=fpropbuf, I=inputs, F=weights, O=out,
alpha=1.0, repeat=1)
def bprop_conv(self, out, weights, deltas, ofmshape, ofmsize, ofmlocs,
ifmshape, links, padding, stride, nifm, ngroups, bpropbuf,
local=False):
"""
Backward propagate the error through a convolutional network layer.
Arguments:
out (GPUTensor): Where to store the backward propagated errors.
weights (GPUTensor): The weight coefficient values for this layer.
deltas (GPUTensor): The error values for this layer
ofmshape (tuple): Dimensions of each output feature map (typically
height and width).
ofmsize (int): Total size of each output feature map.
ofmlocs (GPUTensor): Indices giving the location of each element in
each output feature map stored in out.
ifmshape (tuple): Dimensions of each input feature map (typically
height and width).
links (GPUTensor): Input receptive field indices.
nifm (int): Total number of input feature maps.
padding (int): Number of additional elements to include along each
dimension of each local receptive field during the
convolution operation.
stride (int): Number of neurons to shift the filter at each step.
ngroups (int): Number of groups.
bpropbuf (GPUTensor): Temporary storage buffer used to hold the
backpropagated error for a single receptive
field
local (bool, optional): Whether to do local filtering (True) or
convolution (False, the default)
"""
self.ng.bprop_conv(layer=bpropbuf, F=weights, E=deltas, grad_I=out,
alpha=1.0, repeat=1)
def update_conv(self, out, inputs, weights, deltas, ofmshape, ofmsize,
ofmlocs, ifmshape, links, nifm, padding, stride, ngroups,
fwidth, updatebuf, local=False, layer=None):
"""
Compute the updated gradient for a convolutional network layer.
Arguments:
out (GPUTensor): Where to store the updated gradient value.
inputs (GPUTensor): Will be either the dataset input values (first
layer), or the outputs from the previous layer.
weights (GPUTensor): The weight coefficient values for this layer.
deltas (GPUTensor): The error values for this layer
ofmshape (tuple): Dimensions of each output feature map (typically
height and width).
ofmsize (int): Total size of each output feature map.
ofmlocs (GPUTensor): Indices giving the location of each element in
each output feature map stored in out.
ifmshape (tuple): Dimensions of each input feature map (typically
height and width).
links (GPUTensor): Input receptive field indices.
nifm (int): Total number of input feature maps.
padding (int): Number of additional elements to include along each
dimension of each local receptive field during the
convolution operation.
stride (int): Number of neurons to shift the filter at each step.
ngroups (int): Number of groups.
fwidth (int): Filter width.
updatebuf (GPUTensor): Temporary storage buffer used to hold the
updated gradient for a single receptive
field
local (bool, optional): Whether to do local filtering (True) or
convolution (False, the default)
layer (Layer): The layer object.
"""
self.ng.update_conv(layer=updatebuf, I=inputs, E=deltas, grad_F=out,
alpha=1.0, repeat=1)
def fprop_pool(self, out, inputs, op, ofmshape, ofmsize, ofmlocs, fshape,
ifmshape, links, nifm, padding, stride, fpropbuf):
"""
Forward propagate the inputs of a Pooling network layer to
produce output pre-activations (ready for transformation by an
activation function).
Arguments:
out (GPUTensor): Where to store the forward propagated results.
inputs (GPUTensor): Will be either the dataset input values (first
layer), or the outputs from the previous layer.
op (string): The type of pooling operation to apply. We support
"max", "avg", "l2" currently.
ofmshape (tuple): Dimensions of each output feature map (typically
number of height and width neurons).
ofmsize (int): Total size of each output feature map.
ofmlocs (GPUTensor): Indices giving the location of each element in
each output feature map stored in out.
fshape (tuple): Dimensions of each filter (typically height and
width).
ifmshape (tuple): Dimensions of each input feature map (typically
number of height and width neurons).
links (GPUTensor): Input receptive field indices.
nifm (int): Total number of input feature maps.
padding (int): Number of additional elements to include along each
dimension of each local receptive field during the
pooling operation.
stride (int): Number of neurons to shift the filter at each step.
fpropbuf (GPUTensor): Temporary storage buffer used to hold the
pooled outputs for a single receptive field.
"""
op = op.lower()
if op == "max":
self.ng.fprop_pool(layer=fpropbuf, I=inputs, O=out, repeat=1)
else:
raise AttributeError("unexpected pooling op type: %s", op)
def bprop_pool(self, out, fouts, inputs, deltas, op, ofmshape, ofmsize,
ofmlocs, fshape, fpsize, ifmshape, links, nifm, padding,
stride, bpropbuf):
"""
Backward propagate the error through a pooling network layer.
Arguments:
out (GPUTensor): Where to store the backward propagated errors.
fouts (GPUTensor): Forward propagated outputs from the previous
layer.
inputs (GPUTensor): Will be either the dataset input values (first
layer), or the outputs from the previous layer.
deltas (GPUTensor): The error values for this layer
op (string): The type of pooling operation to apply. We support
"max", "avg", "l2" currently.
ofmshape (tuple): Dimensions of each output feature map (typically
height and width).
ofmsize (int): Total size of each output feature map.
ofmlocs (GPUTensor): Indices giving the location of each element in
each output feature map stored in out.
fshape (tuple): Dimensions of each filter (typically height and
width).
fpsize (int): The size of each filter.
ifmshape (tuple): Dimensions of each input feature map (typically
height and width).
links (GPUTensor): Input receptive field indices.
nifm (int): Total number of input feature maps.
padding (int): Number of additional elements to include along each
dimension of each local receptive field during the
pooling operation.
stride (int): Number of neurons to shift the filter at each step.
bpropbuf (GPUTensor): Temporary storage buffer used to hold the
backpropagated error for a single receptive
field
"""
op = op.lower()
if op == "max":
self.ng.bprop_pool(layer=bpropbuf, I=inputs, E=deltas, grad_I=out,
repeat=1)
else:
raise AttributeError("unexpected pooling op type: %s", op)
def logistic(self, x, out):
"""
Logistic sigmoid nonlinearity, 1/(1+exp(-x))
Arguments:
x (GPUTensor): Input tensor
out (GPUTensor): Output tensor
"""
self.ng.sig(x, out=out)
return out
def rectlin(self, x, out):
"""
Rectified Linear nonlinearity
Arguments:
x (GPUTensor): Input tensor
out (GPUTensor): Output tensor
"""
self.ng.maximum(x, 0., out=out)
return out
def rectleaky(self, x, slope, out):
out[:] = self.ng.maximum(x, x*slope)
def rectleaky_derivative(self, x, slope, out):
out[:] = self.ng.greater(x, 0) * (1.0 - slope) + slope
def sum(self, tsr, axes, out):
"""
Sum
Arguments:
tsr (GPUTensor): Input tensor
axes (int): Axis along which the reduction is performed. If axes
is None, the tensor is flattened and reduced over
both dimensions.
out (GPUTensor): Output tensor
"""
if axes is None:
sze = tsr.shape[0]*tsr.shape[1]
self.ng.sum(tsr.reshape(sze, 1), axis=0, out=out)
else:
self.ng.sum(tsr, axis=axes, out=out)
return out
def mean(self, tsr, axes, out):
"""
Calculates the arithmetic mean of the elements along the specified
axes.
Arguments:
tsr (GPUTensor): Input tensor
axes (int): Axis along which the reduction is performed. If axes
is None, the tensor is flattened and reduced over
both dimensions.
out (GPUTensor): Output tensor
"""
if axes is None:
sze = tsr.shape[0]*tsr.shape[1]
self.ng.mean(tsr.reshape(sze, 1), axis=0, out=out)
else:
self.ng.mean(tsr, axis=axes, out=out)
return out
def min(self, tsr, axes, out):
"""
Calculates the minimum of the elements along the specified
axes.
Arguments:
tsr (GPUTensor): Input tensor
axes (int): Axis along which the reduction is performed. If axes
is None, the tensor is flattened and reduced over
both dimensions.
out (GPUTensor): Output tensor
"""
if axes is None:
sze = tsr.shape[0]*tsr.shape[1]
self.ng.min(tsr.reshape(sze, 1), axis=0, out=out)
else:
self.ng.min(tsr, axis=axes, out=out)
return out
def max(self, tsr, axes, out):
"""
Calculates the maximum of the elements along the specified
axes.
Arguments:
tsr (GPUTensor): Input tensor
axes (int): Axis along which the reduction is performed. If axes
is None, the tensor is flattened and reduced over
both dimensions.
out (GPUTensor): Output tensor
"""
if axes is None:
sze = tsr.shape[0]*tsr.shape[1]
self.ng.max(tsr.reshape(sze, 1), axis=0, out=out)
else:
self.ng.max(tsr, axis=axes, out=out)
return out
def variance(self, tsr, axes, out, mean=None):
"""
Calculates the variance of the elements along the specified
axes.
Arguments:
tsr (GPUTensor): the tensor on which to compute the variance
axes (int, list, optional): the dimension(s) along which to
variance. If set to None, we will
variance over all dimensions.
out (GPUTensor): where the result will be stored.
mean (GPUTensor): the tensor containing mean of tsr
Returns:
GPUTensor: reference to out
"""
if mean is None:
logger.error("GPUTensor requires mean to be specified.")
raise ValueError("mean not specified")
self.ng.mean(self.ng.square(tsr-mean), axis=axes, out=out)
return out
def fabs(self, x, out):
"""
Calculates absolute value of the elements in a tensor
Arguments:
x (GPUTensor): Input tensor
out (GPUTensor): Output tensor
Returns:
GPUTensor: reference to out
"""
self.ng.fabs(x, out=out)
return out
def sqrt(self, x, out):
"""
Calculates square root of the elements in a tensor
Arguments:
x (GPUTensor): Input tensor
out (GPUTensor): Output tensor
Returns:
GPUTensor: reference to out
"""
self.ng.sqrt(x, out=out)
return out
def zeros(self, shape, dtype=default_dtype, persist_values=True):
"""
Allocate a new GPUTensor and fill it with zeros.
Arguments:
shape (tupel): Shape of the desired GPUTensor
dtype (dtype): Optional datatype
persist_values (bool, optional): If set to True (the default), the
values assigned to this Tensor
will persist across multiple begin
and end calls. Setting to False
may provide a performance increase
if values do not need to be
maintained across such calls
Returns:
GPUTensor: output
"""
return self.ng.zeros(shape, dtype=dtype)
def ones(self, shape, dtype=default_dtype, persist_values=True):
"""
Allocate a new GPUTensor and fill it with ones.
Arguments:
shape (tupel): Shape of the desired GPUTensor
dtype (dtype): Optional datatype
persist_values (bool, optional): If set to True (the default), the
values assigned to this Tensor
will persist across multiple begin
and end calls. Setting to False
may provide a performance increase
if values do not need to be
maintained across such calls
Returns:
GPUTensor: output
"""
return self.ng.ones(shape, dtype=dtype)
def empty(self, shape, dtype=default_dtype, persist_values=True):
"""
Allocate a new GPUTensor.
Arguments:
shape (tupel): Shape of the desired GPUTensor
dtype (dtype): Optional datatype
persist_values (bool, optional): If set to True (the default), the
values assigned to this Tensor
will persist across multiple begin
and end calls. Setting to False
may provide a performance increase
if values do not need to be
maintained across such calls
Returns:
GPUTensor: output
"""
return self.ng.empty(shape, dtype=dtype)
def array(self, ary, dtype=default_dtype, persist_values=True, name=None,
allocator=drv.mem_alloc):
"""
Allocate a new GPUTensor and fill it with supplied numpy array.
Arguments:
ary (ndarray): Numpy array with source data
dtype (dtype, optional): Optional datatype
persist_values (bool, optional): If set to True (the default), the
values assigned to this Tensor
will persist across multiple begin
and end calls. Setting to False
may provide a performance increase
if values do not need to be
maintained across such calls
name (string): Name for the GPUTensor
allocator (pycuda): Pycuda memory allocator
Returns:
GPUTensor: output
"""
return GPUTensor(ary.shape, dtype, allocator=allocator, name=name,
rounding=self.ng.round_mode).set(ary)
def add(self, left, right, out):
"""
Elementwise addition
Arguments:
left (GPUTensor, numeric): left-hand side operand.
right (GPUTensor, numeric): right-hand side operand.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
self.ng.add(left, right, out=out)
return out
def subtract(self, left, right, out):
"""
Elementwise subtraction
Arguments:
left (GPUTensor, numeric): left-hand side operand.
right (GPUTensor, numeric): right-hand side operand.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
self.ng.subtract(left, right, out=out)
return out
def multiply(self, left, right, out):
"""
Elementwise multiplication
Arguments:
left (GPUTensor, numeric): left-hand side operand.
right (GPUTensor, numeric): right-hand side operand.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
self.ng.multiply(left, right, out=out)
return out
def divide(self, left, right, out):
"""
Elementwise division
Arguments:
left (GPUTensor, numeric): left-hand side operand.
right (GPUTensor, numeric): right-hand side operand.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
self.ng.divide(left, right, out=out)
return out
def greater(self, left, right, out):
"""
Elementwise greater than testing
Arguments:
left (GPUTensor, numeric): left-hand side operand.
right (GPUTensor, numeric): right-hand side operand.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
self.ng.greater(left, right, out=out)
return out
def equal(self, left, right, out):
"""
Performs element-wise equality testing on each element of left and
right, storing the result in out. Each operand is assumed to be the
same shape (or broadcastable as such).
Arguments:
left (GPUTensor, numeric): left-hand side operand.
right (GPUTensor, numeric): right-hand side operand.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
self.ng.equal(left, right, out=out)
return out
def not_equal(self, left, right, out):
"""
Elementwise not equal testing
Arguments:
left (GPUTensor, numeric): left-hand side operand.
right (GPUTensor, numeric): right-hand side operand.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
self.ng.not_equal(left, right, out=out)
return out
def clip(self, a, a_min, a_max, out):
"""
Elementwise clipping between a range of specified values
Arguments:
a (GPUTensor): input tensor.
a_min (float): floor value.
a_max (float): ceiling value.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
self.ng.clip(a, a_min, a_max, out=out)
return out
def log(self, a, out):
"""
Elementwise base-e logarithm
Arguments:
a (GPUTensor): input tensor.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
self.ng.log(a, out=out)
return out
def tanh(self, a, out):
"""
Elementwise tanh
Arguments:
a (GPUTensor): input tensor.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
self.ng.tanh(a, out=out)
return out
def argmax(self, a, out, axis=0):
"""
Calculates the indices of the maximal element value along the specified
axis. If multiple elements contain the maximum, only the elements of
the first are returned.
Arguments:
tsr (GPUTensor): The GPUTensor on which to find the maximum indices
axis (int): The dimension along which to find the maximum. If set
to None, find the overall maximum index of a flattened
representation of tsr.
out (GPUTensor): Where to store the result. Should be of the
appropriate type and expected shape
Returns:
GPUTensor: reference to out
"""
self.ng.argmax(a, out=out, axis=axis)
return out
def softmax(self, x, out):
"""
Softmax nonlinearity. Computes exp(x-max(x)) / sum_i exp(x_i-max(x_i))
Arguments:
x (GPUTensor): input tensor.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
out[:] = (self.ng.reciprocal(self.ng.sum(
self.ng.exp(x - self.ng.max(x, axis=0)), axis=0)) *
self.ng.exp(x - self.ng.max(x, axis=0)))
return out
def softmax_gradient(self, y, err, out):
"""
Gradient of the softmax nonlinearity.
Arguments:
y (GPUTensor): input tensor.
err (GPUTensor): backpropagated error.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
raise NotImplementedError("Softmax gradient should use shortcut")
return out
def make_binary_mask(self, tsr, keepthresh=0.5, dtype=default_dtype):
"""
Create a binary mask for dropout layers.
Arguments:
tsr (GPUTensor): Output tensor
keepthresh (float): fraction of ones
"""
self.ng.dropout(keep=keepthresh, out=tsr)
def gdm_compound(self, ps_item, us_item, vs_item, momentum_coef,
learning_rate, epoch):
"""
Perform gradient descent update with momentum.
Arguments:
ps_item (GPUTensor): parameter tensor (e.g. a weight matrix)
us_item (GPUTensor): update tensor, contains gradient wrt. weights
vs_item (GPUTensor): velocity tensor.
momentum_coef (float): momentum coefficient.
learning_rate (float): learning rate.
epoch (int): epoch (used in conjunction with diagnostics).
Outputs are written to vs_item (updated velocity)
and ps_item (updated weights)
"""
vs_item[:] = vs_item * momentum_coef - us_item * learning_rate
ps_item[:] = ps_item + vs_item
def gdmwd_compound(self, ps_item, us_item, vs_item, momentum_coef,
learning_rate, wd, epoch):
"""
Perform gradient descent update with momentum and weight decay.
Arguments:
ps_item (GPUTensor): parameter tensor (e.g. a weight matrix)
us_item (GPUTensor): update tensor, contains gradient wrt. weights
vs_item (GPUTensor): velocity tensor.
momentum_coef (float): momentum coefficient.
learning_rate (float): learning rate.
wd (float): weight decay parameter.
epoch (int): epoch (used in conjunction with diagnostics).
Outputs:
ps_item, the updated weights.
vs_item, the updated velocity.
us_item, used as a temp buffer.
"""
vs_item[:] = vs_item * momentum_coef - us_item * \
learning_rate - learning_rate * wd * ps_item
ps_item[:] = ps_item + vs_item
def ada_update(self, ps_item, us_item, gs_item, ds_item, ls_item, ss_item,
rho, epsilon):
"""
Update rule for AdaDelta (Zeiler, http://arxiv.org/abs/1212.5701)
Arguments:
ps_item: weight / parameter (will be updated)
us_item: update
gs_item: expected value of Gradient Squared (will be updated)
ds_item: expected value of Delta Squared (will be updated)
ls_item: learning rate (will be updated)
ss_item: Scratch Space
rho: decay constant (determines window size)
epsilon: small positive constant for numerical stability
"""
# Accumulate E[Grad^2]
gs_item[:] = gs_item * rho + (1.0 - rho) * us_item * us_item
# Calculate Updates
ls_item[:] = self.ng.sqrt((ds_item + epsilon) /
(gs_item + epsilon)) * (-1.0) * us_item
# Accumulate E[Delt^2]
ds_item[:] = ds_item * rho + (1.0 - rho) * ls_item * ls_item
# Final update to the params
ps_item[:] = ps_item + ls_item
def rms_update(self, params, updates, run_squares, velocity, scratch_space,
gamma, epsilon, learning_rate, momentum_coef):
# Update running squares
run_squares[:] = gamma * run_squares + (1. - gamma) * updates * updates
# Now scale the gradient by lr / rms(grad) (with a epsilon term for
# stability) and use it to update the params
if momentum_coef == 0:
params[:] = params - learning_rate * updates * self.ng.reciprocal(
self.ng.sqrt(run_squares) + epsilon)
else:
velocity[:] = velocity * momentum_coef - \
learning_rate * updates * \
self.ng.reciprocal(self.ng.sqrt(run_squares) + epsilon)
params[:] = params + velocity
def fprop_bn_compound(self, inputs, beta, gamma, eps, xvar, xhat, out):
"""
Batch normalization forward pass, compounded to run in 3 kernel calls.
Arguments:
inputs: input data to be normalized
beta: location parameter
gamma: scale parameter
eps: small constant for numerical stability
xvar: variance (updated)
xhat: normalized input (updated)
out: normalized and rescaled input (updated)
"""
xvar[:] = self.ng.reciprocal(self.ng.sqrt(self.ng.var(inputs, axis=1) +
eps))
xhat[:] = xvar * (inputs - self.ng.mean(inputs, axis=1))
out[:] = xhat * gamma + beta
return out
def bprop_bn_compound(self, xhat, error, xvar, gamma,
beta_updates, gamma_updates):
"""
Batch normalization backward pass, compounded to run with 4 kernel
calls.
Arguments:
xhat: normalized input data (updated)
error: backpropagated deltas (updated)
xvar: precomputed variance
gamma: scale parameter
beta_updates: gradient update for beta (updated)
gamma_updates: gradient update for gamma (updated)
"""
gamma_updates[:] = self.ng.sum(xhat * error, axis=1)
beta_updates[:] = self.ng.sum(error, axis=1)
xhat[:] = (xhat * gamma_updates + beta_updates) / float(xhat.shape[1])
error[:] = xvar * gamma * (error - xhat)
# otherwise copy
else:
devA2 = ng.empty(dimA, dtype=np.float32)
devB2 = ng.empty(dimB, dtype=np.float32)
devA2[:] = devA1
devB2[:] = devB1
devC2 = ng.empty(dimC, dtype=np.float32)
devC2[:] = devC1
if op[0] == 't': devA1, devA2 = devA1.T, devA2.T
if op[1] == 't': devB1, devB2 = devB1.T, devB2.T
ng.dot(devA1,
devB1,
devC1,
alpha=alpha,
beta=beta,
repeat=repeat)
cublas_dot(devA2,
devB2,
devC2,
alpha=alpha,
beta=beta,
repeat=repeat)
partial1 = ng.empty((devC1.shape[0], 1), dtype=np.float32)
partial2 = partial1[0:1, 0:1]
diff = ng.max(abs(devC2 - devC1),
partial=partial1,
cpuA = np.random.uniform(-1.0, 1.0,
dimA).astype(np.float32)
cpuB = np.random.uniform(-1.0, 1.0,
dimB).astype(np.float32)
#cpuB = np.identity(n, dtype=dtype)
devA = ng.array(cpuA, dtype=dtype)
devB = ng.array(cpuB, dtype=dtype)
devC = ng.empty(dimC, dtype=dtype)
#repeat = min(int(50.0 * 4096**3 / (m * n * k)), 1000)
if op[0] == 't': cpuA, devA = cpuA.T, devA.T
if op[1] == 't': cpuB, devB = cpuB.T, devB.T
ng.dot(devA, devB, devC, repeat=1)
#context.synchronize()
cpuC = np.dot(cpuA, cpuB)
cpuD = devC.get()
diff = np.absolute(cpuC - cpuD)
max_diff = diff.max()
print(max_diff, cpuD.max())
if max_diff > 0.1 or max_diff != max_diff:
#print(m, n, k, max_diff)
print(cpuD[::max(m // 16, 1), ::max(n // 16, 1)])
print(cpuC[::max(m // 16, 1), ::max(n // 16, 1)])
print(diff[::max(m // 16, 1), ::max(n // 16, 1)])
exit()
devA2 = ng.empty(dimA, dtype=np.float32)
devB2 = ng.empty(dimB, dtype=np.float32)
devA2[:] = devA1
devB2[:] = devB1
devC2 = ng.empty(dimC, dtype=np.float32)
if op[0] == 't': devA1, devA2 = devA1.T, devA2.T
if op[1] == 't': devB1, devB2 = devB1.T, devB2.T
glops16 = 0
glops32 = 0
glops64 = 0
if op == "tn" and dtype is np.float16:
# Experimental 128x16 gemm kernel
glops16 = ng.dot(devA1, devB1, devC1, repeat=repeat, size=16)
if op != 'nt':
glops32 = ng.dot(devA1, devB1, devC1, repeat=repeat, size=32)
glops64 = ng.dot(devA1, devB1, devC1, repeat=repeat, size=64)
glops128 = ng.dot(devA1, devB1, devC1, repeat=repeat, size=128)
glops = max(glops16, glops32, glops64, glops128)
if glops16 == glops:
fastest = 16
elif glops32 == glops:
fastest = 32
elif glops64 == glops:
fastest = 64
else:
fastest = 128
if data_type == "All Ones":
cpuA = np.ones(dimA, dtype=dtype).astype(np.float32)
cpuB = np.ones(dimB, dtype=dtype).astype(np.float32)
#cpuB = np.identity(n, dtype=np.float32)
else:
cpuA = np.random.uniform(-1.0, 1.0, dimA).astype(np.float32)
cpuB = np.random.uniform(-1.0, 1.0, dimB).astype(np.float32)
devA = ng.array(cpuA, dtype=dtype)
devB = ng.array(cpuB, dtype=dtype)
devC = ng.empty(dimC, dtype=dtype)
if op[0] == 't': cpuA, devA = cpuA.T, devA.T
if op[1] == 't': cpuB, devB = cpuB.T, devB.T
ng.dot(devA, devB, devC, repeat=repeat)
if cpu:
cpuC = np.dot(cpuA, cpuB)
cpuD = devC.get()
diff = np.absolute(cpuC - cpuD)
print diff.max()
print cpuD[::max(m//4,1),::max(n//4,1)]
print cpuC[::max(m//4,1),::max(n//4,1)]
print diff[::max(m//4,1),::max(n//4,1)]
# print cpuD
# exit()
devA2[:] = devA1
devB2[:] = devB1
devC2 = ng.empty(dimC, dtype=np.float32)
# devC2 = devC2s.share(dimC, dtype=np.float32)
devC2[:] = devC1
if op[0] == 't': devA1, devA2 = devA1.T, devA2.T
if op[1] == 't': devB1, devB2 = devB1.T, devB2.T
for tile in (32,64,128):
if op == 'nt' and tile != 128:
continue
try:
ng.dot(devA1, devB1, devC1, alpha=alpha, beta=beta, size=tile)
context.synchronize()
cublas_dot(devA2, devB2, devC2, alpha=alpha, beta=beta)
partial1 = ng.empty((devC1.shape[0],1), dtype=np.float32)
partial2 = partial1[0:1,0:1]
if ng.min(ng.finite(devC1), partial=partial1, out=partial2).get()[0,0] == 0.0:
print("Error: NaN KCN: (%d,%d,%d) ab: (%f,%f) dtype: %d" %
(K,C,N, alpha,beta, itemsize))
exit()
diff = ng.max(abs(devC2 - devC1), partial=partial1, out=partial2).get()[0,0]
mean = ng.mean(abs(devC2), partial=partial1, out=partial2).get()[0,0]
pctErr = 100 * diff / mean
cpuA = np.random.uniform(-1.0, 1.0, dimA).astype(np.float32)
cpuB = np.random.uniform(-1.0, 1.0, dimB).astype(np.float32)
# cpuB = np.identity(n, dtype=dtype)
devA = ng.array(cpuA, dtype=dtype)
devB = ng.array(cpuB, dtype=dtype)
devC = ng.empty(dimC, dtype=dtype)
# repeat = min(int(50.0 * 4096**3 / (m * n * k)), 1000)
if op[0] == "t":
cpuA, devA = cpuA.T, devA.T
if op[1] == "t":
cpuB, devB = cpuB.T, devB.T
ng.dot(devA, devB, devC, repeat=1)
# context.synchronize()
cpuC = np.dot(cpuA, cpuB)
cpuD = devC.get()
diff = np.absolute(cpuC - cpuD)
max_diff = diff.max()
print(max_diff, cpuD.max())
if max_diff > 0.1 or max_diff != max_diff:
# print(m, n, k, max_diff)
print(cpuD[:: max(m // 16, 1), :: max(n // 16, 1)])
print(cpuC[:: max(m // 16, 1), :: max(n // 16, 1)])
print(diff[:: max(m // 16, 1), :: max(n // 16, 1)])
exit()
devA2 = devA1
devB2 = devB1
# otherwise copy
else:
devA2 = ng.empty(dimA, dtype=np.float32)
devB2 = ng.empty(dimB, dtype=np.float32)
devA2[:] = devA1
devB2[:] = devB1
devC2 = ng.empty(dimC, dtype=np.float32)
devC2[:] = devC1
if op[0] == 't': devA1, devA2 = devA1.T, devA2.T
if op[1] == 't': devB1, devB2 = devB1.T, devB2.T
ng.dot(devA1, devB1, devC1, alpha=alpha, beta=beta, repeat=repeat)
cublas_dot(devA2, devB2, devC2, alpha=alpha, beta=beta, repeat=repeat)
partial1 = ng.empty((devC1.shape[0],1), dtype=np.float32)
partial2 = partial1[0:1,0:1]
diff = ng.max(abs(devC2 - devC1), partial=partial1, out=partial2).get()[0,0]
mean = ng.mean(abs(devC2), partial=partial1, out=partial2).get()[0,0]
#if diff > .1:
print("Error: %.3f%%" % (100 * diff / mean))
print("--------------------------------------------------------------------------------")
cublas.cublasDestroy(handle)
devA2 = ng.empty(dimA, dtype=np.float32)
devB2 = ng.empty(dimB, dtype=np.float32)
devA2[:] = devA1
devB2[:] = devB1
devC2 = ng.empty(dimC, dtype=np.float32)
if op[0] == 't': devA1, devA2 = devA1.T, devA2.T
if op[1] == 't': devB1, devB2 = devB1.T, devB2.T
glops16 = 0
glops32 = 0
glops64 = 0
if op == "tn" and dtype is np.float16:
# Experimental 128x16 gemm kernel
glops16 = ng.dot(devA1, devB1, devC1, repeat=repeat, size=16)
if op != 'nt':
glops32 = ng.dot(devA1, devB1, devC1, repeat=repeat, size=32)
glops64 = ng.dot(devA1, devB1, devC1, repeat=repeat, size=64)
glops128 = ng.dot(devA1, devB1, devC1, repeat=repeat, size=128)
glops = max(glops16, glops32, glops64, glops128)
if glops16 == glops:
fastest = 16
elif glops32 == glops:
fastest = 32
elif glops64 == glops:
fastest = 64
else:
fastest = 128
class GPU(Backend):
"""
Sets up a NervanaGPU based backend for matrix operations.
Note that some functions defined in the generic Backend class such as are
cross-map pooling and normalization and adaDelta are not implemented for
this backend.
"""
default_dtype = np.float32
def __init__(self, rng_seed, stochastic_round=False, device_id=0):
self.ng = NervanaGPU(stochastic_round=stochastic_round)
logger.info("Initialized NervanaGPU with stochastic_round=%s",
stochastic_round)
self.rng_seed = rng_seed
self.rng_init()
self.device_id = device_id if device_id is not None else 0
def __getstate__(self):
"""
Defines what and how we go about serializing an instance of this class.
Returns:
self.__dict__: The full contents of the backend class instance,
except for the mem_pool which is on device and
cannot be serialized.
"""
if hasattr(self, 'mem_pool') and self.mem_pool is not None:
self.mem_pool_pickle = {'shape': self.mem_pool.shape,
'dtype': np.float32}
self.mem_pool = None
return self.__dict__
def __setstate__(self, state):
"""
Defines how we go about deserializing into an instance of this class.
Arguments:
self.__dict__: The full contents of the backend class instance,
except for the mem_pool which is on device and
cannot be serialized.
"""
self.__dict__.update(state)
self.mem_pool = self.ng.empty(self.mem_pool_pickle['shape'],
dtype=self.mem_pool_pickle['dtype'])
def init_mempool(self, shape, dtype=default_dtype):
"""
Allocates a memory pool for temporary storage
"""
self.mem_pool = self.ng.empty(shape, dtype=dtype)
def alloc_host_mem(self, shape, dtype):
return drv.pagelocked_empty(shape, dtype, order="C", mem_flags=0)
def create_stream(self):
return drv.Stream()
def async_copy(self, dest, src, stream=None):
drv.memcpy_htod_async(dest.gpudata, src, stream)
def rng_init(self):
"""
Initialize and seed the pseudo random number genrator. Random numbers
are generated on the host using numpy, then transfered to device.
"""
seed = None
if 'rng_seed' in self.__dict__:
seed = self.rng_seed
logger.info("Seeding random number generator with: %s", str(seed))
np.random.seed(seed)
def flop_timing_init(self, decorate_fc, decorate_conv, decorate_ew):
"""
Initialize FLOP timing. Wraps the specified MOP calls via a decorator
to record elapsed time and number of operations.
Arguments:
decorate_fc (list): string giving the function names of fully
connected layer forward/backward/update calls
to time.
decorate_conv (list): string giving the function names of
convolutional layer forward/backward/update
calls to time.
decorate_ew (list): string giving the function names of element-wise
calls to time.
Notes:
Must be called prior to first flop_timing_start call
"""
self.start = drv.Event()
self.end = drv.Event()
self.flop_timer = FlopsDecorator(self)
self.flop_timer.decorate(decorate_fc=decorate_fc,
decorate_conv=decorate_conv,
decorate_ew=decorate_ew)
def flop_timinig_start(self):
"""
Start a new FLOP timer.
Returns:
None: dummy value (not used)
"""
return self.start.record()
def flop_timing_finish(self, start_time):
"""
Complete current FLOP timing.
Arguments:
start_time (unused): ignored.
Returns:
float: elapsed time in seconds since prior flop_timing_start call.
"""
self.end.record()
self.end.synchronize()
return self.end.time_since(self.start)
def uniform(self, low=0.0, high=1.0, shape=1, dtype=default_dtype,
persist_values=True, name=None, allocator=drv.mem_alloc):
"""
generate numpy random number and convert to a GPUTensor.
If called with dype=None it will probably explode
"""
ary = np.random.uniform(low, high, shape)
return self.ng.array(ary, dtype=dtype, name=name)
def normal(self, loc=0.0, scale=1.0, size=1, dtype=default_dtype,
persist_values=True, name=None, allocator=drv.mem_alloc):
"""
Gaussian/Normal random number sample generation
"""
ary = np.random.normal(loc, scale, size)
return self.ng.array(ary, dtype=dtype, name=name)
def fprop_fc(self, out, inputs, weights, layer=None):
"""
Forward propagate the inputs of a fully connected network layer to
produce output pre-activations (ready for transformation by an
activation function).
Arguments:
out (GPUTensor): Where to store the forward propagated results.
inputs (GPUTensor): Will be either the dataset input values (first
layer), or the outputs from the previous layer.
weights (GPUTensor): The weight coefficient values for this layer.
layer (Layer): The layer object.
"""
self.ng.dot(weights, inputs, out)
def bprop_fc(self, out, weights, deltas, layer=None):
"""
Backward propagate the error through a fully connected network layer.
Arguments:
out (GPUTensor): Where to store the backward propagated errors.
weights (GPUTensor): The weight coefficient values for this layer.
deltas (GPUTensor): The error values for this layer
layer (Layer): The layer object.
"""
self.ng.dot(weights.T, deltas, out)
def update_fc(self, out, inputs, deltas, layer=None):
"""
Compute the updated gradient for a fully connected network layer.
Arguments:
out (GPUTensor): Where to store the updated gradient value.
inputs (GPUTensor): Will be either the dataset input values (first
layer), or the outputs from the previous layer.
deltas (GPUTensor): The error values for this layer
layer (Layer): The layer object.
"""
self.ng.dot(deltas, inputs.T, out)
def fprop_conv(self, out, inputs, weights, ofmshape, ofmsize, ofmlocs,
ifmshape, links, nifm, padding, stride, ngroups, fpropbuf,
local=False):
"""
Forward propagate the inputs of a convolutional network layer to
produce output pre-activations (ready for transformation by an
activation function).
Arguments:
out (GPUTensor): Where to store the forward propagated results.
inputs (GPUTensor): Will be either the dataset input values (first
layer), or the outputs from the previous layer.
weights (GPUTensor): The weight coefficient values for this layer.
ofmshape (tuple): Dimensions of each output feature map (typically
number of height and width neurons).
ofmsize (int): Total size of each output feature map.
ofmlocs (GPUTensor): Indices giving the location of each element
in each output feature map stored in out.
ifmshape (tuple): Dimensions of each input feature map (typically
number of height and width neurons). For this
backend we expect these values to be square.
links (GPUTensor): Input receptive field indices.
nifm (int): Total number of input feature maps.
padding (int): Number of additional elements to include along each
dimension of each local receptive field during the
convolution operation.
stride (int): Number of neurons to shift the filter at each step.
ngroups (int): Number of groups.
fpropbuf (GPUTensor): Temporary storage buffer used to hold the
convolved outputs for a single receptive
field. Not used for this backend.
local (bool, optional): Whether to do local filtering (True) or
convolution (False, the default)
"""
'''
N: Number of images in mini-batch
C: Number of input feature maps
K: Number of output feature maps
D: Depth of input image
H: Height of input image
W: Width of input image
T: Depth of filter kernel
R: Height of filter kernel
S: Width of filter kernel
'''
self.ng.fprop_conv(layer=fpropbuf, I=inputs, F=weights, O=out,
alpha=1.0, repeat=1)
def bprop_conv(self, out, weights, deltas, ofmshape, ofmsize, ofmlocs,
ifmshape, links, padding, stride, nifm, ngroups, bpropbuf,
local=False):
"""
Backward propagate the error through a convolutional network layer.
Arguments:
out (GPUTensor): Where to store the backward propagated errors.
weights (GPUTensor): The weight coefficient values for this layer.
deltas (GPUTensor): The error values for this layer
ofmshape (tuple): Dimensions of each output feature map (typically
height and width).
ofmsize (int): Total size of each output feature map.
ofmlocs (GPUTensor): Indices giving the location of each element in
each output feature map stored in out.
ifmshape (tuple): Dimensions of each input feature map (typically
height and width).
links (GPUTensor): Input receptive field indices.
nifm (int): Total number of input feature maps.
padding (int): Number of additional elements to include along each
dimension of each local receptive field during the
convolution operation.
stride (int): Number of neurons to shift the filter at each step.
ngroups (int): Number of groups.
bpropbuf (GPUTensor): Temporary storage buffer used to hold the
backpropagated error for a single receptive
field
local (bool, optional): Whether to do local filtering (True) or
convolution (False, the default)
"""
self.ng.bprop_conv(layer=bpropbuf, F=weights, E=deltas, grad_I=out,
alpha=1.0, repeat=1)
def update_conv(self, out, inputs, weights, deltas, ofmshape, ofmsize,
ofmlocs, ifmshape, links, nifm, padding, stride, ngroups,
fwidth, updatebuf, local=False, layer=None):
"""
Compute the updated gradient for a convolutional network layer.
Arguments:
out (GPUTensor): Where to store the updated gradient value.
inputs (GPUTensor): Will be either the dataset input values (first
layer), or the outputs from the previous layer.
weights (GPUTensor): The weight coefficient values for this layer.
deltas (GPUTensor): The error values for this layer
ofmshape (tuple): Dimensions of each output feature map (typically
height and width).
ofmsize (int): Total size of each output feature map.
ofmlocs (GPUTensor): Indices giving the location of each element in
each output feature map stored in out.
ifmshape (tuple): Dimensions of each input feature map (typically
height and width).
links (GPUTensor): Input receptive field indices.
nifm (int): Total number of input feature maps.
padding (int): Number of additional elements to include along each
dimension of each local receptive field during the
convolution operation.
stride (int): Number of neurons to shift the filter at each step.
ngroups (int): Number of groups.
fwidth (int): Filter width.
updatebuf (GPUTensor): Temporary storage buffer used to hold the
updated gradient for a single receptive
field
local (bool, optional): Whether to do local filtering (True) or
convolution (False, the default)
layer (Layer): The layer object.
"""
self.ng.update_conv(layer=updatebuf, I=inputs, E=deltas, grad_F=out,
alpha=1.0, repeat=1)
def fprop_pool(self, out, inputs, op, ofmshape, ofmsize, ofmlocs, fshape,
ifmshape, links, nifm, padding, stride, fpropbuf):
"""
Forward propagate the inputs of a Pooling network layer to
produce output pre-activations (ready for transformation by an
activation function).
Arguments:
out (GPUTensor): Where to store the forward propagated results.
inputs (GPUTensor): Will be either the dataset input values (first
layer), or the outputs from the previous layer.
op (string): The type of pooling operation to apply. We support
"max", "avg", "l2" currently.
ofmshape (tuple): Dimensions of each output feature map (typically
number of height and width neurons).
ofmsize (int): Total size of each output feature map.
ofmlocs (GPUTensor): Indices giving the location of each element in
each output feature map stored in out.
fshape (tuple): Dimensions of each filter (typically height and
width).
ifmshape (tuple): Dimensions of each input feature map (typically
number of height and width neurons).
links (GPUTensor): Input receptive field indices.
nifm (int): Total number of input feature maps.
padding (int): Number of additional elements to include along each
dimension of each local receptive field during the
pooling operation.
stride (int): Number of neurons to shift the filter at each step.
fpropbuf (GPUTensor): Temporary storage buffer used to hold the
pooled outputs for a single receptive field.
"""
op = op.lower()
if op == "max":
self.ng.fprop_pool(layer=fpropbuf, I=inputs, O=out, repeat=1)
else:
raise AttributeError("unexpected pooling op type: %s", op)
def bprop_pool(self, out, fouts, inputs, deltas, op, ofmshape, ofmsize,
ofmlocs, fshape, fpsize, ifmshape, links, nifm, padding,
stride, bpropbuf):
"""
Backward propagate the error through a pooling network layer.
Arguments:
out (GPUTensor): Where to store the backward propagated errors.
fouts (GPUTensor): Forward propagated outputs from the previous
layer.
inputs (GPUTensor): Will be either the dataset input values (first
layer), or the outputs from the previous layer.
deltas (GPUTensor): The error values for this layer
op (string): The type of pooling operation to apply. We support
"max", "avg", "l2" currently.
ofmshape (tuple): Dimensions of each output feature map (typically
height and width).
ofmsize (int): Total size of each output feature map.
ofmlocs (GPUTensor): Indices giving the location of each element in
each output feature map stored in out.
fshape (tuple): Dimensions of each filter (typically height and
width).
fpsize (int): The size of each filter.
ifmshape (tuple): Dimensions of each input feature map (typically
height and width).
links (GPUTensor): Input receptive field indices.
nifm (int): Total number of input feature maps.
padding (int): Number of additional elements to include along each
dimension of each local receptive field during the
pooling operation.
stride (int): Number of neurons to shift the filter at each step.
bpropbuf (GPUTensor): Temporary storage buffer used to hold the
backpropagated error for a single receptive
field
"""
op = op.lower()
if op == "max":
self.ng.bprop_pool(layer=bpropbuf, I=inputs, E=deltas, grad_I=out,
repeat=1)
else:
raise AttributeError("unexpected pooling op type: %s", op)
def logistic(self, x, out):
"""
Logistic sigmoid nonlinearity, 1/(1+exp(-x))
Arguments:
x (GPUTensor): Input tensor
out (GPUTensor): Output tensor
"""
self.ng.sig(x, out=out)
return out
def rectlin(self, x, out):
"""
Rectified Linear nonlinearity
Arguments:
x (GPUTensor): Input tensor
out (GPUTensor): Output tensor
"""
self.ng.maximum(x, 0., out=out)
return out
def rectleaky(self, x, slope, out):
out[:] = self.ng.maximum(x, x*slope)
def rectleaky_derivative(self, x, slope, out):
out[:] = self.ng.greater(x, 0) * (1.0 - slope) + slope
def sum(self, tsr, axes, out):
"""
Sum
Arguments:
tsr (GPUTensor): Input tensor
axes (int): Axis along which the reduction is performed. If axes
is None, the tensor is flattened and reduced over
both dimensions.
out (GPUTensor): Output tensor
"""
if axes is None:
sze = tsr.shape[0]*tsr.shape[1]
self.ng.sum(tsr.reshape(sze, 1), axis=0, out=out)
else:
self.ng.sum(tsr, axis=axes, out=out)
return out
def mean(self, tsr, axes, out):
"""
Calculates the arithmetic mean of the elements along the specified
axes.
Arguments:
tsr (GPUTensor): Input tensor
axes (int): Axis along which the reduction is performed. If axes
is None, the tensor is flattened and reduced over
both dimensions.
out (GPUTensor): Output tensor
"""
if axes is None:
sze = tsr.shape[0]*tsr.shape[1]
self.ng.mean(tsr.reshape(sze, 1), axis=0, out=out)
else:
self.ng.mean(tsr, axis=axes, out=out)
return out
def min(self, tsr, axes, out):
"""
Calculates the minimum of the elements along the specified
axes.
Arguments:
tsr (GPUTensor): Input tensor
axes (int): Axis along which the reduction is performed. If axes
is None, the tensor is flattened and reduced over
both dimensions.
out (GPUTensor): Output tensor
"""
if axes is None:
sze = tsr.shape[0]*tsr.shape[1]
self.ng.min(tsr.reshape(sze, 1), axis=0, out=out)
else:
self.ng.min(tsr, axis=axes, out=out)
return out
def max(self, tsr, axes, out):
"""
Calculates the maximum of the elements along the specified
axes.
Arguments:
tsr (GPUTensor): Input tensor
axes (int): Axis along which the reduction is performed. If axes
is None, the tensor is flattened and reduced over
both dimensions.
out (GPUTensor): Output tensor
"""
if axes is None:
sze = tsr.shape[0]*tsr.shape[1]
self.ng.max(tsr.reshape(sze, 1), axis=0, out=out)
else:
self.ng.max(tsr, axis=axes, out=out)
return out
def variance(self, tsr, axes, out, mean=None):
"""
Calculates the variance of the elements along the specified
axes.
Arguments:
tsr (GPUTensor): the tensor on which to compute the variance
axes (int, list, optional): the dimension(s) along which to
variance. If set to None, we will
variance over all dimensions.
out (GPUTensor): where the result will be stored.
mean (GPUTensor): the tensor containing mean of tsr
Returns:
GPUTensor: reference to out
"""
if mean is None:
logger.error("GPUTensor requires mean to be specified.")
raise ValueError("mean not specified")
self.ng.mean(self.ng.square(tsr-mean), axis=axes, out=out)
return out
def fabs(self, x, out):
"""
Calculates absolute value of the elements in a tensor
Arguments:
x (GPUTensor): Input tensor
out (GPUTensor): Output tensor
Returns:
GPUTensor: reference to out
"""
self.ng.fabs(x, out=out)
return out
def sqrt(self, x, out):
"""
Calculates square root of the elements in a tensor
Arguments:
x (GPUTensor): Input tensor
out (GPUTensor): Output tensor
Returns:
GPUTensor: reference to out
"""
self.ng.sqrt(x, out=out)
return out
def zeros(self, shape, dtype=default_dtype, persist_values=True):
"""
Allocate a new GPUTensor and fill it with zeros.
Arguments:
shape (tupel): Shape of the desired GPUTensor
dtype (dtype): Optional datatype
persist_values (bool, optional): If set to True (the default), the
values assigned to this Tensor
will persist across multiple begin
and end calls. Setting to False
may provide a performance increase
if values do not need to be
maintained across such calls
Returns:
GPUTensor: output
"""
return self.ng.zeros(shape, dtype=dtype)
def ones(self, shape, dtype=default_dtype, persist_values=True):
"""
Allocate a new GPUTensor and fill it with ones.
Arguments:
shape (tupel): Shape of the desired GPUTensor
dtype (dtype): Optional datatype
persist_values (bool, optional): If set to True (the default), the
values assigned to this Tensor
will persist across multiple begin
and end calls. Setting to False
may provide a performance increase
if values do not need to be
maintained across such calls
Returns:
GPUTensor: output
"""
return self.ng.ones(shape, dtype=dtype)
def empty(self, shape, dtype=default_dtype, persist_values=True):
"""
Allocate a new GPUTensor.
Arguments:
shape (tupel): Shape of the desired GPUTensor
dtype (dtype): Optional datatype
persist_values (bool, optional): If set to True (the default), the
values assigned to this Tensor
will persist across multiple begin
and end calls. Setting to False
may provide a performance increase
if values do not need to be
maintained across such calls
Returns:
GPUTensor: output
"""
return self.ng.empty(shape, dtype=dtype)
def array(self, ary, dtype=default_dtype, persist_values=True, name=None,
allocator=drv.mem_alloc):
"""
Allocate a new GPUTensor and fill it with supplied numpy array.
Arguments:
ary (ndarray): Numpy array with source data
dtype (dtype, optional): Optional datatype
persist_values (bool, optional): If set to True (the default), the
values assigned to this Tensor
will persist across multiple begin
and end calls. Setting to False
may provide a performance increase
if values do not need to be
maintained across such calls
name (string): Name for the GPUTensor
allocator (pycuda): Pycuda memory allocator
Returns:
GPUTensor: output
"""
return self.ng.array(ary, dtype=dtype, name=name)
def add(self, left, right, out):
"""
Elementwise addition
Arguments:
left (GPUTensor, numeric): left-hand side operand.
right (GPUTensor, numeric): right-hand side operand.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
self.ng.add(left, right, out=out)
return out
def subtract(self, left, right, out):
"""
Elementwise subtraction
Arguments:
left (GPUTensor, numeric): left-hand side operand.
right (GPUTensor, numeric): right-hand side operand.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
self.ng.subtract(left, right, out=out)
return out
def multiply(self, left, right, out):
"""
Elementwise multiplication
Arguments:
left (GPUTensor, numeric): left-hand side operand.
right (GPUTensor, numeric): right-hand side operand.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
self.ng.multiply(left, right, out=out)
return out
def divide(self, left, right, out):
"""
Elementwise division
Arguments:
left (GPUTensor, numeric): left-hand side operand.
right (GPUTensor, numeric): right-hand side operand.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
self.ng.divide(left, right, out=out)
return out
def greater(self, left, right, out):
"""
Elementwise greater than testing
Arguments:
left (GPUTensor, numeric): left-hand side operand.
right (GPUTensor, numeric): right-hand side operand.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
self.ng.greater(left, right, out=out)
return out
def equal(self, left, right, out):
"""
Performs element-wise equality testing on each element of left and
right, storing the result in out. Each operand is assumed to be the
same shape (or broadcastable as such).
Arguments:
left (GPUTensor, numeric): left-hand side operand.
right (GPUTensor, numeric): right-hand side operand.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
self.ng.equal(left, right, out=out)
return out
def not_equal(self, left, right, out):
"""
Elementwise not equal testing
Arguments:
left (GPUTensor, numeric): left-hand side operand.
right (GPUTensor, numeric): right-hand side operand.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
self.ng.not_equal(left, right, out=out)
return out
def clip(self, a, a_min, a_max, out):
"""
Elementwise clipping between a range of specified values
Arguments:
a (GPUTensor): input tensor.
a_min (float): floor value.
a_max (float): ceiling value.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
self.ng.clip(a, a_min, a_max, out=out)
return out
def log(self, a, out):
"""
Elementwise base-e logarithm
Arguments:
a (GPUTensor): input tensor.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
self.ng.log(a, out=out)
return out
def tanh(self, a, out):
"""
Elementwise tanh
Arguments:
a (GPUTensor): input tensor.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
self.ng.tanh(a, out=out)
return out
def argmax(self, a, out, axis=0):
"""
Calculates the indices of the maximal element value along the specified
axis. If multiple elements contain the maximum, only the elements of
the first are returned.
Arguments:
tsr (GPUTensor): The GPUTensor on which to find the maximum indices
axis (int): The dimension along which to find the maximum. If set
to None, find the overall maximum index of a flattened
representation of tsr.
out (GPUTensor): Where to store the result. Should be of the
appropriate type and expected shape
Returns:
GPUTensor: reference to out
"""
self.ng.argmax(a, out=out, axis=axis)
return out
def softmax(self, x, out):
"""
Softmax nonlinearity. Computes exp(x-max(x)) / sum_i exp(x_i-max(x_i))
Arguments:
x (GPUTensor): input tensor.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
out[:] = (self.ng.reciprocal(self.ng.sum(
self.ng.exp(x - self.ng.max(x, axis=0)), axis=0)) *
self.ng.exp(x - self.ng.max(x, axis=0)))
return out
def softmax_gradient(self, y, err, out):
"""
Gradient of the softmax nonlinearity.
Arguments:
y (GPUTensor): input tensor.
err (GPUTensor): backpropagated error.
out (GPUTensor): where the result will be stored.
Returns:
GPUTensor: reference to out
"""
raise NotImplementedError("Softmax gradient should use shortcut")
return out
def make_binary_mask(self, tsr, keepthresh=0.5, dtype=default_dtype):
"""
Create a binary mask for dropout layers.
Arguments:
tsr (GPUTensor): Output tensor
keepthresh (float): fraction of ones
"""
self.ng.dropout(keep=keepthresh, out=tsr)
def gdm_compound(self, ps_item, us_item, vs_item, momentum_coef,
learning_rate, epoch):
"""
Perform gradient descent update with momentum.
Arguments:
ps_item (GPUTensor): parameter tensor (e.g. a weight matrix)
us_item (GPUTensor): update tensor, contains gradient wrt. weights
vs_item (GPUTensor): velocity tensor.
momentum_coef (float): momentum coefficient.
learning_rate (float): learning rate.
epoch (int): epoch (used in conjunction with diagnostics).
Outputs are written to vs_item (updated velocity)
and ps_item (updated weights)
"""
vs_item[:] = vs_item * momentum_coef - us_item * learning_rate
ps_item[:] = ps_item + vs_item
def gdmwd_compound(self, ps_item, us_item, vs_item, momentum_coef,
learning_rate, wd, epoch):
"""
Perform gradient descent update with momentum and weight decay.
Arguments:
ps_item (GPUTensor): parameter tensor (e.g. a weight matrix)
us_item (GPUTensor): update tensor, contains gradient wrt. weights
vs_item (GPUTensor): velocity tensor.
momentum_coef (float): momentum coefficient.
learning_rate (float): learning rate.
wd (float): weight decay parameter.
epoch (int): epoch (used in conjunction with diagnostics).
Outputs:
ps_item, the updated weights.
vs_item, the updated velocity.
us_item, used as a temp buffer.
"""
vs_item[:] = vs_item * momentum_coef - us_item * \
learning_rate - learning_rate * wd * ps_item
ps_item[:] = ps_item + vs_item
def exp_mavg(self, mavg, newval, rho):
"""
Calculate the exponential moving average
Arguments:
mavg: The running value of the moving average
newval: New sample to be added to the moving average
rho: Interpolation value
"""
mavg[:] = rho * mavg + (1.0 - rho) * newval
def ada_update(self, ps_item, us_item, gs_item, ds_item, ls_item, ss_item,
rho, epsilon):
"""
Update rule for AdaDelta (Zeiler, http://arxiv.org/abs/1212.5701)
Arguments:
ps_item: weight / parameter (will be updated)
us_item: update
gs_item: expected value of Gradient Squared (will be updated)
ds_item: expected value of Delta Squared (will be updated)
ls_item: learning rate (will be updated)
ss_item: Scratch Space
rho: decay constant (determines window size)
epsilon: small positive constant for numerical stability
"""
# Accumulate E[Grad^2]
gs_item[:] = gs_item * rho + (1.0 - rho) * us_item * us_item
# Calculate Updates
ls_item[:] = self.ng.sqrt((ds_item + epsilon) /
(gs_item + epsilon)) * (-1.0) * us_item
# Accumulate E[Delt^2]
ds_item[:] = ds_item * rho + (1.0 - rho) * ls_item * ls_item
# Final update to the params
ps_item[:] = ps_item + ls_item
def rms_update(self, params, updates, run_squares, velocity, scratch_space,
gamma, epsilon, learning_rate, momentum_coef):
# Update running squares
run_squares[:] = gamma * run_squares + (1. - gamma) * updates * updates
# Now scale the gradient by lr / rms(grad) (with a epsilon term for
# stability) and use it to update the params
if momentum_coef == 0:
params[:] = params - learning_rate * updates * self.ng.reciprocal(
self.ng.sqrt(run_squares) + epsilon)
else:
velocity[:] = velocity * momentum_coef - \
learning_rate * updates * \
self.ng.reciprocal(self.ng.sqrt(run_squares) + epsilon)
params[:] = params + velocity
def fprop_bn_compound(self, inputs, beta, gamma, eps, xhat,
xmean, xvar, gmean, gvar, rho, out):
"""
Batch normalization forward pass, compounded to run in 3 kernel calls.
Arguments:
inputs: input data to be normalized
beta: location parameter
gamma: scale parameter
eps: small constant for numerical stability
xvar: variance (updated)
xhat: normalized input (updated)
out: normalized and rescaled input (updated)
"""
xvar[:] = self.ng.var(inputs, axis=1)
xmean[:] = self.ng.mean(inputs, axis=1)
gmean[:] = gmean * rho + (1.0 - rho) * xmean
gvar[:] = gvar * rho + (1.0 - rho) * xvar
xvar[:] = self.ng.reciprocal(self.ng.sqrt(xvar + eps))
xhat[:] = xvar * (inputs - xmean)
out[:] = xhat * gamma + beta
return out
def bprop_bn_compound(self, xhat, error, xvar, gamma,
beta_updates, gamma_updates):
"""
Batch normalization backward pass, compounded to run with 4 kernel
calls.
Arguments:
xhat: normalized input data (updated)
error: backpropagated deltas (updated)
xvar: precomputed variance
gamma: scale parameter
beta_updates: gradient update for beta (updated)
gamma_updates: gradient update for gamma (updated)
"""
gamma_updates[:] = self.ng.sum(xhat * error, axis=1)
beta_updates[:] = self.ng.sum(error, axis=1)
xhat[:] = (xhat * gamma_updates + beta_updates) / float(xhat.shape[1])
error[:] = xvar * gamma * (error - xhat)
if data_type == "All Ones":
cpuA = np.ones(dimA, dtype=dtype).astype(np.float32)
cpuB = np.ones(dimB, dtype=dtype).astype(np.float32)
#cpuB = np.identity(n, dtype=np.float32)
else:
cpuA = np.random.uniform(-1.0, 1.0, dimA).astype(np.float32)
cpuB = np.random.uniform(-1.0, 1.0, dimB).astype(np.float32)
devA = ng.array(cpuA, dtype=dtype)
devB = ng.array(cpuB, dtype=dtype)
devC = ng.empty(dimC, dtype=dtype)
if op[0] == 't': cpuA, devA = cpuA.T, devA.T
if op[1] == 't': cpuB, devB = cpuB.T, devB.T
ng.dot(devA, devB, devC, repeat=repeat)
if cpu:
cpuC = np.dot(cpuA, cpuB)
cpuD = devC.get()
diff = np.absolute(cpuC - cpuD)
print diff.max()
print cpuD[::max(m // 4, 1), ::max(n // 4, 1)]
print cpuC[::max(m // 4, 1), ::max(n // 4, 1)]
print diff[::max(m // 4, 1), ::max(n // 4, 1)]
# print cpuD
# exit()
# devC2 = devC2s.share(dimC, dtype=np.float32)
devC2[:] = devC1
if op[0] == 't':
devA1, devA2 = devA1.T, devA2.T
if op[1] == 't':
devB1, devB2 = devB1.T, devB2.T
for tile in (32, 64, 128):
if op == 'nt' and tile != 128:
continue
try:
ng.dot(devA1,
devB1,
devC1,
alpha=alpha,
beta=beta,
size=tile)
context.synchronize()
cublas_dot(devA2,
devB2,
devC2,
alpha=alpha,
beta=beta)
partial1 = ng.empty(
(devC1.shape[0], 1),
dtype=np.float32)
partial2 = partial1[0:1, 0:1]
class MGPU(GPU):
default_dtype = np.float32
num_dev = 1
is_dist = True
def __init__(self, rng_seed, stochastic_round=False, device_id=0,
num_dev=2):
drv.init()
self.num_dev = num_dev
if device_id == 0:
self.dev_list = range(num_dev)
else:
self.dev_list = device_id
assert len(self.dev_list) == self.num_dev
assert self.num_dev <= drv.Device.count()
self.ctxs = []
self.devs = []
self._strms = []
self._redstrms = []
self._events = []
self._redevents = []
self.async = True
self._nostrms = [None for i in self.dev_list]
for i in self.dev_list:
self.devs.append(drv.Device(i))
for dev in self.devs:
self.ctxs.append(
dev.make_context(drv.ctx_flags.SCHED_BLOCKING_SYNC))
self._strms.append(drv.Stream())
self._redstrms.append(drv.Stream())
self._events.append(drv.Event())
self._redevents.append(drv.Event())
drv.Context.pop()
self.ctxs[0].push()
atexit.register(drv.Context.pop)
MGPUTensor.ctxs = self.ctxs
MGPUTensor.num_dev = num_dev
self.ng = NervanaGPU(stochastic_round=stochastic_round)
logger.info("Initialized %d device NervanaGPU, stochastic_round=%s",
num_dev, stochastic_round)
self.ng.block = None
self.rng_seed = rng_seed
self.rng_init()
# Setup the pairwise contexts
# TODO clean up this code to avoid indexing
for dev1, ctx1 in zip(self.devs, self.ctxs):
ctx1.push()
for dev2, ctx2 in zip(self.devs, self.ctxs):
if dev1 == dev2:
continue
if dev1.can_access_peer(dev2):
ctx1.enable_peer_access(ctx2)
else:
print('Cannot enable peer access between '
'{:d} and {:d}'.format(dev1, dev2))
ctx1.pop()
def make_events(self):
evtlist = []
for ctx in self.ctxs:
ctx.push()
evtlist.append(drv.Event())
ctx.pop()
return evtlist
# These definitions are for performing grouped context commands
# This is experimental and should remove _stack for actual usage
def begin_stack(self, block, identifier):
if block == Block.update:
self.ng.block = Block.update
self.call_stack = []
else:
pass
def end_stack(self, block, identifier):
if block == Block.update:
self.ng.block = None
for idx, ctx in enumerate(self.ctxs):
ctx.push()
self.ng.stream = self.strms[idx]
for method, args, kwargs in self.call_stack:
myargs = [a._tensorlist[idx] if isinstance(
a, MGPUTensor) else a for a in args]
mykwargs = {k: v._tensorlist[idx] if isinstance(
v, MGPUTensor) else v for k, v in kwargs.iteritems()}
getattr(super(MGPU, self), method)(*myargs, **mykwargs)
self.ng.stream = None
ctx.pop()
self.call_stack = None
else:
pass
@property
def strms(self):
return self._strms if self.async else self._nostrms
@property
def redstrms(self):
return self._redstrms if self.async else self._nostrms
def uniform(self, low=0.0, high=1.0, size=1, dtype=default_dtype,
name=None, persist_values=True, ptype='replica'):
"""
generate numpy random number and convert to a GPUTensor.
If called with dtype=None it will probably explode
"""
assert len(size) == 2
result = self.empty(size, dtype=dtype, persist_values=persist_values)
result.ptype = ptype
beshape = size if ptype == 'replica' else (self.num_dev * size[0],
size[1])
ary = np.random.uniform(low, high, beshape).astype(dtype)
self.set(result, ary)
return result
def normal(self, loc=0.0, scale=1.0, size=1, dtype=default_dtype,
name=None, persist_values=True, ptype='replica'):
"""
Gaussian/Normal random number sample generation
"""
assert len(size) == 2
result = self.empty(size, dtype=dtype, persist_values=persist_values)
result.ptype = ptype
beshape = size if ptype == 'replica' else (self.num_dev * size[0],
size[1])
ary = np.random.normal(loc, scale, beshape).astype(dtype)
self.set(result, ary)
return result
def synchronize(self):
if not self.async:
return
for s in self.strms:
s.synchronize()
def redsynchronize(self):
if not self.async:
return
for s in self.redstrms:
s.synchronize()
def allocate_fragment(self, shape, dtype=default_dtype,
persist_values=True):
# TODO: set ptype to be fragment in this case ??
return self.empty((shape[0], shape[1] / self.num_dev), dtype,
persist_values=persist_values)
def zeros_like(self, ary, dtype=default_dtype, persist_values=True,
name=None):
result = self.zeros(ary.shape, dtype=dtype,
persist_values=persist_values)
result.ptype = ary.ptype
return result
def empty_like(self, ary, dtype=default_dtype, persist_values=True,
name=None):
result = self.empty(ary.shape, dtype=dtype,
persist_values=persist_values, name=name)
result.ptype = ary.ptype
return result
def set(self, tensor, data):
assert isinstance(tensor, MGPUTensor)
if tensor.ptype == 'replica':
for dest, strm, ctx in zip(tensor.tlist, self.strms, self.ctxs):
ctx.push()
drv.memcpy_htod_async(dest.ptr, data, strm)
ctx.pop()
# tensor.copy_from(data)
else:
self.scatter(data, tensor)
def scatter(self, hbuf, dbuf):
'''
scatters the array data in hbuf to the mgpu tensor
assumes that dbuf is a M x N and hbuf is M x (Nxk) where k is the
number of replicas
also assumes that dtype of hbuf and dbuf are the same
'''
assert hbuf.size == dbuf.size * dbuf.num_dev
assert isinstance(dbuf, MGPUTensor)
assert hbuf.dtype == dbuf.dtype
ndata = dbuf.size
starts = [i * ndata for i in range(self.num_dev)]
for dest, strm, ctx, doff in zip(dbuf.tlist, self.strms, self.ctxs,
starts):
src = hbuf.reshape((hbuf.size))[doff:(doff + ndata)]
ctx.push()
drv.memcpy_htod_async(dest.ptr, src, strm)
ctx.pop()
self.synchronize()
def fprop_fc(self, out, inputs, weights, layer=None):
"""
In this case, the weights are shards, the acts are replicas
ubuf should be of size nout/num_dev x mbsz
"""
ubuf = layer.mempool[0]
assert ubuf.shape == (weights.shape[0], inputs.shape[1])
if layer.use_biases:
biases = layer.biases.tlist
else:
biases = [None for i in range(self.num_dev)]
for dbuf, ibuf, wt, bs, strm, ctx in zip(ubuf.tlist, inputs.tlist,
weights.tlist, biases,
self.strms, self.ctxs):
ctx.push()
self.ng.stream = strm
self.ng.dot(wt, ibuf, dbuf)
if layer.use_biases:
self.ng.add(dbuf, bs, out=dbuf)
ctx.pop()
# Note, should be safe not to sync because each fragment is computed
# on the same stream that originates the copy
# self.synchronize()
self.fragment_to_replica(ubuf, out)
def bprop_fc(self, out, weights, deltas, layer=None):
"""
Backward propagate the error through a fully connected network layer.
Arguments:
out (GPUTensor): Where to store the backward propagated errors.
weights (GPUTensor): The weight coefficient values for this layer.
deltas (GPUTensor): The error values for this layer
layer (Layer): The layer object.
"""
ubuf = layer.mempool[1]
wtsz = weights.shape[0]
starts = [i * wtsz for i in range(self.num_dev)]
assert out.shape == (weights.shape[1], deltas.shape[1])
assert ubuf.shape == out.shape
for dbuf, ibuf, wt, strm, ctx, off in zip(out.tlist, deltas.tlist,
weights.tlist, self.strms,
self.ctxs, starts):
ctx.push()
self.ng.stream = strm
self.ng.dot(wt.T, ibuf[off:(off + wtsz)], dbuf)
ctx.pop()
# Note, should be safe not to sync because each fragment is computed
# on the same stream that originates the copy
self.synchronize()
self.reduce(out, ubuf)
def update_fc(self, out, inputs, deltas, layer=None):
wtsz = out.shape[0]
starts = [i * wtsz for i in range(self.num_dev)]
for obuf, dbuf, ibuf, strm, ctx, off in zip(out.tlist, deltas.tlist,
inputs.tlist, self.strms,
self.ctxs, starts):
ctx.push()
self.ng.stream = strm
self.ng.dot(dbuf[off:(off + wtsz)], ibuf.T, obuf)
ctx.pop()
# self.synchronize()
def update_fc_bias(self, err, out):
"""
Compute the updated bias gradient for a fully connected network layer.
Arguments:
out (GPUTensor): Where to store the updated gradient value.
err (GPUTensor): backpropagated error
"""
wtsz = out.shape[0]
starts = [i * wtsz for i in range(self.num_dev)]
for ebuf, obuf, strm, ctx, off in zip(err.tlist, out.tlist, self.strms,
self.ctxs, starts):
ctx.push()
self.ng.stream = strm
self.ng.sum(ebuf[off:(off + wtsz)], axis=1, out=obuf)
ctx.pop()
def add_fc_bias(self, inputs, bias):
"""
This is a no-op since we absorb the bias add into the fprop_fc call
"""
pass
def reduce_tensor(self, ary, async=True):
'''
This is the case for the scalar tensor
'''
assert ary.size == 1
if ary.ptype == 'replica':
self.ctxs[0].push()
result = ary.tlist[0].get()
self.ctxs[0].pop()
return result
result = np.zeros((self.num_dev, 1), ary.dtype)
for i, (ctx, src_buf, strm) in enumerate(zip(
self.ctxs, ary.tlist, self.strms)):
ctx.push()
drv.memcpy_dtoh_async(result[i], src_buf.ptr, strm)
ctx.pop()
self.synchronize()
return result.sum()
class MGPU(GPU):
default_dtype = np.float32
num_dev = 1
is_dist = True
def __init__(self,
rng_seed,
stochastic_round=False,
device_id=0,
num_dev=2):
drv.init()
self.num_dev = num_dev
if device_id == 0:
self.dev_list = range(num_dev)
else:
self.dev_list = device_id
assert len(self.dev_list) == self.num_dev
assert self.num_dev <= drv.Device.count()
self.ctxs = []
self.devs = []
self._strms = []
self._redstrms = []
self._events = []
self._redevents = []
self. async = True
self._nostrms = [None for i in self.dev_list]
for i in self.dev_list:
self.devs.append(drv.Device(i))
for dev in self.devs:
self.ctxs.append(
dev.make_context(drv.ctx_flags.SCHED_BLOCKING_SYNC))
self._strms.append(drv.Stream())
self._redstrms.append(drv.Stream())
self._events.append(drv.Event())
self._redevents.append(drv.Event())
drv.Context.pop()
self.ctxs[0].push()
atexit.register(drv.Context.pop)
MGPUTensor.ctxs = self.ctxs
MGPUTensor.num_dev = num_dev
self.ng = NervanaGPU(stochastic_round=stochastic_round)
logger.info("Initialized %d device NervanaGPU, stochastic_round=%s",
num_dev, stochastic_round)
self.ng.block = None
self.rng_seed = rng_seed
self.rng_init()
# Setup the pairwise contexts
# TODO clean up this code to avoid indexing
for dev1, ctx1 in zip(self.devs, self.ctxs):
ctx1.push()
for dev2, ctx2 in zip(self.devs, self.ctxs):
if dev1 == dev2:
continue
if dev1.can_access_peer(dev2):
ctx1.enable_peer_access(ctx2)
else:
print('Cannot enable peer access between '
'{:d} and {:d}'.format(dev1, dev2))
ctx1.pop()
def make_events(self):
evtlist = []
for ctx in self.ctxs:
ctx.push()
evtlist.append(drv.Event())
ctx.pop()
return evtlist
# These definitions are for performing grouped context commands
# This is experimental and should remove _stack for actual usage
def begin_stack(self, block, identifier):
if block == Block.update:
self.ng.block = Block.update
self.call_stack = []
else:
pass
def end_stack(self, block, identifier):
if block == Block.update:
self.ng.block = None
for idx, ctx in enumerate(self.ctxs):
ctx.push()
self.ng.stream = self.strms[idx]
for method, args, kwargs in self.call_stack:
myargs = [
a._tensorlist[idx] if isinstance(a, MGPUTensor) else a
for a in args
]
mykwargs = {
k:
v._tensorlist[idx] if isinstance(v, MGPUTensor) else v
for k, v in kwargs.iteritems()
}
getattr(super(MGPU, self), method)(*myargs, **mykwargs)
self.ng.stream = None
ctx.pop()
self.call_stack = None
else:
pass
@property
def strms(self):
return self._strms if self. async else self._nostrms
@property
def redstrms(self):
return self._redstrms if self. async else self._nostrms
def uniform(self,
low=0.0,
high=1.0,
size=1,
dtype=default_dtype,
name=None,
persist_values=True,
ptype='replica'):
"""
generate numpy random number and convert to a GPUTensor.
If called with dtype=None it will probably explode
"""
assert len(size) == 2
result = self.empty(size, dtype=dtype, persist_values=persist_values)
result.ptype = ptype
beshape = size if ptype == 'replica' else (self.num_dev * size[0],
size[1])
ary = np.random.uniform(low, high, beshape).astype(dtype)
self.set(result, ary)
return result
def normal(self,
loc=0.0,
scale=1.0,
size=1,
dtype=default_dtype,
name=None,
persist_values=True,
ptype='replica'):
"""
Gaussian/Normal random number sample generation
"""
assert len(size) == 2
result = self.empty(size, dtype=dtype, persist_values=persist_values)
result.ptype = ptype
beshape = size if ptype == 'replica' else (self.num_dev * size[0],
size[1])
ary = np.random.normal(loc, scale, beshape).astype(dtype)
self.set(result, ary)
return result
def synchronize(self):
if not self. async:
return
for s in self.strms:
s.synchronize()
def redsynchronize(self):
if not self. async:
return
for s in self.redstrms:
s.synchronize()
def allocate_fragment(self,
shape,
dtype=default_dtype,
persist_values=True):
# TODO: set ptype to be fragment in this case ??
return self.empty((shape[0], shape[1] / self.num_dev),
dtype,
persist_values=persist_values)
def zeros_like(self,
ary,
dtype=default_dtype,
persist_values=True,
name=None):
result = self.zeros(ary.shape,
dtype=dtype,
persist_values=persist_values)
result.ptype = ary.ptype
return result
def empty_like(self,
ary,
dtype=default_dtype,
persist_values=True,
name=None):
result = self.empty(ary.shape,
dtype=dtype,
persist_values=persist_values,
name=name)
result.ptype = ary.ptype
return result
def set(self, tensor, data):
assert isinstance(tensor, MGPUTensor)
if tensor.ptype == 'replica':
for dest, strm, ctx in zip(tensor.tlist, self.strms, self.ctxs):
ctx.push()
drv.memcpy_htod_async(dest.ptr, data, strm)
ctx.pop()
# tensor.copy_from(data)
else:
self.scatter(data, tensor)
def scatter(self, hbuf, dbuf):
'''
scatters the array data in hbuf to the mgpu tensor
assumes that dbuf is a M x N and hbuf is M x (Nxk) where k is the
number of replicas
also assumes that dtype of hbuf and dbuf are the same
'''
assert hbuf.size == dbuf.size * dbuf.num_dev
assert isinstance(dbuf, MGPUTensor)
assert hbuf.dtype == dbuf.dtype
ndata = dbuf.size
starts = [i * ndata for i in range(self.num_dev)]
for dest, strm, ctx, doff in zip(dbuf.tlist, self.strms, self.ctxs,
starts):
src = hbuf.reshape((hbuf.size))[doff:(doff + ndata)]
ctx.push()
drv.memcpy_htod_async(dest.ptr, src, strm)
ctx.pop()
self.synchronize()
def fprop_fc(self, out, inputs, weights, layer=None):
"""
In this case, the weights are shards, the acts are replicas
ubuf should be of size nout/num_dev x mbsz
"""
ubuf = layer.mempool[0]
assert ubuf.shape == (weights.shape[0], inputs.shape[1])
if layer.use_biases:
biases = layer.biases.tlist
else:
biases = [None for i in range(self.num_dev)]
for dbuf, ibuf, wt, bs, strm, ctx in zip(ubuf.tlist, inputs.tlist,
weights.tlist, biases,
self.strms, self.ctxs):
ctx.push()
self.ng.stream = strm
self.ng.dot(wt, ibuf, dbuf)
if layer.use_biases:
self.ng.add(dbuf, bs, out=dbuf)
ctx.pop()
# Note, should be safe not to sync because each fragment is computed
# on the same stream that originates the copy
# self.synchronize()
self.fragment_to_replica(ubuf, out)
def bprop_fc(self, out, weights, deltas, layer=None):
"""
Backward propagate the error through a fully connected network layer.
Arguments:
out (GPUTensor): Where to store the backward propagated errors.
weights (GPUTensor): The weight coefficient values for this layer.
deltas (GPUTensor): The error values for this layer
layer (Layer): The layer object.
"""
ubuf = layer.mempool[1]
wtsz = weights.shape[0]
starts = [i * wtsz for i in range(self.num_dev)]
assert out.shape == (weights.shape[1], deltas.shape[1])
assert ubuf.shape == out.shape
for dbuf, ibuf, wt, strm, ctx, off in zip(out.tlist, deltas.tlist,
weights.tlist, self.strms,
self.ctxs, starts):
ctx.push()
self.ng.stream = strm
self.ng.dot(wt.T, ibuf[off:(off + wtsz)], dbuf)
ctx.pop()
# Note, should be safe not to sync because each fragment is computed
# on the same stream that originates the copy
self.synchronize()
self.reduce(out, ubuf)
def update_fc(self, out, inputs, deltas, layer=None):
wtsz = out.shape[0]
starts = [i * wtsz for i in range(self.num_dev)]
for obuf, dbuf, ibuf, strm, ctx, off in zip(out.tlist, deltas.tlist,
inputs.tlist, self.strms,
self.ctxs, starts):
ctx.push()
self.ng.stream = strm
self.ng.dot(dbuf[off:(off + wtsz)], ibuf.T, obuf)
ctx.pop()
# self.synchronize()
def update_fc_bias(self, err, out):
"""
Compute the updated bias gradient for a fully connected network layer.
Arguments:
out (GPUTensor): Where to store the updated gradient value.
err (GPUTensor): backpropagated error
"""
wtsz = out.shape[0]
starts = [i * wtsz for i in range(self.num_dev)]
for ebuf, obuf, strm, ctx, off in zip(err.tlist, out.tlist, self.strms,
self.ctxs, starts):
ctx.push()
self.ng.stream = strm
self.ng.sum(ebuf[off:(off + wtsz)], axis=1, out=obuf)
ctx.pop()
def add_fc_bias(self, inputs, bias):
"""
This is a no-op since we absorb the bias add into the fprop_fc call
"""
pass
def reduce_tensor(self, ary, async=True):
'''
This is the case for the scalar tensor
'''
assert ary.size == 1
if ary.ptype == 'replica':
self.ctxs[0].push()
result = ary.tlist[0].get()
self.ctxs[0].pop()
return result
result = np.zeros((self.num_dev, 1), ary.dtype)
for i, (ctx, src_buf,
strm) in enumerate(zip(self.ctxs, ary.tlist, self.strms)):
ctx.push()
drv.memcpy_dtoh_async(result[i], src_buf.ptr, strm)
ctx.pop()
self.synchronize()
return result.sum()
devA2[:] = devA1
devB2[:] = devB1
devC2 = ng.empty(dimC, dtype=np.float32)
if op[0] == 't': devA1, devA2 = devA1.T, devA2.T
if op[1] == 't': devB1, devB2 = devB1.T, devB2.T
glops32x128 = 0
glops128x32 = 0
glops128x64 = 0
if op != 'tn':
glops32x128 = ng.dot(devA1,
devB1,
devC1,
repeat=repeat,
size='32x128')
if op != 'nt':
glops128x32 = ng.dot(devA1,
devB1,
devC1,
repeat=repeat,
size='128x32')
glops128x64 = ng.dot(devA1,
devB1,
devC1,
repeat=repeat,
size='128x64')
glops128x128 = ng.dot(devA1,
devB1,
