def AccumulateConvDeriv(self, edge, deriv):
"""Accumulate the derivative w.r.t the outputs of this layer.
Each layer needs to compute derivatives w.r.t its outputs. These outputs may
have been connected to lots of other nodes through outgoing edges.
This method adds up the derivatives contributed by each outgoing edge.
It gets derivatives w.r.t the inputs at the other end of an outgoing edge.
Args:
edge: The edge which is sending the derivative.
deriv: The derivative w.r.t the inputs at the other end of this edge.
"""
if self.dirty: # If some derivatives have already been received.
raise Exception('Not implemented.')
self.dirty = True
w = edge.params['weight']
conv = edge.conv_params
size = conv.size
stride = conv.stride
padding = conv.padding
num_filters = conv.num_filters
num_colors = conv.num_colors
f, numdims = w.shape
assert f == num_filters, 'f is %d but num_filters is %d' % (
f, num_filters)
assert numdims == size**2 * num_colors
input_t = edge.input_t
numimages, numdims = input_t.shape
assert numdims % num_colors == 0
x = int(np.sqrt(numdims / num_colors))
assert x**2 == numdims / num_colors
n_locs = (x + 2 * padding - size) / stride + 1
if conv.max_pool:
deriv.transpose(edge.output_t2)
n_pool_locs = (n_locs + 2 * padding -
conv.pool_size) / conv.pool_stride + 1
cc.MaxPoolUndo(edge.unpooled_layer, edge.unpooled_layer,
edge.output_t2, edge.output_t, conv.pool_size, 0,
conv.pool_stride, n_pool_locs)
else:
deriv.transpose(edge.output_t)
if self.is_input:
return
if conv.max_pool:
output_t = edge.unpooled_layer
else:
output_t = edge.output_t
cc.convDown(output_t, w, input_t, n_locs, stride, size, x, num_colors)
input_t.transpose(self.deriv)
def AccumulateConvDeriv(self, edge, deriv):
"""Accumulate the derivative w.r.t the outputs of this layer.
Each layer needs to compute derivatives w.r.t its outputs. These outputs may
have been connected to lots of other nodes through outgoing edges.
This method adds up the derivatives contributed by each outgoing edge.
It gets derivatives w.r.t the inputs at the other end of an outgoing edge.
Args:
edge: The edge which is sending the derivative.
deriv: The derivative w.r.t the inputs at the other end of this edge.
"""
if self.dirty: # If some derivatives have already been received.
raise Exception('Not implemented.')
self.dirty = True
w = edge.params['weight']
conv = edge.conv_params
size = conv.size
stride = conv.stride
padding = conv.padding
num_filters = conv.num_filters
num_colors = conv.num_colors
f, numdims = w.shape
assert f == num_filters, 'f is %d but num_filters is %d' % (f, num_filters)
assert numdims == size**2 * num_colors
input_t = edge.input_t
numimages, numdims = input_t.shape
assert numdims % num_colors == 0
x = int(np.sqrt(numdims / num_colors))
assert x**2 == numdims/num_colors
n_locs = (x + 2 * padding - size) / stride + 1
if conv.max_pool:
deriv.transpose(edge.output_t2)
n_pool_locs = (n_locs + 2 * padding - conv.pool_size) / conv.pool_stride + 1
cc.MaxPoolUndo(edge.unpooled_layer, edge.unpooled_layer, edge.output_t2,
edge.output_t, conv.pool_size, 0, conv.pool_stride, n_pool_locs)
else:
deriv.transpose(edge.output_t)
if self.is_input:
return
if conv.max_pool:
output_t = edge.unpooled_layer
else:
output_t = edge.output_t
cc.convDown(output_t, w, input_t, n_locs, stride, size, x, num_colors)
input_t.transpose(self.deriv)
def AccumulateConvDeriv(self, edge, deriv):
"""Accumulate the derivative w.r.t the outputs of this layer.
Each layer needs to compute derivatives w.r.t its outputs. These outputs may
have been connected to lots of other nodes through outgoing edges.
This method adds up the derivatives contributed by each outgoing edge.
It gets derivatives w.r.t the inputs at the other end of an outgoing edge.
Args:
edge: The edge which is sending the derivative.
deriv: The derivative w.r.t the inputs at the other end of this edge.
"""
if self.dirty: # If some derivatives have already been received.
raise Exception("Not implemented.")
self.dirty = True
w = edge.params["weight"]
conv = edge.conv_params
size = conv.size
stride = conv.stride
padding = conv.padding
num_filters = conv.num_filters
num_colors = conv.num_colors
f, numdims = w.shape
assert f == num_filters, "f is %d but num_filters is %d" % (f, num_filters)
if edge.conv:
assert numdims == size ** 2 * num_colors
input_t = edge.input_t
numImages, numdims = input_t.shape
assert numdims % num_colors == 0
x = int(np.sqrt(numdims / num_colors))
assert x ** 2 == numdims / num_colors
n_locs = (x + 2 * padding - size) / stride + 1
# pdb.set_trace()
# Incoming gradient.
deriv.transpose(edge.output_t2)
input_grads = edge.output_t2
# Output activation (after conv + pool? + norm?)
output_acts = edge.output_t
if conv.rnorm:
# ResponseNormUndo overwrites input_acts, so make a copy.
input_acts = edge.rnorm_temp1
input_acts.assign(edge.unrnormalized_layer)
output_grads = edge.rnorm_temp2
denoms = edge.denoms
sizeX = conv.norm_size
pow_scale = conv.pow_scale
add_scale = conv.add_scale
cc.ResponseNormUndo(
input_grads, denoms, output_acts, input_acts, output_grads, num_filters, sizeX, add_scale, pow_scale
)
input_grads = output_grads
output_acts = edge.unrnormalized_layer
if conv.max_pool:
input_acts = edge.unpooled_layer
output_grads = edge.unpooled_layer
# It's OK to overwrite input_acts because we don't need it later.
n_pool_locs = (n_locs - conv.pool_size) / conv.pool_stride + 1
sizeX = conv.pool_size
strideX = conv.pool_stride
cc.MaxPoolUndo(output_grads, input_acts, input_grads, output_acts, sizeX, 0, strideX, n_pool_locs)
input_grads = output_grads
output_acts = input_acts
# pdb.set_trace()
if self.is_input:
return
output_grads = edge.input_t2
if edge.conv:
cc.convDown(input_grads, w, output_grads, n_locs, padding, stride, size, x, num_colors)
else:
cc.localDown(input_grads, w, output_grads, n_locs, padding, stride, size, x, num_colors)
output_grads.transpose(self.deriv)
def AccumulateConvDeriv(layer, edge, deriv):
"""Accumulate the derivative w.r.t the outputs of this layer.
Each layer needs to compute derivatives w.r.t its outputs. These outputs may
have been connected to lots of other nodes through outgoing edges.
This method adds up the derivatives contributed by each outgoing edge.
It gets derivatives w.r.t the inputs at the other end of an outgoing edge.
Args:
edge: The edge which is sending the derivative.
deriv: The derivative w.r.t the inputs at the other end of this edge.
"""
if layer.dirty: # If some derivatives have already been received.
raise Exception('Not implemented.')
layer.dirty = True
w = edge.params['weight']
conv = edge.conv_params
size = conv.size
stride = conv.stride
padding = conv.padding
num_filters = conv.num_filters
num_colors = conv.num_colors
input_t = edge.input_t
numImages, numdims = input_t.shape
assert numdims % num_colors == 0
x = int(math.sqrt(numdims / num_colors))
assert x**2 == numdims/num_colors
n_locs = (x + 2 * padding - size) / stride + 1
# Incoming gradient.
deriv.transpose(edge.output_t2)
input_grads = edge.output_t2
# Output activation (after conv + pool? + norm?)
output_acts = edge.output_t
if conv.rnorm:
# ResponseNormUndo overwrites input_acts, so make a copy.
input_acts = edge.rnorm_temp1
input_acts.assign(edge.unrnormalized_layer)
output_grads = edge.rnorm_temp2
denoms = edge.denoms
sizeX = conv.norm_size
pow_scale = conv.pow_scale
add_scale = conv.add_scale
cc.ResponseNormUndo(input_grads, denoms, output_acts, input_acts,
output_grads, num_filters, sizeX, add_scale,
pow_scale)
input_grads = output_grads
output_acts = edge.unrnormalized_layer
if conv.max_pool:
input_acts = edge.unpooled_layer
output_grads = edge.unpooled_layer
# It's OK to overwrite input_acts because we don't need it later.
n_pool_locs = (n_locs - conv.pool_size) / conv.pool_stride + 1
sizeX = conv.pool_size
strideX = conv.pool_stride
cc.MaxPoolUndo(output_grads, input_acts, input_grads, output_acts, sizeX,
0, strideX, n_pool_locs)
input_grads = output_grads
output_acts = input_acts
if layer.is_input:
return
output_grads = edge.input_t2
if edge.conv:
cc.convDown(input_grads, w, output_grads, n_locs, padding, stride, size, x, num_colors)
else:
cc.localDown(input_grads, w, output_grads, n_locs, padding, stride, size, x, num_colors)
output_grads.transpose(layer.deriv)