def grad(self, inputs, dout):
print "here"
inputs, filters = inputs
# if 'Cuda' not in str(type(images)):
# raise TypeError("inputs must be cuda")
# if 'Cuda' not in str(type(filters)):
# raise TypeError("filters must be cuda")
dout, = dout
#dout = cuda.basic_ops.gpu_contiguous(
# cuda.basic_ops.as_cuda_ndarray_variable(dout))
#dout = gpu_contiguous(dout)
# if 'Cuda' not in str(type(dout)):
# raise TypeError("output gradients must be cuda")
filters = filters.dimshuffle(1, 0, 2, 3, 4)
#inputs = inputs.dimshuffle(1, 0, 2, 3, 4)
#print inputs.shape, filters.shape, dout.shape
d_inputs = conv.conv3d_fft(dout, inputs, dout.shape, inputs.shape)
d_filters = conv.conv3d_fft(dout, filters, dout.shape, filters.shape)
return d_inputs, d_filters
def thunk():
inp = inputs[0][0]
filters = inputs[1][0]
# output_shape = self.input_shape
# output_shape[-1] = (output_shape[-1] - 1) * 2 # restore full signal length
# output_shape = tuple(output_shape)
z = outputs[0]
# batch size, input channels, input dim 0, input dim 1
b, ic, i0, i1, i2 = self.input_shape
# output channels, input channels, filter dim 0, filter dim 1
oc, ic_, f0, f1, f2 = self.filter_shape
# Output shape
output_shape = [b, oc, i0 - f0 + 1, i1 - f1 + 1, i2 - f2 + 1]
# only allocate if there is no previous allocation of the right size.
if z[0] is None or z[0].shape != output_shape:
z[0] = cuda.CudaNdarray.zeros(output_shape)
output_pycuda = to_gpuarray(z[0])
print "Perform Conv"
output_pycuda = conv.conv3d_fft(inp, filters,
output_pycuda,
self.input_shape,
self.filter_shape)
print "End of conv Conv"
def thunk():
inp = inputs[0][0]
filters = inputs[1][0]
# output_shape = self.input_shape
# output_shape[-1] = (output_shape[-1] - 1) * 2 # restore full signal length
# output_shape = tuple(output_shape)
z = outputs[0]
# batch size, input channels, input dim 0, input dim 1
b, ic, i0, i1, i2 = self.input_shape
# output channels, input channels, filter dim 0, filter dim 1
oc, ic_, f0, f1, f2 = self.filter_shape
# Output shape
output_shape = [b, oc, i0 - f0 + 1, i1 - f1 + 1, i2 - f2 + 1]
# only allocate if there is no previous allocation of the right size.
if z[0] is None or z[0].shape != output_shape:
z[0] = cuda.CudaNdarray.zeros(output_shape)
output_pycuda = to_gpuarray(z[0])
print "Perform Conv"
output_pycuda = conv.conv3d_fft(inp, filters, output_pycuda,
self.input_shape,
self.filter_shape)
print "End of conv Conv"
def grad(self, inputs, dout):
print "here"
inputs, filters = inputs
# if 'Cuda' not in str(type(images)):
# raise TypeError("inputs must be cuda")
# if 'Cuda' not in str(type(filters)):
# raise TypeError("filters must be cuda")
dout, = dout
#dout = cuda.basic_ops.gpu_contiguous(
# cuda.basic_ops.as_cuda_ndarray_variable(dout))
#dout = gpu_contiguous(dout)
# if 'Cuda' not in str(type(dout)):
# raise TypeError("output gradients must be cuda")
filters = filters.dimshuffle(1, 0, 2, 3, 4)
#inputs = inputs.dimshuffle(1, 0, 2, 3, 4)
#print inputs.shape, filters.shape, dout.shape
d_inputs = conv.conv3d_fft(dout, inputs, dout.shape, inputs.shape)
d_filters = conv.conv3d_fft(dout, filters, dout.shape, filters.shape)
return d_inputs, d_filters
