def numpy_err_input_update(self):
"""Backpropagate error (will compute err_input).
"""
if not self.need_err_input:
return
self.err_input.map_invalidate()
self.err_output.map_read()
self.weights.map_read()
err_output = reshape(self.err_output.mem, [self.err_output.shape[0], self.err_output.sample_size])
err_input = reshape(self.err_input.mem, [self.err_input.shape[0], self.err_input.sample_size])
if self.weights_transposed:
numpy.dot(err_output, self.weights.mem.transpose(), err_input)
else:
numpy.dot(err_output, self.weights.mem, err_input)
def numpy_weights_update(self):
self.input.map_read()
self.output.map_read()
self.err_output.map_write()
err_output = reshape(self.err_output.mem, [self.err_output.shape[0], self.err_output.sample_size])
inp = reshape(self.input.mem, [self.input.shape[0], self.input.sample_size])
self.gradient_weights.map_write()
if self.weights_transposed:
numpy.dot(inp.transpose(), err_output, self.gradient_weights.mem)
else:
numpy.dot(err_output.transpose(), inp, self.gradient_weights.mem)
self.numpy_update("weights")
def numpy_err_input_update(self):
"""Backpropagate error (will compute err_input).
"""
if not self.need_err_input:
return
self.err_input.map_invalidate()
self.err_output.map_read()
self.weights.map_read()
err_output = reshape(
self.err_output.mem,
[self.err_output.shape[0], self.err_output.sample_size])
err_input = reshape(
self.err_input.mem,
[self.err_input.shape[0], self.err_input.sample_size])
if self.weights_transposed:
numpy.dot(err_output, self.weights.mem.transpose(), err_input)
else:
numpy.dot(err_output, self.weights.mem, err_input)
def numpy_weights_update(self):
self.input.map_read()
self.output.map_read()
self.err_output.map_write()
err_output = reshape(
self.err_output.mem,
[self.err_output.shape[0], self.err_output.sample_size])
inp = reshape(self.input.mem,
[self.input.shape[0], self.input.sample_size])
self.gradient_weights.map_write()
if self.weights_transposed:
numpy.dot(inp.transpose(), err_output, self.gradient_weights.mem)
else:
numpy.dot(err_output.transpose(), inp, self.gradient_weights.mem)
self.numpy_update('weights')
def numpy_run(self):
"""Forward propagation from batch on CPU only.
"""
self.err_input.map_invalidate()
self.err_output.map_read()
out = reshape(self.err_output.mem, self.output_shape)
inp = self.err_input.mem
inp[:] = 0
inp[:, self.padding[1]:self.padding[1] + self.output_shape[1],
self.padding[0]:self.padding[0] + self.output_shape[2], :] = out
def numpy_run(self):
"""Forward propagation from batch on CPU only.
"""
self.err_input.map_invalidate()
self.err_output.map_read()
out = reshape(self.err_output.mem, self.output_shape)
inp = self.err_input.mem
inp[:] = 0
inp[:, self.padding[1]:self.padding[1] + self.output_shape[1],
self.padding[0]:self.padding[0] + self.output_shape[2], :] = out
def numpy_apply_exp(self):
self.output.map_write()
self.max_idx.map_invalidate()
out = self.output.mem
out = reshape(out, (out.shape[0], out.size // out.shape[0]))
for i, sample in enumerate(out):
im = sample.argmax()
self.max_idx[i] = im
m = sample[im]
sample -= m
numpy.exp(sample, sample)
smm = sample.sum()
sample /= smm
def numpy_run(self):
"""Forward propagation from batch on CPU only.
"""
self.output.map_invalidate()
self.input.map_read()
self.weights.map_read()
self.bias.map_read()
mem = numpy.dot(self.input.matrix,
self.weights.mem if self.weights_transposed
else self.weights.mem.transpose())
if self.include_bias:
mem += self.bias.mem
reshape(self.output.mem, mem.shape)[:] = mem[:]
def numpy_apply_exp(self):
self.output.map_write()
self.max_idx.map_invalidate()
out = self.output.mem
out = reshape(out, (out.shape[0], out.size // out.shape[0]))
for i, sample in enumerate(out):
im = sample.argmax()
self.max_idx[i] = im
m = sample[im]
sample -= m
numpy.exp(sample, sample)
smm = sample.sum()
sample /= smm
def numpy_run(self):
"""Forward propagation from batch on CPU only.
"""
self.output.map_invalidate()
self.input.map_read()
self.weights.map_read()
self.bias.map_read()
mem = numpy.dot(
self.input.matrix, self.weights.mem
if self.weights_transposed else self.weights.mem.transpose())
if self.include_bias:
mem += self.bias.mem
reshape(self.output.mem, mem.shape)[:] = mem[:]
