def conv2d_forward_batch(self, inputs, weights, bias, outputs,
padding, stride):
num_filters = weights.shape[0]
num_images, input_rows, input_cols, num_input_maps = inputs.shape
kernel_shape = weights.shape[1:]
num_output_pixels = outputs.shape[1] * outputs.shape[2]
num_kernel_params = np.prod(kernel_shape)
out_shape = (num_output_pixels, num_filters)
for i in range(num_images):
col = np.zeros((num_output_pixels, num_kernel_params),
dtype=self.dtype)
_cpuop.im2col(inputs[i].reshape(inputs[i].size),
input_rows, input_cols, num_input_maps,
kernel_shape[0], kernel_shape[1],
padding, padding, padding, padding, stride[0],
stride[1], col.reshape(col.size))
reshaped_params = weights.reshape(num_filters, num_kernel_params)
np.dot(col, reshaped_params.T, out=outputs[i].reshape(out_shape))
outputs += bias.reshape((1, 1, 1, num_filters))
def conv2d_backward_batch(self, inputs, params, padding, stride,
in_deltas, out_deltas, dparams, dbias):
num_filters = params.shape[0]
num_images, input_rows, input_cols, num_input_maps = inputs.shape
kernel_shape = params.shape[1:]
num_output_pixels = out_deltas.shape[1] * out_deltas.shape[2]
num_kernel_params = np.prod(kernel_shape)
dparams.fill(0.0)
dbias.fill(0.0)
col = np.zeros((num_output_pixels, num_kernel_params),
dtype=self.dtype)
for i in range(num_images):
_cpuop.im2col(inputs[i].reshape(inputs[i].size),
input_rows, input_cols, num_input_maps,
kernel_shape[0], kernel_shape[1],
padding, padding, padding, padding,
stride[0], stride[1], col.reshape(col.size))
# Compute gradients
reshaped_dparams = dparams.reshape(num_filters, num_kernel_params)
reshaped_out_deltas = out_deltas[i].reshape((num_output_pixels,
num_filters))
self.dot_add_mm(reshaped_out_deltas, col, out=reshaped_dparams,
transa=True)
dbias += np.sum(reshaped_out_deltas, axis=0)
# Compute in_deltas
reshaped_params = params.reshape((num_filters, num_kernel_params))
np.dot(reshaped_out_deltas, reshaped_params, out=col)
_cpuop.col2im(col.reshape(col.size),
input_rows, input_cols, num_input_maps,
kernel_shape[0], kernel_shape[1],
padding, padding, padding, padding,
stride[0], stride[1],
in_deltas[i].reshape(in_deltas[i].size))
