def conv1d_sd(input, filters, image_shape, filter_shape, border_mode='valid',
subsample=(1,), filter_flip=True):
"""
using a single dot product
"""
if border_mode not in ('valid', 0, (0,)):
raise RuntimeError("Unsupported border_mode for conv1d_sd: "
"%s" % border_mode)
batch_size, num_input_channels, input_length = image_shape
num_filters, num_input_channels_, filter_length = filter_shape
stride = subsample[0]
if filter_length % stride > 0:
raise RuntimeError("Filter length (%d) is not a multiple of the "
"stride (%d)" % (filter_length, stride))
num_steps = filter_length // stride
output_length = (input_length - filter_length + stride) // stride
# pad the input so all the shifted dot products fit inside.
# shape is (b, c, l)
padded_length = ((input_length // filter_length) * filter_length +
(num_steps - 1) * stride)
# at this point, it is possible that the padded_length is SMALLER than the
# input size. so then we have to truncate first.
truncated_length = min(input_length, padded_length)
input_truncated = input[:, :, :truncated_length]
input_padded_shape = (batch_size, num_input_channels, padded_length)
input_padded = T.zeros(input_padded_shape)
input_padded = T.set_subtensor(input_padded[:, :, :truncated_length],
input_truncated)
inputs = []
for num in range(num_steps):
shift = num * stride
length = (padded_length - shift) // filter_length
r_input_shape = (batch_size, num_input_channels, length, filter_length)
r_input = input_padded[
:, :, shift:length * filter_length + shift].reshape(r_input_shape)
inputs.append(r_input)
inputs_stacked = T.stack(*inputs) # shape is (n, b, c, w, f)
filters_flipped = filters[:, :, ::-1] if filter_flip else filters
r_conved = T.tensordot(inputs_stacked, filters_flipped,
np.asarray([[2, 4], [1, 2]]))
# resulting shape is (n, b, w, n_filters)
# output needs to be (b, n_filters, w * n)
r_conved = r_conved.dimshuffle(1, 3, 2, 0) # (b, n_filters, w, n)
conved = r_conved.reshape((r_conved.shape[0], r_conved.shape[1],
r_conved.shape[2] * r_conved.shape[3]))
# result is (b, n_f, l)
# remove padding
return conved[:, :, :output_length]
def unroll_scan(fn, sequences, outputs_info, non_sequences, n_steps,
go_backwards=False):
"""
Helper function to unroll for loops. Can be used to unroll theano.scan.
The parameter names are identical to theano.scan, please refer to here
for more information.
Note that this function does not support the truncate_gradient
setting from theano.scan.
Parameters
----------
fn : function
Function that defines calculations at each step.
sequences : TensorVariable or list of TensorVariables
List of TensorVariable with sequence data. The function iterates
over the first dimension of each TensorVariable.
outputs_info : list of TensorVariables
List of tensors specifying the initial values for each recurrent
value.
non_sequences: list of TensorVariables
List of theano.shared variables that are used in the step function.
n_steps: int
Number of steps to unroll.
go_backwards: bool
If true the recursion starts at sequences[-1] and iterates
backwards.
Returns
-------
List of TensorVariables. Each element in the list gives the recurrent
values at each time step.
"""
if not isinstance(sequences, (list, tuple)):
sequences = [sequences]
# When backwards reverse the recursion direction
counter = range(n_steps)
if go_backwards:
counter = counter[::-1]
output = []
prev_vals = outputs_info
for i in counter:
step_input = [s[i] for s in sequences] + prev_vals + non_sequences
out_ = fn(*step_input)
# The returned values from step can be either a TensorVariable,
# a list, or a tuple. Below, we force it to always be a list.
if isinstance(out_, T.TensorVariable):
out_ = [out_]
if isinstance(out_, tuple):
out_ = list(out_)
output.append(out_)
prev_vals = output[-1]
# iterate over each scan output and convert it to same format as scan:
# [[output11, output12,...output1n],
# [output21, output22,...output2n],...]
output_scan = []
for i in range(len(output[0])):
l = map(lambda x: x[i], output)
output_scan.append(T.stack(*l))
return output_scan
