def testCausalConv2DLayerStridedWithPaddingFPropV2(self, seq_len):
"""Check strided convs get the same values for different length dim."""
with self.session(use_gpu=True):
batch_size = 5
expected_seq_len = 3
params = conv_layers.CausalConv2DLayerWithPadding.Params()
params.v2_padding = True
params.weight_norm = False
params.filter_stride = [2, 2]
params.name = 'conv'
params.filter_shape = [3, 1, 1, 1]
params.params_init = py_utils.WeightInit.Constant(1.0)
conv_layer = params.Instantiate()
# Set up the padding for the sequence length. (starting at 5).
in_padding = tf.constant([
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 1],
[0, 0, 0, 1, 1],
[0, 0, 1, 1, 1],
[0, 1, 1, 1, 1],
], tf.float32)
in_padding = tf.pad(
in_padding, [[0, 0], [0, seq_len - 5]], constant_values=1.0)
inputs = 1.0 + tf.tile(
tf.reshape(tf.range(seq_len, dtype=tf.float32), [1, seq_len, 1, 1]),
[batch_size, 1, 3, 1])
inputs = py_utils.ApplyPadding(
tf.reshape(in_padding, [batch_size, seq_len, 1, 1]), inputs)
inputs = py_utils.Debug(inputs)
output, out_padding = conv_layer.FPropDefaultTheta(inputs, in_padding)
output = py_utils.Debug(output)
out_padding = py_utils.Debug(out_padding)
self.evaluate(tf.global_variables_initializer())
output, out_padding = self.evaluate([output, out_padding])
self.assertEqual((batch_size, expected_seq_len, 2, 1), output.shape)
self.assertAllClose([
[0, 0, 0],
[0, 0, 1],
[0, 0, 1],
[0, 1, 1],
[0, 1, 1],
], out_padding)
self.assertAllClose(
[
[[[1], [1]], [[6], [6]], [[12], [12]]],
[[[1], [1]], [[6], [6]], [[7], [7]]],
[[[1], [1]], [[6], [6]], [[3], [3]]], # NOTE: not padded.
[[[1], [1]], [[3], [3]], [[0], [0]]],
[[[1], [1]], [[1], [1]], [[0], [0]]],
],
output)
def testComputeConvOutputPaddingV2(self, padding, expected_padding, stride,
padding_algorithm):
"""Test Convolution padding computation."""
padding = tf.constant([padding], tf.float32)
expected_padding = tf.constant([expected_padding], tf.float32)
with self.session(use_gpu=True):
conv_padding = conv_layers._ComputeConvOutputPaddingV2(
padding, window=3, stride=stride, padding_algorithm=padding_algorithm)
self.evaluate(tf.global_variables_initializer())
conv_padding = py_utils.Debug(conv_padding)
conv_padding = self.evaluate(conv_padding)
tf.logging.info('expected_padding {expected_padding}')
self.assertAllClose(expected_padding, conv_padding)
def testConv2DLayerStridedWithPaddingFProp(self, seq_len):
"""Check strided convs get the same values for different length dim."""
# TODO(isaace): THIS TEST SHOWS THAT THERE IS A BUG IN THE CODE.
with self.session(use_gpu=True):
batch_size = 3
expected_seq_len = 3
params = conv_layers.Conv2DLayerWithPadding.Params()
params.weight_norm = False
params.filter_stride = [2, 2]
params.name = 'conv'
params.filter_shape = [3, 3, 1, 1]
params.params_init = py_utils.WeightInit.Constant(1.0)
conv_layer = params.Instantiate()
# Set up the padding for the sequence length. (starting at 5).
in_padding = tf.constant([
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 1],
[0, 0, 0, 1, 1],
], tf.float32)
in_padding = tf.pad(
in_padding, [[0, 0], [0, seq_len - 5]], constant_values=1.0)
inputs = 1.0 + tf.tile(
tf.reshape(tf.range(seq_len, dtype=tf.float32), [1, seq_len, 1, 1]),
[batch_size, 1, 3, 1])
inputs = py_utils.ApplyPadding(
tf.reshape(in_padding, [batch_size, seq_len, 1, 1]), inputs)
inputs = py_utils.Debug(inputs)
output, out_padding = conv_layer.FPropDefaultTheta(inputs, in_padding)
output = py_utils.Debug(output)
out_padding = py_utils.Debug(out_padding)
self.evaluate(tf.global_variables_initializer())
output, out_padding = self.evaluate([output, out_padding])
self.assertEqual((batch_size, expected_seq_len, 2, 1), output.shape)
self.assertAllClose([
[0, 0, 1],
[0, 0, 1],
[0, 1, 1],
], out_padding)
# This here shows a bug in the implementation; the output should be the
# same. Also there are bugs with the output not having the correct
# padding.
if seq_len == 5:
self.assertAllClose([
[[[6], [6]], [[18], [18]], [[18], [18]]],
[[[6], [6]], [[18], [18]], [[8], [8]]],
[[[6], [6]], [[10], [10]], [[0], [0]]],
], output)
elif seq_len == 6:
self.assertAllClose([
[[[12], [12]], [[24], [24]], [[10], [10]]],
[[[12], [12]], [[14], [14]], [[0], [0]]],
[[[12], [12]], [[6], [6]], [[0], [0]]],
], output)
else:
raise ValueError('Test does not handle length {seq_len}')
