diff --git a/util/feeding.py b/util/feeding.py index 817338cf..829d2ffe 100644 --- a/util/feeding.py +++ b/util/feeding.py @@ -16,7 +16,7 @@ from util.config import Config from util.logging import log_error from util.text import text_to_char_array from util.flags import FLAGS -from util.spectrogram_augmentations import augment_sparse_deform, augment_freq_time_mask, augment_dropout, augment_pitch_and_tempo, augment_speed_up +from util.spectrogram_augmentations import augment_freq_time_mask, augment_dropout, augment_pitch_and_tempo, augment_speed_up def read_csvs(csv_files): source_data = None @@ -40,11 +40,6 @@ def samples_to_mfccs(samples, sample_rate, train_phase=False): # Data Augmentations if train_phase: - if FLAGS.augmention_sparse_deform: - spectrogram = augment_sparse_deform(spectrogram, - time_warping_para=FLAGS.augmentation_time_warp_max_warping, - normal_around_warping_std=FLAGS.augmentation_sparse_deform_std_warp) - if FLAGS.augmentation_spec_dropout_keeprate < 1: spectrogram = augment_dropout(spectrogram, keep_prob=FLAGS.augmentation_spec_dropout_keeprate) diff --git a/util/flags.py b/util/flags.py index 8ec0ce60..faa59f7a 100644 --- a/util/flags.py +++ b/util/flags.py @@ -27,13 +27,9 @@ def create_flags(): f.DEFINE_float('data_aug_features_additive', 0, 'std of the Gaussian additive noise') f.DEFINE_float('data_aug_features_multiplicative', 0, 'std of normal distribution around 1 for multiplicative noise') - f.DEFINE_integer('augmention_sparse_deform', 0, 'whether to use time-warping augmentation') - f.DEFINE_integer('augmentation_time_warp_max_warping', 12, 'max value for warping') - f.DEFINE_float('augmentation_sparse_deform_std_warp', 0.5, 'std for warping different values to different frequencies') - f.DEFINE_float('augmentation_spec_dropout_keeprate', 1, 'keep rate of dropout augmentation on spectrogram (if 1, no dropout will be performed on spectrogram)') - f.DEFINE_integer('augmentation_freq_and_time_masking', 0, 'whether to use frequency and time masking augmentation') + f.DEFINE_boolean('augmentation_freq_and_time_masking', False, 'whether to use frequency and time masking augmentation') f.DEFINE_integer('augmentation_freq_and_time_masking_freq_mask_range', 5, 'max range of masks in the frequency domain when performing freqtime-mask augmentation') f.DEFINE_integer('augmentation_freq_and_time_masking_number_freq_masks', 3, 'number of masks in the frequency domain when performing freqtime-mask augmentation') f.DEFINE_integer('augmentation_freq_and_time_masking_time_mask_range', 2, 'max range of masks in the time domain when performing freqtime-mask augmentation') @@ -41,7 +37,7 @@ def create_flags(): f.DEFINE_float('augmentation_speed_up_std', 0, 'std for speeding-up tempo. If std is 0, this augmentation is not performed') - f.DEFINE_integer('augmentation_pitch_and_tempo_scaling', 0, 'whether to use spectrogram speed and tempo scaling') + f.DEFINE_boolean('augmentation_pitch_and_tempo_scaling', False, 'whether to use spectrogram speed and tempo scaling') f.DEFINE_float('augmentation_pitch_and_tempo_scaling_min_pitch', 0.95, 'min value of pitch scaling') f.DEFINE_float('augmentation_pitch_and_tempo_scaling_max_pitch', 1.2, 'max value of pitch scaling') f.DEFINE_float('augmentation_pitch_and_tempo_scaling_max_tempo', 1.2, 'max vlaue of tempo scaling') diff --git a/util/sparse_image_warp.py b/util/sparse_image_warp.py deleted file mode 100644 index 2bd69d45..00000000 --- a/util/sparse_image_warp.py +++ /dev/null @@ -1,177 +0,0 @@ -## Implementation of sparse_image_warp that handles dynamic shapes -from tensorflow.contrib.image.python.ops import dense_image_warp -from tensorflow.contrib.image.python.ops import interpolate_spline - -from tensorflow.python.framework import dtypes -from tensorflow.python.framework import ops -from tensorflow.python.framework import tensor_shape -from tensorflow.python.ops import array_ops -from tensorflow.python.ops import math_ops - - -def _get_grid_locations(image_height, image_width): - """Wrapper for array_ops.meshgrid.""" - - y_range = math_ops.linspace(0.0, math_ops.to_float(image_height) - 1, - image_height) - x_range = math_ops.linspace(0.0, math_ops.to_float(image_width) - 1, - image_width) - y_grid, x_grid = array_ops.meshgrid(y_range, x_range, indexing='ij') - return array_ops.stack((y_grid, x_grid), -1) - - -def _expand_to_minibatch(array, batch_size): - """Tile arbitrarily-sized array to include new batch dimension.""" - batch_size = array_ops.expand_dims(batch_size, 0) - array_ones = array_ops.ones((array_ops.rank(array)), dtype=dtypes.int32) - tiles = array_ops.concat([batch_size, array_ones], axis=0) - return array_ops.tile(array_ops.expand_dims(array, 0), tiles) - - -def _get_boundary_locations(image_height, image_width, num_points_per_edge): - """Compute evenly-spaced indices along edge of image.""" - image_height = math_ops.to_float(image_height) - image_width = math_ops.to_float(image_width) - y_range = math_ops.linspace(0.0, image_height - 1, num_points_per_edge + 2) - x_range = math_ops.linspace(0.0, image_width - 1, num_points_per_edge + 2) - ys, xs = array_ops.meshgrid(y_range, x_range, indexing='ij') - is_boundary = math_ops.logical_or( - math_ops.logical_or(math_ops.equal(xs, 0), # pylint: disable=bad-continuation - math_ops.equal(xs, image_width - 1)), - math_ops.logical_or(math_ops.equal(ys, 0), # pylint: disable=bad-continuation - math_ops.equal(ys, image_height - 1))) - return array_ops.stack([array_ops.boolean_mask(ys, is_boundary), - array_ops.boolean_mask(xs, is_boundary)], axis=-1) - - -def _add_zero_flow_controls_at_boundary(control_point_locations, - control_point_flows, image_height, - image_width, boundary_points_per_edge): - """Add control points for zero-flow boundary conditions. - Augment the set of control points with extra points on the - boundary of the image that have zero flow. - Args: - control_point_locations: input control points - control_point_flows: their flows - image_height: image height - image_width: image width - boundary_points_per_edge: number of points to add in the middle of each - edge (not including the corners). - The total number of points added is - 4 + 4*(boundary_points_per_edge). - Returns: - merged_control_point_locations: augmented set of control point locations - merged_control_point_flows: augmented set of control point flows - """ - - batch_size = tensor_shape.dimension_value(control_point_locations.shape[0]) - - boundary_point_locations = _get_boundary_locations(image_height, image_width, - boundary_points_per_edge) - - boundary_point_flows = array_ops.zeros([array_ops.shape(boundary_point_locations)[0], 2]) - - boundary_point_locations = _expand_to_minibatch(boundary_point_locations, - batch_size) - - boundary_point_flows = _expand_to_minibatch(boundary_point_flows, batch_size) - - merged_control_point_locations = array_ops.concat([control_point_locations, boundary_point_locations], 1) - - merged_control_point_flows = array_ops.concat([control_point_flows, boundary_point_flows], 1) - - return merged_control_point_locations, merged_control_point_flows - - -def sparse_image_warp(image, - source_control_point_locations, - dest_control_point_locations, - interpolation_order=2, - regularization_weight=0.0, - num_boundary_points=0, - name='sparse_image_warp'): - """Image warping using correspondences between sparse control points. - Apply a non-linear warp to the image, where the warp is specified by - the source and destination locations of a (potentially small) number of - control points. First, we use a polyharmonic spline - (`tf.contrib.image.interpolate_spline`) to interpolate the displacements - between the corresponding control points to a dense flow field. - Then, we warp the image using this dense flow field - (`tf.contrib.image.dense_image_warp`). - Let t index our control points. For regularization_weight=0, we have: - warped_image[b, dest_control_point_locations[b, t, 0], - dest_control_point_locations[b, t, 1], :] = - image[b, source_control_point_locations[b, t, 0], - source_control_point_locations[b, t, 1], :]. - For regularization_weight > 0, this condition is met approximately, since - regularized interpolation trades off smoothness of the interpolant vs. - reconstruction of the interpolant at the control points. - See `tf.contrib.image.interpolate_spline` for further documentation of the - interpolation_order and regularization_weight arguments. - Args: - image: `[batch, height, width, channels]` float `Tensor` - source_control_point_locations: `[batch, num_control_points, 2]` float - `Tensor` - dest_control_point_locations: `[batch, num_control_points, 2]` float - `Tensor` - interpolation_order: polynomial order used by the spline interpolation - regularization_weight: weight on smoothness regularizer in interpolation - num_boundary_points: How many zero-flow boundary points to include at - each image edge.Usage: - num_boundary_points=0: don't add zero-flow points - num_boundary_points=1: 4 corners of the image - num_boundary_points=2: 4 corners and one in the middle of each edge - (8 points total) - num_boundary_points=n: 4 corners and n-1 along each edge - name: A name for the operation (optional). - Note that image and offsets can be of type tf.half, tf.float32, or - tf.float64, and do not necessarily have to be the same type. - Returns: - warped_image: `[batch, height, width, channels]` float `Tensor` with same - type as input image. - flow_field: `[batch, height, width, 2]` float `Tensor` containing the dense - flow field produced by the interpolation. - """ - - image = ops.convert_to_tensor(image) - source_control_point_locations = ops.convert_to_tensor( - source_control_point_locations) - dest_control_point_locations = ops.convert_to_tensor( - dest_control_point_locations) - - control_point_flows = ( - dest_control_point_locations - source_control_point_locations) - - clamp_boundaries = num_boundary_points > 0 - boundary_points_per_edge = num_boundary_points - 1 - - with ops.name_scope(name): - batch_size, image_height, image_width = (array_ops.shape(image)[0], - array_ops.shape(image)[1], - array_ops.shape(image)[2]) - # This generates the dense locations where the interpolant - # will be evaluated. - grid_locations = _get_grid_locations(image_height, image_width) - - flattened_grid_locations = array_ops.reshape(grid_locations, - [image_height*image_width, 2]) - - flattened_grid_locations = _expand_to_minibatch(flattened_grid_locations, - batch_size) - - if clamp_boundaries: - (dest_control_point_locations, - control_point_flows) = _add_zero_flow_controls_at_boundary(dest_control_point_locations, - control_point_flows, image_height, - image_width, boundary_points_per_edge) - - flattened_flows = interpolate_spline.interpolate_spline(dest_control_point_locations, control_point_flows, - flattened_grid_locations, interpolation_order, - regularization_weight) - - dense_flows = array_ops.reshape(flattened_flows, - [batch_size, image_height, image_width, 2]) - - warped_image = dense_image_warp.dense_image_warp(image, dense_flows) - - return warped_image, dense_flows \ No newline at end of file diff --git a/util/spectrogram_augmentations.py b/util/spectrogram_augmentations.py index 98159826..9cf36a24 100644 --- a/util/spectrogram_augmentations.py +++ b/util/spectrogram_augmentations.py @@ -1,33 +1,4 @@ import tensorflow as tf -from util.sparse_image_warp import sparse_image_warp - -def augment_sparse_deform(mel_spectrogram, - time_warping_para=12, - normal_around_warping_std=0.5): - mel_spectrogram = tf.expand_dims(mel_spectrogram, -1) - freq_max = tf.shape(mel_spectrogram)[1] - time_max = tf.shape(mel_spectrogram)[2] - center_freq = tf.cast(freq_max, tf.float32)/2.0 - random_time_point = tf.random.uniform(shape=(), minval=time_warping_para, maxval=tf.cast(time_max, tf.float32) - time_warping_para) - chosen_warping = tf.random.uniform(shape=(), minval=0, maxval=time_warping_para) - #add different warping values to different frequencies - normal_around_warping = tf.random.normal(mean=chosen_warping, stddev=normal_around_warping_std, shape=(3,)) - - control_point_freqs = tf.stack([0.0, center_freq, tf.cast(freq_max, tf.float32)], axis=0) - control_point_times_src = tf.stack([random_time_point, random_time_point, random_time_point], axis=0) - control_point_times_dst = control_point_times_src+normal_around_warping - - control_src = tf.expand_dims(tf.stack([control_point_freqs, control_point_times_src], axis=-1), 0) - control_dst = tf.expand_dims(tf.stack([control_point_freqs, control_point_times_dst], axis=1), 0) - warped_mel_spectrogram, _ = sparse_image_warp(mel_spectrogram, - source_control_point_locations=control_src, - dest_control_point_locations=control_dst, - interpolation_order=2, - regularization_weight=0, - num_boundary_points=1 - ) - warped_mel_spectrogram = warped_mel_spectrogram[:, :, :, 0] - return warped_mel_spectrogram def augment_freq_time_mask(mel_spectrogram, frequency_masking_para=30,