def word_error_rate(raw_predictions, labels, lookup=None, weights_fn=common_layers.weights_nonzero): """Calculate word error rate. Args: raw_predictions: The raw predictions. labels: The actual labels. lookup: A tf.constant mapping indices to output tokens. weights_fn: Weighting function. Returns: The word error rate. """ def from_tokens(raw, lookup_): gathered = tf.gather(lookup_, tf.cast(raw, tf.int32)) joined = tf.regex_replace(tf.reduce_join(gathered, axis=1), b"<EOS>.*", b"") cleaned = tf.regex_replace(joined, b"_", b" ") tokens = tf.string_split(cleaned, " ") return tokens def from_characters(raw, lookup_): """Convert ascii+2 encoded codes to string-tokens.""" corrected = tf.bitcast(tf.clip_by_value(tf.subtract(raw, 2), 0, 255), tf.uint8) gathered = tf.gather(lookup_, tf.cast(corrected, tf.int32))[:, :, 0] joined = tf.reduce_join(gathered, axis=1) cleaned = tf.regex_replace(joined, b"\0", b"") tokens = tf.string_split(cleaned, " ") return tokens if lookup is None: lookup = tf.constant([chr(i) for i in range(256)]) convert_fn = from_characters else: convert_fn = from_tokens if weights_fn is not common_layers.weights_nonzero: raise ValueError("Only weights_nonzero can be used for this metric.") with tf.variable_scope("word_error_rate", values=[raw_predictions, labels]): raw_predictions = tf.squeeze(tf.argmax(raw_predictions, axis=-1), axis=(2, 3)) labels = tf.squeeze(labels, axis=(2, 3)) reference = convert_fn(labels, lookup) predictions = convert_fn(raw_predictions, lookup) distance = tf.reduce_sum( tf.edit_distance(predictions, reference, normalize=False)) reference_length = tf.cast(tf.size(reference.values, out_type=tf.int32), dtype=tf.float32) return distance / reference_length, reference_length
def sequence_edit_distance(predictions, labels, weights_fn=common_layers.weights_nonzero): """Average edit distance, ignoring padding 0s. The score returned is the edit distance divided by the total length of reference truth and the weight returned is the total length of the truth. Args: predictions: Tensor of shape [`batch_size`, `length`, 1, `num_classes`] and type tf.float32 representing the logits, 0-padded. labels: Tensor of shape [`batch_size`, `length`, 1, 1] and type tf.int32 representing the labels of same length as logits and 0-padded. weights_fn: ignored. The weights returned are the total length of the ground truth labels, excluding 0-paddings. Returns: (edit distance / reference length, reference length) Raises: ValueError: if weights_fn is not common_layers.weights_nonzero. """ if weights_fn is not common_layers.weights_nonzero: raise ValueError("Only weights_nonzero can be used for this metric.") with tf.variable_scope("edit_distance", values=[predictions, labels]): # Transform logits into sequence classes by taking max at every step. predictions = tf.to_int32( tf.squeeze(tf.argmax(predictions, axis=-1), axis=(2, 3))) nonzero_idx = tf.where(tf.not_equal(predictions, 0)) sparse_outputs = tf.SparseTensor( nonzero_idx, tf.gather_nd(predictions, nonzero_idx), tf.shape(predictions, out_type=tf.int64)) labels = tf.squeeze(labels, axis=(2, 3)) nonzero_idx = tf.where(tf.not_equal(labels, 0)) label_sparse_outputs = tf.SparseTensor( nonzero_idx, tf.gather_nd(labels, nonzero_idx), tf.shape(labels, out_type=tf.int64)) distance = tf.reduce_sum( tf.edit_distance(sparse_outputs, label_sparse_outputs, normalize=False)) reference_length = tf.to_float( common_layers.shape_list(nonzero_idx)[0]) return distance / reference_length, reference_length
targets = tf.SparseTensor(indices, values, shape) # Compute Loss losses = tf.nn.ctc_loss(targets, logits, SeqLens) loss = tf.reduce_mean(losses) TrainLoss_s = tf.summary.scalar('TrainLoss', loss) # CTC Beam Search Decoder to decode pred string from the prob map decoded, log_prob = tf.nn.ctc_beam_search_decoder(logits, SeqLens) predicted = tf.to_int32(decoded[0]) error_rate = tf.reduce_sum( tf.edit_distance(predicted, targets, normalize=False)) / tf.to_float( tf.size(targets.values)) TrainError_s = tf.summary.scalar('TrainError', error_rate) tvars = tf.trainable_variables() grad, _ = tf.clip_by_global_norm(tf.gradients(loss, tvars), cfg.MaxGradientNorm) optimizer = tf.train.AdamOptimizer(learning_rate=cfg.LearningRate) train_step = optimizer.apply_gradients(zip(grad, tvars)) #These values are used to draw performance graphs. Updated after each epoch. OverallTrainingLoss = tf.Variable(0,