Convert tensor to SparseTensor for ctc_loss

Question

Is there a way to convert a dense tensor to a rarefied tensor? Apparently, Tensorflow Estimator.fit does not accept SparseTensors as tags. One of the reasons I would like to pass SparseTensors to Tensorflow Estimator.fit is to use tenorflow ctc_loss. Here is the code:

import dataset_utils import tensorflow as tf import numpy as np from tensorflow.contrib import grid_rnn, learn, layers, framework def grid_rnn_fn(features, labels, mode): input_layer = tf.reshape(features["x"], [-1, 48, 1596]) indices = tf.where(tf.not_equal(labels, tf.constant(0, dtype=tf.int32))) values = tf.gather_nd(labels, indices) sparse_labels = tf.SparseTensor(indices, values, dense_shape=tf.shape(labels, out_type=tf.int64)) cell_fw = grid_rnn.Grid2LSTMCell(num_units=128) cell_bw = grid_rnn.Grid2LSTMCell(num_units=128) bidirectional_grid_rnn = tf.nn.bidirectional_dynamic_rnn(cell_fw, cell_bw, input_layer, dtype=tf.float32) outputs = tf.reshape(bidirectional_grid_rnn[0], [-1, 256]) W = tf.Variable(tf.truncated_normal([256, 80], stddev=0.1, dtype=tf.float32), name='W') b = tf.Variable(tf.constant(0., dtype=tf.float32, shape=[80], name='b')) logits = tf.matmul(outputs, W) + b logits = tf.reshape(logits, [tf.shape(input_layer)[0], -1, 80]) logits = tf.transpose(logits, (1, 0, 2)) loss = None train_op = None if mode != learn.ModeKeys.INFER: #Error occurs here loss = tf.nn.ctc_loss(inputs=logits, labels=sparse_labels, sequence_length=320) ... # returning ModelFnOps def main(_): image_paths, labels = dataset_utils.read_dataset_list('../test/dummy_labels_file.txt') data_dir = "../test/dummy_data/" images = dataset_utils.read_images(data_dir=data_dir, image_paths=image_paths, image_extension='png') print('Done reading images') images = dataset_utils.resize(images, (1596, 48)) images = dataset_utils.transpose(images) labels = dataset_utils.encode(labels) x_train, x_test, y_train, y_test = dataset_utils.split(features=images, test_size=0.5, labels=labels) train_input_fn = tf.estimator.inputs.numpy_input_fn( x={"x": np.array(x_train)}, y=np.array(y_train), num_epochs=1, shuffle=True, batch_size=1 ) classifier = learn.Estimator(model_fn=grid_rnn_fn, model_dir="/tmp/grid_rnn_ocr_model") classifier.fit(input_fn=train_input_fn) 

UPDATE

It turns out that this solution from here converts the dense tensor to sparse:

 indices = tf.where(tf.not_equal(labels, tf.constant(0, dtype=tf.int32))) values = tf.gather_nd(labels, indices) sparse_labels = tf.SparseTensor(indices, values, dense_shape=tf.shape(labels, out_type=tf.int64)) 

However, I ran into this error now raised by ctc_loss:

 ValueError: Shape must be rank 1 but is rank 0 for 'CTCLoss' (op: 'CTCLoss') with input shapes: [?,?,80], [?,2], [?], []. 

I have this code that converts dense labels to sparse ones:

 def convert_to_sparse(labels, dtype=np.int32): indices = [] values = [] for n, seq in enumerate(labels): indices.extend(zip([n] * len(seq), range(len(seq)))) values.extend(seq) indices = np.asarray(indices, dtype=dtype) values = np.asarray(values, dtype=dtype) shape = np.asarray([len(labels), np.asarray(indices).max(0)[1] + 1], dtype=dtype) return indices, values, shape 

I converted y_train to sparse labels and placed the values ​​inside SparseTensor :

 sparse_y_train = convert_to_sparse(y_train) print(tf.SparseTensor( indices=sparse_y_train[0], values=sparse_y_train[1], dense_shape=sparse_y_train )) 

And compared it to the SparseTensor created inside grid_rnn_fn:

 indices = tf.where(tf.not_equal(labels, tf.constant(0, dtype=tf.int32))) values = tf.gather_nd(labels, indices) sparse_labels = tf.SparseTensor(indices, values, dense_shape=tf.shape(labels, out_type=tf.int64)) 

Here is what I got:

For sparse_y_train :

 SparseTensor(indices=Tensor("SparseTensor/indices:0", shape=(33, 2), dtype=int64), values=Tensor("SparseTensor/values:0", shape=(33,), dtype=int32), dense_shape=Tensor("SparseTensor/dense_shape:0", shape=(2,), dtype=int64)) 

For sparse_labels :

 SparseTensor(indices=Tensor("Where:0", shape=(?, 2), dtype=int64), values=Tensor("GatherNd:0", shape=(?,), dtype=int32), dense_shape=Tensor("Shape:0", shape=(2,), dtype=int64)) 

This makes me think that ctc_loss cannot handle SparseTensors as labels with dynamic shapes.

Answer 1

Yes. You can convert the tensor to a sparse tensor and vice versa:

Let sparse be a sparse tensor and dense be a dense tensor.

From rare to dense:

  dense = tf.sparse_to_dense(sparse.indices, sparse.shape, sparse.values) 

From dense to sparse:

 zero = tf.constant(0, dtype=tf.float32) where = tf.not_equal(dense, zero) indices = tf.where(where) values = tf.gather_nd(dense, indices) sparse = tf.SparseTensor(indices, values, dense.shape)