import os import sys from typing import List, Tuple import tf_keras as keras import numpy as np from dataloader_iam import Batch import tensorflow.compat.v1 as tf tf.compat.v1.disable_v2_behavior # Disable eager mode tf.compat.v1.disable_eager_execution class DecoderType: """ CTC decoder types. """ BestPath = 0 BeamSearch = 1 WordBeamSearch = 2 class Model: """ Minimalistic TF model for HTR. """ def __init__(self, char_list: List[str], model_dir: str, decoder_type: str = DecoderType.BestPath, must_restore: bool = False, dump: bool = False) -> None: """ Init model: add CNN, RNN and CTC and initialize TF. """ self.dump = dump self.char_list = char_list self.decoder_type = decoder_type self.must_restore = must_restore self.snap_ID = 0 self.model_dir = model_dir tf.compat.v1.disable_eager_execution() # Whether to use normalization over a batch or a population self.is_train = tf.compat.v1.placeholder(tf.bool, name='is_train') # input image batch self.input_imgs = tf.compat.v1.placeholder(tf.float32, shape=(None, None, None)) # setup CNN, RNN and CTC self.setup_cnn() self.setup_rnn() self.setup_ctc() # setup optimizer to train NN self.batches_trained = 0 self.update_ops = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.UPDATE_OPS) with tf.control_dependencies(self.update_ops): self.optimizer = tf.compat.v1.train.AdamOptimizer().minimize(self.loss) # initialize TF self.sess, self.saver = self.setup_tf() def setup_cnn(self) -> None: """ Create CNN layers. """ cnn_in4d = tf.expand_dims(input=self.input_imgs, axis=3) # list of parameters for the layers kernel_vals = [5, 5, 3, 3, 3] feature_vals = [1, 32, 64, 128, 128, 256] stride_vals = pool_vals = [(2, 2), (2, 2), (1, 2), (1, 2), (1, 2)] num_layers = len(stride_vals) # create layers pool = cnn_in4d # input to first CNN layer for i in range(num_layers): kernel = tf.Variable( tf.random.truncated_normal([kernel_vals[i], kernel_vals[i], feature_vals[i], feature_vals[i + 1]], stddev=0.1)) conv = tf.nn.conv2d(input=pool, filters=kernel, padding='SAME', strides=(1, 1, 1, 1)) conv_norm = tf.keras.layers.BatchNormalization()(conv, training=self.is_train) relu = tf.nn.relu(conv_norm) pool = tf.nn.max_pool2d(input=relu, ksize=(1, pool_vals[i][0], pool_vals[i][1], 1), strides=(1, stride_vals[i][0], stride_vals[i][1], 1), padding='VALID') self.cnn_out_4d = pool def setup_rnn(self) -> None: """ Create RNN layers. """ rnn_in3d = tf.squeeze(self.cnn_out_4d, axis=[2]) # basic cells which is used to build RNN num_hidden = 256 cells = [tf.compat.v1.nn.rnn_cell.LSTMCell(num_units=num_hidden, state_is_tuple=True) for _ in range(2)] # 2 layers # stack basic cells stacked = tf.compat.v1.nn.rnn_cell.MultiRNNCell(cells, state_is_tuple=True) # bidirectional RNN # BxTxF -> BxTx2H (fw, bw), _ = tf.compat.v1.nn.bidirectional_dynamic_rnn(cell_fw=stacked, cell_bw=stacked, inputs=rnn_in3d, dtype=rnn_in3d.dtype) # BxTxH + BxTxH -> BxTx2H -> BxTx1X2H concat = tf.expand_dims(tf.concat([fw, bw], 2), 2) # project output to chars (including blank): BxTx1x2H -> BxTx1xC -> BxTxC kernel = tf.Variable(tf.random.truncated_normal([1, 1, num_hidden * 2, len(self.char_list) + 1], stddev=0.1)) self.rnn_out_3d = tf.squeeze(tf.nn.atrous_conv2d(value=concat, filters=kernel, rate=1, padding='SAME'), axis=[2]) def setup_ctc(self) -> None: """ Create CTC loss and decoder. """ # BxTxC -> TxBxC self.ctc_in_3d_tbc = tf.transpose(a=self.rnn_out_3d, perm=[1, 0, 2]) # ground truth text as sparse tensor self.gt_texts = tf.SparseTensor(tf.compat.v1.placeholder(tf.int64, shape=[None, 2]), tf.compat.v1.placeholder(tf.int32, [None]), tf.compat.v1.placeholder(tf.int64, [2])) # calc loss for batch self.seq_len = tf.compat.v1.placeholder(tf.int32, [None]) self.loss = tf.reduce_mean( input_tensor=tf.compat.v1.nn.ctc_loss(labels=self.gt_texts, inputs=self.ctc_in_3d_tbc, sequence_length=self.seq_len, ctc_merge_repeated=True)) # calc loss for each element to compute label probability self.saved_ctc_input = tf.compat.v1.placeholder(tf.float32, shape=[None, None, len(self.char_list) + 1]) self.loss_per_element = tf.compat.v1.nn.ctc_loss(labels=self.gt_texts, inputs=self.saved_ctc_input, sequence_length=self.seq_len, ctc_merge_repeated=True) # best path decoding or beam search decoding if self.decoder_type == DecoderType.BestPath: self.decoder = tf.nn.ctc_greedy_decoder(inputs=self.ctc_in_3d_tbc, sequence_length=self.seq_len) elif self.decoder_type == DecoderType.BeamSearch: self.decoder = tf.nn.ctc_beam_search_decoder(inputs=self.ctc_in_3d_tbc, sequence_length=self.seq_len, beam_width=50) # word beam search decoding (see https://github.com/githubharald/CTCWordBeamSearch) elif self.decoder_type == DecoderType.WordBeamSearch: # prepare information about language (dictionary, characters in dataset, characters forming words) chars = ''.join(self.char_list) word_chars = open('../model/wordCharList.txt').read().splitlines()[0] corpus = open('../data/corpus.txt').read() # decode using the "Words" mode of word beam search from word_beam_search import WordBeamSearch self.decoder = WordBeamSearch(50, 'Words', 0.0, corpus.encode('utf8'), chars.encode('utf8'), word_chars.encode('utf8')) # the input to the decoder must have softmax already applied self.wbs_input = tf.nn.softmax(self.ctc_in_3d_tbc, axis=2) def setup_tf(self) -> Tuple[tf.compat.v1.Session, tf.compat.v1.train.Saver]: """ Initialize TF. """ print('Python: ' + sys.version) print('Tensorflow: ' + tf.__version__) sess = tf.compat.v1.Session() # TF session saver = tf.compat.v1.train.Saver(max_to_keep=1) # saver saves model to file latest_snapshot = tf.train.latest_checkpoint(self.model_dir ) # is there a saved model? # if model must be restored (for inference), there must be a snapshot if self.must_restore and not latest_snapshot: raise Exception('No saved model found in: ' + model_dir) # load saved model if available if latest_snapshot: print('Init with stored values from ' + latest_snapshot) saver.restore(sess, latest_snapshot) else: print('Init with new values') sess.run(tf.compat.v1.global_variables_initializer()) return sess, saver def to_sparse(self, texts: List[str]) -> Tuple[List[List[int]], List[int], List[int]]: """ Put ground truth texts into sparse tensor for ctc_loss. """ indices = [] values = [] shape = [len(texts), 0] # last entry must be max(labelList[i]) # go over all texts for batchElement, text in enumerate(texts): # convert to string of label (i.e. class-ids) label_str = [self.char_list.index(c) for c in text] # sparse tensor must have size of max. label-string if len(label_str) > shape[1]: shape[1] = len(label_str) # put each label into sparse tensor for i, label in enumerate(label_str): indices.append([batchElement, i]) values.append(label) return indices, values, shape def decoder_output_to_text(self, ctc_output: tuple, batch_size: int) -> List[str]: """ Extract texts from output of CTC decoder. """ # word beam search: already contains label strings if self.decoder_type == DecoderType.WordBeamSearch: label_strs = ctc_output # TF decoders: label strings are contained in sparse tensor else: # ctc returns tuple, first element is SparseTensor decoded = ctc_output[0][0] # contains string of labels for each batch element label_strs = [[] for _ in range(batch_size)] # go over all indices and save mapping: batch -> values for (idx, idx2d) in enumerate(decoded.indices): label = decoded.values[idx] batch_element = idx2d[0] # index according to [b,t] label_strs[batch_element].append(label) # map labels to chars for all batch elements return [''.join([self.char_list[c] for c in labelStr]) for labelStr in label_strs] def train_batch(self, batch: Batch) -> float: """ Feed a batch into the NN to train it. """ num_batch_elements = len(batch.imgs) max_text_len = batch.imgs[0].shape[0] // 4 sparse = self.to_sparse(batch.gt_texts) eval_list = [self.optimizer, self.loss] feed_dict = {self.input_imgs: batch.imgs, self.gt_texts: sparse, self.seq_len: [max_text_len] * num_batch_elements, self.is_train: True} _, loss_val = self.sess.run(eval_list, feed_dict) self.batches_trained += 1 return loss_val @staticmethod def dump_nn_output(rnn_output: np.ndarray) -> None: """ Dump the output of the NN to CSV file(s). """ dump_dir = '../dump/' if not os.path.isdir(dump_dir): os.mkdir(dump_dir) # iterate over all batch elements and create a CSV file for each one max_t, max_b, max_c = rnn_output.shape for b in range(max_b): csv = '' for t in range(max_t): for c in range(max_c): csv += str(rnn_output[t, b, c]) + ';' csv += '\n' fn = dump_dir + 'rnnOutput_' + str(b) + '.csv' print('Write dump of NN to file: ' + fn) with open(fn, 'w') as f: f.write(csv) def infer_batch(self, batch: Batch, calc_probability: bool = False, probability_of_gt: bool = False): """ Feed a batch into the NN to recognize the texts. """ # decode, optionally save RNN output num_batch_elements = len(batch.imgs) # put tensors to be evaluated into list eval_list = [] if self.decoder_type == DecoderType.WordBeamSearch: eval_list.append(self.wbs_input) else: eval_list.append(self.decoder) if self.dump or calc_probability: eval_list.append(self.ctc_in_3d_tbc) # sequence length depends on input image size (model downsizes width by 4) max_text_len = batch.imgs[0].shape[0] // 4 # dict containing all tensor fed into the model feed_dict = {self.input_imgs: batch.imgs, self.seq_len: [max_text_len] * num_batch_elements, self.is_train: False} # evaluate model eval_res = self.sess.run(eval_list, feed_dict) # TF decoders: decoding already done in TF graph if self.decoder_type != DecoderType.WordBeamSearch: decoded = eval_res[0] # word beam search decoder: decoding is done in C++ function compute() else: decoded = self.decoder.compute(eval_res[0]) # map labels (numbers) to character string texts = self.decoder_output_to_text(decoded, num_batch_elements) # feed RNN output and recognized text into CTC loss to compute labeling probability probs = None if calc_probability: sparse = self.to_sparse(batch.gt_texts) if probability_of_gt else self.to_sparse(texts) ctc_input = eval_res[1] eval_list = self.loss_per_element feed_dict = {self.saved_ctc_input: ctc_input, self.gt_texts: sparse, self.seq_len: [max_text_len] * num_batch_elements, self.is_train: False} loss_vals = self.sess.run(eval_list, feed_dict) probs = np.exp(-loss_vals) # dump the output of the NN to CSV file(s) if self.dump: self.dump_nn_output(eval_res[1]) return texts, probs def save(self) -> None: """ Save model to file. """ self.snap_ID += 1 self.saver.save(self.sess, '../model/snapshot', global_step=self.snap_ID)