def main(unused_argv): vocab, pretrained_matrix = load_glove(vocab_size=100000, embedding_size=300) with tf.Graph().as_default(): image_id, image_features, indicator, word_ids, pointer_ids = import_mscoco( mode="train", batch_size=FLAGS.batch_size, num_epochs=FLAGS.num_epochs, is_mini=FLAGS.is_mini) lengths = tf.reduce_sum(indicator, [1]) show_and_tell_cell = ShowAndTellCell(300) best_first_image_captioner = BestFirstImageCaptioner(show_and_tell_cell, vocab, pretrained_matrix) word_logits, wids, pointer_logits, pids, ids, _lengths = best_first_image_captioner( mean_image_features=image_features, word_ids=word_ids, pointer_ids=pointer_ids, lengths=lengths) tf.losses.sparse_softmax_cross_entropy(pointer_ids, pointer_logits) tf.losses.sparse_softmax_cross_entropy(word_ids, word_logits) loss = tf.losses.get_total_loss() global_step = tf.train.get_or_create_global_step() optimizer = tf.train.AdamOptimizer() learning_step = optimizer.minimize(loss, var_list=best_first_image_captioner.variables, global_step=global_step) captioner_saver = tf.train.Saver(var_list=best_first_image_captioner.variables + [global_step]) captioner_ckpt, captioner_ckpt_name = get_best_first_checkpoint() with tf.Session() as sess: sess.run(tf.variables_initializer(optimizer.variables())) if captioner_ckpt is not None: captioner_saver.restore(sess, captioner_ckpt) else: sess.run(tf.variables_initializer(best_first_image_captioner.variables + [global_step])) captioner_saver.save(sess, captioner_ckpt_name, global_step=global_step) last_save = time.time() for i in itertools.count(): time_start = time.time() try: twids, tpids, _ids, _lengths, _loss, _learning_step = sess.run([ word_ids, pointer_ids, ids, lengths, loss, learning_step]) except: break iteration = sess.run(global_step) insertion_sequence = insertion_sequence_to_array(twids, tpids, _lengths, vocab) print(PRINT_STRING.format( iteration, _loss, list_of_ids_to_string(insertion_sequence[0], vocab), list_of_ids_to_string(twids[0, :].tolist(), vocab), FLAGS.batch_size / (time.time() - time_start))) new_save = time.time() if new_save - last_save > 3600: # save the model every hour captioner_saver.save(sess, captioner_ckpt_name, global_step=global_step) last_save = new_save captioner_saver.save(sess, captioner_ckpt_name, global_step=global_step) print("Finishing training.")
def main(unused_argv): vocab, pretrained_matrix = load_glove(vocab_size=100000, embedding_size=300) with tf.Graph().as_default(): image_id, running_ids, indicator, previous_id, next_id, pointer, image_features = ( import_mscoco(mode="train", batch_size=BATCH_SIZE, num_epochs=100, is_mini=True)) best_first_module = BestFirstModule(pretrained_matrix) pointer_logits, word_logits = best_first_module( image_features, running_ids, previous_id, indicators=indicator, pointer_ids=pointer) tf.losses.sparse_softmax_cross_entropy(pointer, pointer_logits) tf.losses.sparse_softmax_cross_entropy(next_id, word_logits) loss = tf.losses.get_total_loss() ids = tf.argmax(word_logits, axis=-1, output_type=tf.int32) global_step = tf.train.get_or_create_global_step() learning_rate = LEARNING_RATE learning_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, var_list=best_first_module.variables, global_step=global_step) captioner_saver = tf.train.Saver(var_list=best_first_module.variables + [global_step]) captioner_ckpt, captioner_ckpt_name = get_best_first_checkpoint() with tf.Session() as sess: if captioner_ckpt is not None: captioner_saver.restore(sess, captioner_ckpt) else: sess.run(tf.variables_initializer(best_first_module.variables + [global_step])) captioner_saver.save(sess, captioner_ckpt_name, global_step=global_step) last_save = time.time() _ids, _loss, _learning_step = sess.run([ids, loss, learning_step]) for i in itertools.count(): time_start = time.time() try: _ids, _loss, _learning_step = sess.run([ids, loss, learning_step]) except: break iteration = sess.run(global_step) print(PRINT_STRING.format( iteration, _loss, list_of_ids_to_string(_ids.tolist(), vocab), BATCH_SIZE / (time.time() - time_start))) new_save = time.time() if new_save - last_save > 3600: # save the model every hour captioner_saver.save(sess, captioner_ckpt_name, global_step=global_step) last_save = new_save captioner_saver.save(sess, captioner_ckpt_name, global_step=global_step) print("Finishing training.")
def main(unused_argv): vocab, pretrained_matrix = load_glove(vocab_size=100000, embedding_size=300) with tf.Graph().as_default(): image_id, spatial_features, input_seq, target_seq, indicator = ( import_mscoco(mode="train", batch_size=BATCH_SIZE, num_epochs=100, is_mini=True)) visual_sentinel_cell = VisualSentinelCell(300) image_captioner = ImageCaptioner(visual_sentinel_cell, vocab, pretrained_matrix) logits, ids = image_captioner(lengths=tf.reduce_sum(indicator, axis=1), spatial_image_features=spatial_features, seq_inputs=input_seq) tf.losses.sparse_softmax_cross_entropy(target_seq, logits, weights=indicator) loss = tf.losses.get_total_loss() learning_step = tf.train.GradientDescentOptimizer(1.0).minimize( loss, var_list=image_captioner.variables) captioner_saver = tf.train.Saver(var_list=image_captioner.variables) captioner_ckpt, captioner_ckpt_name = get_visual_sentinel_checkpoint() with tf.Session() as sess: sess.run(tf.variables_initializer(image_captioner.variables)) if captioner_ckpt is not None: captioner_saver.restore(sess, captioner_ckpt) captioner_saver.save(sess, captioner_ckpt_name) last_save = time.time() for i in itertools.count(): time_start = time.time() try: _ids, _loss, _learning_step = sess.run( [ids, loss, learning_step]) except: break print( PRINT_STRING.format( i, _loss, list_of_ids_to_string(_ids[0, :].tolist(), vocab), BATCH_SIZE / (time.time() - time_start))) new_save = time.time() if new_save - last_save > 3600: # save the model every hour captioner_saver.save(sess, captioner_ckpt_name) last_save = new_save captioner_saver.save(sess, captioner_ckpt_name) print("Finishing training.")
def main(unused_argv): vocab, pretrained_matrix = load_glove(vocab_size=100000, embedding_size=300) attribute_map = get_visual_attributes() attribute_to_word_lookup_table = vocab.word_to_id( attribute_map.reverse_vocab) with tf.Graph().as_default(): (image_id, image_features, object_features, input_seq, target_seq, indicator, attributes) = import_mscoco(mode="train", batch_size=FLAGS.batch_size, num_epochs=FLAGS.num_epochs, is_mini=FLAGS.is_mini) attribute_detector = AttributeDetector(1000) _, image_attributes, object_attributes = attribute_detector( image_features, object_features) grounded_attribute_cell = GroundedAttributeCell(1024) attribute_captioner = AttributeCaptioner( grounded_attribute_cell, vocab, pretrained_matrix, attribute_to_word_lookup_table) logits, ids = attribute_captioner(lengths=tf.reduce_sum(indicator, axis=1), mean_image_features=image_features, mean_object_features=object_features, seq_inputs=input_seq, image_attributes=image_attributes, object_attributes=object_attributes) tf.losses.sparse_softmax_cross_entropy(target_seq, logits, weights=indicator) loss = tf.losses.get_total_loss() global_step = tf.train.get_or_create_global_step() learning_rate = tf.train.exponential_decay(5e-4, global_step, 3 * 586363 // FLAGS.batch_size, 0.8, staircase=True) optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate) learning_step = optimizer.minimize( loss, var_list=attribute_captioner.variables, global_step=global_step) detector_saver = tf.train.Saver(var_list=attribute_detector.variables + [global_step]) detector_ckpt, detector_ckpt_name = get_attribute_detector_checkpoint() captioner_saver = tf.train.Saver( var_list=attribute_captioner.variables + [global_step]) captioner_ckpt, captioner_ckpt_name = get_grounded_attribute_checkpoint( ) with tf.Session() as sess: sess.run(tf.variables_initializer(optimizer.variables())) if detector_ckpt is not None: detector_saver.restore(sess, detector_ckpt) else: sess.run( tf.variables_initializer(attribute_detector.variables + [global_step])) if captioner_ckpt is not None: captioner_saver.restore(sess, captioner_ckpt) else: sess.run( tf.variables_initializer(attribute_captioner.variables + [global_step])) captioner_saver.save(sess, captioner_ckpt_name, global_step=global_step) last_save = time.time() for i in itertools.count(): time_start = time.time() try: _target, _ids, _loss, _learning_step = sess.run( [target_seq, ids, loss, learning_step]) except: break iteration = sess.run(global_step) print( PRINT_STRING.format( iteration, _loss, list_of_ids_to_string(_ids[0, :].tolist(), vocab), list_of_ids_to_string(_target[0, :].tolist(), vocab), FLAGS.batch_size / (time.time() - time_start))) new_save = time.time() if new_save - last_save > 3600: # save the model every hour captioner_saver.save(sess, captioner_ckpt_name, global_step=global_step) last_save = new_save captioner_saver.save(sess, captioner_ckpt_name, global_step=global_step) print("Finishing training.")
def main(unused_argv): vocab, pretrained_matrix = load_glove(vocab_size=100000, embedding_size=300) with tf.Graph().as_default(): image_id, spatial_features, input_seq, target_seq, indicator = import_mscoco( mode="train", batch_size=FLAGS.batch_size, num_epochs=FLAGS.num_epochs, is_mini=FLAGS.is_mini) image_captioner = ImageCaptioner(SpatialAttentionCell(300), vocab, pretrained_matrix) logits, ids = image_captioner(lengths=tf.reduce_sum(indicator, axis=1), spatial_image_features=spatial_features, seq_inputs=input_seq) tf.losses.sparse_softmax_cross_entropy(target_seq, logits, weights=indicator) loss = tf.losses.get_total_loss() global_step = tf.train.get_or_create_global_step() optimizer = tf.train.AdamOptimizer() learning_step = optimizer.minimize(loss, var_list=image_captioner.variables, global_step=global_step) captioner_saver = tf.train.Saver(var_list=image_captioner.variables + [global_step]) captioner_ckpt, captioner_ckpt_name = get_spatial_attention_checkpoint( ) with tf.Session() as sess: sess.run(tf.variables_initializer(optimizer.variables())) if captioner_ckpt is not None: captioner_saver.restore(sess, captioner_ckpt) else: sess.run( tf.variables_initializer(image_captioner.variables + [global_step])) captioner_saver.save(sess, captioner_ckpt_name, global_step=global_step) last_save = time.time() for i in itertools.count(): time_start = time.time() try: _target, _ids, _loss, _learning_step = sess.run( [target_seq, ids, loss, learning_step]) except: break iteration = sess.run(global_step) print( PRINT_STRING.format( iteration, _loss, list_of_ids_to_string(_ids[0, :].tolist(), vocab), list_of_ids_to_string(_target[0, :].tolist(), vocab), FLAGS.batch_size / (time.time() - time_start))) new_save = time.time() if new_save - last_save > 3600: captioner_saver.save(sess, captioner_ckpt_name, global_step=global_step) last_save = new_save captioner_saver.save(sess, captioner_ckpt_name, global_step=global_step) print("Finishing training.")