training/utils/preprocess_questions.py [239:365]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - for j in range(len(action_labels[i])): action_labels_mat[i][j] = action_labels[i][j] # pad encoded questions maxQLength = max(len(x) for x in encoded_questions) for qe in encoded_questions: while len(qe) < maxQLength: qe.append(vocab['questionTokenToIdx']['']) # make train/test splits inds = list(range(0, len(idx))) random.shuffle(inds) train_envs = data['splits']['train'] val_envs = data['splits']['val'] test_envs = data['splits']['test'] assert any([x in train_envs for x in test_envs]) == False assert any([x in train_envs for x in val_envs]) == False train_inds = [i for i in inds if envs[i] in train_envs] val_inds = [i for i in inds if envs[i] in val_envs] test_inds = [i for i in inds if envs[i] in test_envs] # TRAIN train_idx = [idx[i] for i in train_inds] train_encoded_questions = [encoded_questions[i] for i in train_inds] train_question_types = [question_types[i] for i in train_inds] train_answers = [answers[i] for i in train_inds] train_envs = [envs[i] for i in train_inds] train_pos_queue = [pos_queue[i] for i in train_inds] train_boxes = [boxes[i] for i in train_inds] train_action_labels = action_labels_mat[train_inds] train_action_lengths = [action_lengths[i] for i in train_inds] # VAL val_idx = [idx[i] for i in val_inds] val_encoded_questions = [encoded_questions[i] for i in val_inds] val_question_types = [question_types[i] for i in val_inds] val_answers = [answers[i] for i in val_inds] val_envs = [envs[i] for i in val_inds] val_pos_queue = [pos_queue[i] for i in val_inds] val_boxes = [boxes[i] for i in val_inds] val_action_labels = action_labels_mat[val_inds] val_action_lengths = [action_lengths[i] for i in val_inds] # TEST test_idx = [idx[i] for i in test_inds] test_encoded_questions = [encoded_questions[i] for i in test_inds] test_question_types = [question_types[i] for i in test_inds] test_answers = [answers[i] for i in test_inds] test_envs = [envs[i] for i in test_inds] test_pos_queue = [pos_queue[i] for i in test_inds] test_boxes = [boxes[i] for i in test_inds] test_action_labels = action_labels_mat[test_inds] test_action_lengths = [action_lengths[i] for i in test_inds] # parse envs all_envs = list(set(envs)) train_env_idx = [all_envs.index(x) for x in train_envs] val_env_idx = [all_envs.index(x) for x in val_envs] test_env_idx = [all_envs.index(x) for x in test_envs] # write h5 files print('Writing hdf5') train_encoded_questions = np.asarray( train_encoded_questions, dtype=np.int16) print('Train', train_encoded_questions.shape) with h5py.File(args.output_train_h5, 'w') as f: f.create_dataset('idx', data=np.asarray(train_idx)) f.create_dataset('questions', data=train_encoded_questions) f.create_dataset('answers', data=np.asarray(train_answers)) f.create_dataset( 'action_labels', data=np.asarray(train_action_labels), dtype=np.int16) f.create_dataset( 'action_lengths', data=np.asarray(train_action_lengths), dtype=np.int16) val_encoded_questions = np.asarray(val_encoded_questions, dtype=np.int16) print('Val', val_encoded_questions.shape) with h5py.File(args.output_val_h5, 'w') as f: f.create_dataset('idx', data=np.asarray(val_idx)) f.create_dataset('questions', data=val_encoded_questions) f.create_dataset('answers', data=np.asarray(val_answers)) f.create_dataset( 'action_labels', data=np.asarray(val_action_labels), dtype=np.int16) f.create_dataset( 'action_lengths', data=np.asarray(val_action_lengths), dtype=np.int16) test_encoded_questions = np.asarray(test_encoded_questions, dtype=np.int16) print('Test', test_encoded_questions.shape) with h5py.File(args.output_test_h5, 'w') as f: f.create_dataset('idx', data=np.asarray(test_idx)) f.create_dataset('questions', data=test_encoded_questions) f.create_dataset('answers', data=np.asarray(test_answers)) f.create_dataset( 'action_labels', data=np.asarray(test_action_labels), dtype=np.int16) f.create_dataset( 'action_lengths', data=np.asarray(test_action_lengths), dtype=np.int16) json.dump({ 'envs': all_envs, 'train_env_idx': train_env_idx, 'val_env_idx': val_env_idx, 'test_env_idx': test_env_idx, 'train_pos_queue': train_pos_queue, 'val_pos_queue': val_pos_queue, 'test_pos_queue': test_pos_queue, 'train_boxes': train_boxes, 'val_boxes': val_boxes, 'test_boxes': test_boxes }, open(args.output_data_json, 'w')) - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - training/utils/preprocess_questions_pkl.py [241:367]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - for j in range(len(action_labels[i])): action_labels_mat[i][j] = action_labels[i][j] # pad encoded questions maxQLength = max(len(x) for x in encoded_questions) for qe in encoded_questions: while len(qe) < maxQLength: qe.append(vocab['questionTokenToIdx']['']) # make train/test splits inds = list(range(0, len(idx))) random.shuffle(inds) train_envs = data['splits']['train'] val_envs = data['splits']['val'] test_envs = data['splits']['test'] assert any([x in train_envs for x in test_envs]) == False assert any([x in train_envs for x in val_envs]) == False train_inds = [i for i in inds if envs[i] in train_envs] val_inds = [i for i in inds if envs[i] in val_envs] test_inds = [i for i in inds if envs[i] in test_envs] # TRAIN train_idx = [idx[i] for i in train_inds] train_encoded_questions = [encoded_questions[i] for i in train_inds] train_question_types = [question_types[i] for i in train_inds] train_answers = [answers[i] for i in train_inds] train_envs = [envs[i] for i in train_inds] train_pos_queue = [pos_queue[i] for i in train_inds] train_boxes = [boxes[i] for i in train_inds] train_action_labels = action_labels_mat[train_inds] train_action_lengths = [action_lengths[i] for i in train_inds] # VAL val_idx = [idx[i] for i in val_inds] val_encoded_questions = [encoded_questions[i] for i in val_inds] val_question_types = [question_types[i] for i in val_inds] val_answers = [answers[i] for i in val_inds] val_envs = [envs[i] for i in val_inds] val_pos_queue = [pos_queue[i] for i in val_inds] val_boxes = [boxes[i] for i in val_inds] val_action_labels = action_labels_mat[val_inds] val_action_lengths = [action_lengths[i] for i in val_inds] # TEST test_idx = [idx[i] for i in test_inds] test_encoded_questions = [encoded_questions[i] for i in test_inds] test_question_types = [question_types[i] for i in test_inds] test_answers = [answers[i] for i in test_inds] test_envs = [envs[i] for i in test_inds] test_pos_queue = [pos_queue[i] for i in test_inds] test_boxes = [boxes[i] for i in test_inds] test_action_labels = action_labels_mat[test_inds] test_action_lengths = [action_lengths[i] for i in test_inds] # parse envs all_envs = list(set(envs)) train_env_idx = [all_envs.index(x) for x in train_envs] val_env_idx = [all_envs.index(x) for x in val_envs] test_env_idx = [all_envs.index(x) for x in test_envs] # write h5 files print('Writing hdf5') train_encoded_questions = np.asarray( train_encoded_questions, dtype=np.int16) print('Train', train_encoded_questions.shape) with h5py.File(args.output_train_h5, 'w') as f: f.create_dataset('idx', data=np.asarray(train_idx)) f.create_dataset('questions', data=train_encoded_questions) f.create_dataset('answers', data=np.asarray(train_answers)) f.create_dataset( 'action_labels', data=np.asarray(train_action_labels), dtype=np.int16) f.create_dataset( 'action_lengths', data=np.asarray(train_action_lengths), dtype=np.int16) val_encoded_questions = np.asarray(val_encoded_questions, dtype=np.int16) print('Val', val_encoded_questions.shape) with h5py.File(args.output_val_h5, 'w') as f: f.create_dataset('idx', data=np.asarray(val_idx)) f.create_dataset('questions', data=val_encoded_questions) f.create_dataset('answers', data=np.asarray(val_answers)) f.create_dataset( 'action_labels', data=np.asarray(val_action_labels), dtype=np.int16) f.create_dataset( 'action_lengths', data=np.asarray(val_action_lengths), dtype=np.int16) test_encoded_questions = np.asarray(test_encoded_questions, dtype=np.int16) print('Test', test_encoded_questions.shape) with h5py.File(args.output_test_h5, 'w') as f: f.create_dataset('idx', data=np.asarray(test_idx)) f.create_dataset('questions', data=test_encoded_questions) f.create_dataset('answers', data=np.asarray(test_answers)) f.create_dataset( 'action_labels', data=np.asarray(test_action_labels), dtype=np.int16) f.create_dataset( 'action_lengths', data=np.asarray(test_action_lengths), dtype=np.int16) json.dump({ 'envs': all_envs, 'train_env_idx': train_env_idx, 'val_env_idx': val_env_idx, 'test_env_idx': test_env_idx, 'train_pos_queue': train_pos_queue, 'val_pos_queue': val_pos_queue, 'test_pos_queue': test_pos_queue, 'train_boxes': train_boxes, 'val_boxes': val_boxes, 'test_boxes': test_boxes }, open(args.output_data_json, 'w')) - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -