constants_crunchy.py

import os
import pdb

def mkdir_if_not_exists(dir_name):
    if not os.path.exists(dir_name):
        os.mkdir(dir_name)
        
experiment_name = 'QA_classifier_wordvec_xform' #'QA_joint_pretrain_genome_split'

# Global output directory (all subexperiments will be saved here)
global_output_dir = '/home/tanmay/Code/GenVQA/Exp_Results/VQA'

global_experiment_dir = os.path.join(
    global_output_dir,
    experiment_name)

tb_log_dir = os.path.join(
    global_experiment_dir,
    'tensorboard_logdir')

mkdir_if_not_exists(global_output_dir)
mkdir_if_not_exists(global_experiment_dir)
mkdir_if_not_exists(tb_log_dir)

#height and width to which images are resized before feeding into networks
image_size = (224, 224) 

# Token to be used if object or attribute variable is unknown
unknown_token = 'UNK'

# Genome Data paths
data_absolute_path = '/home/ssd/VisualGenome'

image_dir = os.path.join(data_absolute_path, 'cropped_regions_large')
genome_resnet_feat_dir = os.path.join(
    data_absolute_path,
    'cropped_regions_large_resnet_features')

object_labels_json = os.path.join(
    data_absolute_path,
    'restructured/object_labels.json')

attribute_labels_json = os.path.join(
    data_absolute_path,
    'restructured/attribute_labels.json')

regions_json = os.path.join(
    data_absolute_path,
    'restructured/region_with_labels.json')

mean_image_filename = os.path.join(
    data_absolute_path,
    'restructured/mean_image.jpg')

vocab_json = os.path.join(
    data_absolute_path,
    'restructured/vocab_subset.json')

genome_train_subset_region_ids = os.path.join(
    data_absolute_path,
    'restructured/train_subset_region_ids.json')

genome_train_held_out_region_ids = os.path.join(
    data_absolute_path,
    'restructured/train_held_out_region_ids.json')

genome_test_region_ids = os.path.join(
    data_absolute_path,
    'restructured/test_region_ids.json')

num_object_labels = 1000
num_attribute_labels = 1000

# Regions data partition
# First 80% meant to be used for training
# Next 10% is set aside for validation
# Last 10% is to be used for testing
# num_total_regions = 1951768
# num_train_regions = 1561416 # First 80%
# num_val_regions = 195176 # Next 10%
# num_test_regions = num_total_regions \
#                    - num_train_regions \
#                    - num_val_regions 

# Pretrained resnet ckpt
resnet_ckpt = '/home/tanmay/Downloads/pretrained_networks/' + \
              'Resnet/tensorflow-resnet-pretrained-20160509/' + \
              'ResNet-L50.ckpt'

# Pretrained word vectors
word2vec_binary = '/home/tanmay/Code/word2vec/word2vec-api-master/' + \
                  'GoogleNews-vectors-negative300.bin'

word_vector_size = 300
resnet_feat_dim = 2048

# Numpy matrix storing vocabulary word vectors
pretrained_vocab_word_vectors_npy = os.path.join(
    data_absolute_path,
    'restructured/pretrained_vocab_word_vectors.npy')

# Object Attribute Classifier Training Params
region_batch_size = 200
region_num_epochs = 6
region_queue_size = 400
region_regularization_coeff = 1e-5
region_lr = 1e-3
region_log_every_n_iter = 500
region_output_dir = os.path.join(
    global_experiment_dir,
    'object_attribute_classifiers')
    
mkdir_if_not_exists(region_output_dir)

region_model = os.path.join(
    region_output_dir,
    'model')                    

# Object Attribute Finetuning Params
region_fine_tune_from_iter = 50500
region_fine_tune_from = region_model + '-' + str(region_fine_tune_from_iter)

# Object Attribute Model Selection
region_start_model = 8000
region_step_size = 8000
region_model_accuracies_txt = os.path.join(
    region_output_dir,
    'model_accuracies.txt')

# Object Attribute Classifier Evaluation Params
region_eval_on = 'train_held_out' # One of {'test','train_held_out','train_subset'}
region_model_to_eval = region_model + '-' + '102000'

region_attribute_scores_dirname = os.path.join(
    region_output_dir,
    'attribute_scores')

mkdir_if_not_exists(region_attribute_scores_dirname)

# Answer prediction
num_region_proposals = 100
num_mcq_candidates = 18
num_negative_answers = num_mcq_candidates - 1

# VQA data paths
vqa_basedir = '/home/ssd/VQA/'

vqa_train_image_dir = os.path.join(
    vqa_basedir,
    'train2014_cropped_large')
vqa_train_resnet_feat_dir = os.path.join(
    vqa_basedir,
    'train2014_cropped_large_resnet_features')
vqa_train_anno = os.path.join(
    vqa_basedir,
    'mscoco_train2014_annotations_with_parsed_questions.json')
vqa_train_subset_qids = os.path.join(
    vqa_basedir,
    'train_subset_qids.json')
vqa_train_held_out_qids = os.path.join(
    vqa_basedir,
    'train_held_out_qids.json')

vqa_val_image_dir = os.path.join(
    vqa_basedir,
    'val2014_cropped_large')
vqa_val_resnet_feat_dir = os.path.join(
    vqa_basedir,
    'val2014_cropped_large_resnet_features')
vqa_val_anno = os.path.join(
    vqa_basedir,
    'mscoco_val2014_annotations_with_parsed_questions.json')
vqa_val_qids = os.path.join(
    vqa_basedir,
    'val_qids.json')

vqa_answer_vocab_json = os.path.join(
    vqa_basedir,
    'answer_vocab.json')

# VQA dataset params
# num_train_questions = 248349
# num_train_held_out_questions = 12500
# num_train_subset_questions = num_train_questions - num_train_held_out_questions
# num_val_questions = 121512
# num_val_subset_questions = 10000

# num_test_questions = 0

# Answer classifier training params
answer_batch_size = 50
answer_num_epochs = 6
answer_offset = 0
answer_obj_atr_loss_wt = 0.1
answer_regularization_coeff = 1e-5
answer_queue_size = 500
answer_embedding_dim = 600
answer_lr = 1e-4
answer_log_every_n_iter = 500
answer_output_dir = os.path.join(
    global_experiment_dir,
    'answer_classifiers')
    
mkdir_if_not_exists(answer_output_dir)

pretrained_model = '/home/tanmay/Code/GenVQA/Exp_Results/VQA/' + \
    'object_attribute_classifier_wordvec_xform/' + \
    'object_attribute_classifiers/model-102000'

answer_model = os.path.join(
    answer_output_dir,
    'model')

# Answer classifier additional joint training params
num_regions_with_labels = 100

# Answer fine tune params
answer_fine_tune_from_iter = 13000
answer_fine_tune_from = answer_model + '-' + str(answer_fine_tune_from_iter)

# Answer eval params
answer_eval_on = 'val'
answer_model_to_eval = answer_model + '-42000'

vqa_results_dir = os.path.join(
    answer_output_dir,
    'Results')

mkdir_if_not_exists(vqa_results_dir)

answer_eval_data_json = os.path.join(
    vqa_results_dir,
    'eval_' + answer_eval_on + '_data.json')

answer_eval_results_json = os.path.join(
    vqa_results_dir,
    'eval_' + answer_eval_on + '_results.json')

# Select best model
models_dir = answer_output_dir
start_model = 40000
step_size = 2000
model_accuracies_txt = os.path.join(
    answer_output_dir,
    'model_accuracies.txt')

# Fine Grained Evaluation File paths
raw_vqa_val_ques_json = os.path.join(
    vqa_basedir,
    'MultipleChoice_mscoco_val2014_questions.json')

raw_vqa_val_anno_json = os.path.join(
    vqa_basedir,
    'mscoco_val2014_annotations.json')