def prepare_train_data_set(**data_cofig):
data_root_dir = data_cofig['data_root_dir']
vocab_layout_file = os.path.join(data_root_dir,
data_cofig['vocab_layout_file'])
assembler = Assembler(vocab_layout_file)
imdb_file_trn = os.path.join(data_root_dir, 'imdb',
data_cofig['imdb_file_trn'])
image_feat_dir = os.path.join(data_root_dir,
data_cofig['preprocess_model'], 'train')
vocab_question_file = os.path.join(data_root_dir,
data_cofig['vocab_question_file'])
vocab_answer_file = os.path.join(data_root_dir,
data_cofig['vocab_answer_file'])
prune_filter_module = data_cofig['prune_filter_module']
N = data_cofig['N']
T_encoder = data_cofig['T_encoder']
T_decoder = data_cofig['T_decoder']
data_reader_trn = DataReader(imdb_file_trn,
image_feat_dir,
shuffle=False,
one_pass=True,
batch_size=N,
T_encoder=T_encoder,
T_decoder=T_decoder,
assembler=assembler,
vocab_question_file=vocab_question_file,
vocab_answer_file=vocab_answer_file,
prune_filter_module=prune_filter_module)
num_vocab_txt = data_reader_trn.batch_loader.vocab_dict.num_vocab
num_vocab_nmn = len(assembler.module_names)
num_choices = data_reader_trn.batch_loader.answer_dict.num_vocab
return data_reader_trn, num_vocab_txt, num_choices, num_vocab_nmn, assembler
def prepare_train_data_set(**data_cofig):
data_root_dir = data_cofig['data_root_dir']
vocab_layout_file = os.path.join(data_root_dir,
data_cofig['vocab_layout_file'])
assembler = Assembler(vocab_layout_file)
imdb_file_trn = os.path.join(data_root_dir, 'imdb',
data_cofig['imdb_file_trn'])
image_feat_dir = os.path.join(data_root_dir,
data_cofig['preprocess_model'], 'train')
vocab_question_file = os.path.join(data_root_dir,
data_cofig['vocab_question_file'])
vocab_answer_file = os.path.join(data_root_dir,
data_cofig['vocab_answer_file'])
prune_filter_module = data_cofig['prune_filter_module']
N = data_cofig['N']
T_encoder = data_cofig['T_encoder']
T_decoder = data_cofig['T_decoder']
image_depth_first = data_cofig['image_depth_first']
vqa_train_dataset = vqa_dataset(imdb_file=imdb_file_trn,
image_feat_directory=image_feat_dir,
T_encoder=T_encoder,
T_decoder=T_decoder,
assembler=assembler,
vocab_question_file=vocab_question_file,
vocab_answer_file=vocab_answer_file,
prune_filter_module=prune_filter_module,
image_depth_first=image_depth_first)
data_reader_trn = DataLoader(dataset=vqa_train_dataset,
batch_size=N,
shuffle=True)
num_vocab_txt = vqa_train_dataset.vocab_dict.num_vocab
num_vocab_nmn = len(assembler.module_names)
num_choices = vqa_train_dataset.answer_dict.num_vocab
return data_reader_trn, num_vocab_txt, num_choices, num_vocab_nmn, assembler
T_decoder = 20
N = 64
prune_filter_module = True
imdb_file_tst = './exp_clevr/data/imdb/imdb_%s.npy' % tst_image_set
snapshot_file = './exp_clevr/tfmodel/%s/%s' % (exp_name, snapshot_name)
save_file = './exp_clevr/results/%s/%s.%s.txt' % (exp_name, snapshot_name,
tst_image_set)
os.makedirs(os.path.dirname(save_file), exist_ok=True)
eval_output_file = './exp_clevr/eval_outputs/%s/%s.%s.txt' % (
exp_name, snapshot_name, tst_image_set)
os.makedirs(os.path.dirname(eval_output_file), exist_ok=True)
assembler = Assembler(vocab_layout_file)
data_reader_tst = DataReader(imdb_file_tst,
shuffle=False,
one_pass=True,
batch_size=N,
T_encoder=T_encoder,
T_decoder=T_decoder,
assembler=assembler,
vocab_question_file=vocab_question_file,
vocab_answer_file=vocab_answer_file,
prune_filter_module=prune_filter_module)
print('Running test ...')
answer_correct_total = 0
layout_correct_total = 0
layout_valid_total = 0
def run_eval(exp_name,
snapshot_name,
tst_image_set,
data_dir,
image_feat_dir,
tf_model_dir,
print_log=False):
vocab_question_file = os.path.join(data_dir, "vocabulary_clevr.txt")
vocab_layout_file = os.path.join(data_dir, "vocabulary_layout.txt")
vocab_answer_file = os.path.join(data_dir, "answers_clevr.txt")
imdb_file_tst_base_name = 'imdb_%s.npy' % tst_image_set
imdb_file_tst = os.path.join(data_dir, "imdb", imdb_file_tst_base_name)
image_feat_dir_tst = os.path.join(image_feat_dir, tst_image_set)
#module_snapshot_file = './exp_clevr/tfmodel/%s/%s' % (exp_name, "model_"+snapshot_name)
module_snapshot_file = os.path.join(tf_model_dir, exp_name,
"model_" + snapshot_name)
assembler = Assembler(vocab_layout_file)
data_reader_tst = DataReader(imdb_file_tst,
image_feat_dir_tst,
shuffle=False,
one_pass=True,
batch_size=N,
T_encoder=T_encoder,
T_decoder=T_decoder,
assembler=assembler,
vocab_question_file=vocab_question_file,
vocab_answer_file=vocab_answer_file,
prune_filter_module=prune_filter_module)
if data_reader_tst is not None:
print('Running test ...')
answer_correct_total = 0
layout_correct_total = 0
layout_valid_total = 0
num_questions_total = 0
##load my model
myModel = torch.load(module_snapshot_file)
for i, batch in enumerate(data_reader_tst.batches()):
_, batch_size = batch['input_seq_batch'].shape
input_text_seq_lens = batch['seq_length_batch']
input_text_seqs = batch['input_seq_batch']
input_layouts = batch['gt_layout_batch']
input_images = batch['image_feat_batch']
input_answers = batch['answer_label_batch']
num_questions_total += batch_size
input_txt_variable = Variable(torch.LongTensor(input_text_seqs))
input_txt_variable = input_txt_variable.cuda(
) if use_cuda else input_txt_variable
input_layout_variable = None
_, _, myAnswer, predicted_layouts, expr_validity_array, _ = myModel(
input_txt_variable=input_txt_variable,
input_text_seq_lens=input_text_seq_lens,
input_layout_variable=input_layout_variable,
input_answers=None,
input_images=input_images,
sample_token=False)
layout_correct_total += np.sum(
np.all(predicted_layouts == input_layouts, axis=0))
answer_correct_total += np.sum(
np.logical_and(expr_validity_array, myAnswer == input_answers))
layout_valid_total += np.sum(expr_validity_array)
##current accuracy
layout_accuracy = layout_correct_total / num_questions_total
answer_accuracy = answer_correct_total / num_questions_total
layout_validity = layout_valid_total / num_questions_total
if (i + 1) % 100 == 0 and print_log:
print(
"iter:",
i + 1,
" layout_accuracy=%.4f" % layout_accuracy,
" answer_accuracy=%.4f" % answer_accuracy,
" layout_validity=%.4f" % layout_validity,
)
return layout_accuracy, layout_correct_total, num_questions_total, answer_accuracy