def main(_):
# Load model configuration
cu = CommonUtiler()
config_path = os.path.join('./model_conf', FLAGS.model_name + '.py')
config = cu.load_config(config_path)
# Evaluate trained models on val
decoder = mRNNDecoder(config,
FLAGS.model_name,
FLAGS.vocab_path,
gpu_memory_fraction=FLAGS.gpu_memory_fraction)
start, stop, step = [int(x) for x in FLAGS.eval_stat.split()]
for i in range(start, stop, step):
model_path = os.path.join(FLAGS.model_root, FLAGS.model_name,
'variables', 'model_%d.ckpt' % i)
while not os.path.exists(model_path + ".meta"):
logger.warning('Cannot load model file, sleep 1 hour to retry')
time.sleep(3600)
decoder.load_model(model_path)
num_decode = 0
pred_sentences = []
for anno_file_path in FLAGS.anno_files_path.split(':'):
annos = np.load(anno_file_path).tolist()
for anno in annos:
feat_path = os.path.join(
FLAGS.vf_dir, anno['file_path'],
anno['file_name'].split('.')[0] + '.txt')
visual_features = np.loadtxt(feat_path)
sentences = decoder.decode(visual_features, FLAGS.beam_size)
sentence_coco = {
'image_id': anno['id'],
'caption': ' '.join(sentences[0]['words'])
}
pred_sentences.append(sentence_coco)
num_decode += 1
if num_decode % 100 == 0:
logger.info('%d images are decoded' % num_decode)
pred_path = os.path.join(FLAGS.model_root, FLAGS.model_name,
'decode_val_result', 'generated_%d.json' % i)
result_path = os.path.join(FLAGS.model_root, FLAGS.model_name,
'decode_val_result', 'result_%d.txt' % i)
cu.create_dir_if_not_exists(os.path.dirname(pred_path))
with open(pred_path, 'w') as fout:
json.dump(pred_sentences, fout)
cu.coco_val_eval(pred_path, result_path)
def main(_):
# Load model configuration
cu = CommonUtiler()
config_path = os.path.join('./model_conf', FLAGS.model_name + '.py')
config = cu.load_config(config_path)
# Start model training
with tf.Graph().as_default(), tf.Session(config=tf.ConfigProto(
intra_op_parallelism_threads=FLAGS.ses_threads)) as session:
initializer = tf.random_uniform_initializer(-config.init_scale,
config.init_scale)
assert len(config.buckets) >= 1
assert config.buckets[-1] == config.max_num_steps
models = []
with tf.variable_scope("mRNNmodel",
reuse=None,
initializer=initializer):
m = mRNNModel(is_training=True,
num_steps=config.buckets[0],
config=config,
model_name=FLAGS.model_name,
flag_with_saver=True,
model_root=FLAGS.model_root)
models.append(m)
with tf.variable_scope("mRNNmodel", reuse=True):
for bucket in config.buckets[1:]:
m = mRNNModel(is_training=True,
num_steps=bucket,
config=config,
model_name=FLAGS.model_name,
model_root=FLAGS.model_root)
models.append(m)
hdlr = logging.FileHandler(os.path.join(m.model_dir, 'log.txt'))
hdlr.setLevel(logging.INFO)
hdlr.setFormatter(logging.Formatter(formatter_log))
logger.addHandler(hdlr)
if FLAGS.pre_trained_model_path:
models[0].saver.restore(session, FLAGS.pre_trained_model_path)
logger.info('Continue to train from %s',
FLAGS.pre_trained_model_path)
else:
tf.global_variables_initializer()
iters_done = 0
data_provider = mRNNCocoBucketDataProvider(
FLAGS.anno_files_path.split(':'), FLAGS.vocab_path,
config.vocab_size, FLAGS.vf_dir, config.vf_size)
for i in range(config.num_epoch):
train_cost, iters_done = run_epoch(session,
iters_done,
config,
models,
data_provider,
verbose=True)
logger.info("Train cost for epoch %d is %.3f" % (i, train_cost))
# Save final copy of the model
models[0].saver.save(
session, os.path.join(m.variable_dir,
'model_%d.ckpt' % iters_done))
pylab.rcParams['figure.figsize'] = (6.0, 4.0)
#sys.path.append('./py_lib/')
from py_lib.common_utils import CommonUtiler
from py_lib.tf_mrnn_decoder import mRNNDecoder
from py_lib.vision import ImageFeatureExtractor
# set up paths
mrnn_model_path = './trained_models/coco_caption/mrnn_GRU_570K.ckpt'
mrnn_config_path = './model_conf/mrnn_GRU_conf.py'
mrnn_vocab_path = './trained_models/coco_caption/mscoco_mc3_vocab'
img_model_path = './external/tf_cnn_models/inception_v3.pb'
# initilize feature extractor and sentence decoder
cu = CommonUtiler()
config = cu.load_config(mrnn_config_path)
ife = ImageFeatureExtractor(img_model_path)
decoder = mRNNDecoder(config, 'demo', mrnn_vocab_path)
demo_image_path = 'demo_image.jpg'
beam_size = 3
# extract visual feature for the image
visual_features = ife.extract_features(demo_image_path, flag_from_file=True)
# generate sentences
decoder.load_model(mrnn_model_path)
sentences = decoder.decode(visual_features, beam_size)
# Visualize the result
print('Top generated sentences and their log-likelihood:')
for (ind_s, sentence) in enumerate(sentences):
print(' %d (%.2f): %s' %