def main(argv):
parser = argparse.ArgumentParser()
add_args.for_general(parser)
add_args.for_inference(parser)
add_args.for_evaluation(parser)
add_args.for_feature(parser)
add_args.for_lstm(parser)
args = parser.parse_args()
config = tf.estimator.RunConfig(save_summary_steps=float('inf'),
log_step_count_steps=10)
params = {
'image_size': args.image_size,
'gazemap_size': args.gazemap_size,
'feature_map_size': args.feature_map_size,
'model_dir': args.model_dir
}
model = tf.estimator.Estimator(model_fn=model_fn,
model_dir=args.model_dir,
config=config,
params=params)
#determine which checkpoint to restore
if args.model_iteration is None:
best_ckpt_dir = os.path.join(args.model_dir, 'best_ckpt')
if os.path.isdir(best_ckpt_dir):
ckpt_name = [
f.split('.index')[0] for f in os.listdir(best_ckpt_dir)
if f.endswith('.index')
][0]
ckpt_path = os.path.join(best_ckpt_dir, ckpt_name)
args.model_iteration = ckpt_name.split('-')[1]
else:
ckpt_name = 'model.ckpt-' + model_iteration
ckpt_path = os.path.join(args.model_dir, ckpt_name)
predict_generator = model.predict(
input_fn=lambda: input_fn('validation',
batch_size=1,
n_steps=None,
shuffle=False,
include_labels=False,
n_epochs=1,
args=args),
checkpoint_path=ckpt_path)
output_dir = os.path.join(args.model_dir,
'prediction_iter_' + args.model_iteration)
if not os.path.isdir(output_dir):
os.makedirs(output_dir)
for res in predict_generator:
output_path = os.path.join(
output_dir,
str(res['video_id']) + '_' +
str(res['predicted_time_points'][0]).zfill(5) + '.jpg')
gazemap = np.reshape(res['ps'], args.gazemap_size)
misc.imsave(output_path, gazemap)
def main(argv):
parser = argparse.ArgumentParser()
add_args.for_general(parser)
add_args.for_inference(parser)
add_args.for_evaluation(parser)
add_args.for_feature(parser)
add_args.for_lstm(parser)
args = parser.parse_args()
config = tf.estimator.RunConfig(save_summary_steps=float('inf'),
log_step_count_steps=10)
params = {
'image_size': args.image_size,
'gazemap_size': args.gazemap_size,
'model_dir': args.model_dir
}
model = tf.estimator.Estimator(
model_fn=model_fn,
model_dir=args.model_dir,
config=config,
params=params)
#determine which checkpoint to restore
'''
if args.model_iteration is None:
print("model iteration is none")
best_ckpt_dir = 'F:\vinay\driver_attention_prediction\pretrained_models\model_for_inference\best_ckpt'
if os.path.isdir(best_ckpt_dir):
ckpt_name = [f.split('.index')[0] for f in os.listdir(best_ckpt_dir) if f.endswith('.index')][0]
ckpt_path = 'F:\vinay\driver_attention_prediction\pretrained_models\model_for_inference\best_ckpt'
args.model_iteration = 'model.ckpt'
'''
if args.model_iteration is None:
print('model iteration is none')
best_ckpt_dir = os.path.join(args.model_dir, 'best_ckpt')
if os.path.isdir(best_ckpt_dir):
print('second loop model iteration none')
ckpt_name = [f.split('.index')[0] for f in os.listdir(best_ckpt_dir) if f.endswith('.index')][0]
print(ckpt_name)
ckpt_path = os.path.join(best_ckpt_dir, ckpt_name)
print(ckpt_path)
args.model_iteration = ckpt_name.split('-')[1]
print("*************")
print(args.model_iteration)
else:
print('model iteration is not none')
ckpt_name = 'model.ckpt-'+args.model_iteration
ckpt_path = os.path.join(args.model_dir, ckpt_name)
K.clear_session()
predict_generator = model.predict(
input_fn = lambda: input_fn(None,
batch_size=1, n_steps=None,
shuffle=False,
n_epochs=1, args=args),
checkpoint_path=ckpt_path)
output_dir = os.path.join(args.data_dir, 'image_features_'+args.feature_name)
if not os.path.isdir(output_dir):
os.makedirs(output_dir)
previous_video_id = None
for res in predict_generator:
if previous_video_id is None:
print('Start inference for video: %s' % res['video_id'])
previous_video_id = res['video_id']
elif res['video_id'] != previous_video_id:
print('Start inference for video: %s' % res['video_id'])
previous_video_id = res['video_id']
output_path = os.path.join(output_dir,
str(res['video_id'])+'_'+str(res['predicted_time_points'][0]).zfill(5)+'.npy')
feature_map = res['feature_maps']
np.save(output_path, feature_map)
def main(argv):
parser = argparse.ArgumentParser()
add_args.for_general(parser)
add_args.for_inference(parser)
add_args.for_feature(parser)
add_args.for_training(parser)
add_args.for_lstm(parser)
args = parser.parse_args()
config = tf.estimator.RunConfig(save_summary_steps=float('inf'),
log_step_count_steps=10)
params = {
'image_size': args.image_size,
'gazemap_size': args.gazemap_size,
'feature_map_size': args.feature_map_size,
'model_dir': args.model_dir,
'weight_data': args.weight_data,
'epsilon': 1e-12,
'learning_rate': args.learning_rate,
'quick_summary_period': args.quick_summary_period,
'slow_summary_period': args.slow_summary_period,
}
model = tf.estimator.Estimator(model_fn=model_fn,
model_dir=args.model_dir,
config=config,
params=params)
# set up the directory to save the best checkpoint
best_ckpt_dir = os.path.join(args.model_dir, 'best_ckpt')
if not os.path.isdir(best_ckpt_dir) or len(os.listdir(best_ckpt_dir)) == 0:
smallest_loss = float('Inf')
if not os.path.isdir(best_ckpt_dir):
os.makedirs(best_ckpt_dir)
else:
smallest_loss = [
float(f.split('_')[1]) for f in os.listdir(best_ckpt_dir)
if f.startswith('loss_')
][0]
for _ in range(args.train_epochs // args.epochs_before_validation):
# Train the model.
K.clear_session()
model.train(
input_fn=lambda: input_fn('training',
args.batch_size,
args.n_steps,
shuffle=True,
include_labels=True,
n_epochs=args.epochs_before_validation,
args=args,
weight_data=args.weight_data))
# validate the model
K.clear_session()
valid_results = model.evaluate(
input_fn=lambda: input_fn('validation',
batch_size=1,
n_steps=None,
shuffle=False,
include_labels=True,
n_epochs=1,
args=args,
weight_data=False))
print(valid_results)
if -valid_results['custom_cc'] < smallest_loss:
smallest_loss = -valid_results['custom_cc']
# delete best_ckpt_dir
shutil.rmtree(best_ckpt_dir)
# re-make best_ckpt_dir as empty
os.makedirs(best_ckpt_dir)
# note down the new smallest loss
open(os.path.join(best_ckpt_dir, 'loss_%f' % smallest_loss),
'a').close()
# copy the checkpoint
files_to_copy = [
f for f in os.listdir(args.model_dir)
if f.startswith('model.ckpt-' +
str(valid_results['global_step']))
]
for f in files_to_copy:
shutil.copyfile(os.path.join(args.model_dir, f),
os.path.join(best_ckpt_dir, f))
def main(argv):
parser = argparse.ArgumentParser()
add_args.for_general(parser)
add_args.for_inference(parser)
add_args.for_evaluation(parser)
add_args.for_feature(parser)
add_args.for_lstm(parser)
args = parser.parse_args()
# if 'session' in locals() and session is not None:
# print('Close interactive session')
# session.close()
config = tf.estimator.RunConfig(save_summary_steps=float('inf'),
log_step_count_steps=10)
params = {
'image_size': args.image_size,
'gazemap_size': args.gazemap_size,
'model_dir': args.model_dir
}
model = tf.estimator.Estimator(model_fn=model_fn,
model_dir=args.model_dir,
config=config,
params=params)
#determine which checkpoint to restore
if args.model_iteration is None:
best_ckpt_dir = os.path.join(args.model_dir, 'best_ckpt')
if os.path.isdir(best_ckpt_dir):
ckpt_name = [
f.split('.index')[0] for f in os.listdir(best_ckpt_dir)
if f.endswith('.index')
][0]
ckpt_path = os.path.join(best_ckpt_dir, ckpt_name)
args.model_iteration = ckpt_name.split('-')[1]
else:
ckpt_name = 'model.ckpt-' + args.model_iteration
ckpt_path = os.path.join(args.model_dir, ckpt_name)
predict_generator = model.predict(input_fn=lambda: input_fn('inference',
batch_size=1,
n_steps=None,
shuffle=False,
n_epochs=1,
args=args),
checkpoint_path=ckpt_path)
output_dir = os.path.join(args.model_dir,
'prediction_iter_' + args.model_iteration)
if not os.path.isdir(output_dir):
os.makedirs(output_dir)
previous_video_id = None
for res in predict_generator:
if previous_video_id is None:
print('Start inference for video: %s' % res['video_id'])
previous_video_id = res['video_id']
elif res['video_id'] != previous_video_id:
print('Start inference for video: %s' % res['video_id'])
previous_video_id = res['video_id']
output_path = os.path.join(
output_dir,
str(res['video_id']) + '_' +
str(res['predicted_time_points'][0]).zfill(5) + '.jpg')
gazemap = np.reshape(res['ps'], args.gazemap_size)
# misc.imsave(output_path, gazemap)
imageio.imwrite(output_path, normalize_map(gazemap).astype(np.uint8))
def main(argv):
parser = argparse.ArgumentParser()
add_args.for_general(parser)
add_args.for_inference(parser)
add_args.for_evaluation(parser)
add_args.for_feature(parser)
add_args.for_lstm(parser)
args = parser.parse_args()
config = tf.estimator.RunConfig(save_summary_steps=float('inf'))
params = {
'camera_size': args.camera_size,
'small_camera_size': args.small_camera_size,
'visual_size': args.visual_size,
'model_dir': args.model_dir,
'use_foveal': args.use_foveal,
'foveal_only': args.foveal_only,
'attention_model_dir': args.attention_model_dir,
'weight_data': args.weight_data,
'epsilon': 1e-12,
'readout': args.readout,
'output_details': True,
'sample_fovea': args.sample_fovea,
'attention_logit_factor': args.attention_logit_factor,
'premade_attention_maps': args.premade_attention_maps,
'premade_features': args.premade_features,
'feature_map_size': args.feature_map_size,
'gazemap_size': args.gazemap_size,
'random_fovea': args.random_fovea,
}
if args.visible_gpus is not None:
os.environ["CUDA_VISIBLE_DEVICES"] = args.visible_gpus
model = tf.estimator.Estimator(model_fn=model_fn,
model_dir=args.model_dir,
config=config,
params=params)
#determine which checkpoint to restore
if args.model_iteration is None:
best_ckpt_dir = os.path.join(args.model_dir, 'best_ckpt')
if os.path.isdir(best_ckpt_dir):
ckpt_name = [
f.split('.index')[0] for f in os.listdir(best_ckpt_dir)
if f.endswith('.index')
][0]
ckpt_path = os.path.join(best_ckpt_dir, ckpt_name)
args.model_iteration = ckpt_name.split('-')[1]
else:
ckpt_name = 'model.ckpt-' + model_iteration
ckpt_path = os.path.join(args.model_dir, ckpt_name)
predict_generator = model.predict(
input_fn=lambda: input_fn('test',
batch_size=args.batch_size,
n_steps=args.n_steps,
shuffle=False,
include_labels=True,
n_epochs=1,
args=args),
checkpoint_path=ckpt_path)
output_dir = os.path.join(args.model_dir,
'prediction_iter_' + args.model_iteration)
if not os.path.isdir(output_dir):
os.makedirs(output_dir)
dfs = []
video_ids = []
for res in tqdm(predict_generator):
n_steps = len(res['predicted_time_points'])
video_id = res['video_id'].decode("utf-8")
if '/' in video_id:
video_id = video_id.split('/')[-1].split('.')[0]
df = pd.DataFrame.from_dict({
'video_key': [len(video_ids)] * n_steps, # start from 0 but not 1
'time_point': res['predicted_time_points'],
'speed_x': res['speed'][:, 0],
'speed_y': res['speed'][:, 1],
'output_speed': res['output_speeds'][:, 0],
'likelihood': res['likelihood'],
'overlap': res['overlap'],
})
dfs.append(df)
video_ids.append(video_id)
output_df = pd.concat(dfs)
feather.write_dataframe(output_df,
os.path.join(output_dir, 'outputs.feather'))
video_df = pd.DataFrame(data={
'video_key': range(len(video_ids)),
'video_id': video_ids,
})
feather.write_dataframe(video_df, os.path.join(output_dir,
'videos.feather'))
def main(argv):
parser = argparse.ArgumentParser()
add_args.for_general(parser)
add_args.for_inference(parser)
add_args.for_evaluation(parser)
add_args.for_feature(parser)
add_args.for_lstm(parser)
parser.add_argument('--max_length',
default=310,
type=int,
help="maximum length of one tfrecord")
parser.add_argument(
'--data_subset',
default='test',
type=str,
help=
"the tfrecords of which subset to divide, i.e., training, validation or test"
)
args = parser.parse_args()
MAX_LENGTH = args.max_length
if args.visible_gpus is not None:
os.environ["CUDA_VISIBLE_DEVICES"] = args.visible_gpus
sess = tf.Session()
dataset = args.data_subset
output_dir = args.data_dir + '/' + 'tfrecords_segments' + '/' + dataset
if not os.path.isdir(output_dir):
os.makedirs(output_dir)
this_input_fn = lambda: input_fn(
dataset, args.validation_batch_size, n_epochs=1, args=args)
ds = this_input_fn()
iterator = ds.make_one_shot_iterator()
next_element = iterator.get_next()
while True:
try:
res = sess.run(next_element)
video_path = res['video_id'].decode("utf-8")
video_id = video_path.split('/')[-1].split('.')[0]
#premade_feature_dir = '/data/alexnet_features/' + dataset
#premade_features = np.load(os.path.join(premade_feature_dir, video_id+'.npy'))
length = len(res['cameras'])
for i in range(math.ceil(float(length) / MAX_LENGTH)):
startIdx = i * MAX_LENGTH
endIdx = min((i + 1) * MAX_LENGTH, length)
example = tf.train.Example(features=tf.train.Features(
feature={
'image/class/video_name':
_bytes_feature([video_path.encode('utf-8')]),
'image/encoded':
_bytes_feature(res['cameras'][startIdx:endIdx]),
'image/low_res':
_bytes_feature(res['low_res_cameras']
[startIdx:endIdx]),
'image/speeds':
_float_feature(res['speed'][startIdx:endIdx].ravel().
tolist()), # ravel l*2 into list
'time_points':
_int64_feature(res['time_points']
[startIdx:endIdx].tolist()),
#'image/premade_features': _float_feature(premade_features[startIdx:endIdx].ravel().tolist()),
}))
writer = tf.python_io.TFRecordWriter(
os.path.join(
output_dir,
video_id + '_' + str(i).zfill(2) + '.tfrecords'))
writer.write(example.SerializeToString())
writer.close()
print(video_id)
except tf.errors.OutOfRangeError:
break
def main(argv):
parser = argparse.ArgumentParser()
add_args.for_general(parser)
add_args.for_inference(parser)
add_args.for_feature(parser)
add_args.for_training(parser)
add_args.for_lstm(parser)
args = parser.parse_args()
config = tf.estimator.RunConfig(
save_summary_steps=float('inf'),
log_step_count_steps=args.quick_summary_period,
keep_checkpoint_max=60)
params = {
'camera_size': args.camera_size,
'small_camera_size': args.small_camera_size,
'visual_size': args.visual_size,
'model_dir': args.model_dir,
'use_foveal': args.use_foveal,
'random_fovea': args.random_fovea,
'foveal_only': args.foveal_only,
'attention_model_dir': args.attention_model_dir,
'weight_data': args.weight_data,
'epsilon': 1e-12,
'learning_rate': args.learning_rate,
'quick_summary_period': args.quick_summary_period,
'slow_summary_period': args.slow_summary_period,
'readout': args.readout,
'augment_data': args.augment_data,
'feature_map_size': args.feature_map_size,
'gazemap_size': args.gazemap_size,
'stability_loss_weight': args.stability_loss_weight,
'sample_fovea': args.sample_fovea,
'attention_logit_factor': args.attention_logit_factor,
'premade_attention_maps': args.premade_attention_maps,
'premade_features': args.premade_features,
}
if args.visible_gpus is not None:
os.environ["CUDA_VISIBLE_DEVICES"] = args.visible_gpus
model = tf.estimator.Estimator(model_fn=model_fn,
model_dir=args.model_dir,
config=config,
params=params)
# set up the directory to save the best checkpoint
best_ckpt_dir = os.path.join(args.model_dir, 'best_ckpt')
if not os.path.isdir(best_ckpt_dir) or len(os.listdir(best_ckpt_dir)) == 0:
smallest_loss = float('Inf')
if not os.path.isdir(best_ckpt_dir):
os.makedirs(best_ckpt_dir)
else:
smallest_loss = [
float(f.split('_')[1]) for f in os.listdir(best_ckpt_dir)
if f.startswith('loss_')
][0]
for _ in range(args.train_epochs // args.epochs_before_validation):
# Train the model.
K.clear_session()
model.train(
input_fn=lambda: input_fn('training',
args.batch_size,
args.n_steps,
shuffle=True,
include_labels=True,
n_epochs=args.epochs_before_validation,
args=args,
weight_data=args.weight_data,
augment_data=args.augment_data))
# validate the model
K.clear_session()
valid_results = model.evaluate(
input_fn=lambda: input_fn('validation',
batch_size=args.validation_batch_size,
n_steps=args.validation_n_steps,
shuffle=False,
include_labels=True,
n_epochs=1,
args=args,
weight_data=False))
print(valid_results)
if valid_results['mae'] < smallest_loss:
smallest_loss = valid_results['mae']
# delete best_ckpt_dir
shutil.rmtree(best_ckpt_dir)
# re-make best_ckpt_dir as empty
os.makedirs(best_ckpt_dir)
# note down the new smallest loss
open(os.path.join(best_ckpt_dir, 'loss_%f' % smallest_loss),
'a').close()
# copy the checkpoint
files_to_copy = [
f for f in os.listdir(args.model_dir)
if f.startswith('model.ckpt-' +
str(valid_results['global_step']))
]
for f in files_to_copy:
shutil.copyfile(os.path.join(args.model_dir, f),
os.path.join(best_ckpt_dir, f))