def from_file(cls, filename, logger=None, optional_dict2=None):
# override TrainTestModel.from_file
with open(filename, 'rb') as file:
info_loaded = pickle.load(file)
train_test_model = cls(param_dict={},
logger=logger,
optional_dict2=optional_dict2)
train_test_model.param_dict = info_loaded['param_dict']
train_test_model.model_dict = info_loaded['model_dict']
# == special handling of tensorflow: load .model differently ==
input_image_batch, logits, y_, y_p, W_conv0, W_conv1, loss, train_step \
= cls.create_tf_variables(train_test_model.param_dict)
saver = tf.train.Saver()
sess = tf.Session()
checkpoint_dir = get_dir_without_last_slash(filename)
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
else:
assert False
model = {
'sess': sess,
'y_p': y_p,
'input_image_batch': input_image_batch,
}
train_test_model.model_dict['model'] = model
return train_test_model
def from_file(cls, filename, logger=None, optional_dict2=None):
# override TrainTestModel.from_file
with open(filename, 'rb') as file:
info_loaded = pickle.load(file)
train_test_model = cls(param_dict={}, logger=logger,
optional_dict2=optional_dict2)
train_test_model.param_dict = info_loaded['param_dict']
train_test_model.model_dict = info_loaded['model_dict']
# == special handling of tensorflow: load .model differently ==
input_image_batch, logits, y_, y_p, W_conv0, W_conv1, loss, train_step \
= cls.create_tf_variables(train_test_model.param_dict)
saver = tf.train.Saver()
sess = tf.Session()
checkpoint_dir = get_dir_without_last_slash(filename)
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
else:
assert False
model = {
'sess': sess,
'y_p': y_p,
'input_image_batch': input_image_batch,
}
train_test_model.model_dict['model'] = model
return train_test_model
def delete(filename):
"""
override TrainTestModel.delete
"""
if os.path.exists(filename):
os.remove(filename)
if os.path.exists(filename + '.model'):
os.remove(filename + '.model')
if os.path.exists(filename + '.model.meta'):
os.remove(filename + '.model.meta')
filedir = get_dir_without_last_slash(filename)
if os.path.exists(filedir + '/checkpoint'):
os.remove(filedir + '/checkpoint')
def delete(filename):
"""
override TrainTestModel.delete(filename)
"""
if os.path.exists(filename):
os.remove(filename)
if os.path.exists(filename + '.model'):
os.remove(filename + '.model')
if os.path.exists(filename + '.model.meta'):
os.remove(filename + '.model.meta')
filedir = get_dir_without_last_slash(filename)
if os.path.exists(filedir + '/checkpoint'):
os.remove(filedir + '/checkpoint')
def _run_on_asset(self, asset):
# Override Executor._run_on_asset to skip working on ref video
if self.result_store:
result = self.result_store.load(asset, self.executor_id)
else:
result = None
# if result can be retrieved from result_store, skip log file
# generation and reading result from log file, but directly return
# return the retrieved result
if result is not None:
if self.logger:
self.logger.info('{id} result exists. Skip {id} run.'.
format(id=self.executor_id))
else:
if self.logger:
self.logger.info('{id} result does\'t exist. Perform {id} '
'calculation.'.format(id=self.executor_id))
# at this stage, it is certain that asset.ref_path and
# asset.dis_path will be used. must early determine that
# they exists
self._assert_paths(asset)
# if no rescaling is involved, directly work on ref_path/dis_path,
# instead of opening workfiles
self._set_asset_use_path_as_workpath(asset)
# remove workfiles if exist (do early here to avoid race condition
# when ref path and dis path have some overlap)
if asset.use_path_as_workpath:
# do nothing
pass
else:
self._close_dis_workfile(asset)
log_file_path = self._get_log_file_path(asset)
make_parent_dirs_if_nonexist(log_file_path)
if asset.use_path_as_workpath:
# do nothing
pass
else:
if self.fifo_mode:
dis_p = multiprocessing.Process(target=self._open_dis_workfile,
args=(asset, True))
dis_p.start()
self._wait_for_workfiles(asset)
else:
self._open_dis_workfile(asset, fifo_mode=False)
self._prepare_log_file(asset)
self._generate_result(asset)
# clean up workfiles
if self.delete_workdir:
if asset.use_path_as_workpath:
# do nothing
pass
else:
self._close_dis_workfile(asset)
if self.logger:
self.logger.info("Read {id} log file, get scores...".
format(type=self.executor_id))
# collect result from each asset's log file
result = self._read_result(asset)
# save result
if self.result_store:
self.result_store.save(result)
# clean up workdir and log files in it
if self.delete_workdir:
# remove log file
self._remove_log(asset)
# remove dir
log_file_path = self._get_log_file_path(asset)
log_dir = get_dir_without_last_slash(log_file_path)
try:
os.rmdir(log_dir)
except OSError as e:
if e.errno == 39: # [Errno 39] Directory not empty
# VQM could generate an error file with non-critical
# information like: '3 File is longer than 15 seconds.
# Results will be calculated using first 15 seconds
# only.' In this case, want to keep this
# informational file and pass
pass
result = self._post_process_result(result)
return result
def _run_on_asset(self, asset):
# Wraper around the essential function _generate_result, to
# do housekeeping work including 1) asserts of asset, 2) skip run if
# log already exist, 3) creating fifo, 4) delete work file and dir
# asserts
self._assert_an_asset(asset)
if self.result_store:
result = self.result_store.load(asset, self.executor_id)
else:
result = None
# if result can be retrieved from result_store, skip log file
# generation and reading result from log file, but directly return
# return the retrieved result
if result is not None:
if self.logger:
self.logger.info('{id} result exists. Skip {id} run.'.
format(id=self.executor_id))
else:
if self.logger:
self.logger.info('{id} result does\'t exist. Perform {id} '
'calculation.'.format(id=self.executor_id))
# at this stage, it is certain that asset.ref_path and
# asset.dis_path will be used. must early determine that
# they exists
self._assert_paths(asset)
# if no rescaling is involved, directly work on ref_path/dis_path,
# instead of opening workfiles
self._set_asset_use_path_as_workpath(asset)
# remove workfiles if exist (do early here to avoid race condition
# when ref path and dis path have some overlap)
if asset.use_path_as_workpath:
# do nothing
pass
else:
self._close_ref_workfile(asset)
self._close_dis_workfile(asset)
log_file_path = self._get_log_file_path(asset)
make_parent_dirs_if_nonexist(log_file_path)
if asset.use_path_as_workpath:
# do nothing
pass
else:
if self.fifo_mode:
ref_p = multiprocessing.Process(target=self._open_ref_workfile,
args=(asset, True))
dis_p = multiprocessing.Process(target=self._open_dis_workfile,
args=(asset, True))
ref_p.start()
dis_p.start()
self._wait_for_workfiles(asset)
else:
self._open_ref_workfile(asset, fifo_mode=False)
self._open_dis_workfile(asset, fifo_mode=False)
self._prepare_log_file(asset)
self._generate_result(asset)
# clean up workfiles
if self.delete_workdir:
if asset.use_path_as_workpath:
# do nothing
pass
else:
self._close_ref_workfile(asset)
self._close_dis_workfile(asset)
if self.logger:
self.logger.info("Read {id} log file, get scores...".
format(type=self.executor_id))
# collect result from each asset's log file
result = self._read_result(asset)
# save result
if self.result_store:
self.result_store.save(result)
# clean up workdir and log files in it
if self.delete_workdir:
# remove log file
self._remove_log(asset)
# remove dir
log_file_path = self._get_log_file_path(asset)
log_dir = get_dir_without_last_slash(log_file_path)
try:
os.rmdir(log_dir)
except OSError as e:
if e.errno == 39: # [Errno 39] Directory not empty
# VQM could generate an error file with non-critical
# information like: '3 File is longer than 15 seconds.
# Results will be calculated using first 15 seconds
# only.' In this case, want to keep this
# informational file and pass
pass
result = self._post_process_result(result)
return result