def from_raw_file(cls, model_filename, additional_model_dict, logger):
"""
Construct from raw libsvm model file.
:param model_filename:
:param additional_model_dict: must contain keys feature_names, norm_type
and optional slopes and intercepts
:param logger:
:return:
"""
# assert additional_model_dict
assert 'feature_names' in additional_model_dict
assert 'norm_type' in additional_model_dict
norm_type = additional_model_dict['norm_type']
assert norm_type == 'none' or norm_type == 'linear_rescale'
if norm_type == 'linear_rescale':
assert 'slopes' in additional_model_dict
assert 'intercepts' in additional_model_dict
train_test_model = cls(param_dict={}, logger=logger)
train_test_model.model_dict.update(additional_model_dict)
model = svmutil.svm_load_model(model_filename)
train_test_model.model_dict['model'] = model
return train_test_model
def from_raw_file(cls, model_filename, additional_model_dict, logger):
"""
Construct from raw libsvm model file.
:param model_filename:
:param additional_model_dict: must contain keys feature_names, norm_type
and optional slopes and intercepts
:param logger:
:return:
"""
# assert additional_model_dict
assert 'feature_names' in additional_model_dict
assert 'norm_type' in additional_model_dict
norm_type = additional_model_dict['norm_type']
assert norm_type == 'none' or norm_type == 'linear_rescale'
if norm_type == 'linear_rescale':
assert 'slopes' in additional_model_dict
assert 'intercepts' in additional_model_dict
train_test_model = cls(param_dict={}, logger=logger)
train_test_model.model_dict.update(additional_model_dict)
model = svmutil.svm_load_model(model_filename)
train_test_model.model_dict['model'] = model
return train_test_model
def _from_info_loaded(cls, info_loaded, filename, logger, optional_dict2):
# override TrainTestModel._from_info_loaded
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']
if issubclass(cls, LibsvmNusvrTrainTestModel):
# == special handling of libsvmnusvr: load .model differently ==
model = svmutil.svm_load_model(filename + '.model')
train_test_model.model_dict['model'] = model
return train_test_model
def from_file(cls, filename, logger=None, optional_dict2=None):
with open(filename, 'rb') as file:
info_loaded = pickle.load(file)
model_type = info_loaded['model_dict']['model_type']
model_class = TrainTestModel.find_subclass(model_type)
train_test_model = model_class(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']
if issubclass(model_class, LibsvmNusvrTrainTestModel):
# == special handling of libsvmnusvr: load .model differently ==
model = svmutil.svm_load_model(filename + '.model')
train_test_model.model_dict['model'] = model
return train_test_model
def _run_on_asset(self, asset):
# Override Executor._run_on_asset(self, asset), which runs a
# FeatureAssembler, collect a feature vector, run
# TrainTestModel.predict() on it, and return a Result object
# (in this case, both Executor._run_on_asset(self, asset) and
# QualityRunner._read_result(self, asset) get bypassed.
vmaf_fassembler = self._get_vmaf_feature_assembler_instance(asset)
vmaf_fassembler.run()
feature_result = vmaf_fassembler.results[0]
# =====================================================================
try:
svmutil
except NameError:
from vmaf import svmutil
# SVR predict
model = svmutil.svm_load_model(self.SVM_MODEL_FILE)
ordered_scaled_scores_list = []
for scores_key in self.SVM_MODEL_ORDERED_SCORES_KEYS:
scaled_scores = self._rescale(
feature_result[scores_key],
self.FEATURE_RESCALE_DICT[scores_key])
ordered_scaled_scores_list.append(scaled_scores)
scores = []
for score_vector in zip(*ordered_scaled_scores_list):
vif, adm, ansnr, motion = score_vector
xs = [[vif, adm, ansnr, motion]]
score = svmutil.svm_predict([0], xs, model)[0][0]
score = self._post_correction(motion, score)
scores.append(score)
result_dict = {}
# add all feature result
result_dict.update(feature_result.result_dict)
# add quality score
result_dict[self.get_scores_key()] = scores
return Result(asset, self.executor_id, result_dict)
def from_file(cls, filename, logger=None, optional_dict2=None):
with open(filename, 'rb') as file:
info_loaded = pickle.load(file)
model_type = info_loaded['model_dict']['model_type']
if model_type == LibsvmNusvrTrainTestModel.TYPE:
train_test_model = LibsvmNusvrTrainTestModel(
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 libsvmnusvr: load .model differently ==
model = svmutil.svm_load_model(filename + '.model')
train_test_model.model_dict['model'] = model
else:
model_class = TrainTestModel.find_subclass(model_type)
train_test_model = model_class(
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']
return train_test_model
def _run_on_asset(self, asset):
# Override Executor._run_on_asset(self, asset), which runs a
# FeatureAssembler, collect a feature vector, run
# TrainTestModel.predict() on it, and return a Result object
# (in this case, both Executor._run_on_asset(self, asset) and
# QualityRunner._read_result(self, asset) get bypassed.
vmaf_fassembler = self._get_vmaf_feature_assembler_instance(asset)
vmaf_fassembler.run()
feature_result = vmaf_fassembler.results[0]
# =====================================================================
# SVR predict
model = svmutil.svm_load_model(self.SVM_MODEL_FILE)
ordered_scaled_scores_list = []
for scores_key in self.SVM_MODEL_ORDERED_SCORES_KEYS:
scaled_scores = self._rescale(feature_result[scores_key],
self.FEATURE_RESCALE_DICT[scores_key])
ordered_scaled_scores_list.append(scaled_scores)
scores = []
for score_vector in zip(*ordered_scaled_scores_list):
vif, adm, ansnr, motion = score_vector
xs = [[vif, adm, ansnr, motion]]
score = svmutil.svm_predict([0], xs, model)[0][0]
score = self._post_correction(motion, score)
scores.append(score)
result_dict = {}
# add all feature result
result_dict.update(feature_result.result_dict)
# add quality score
result_dict[self.get_scores_key()] = scores
return Result(asset, self.executor_id, result_dict)
