def show_local_explanations(cls, results, indexs=None):
"""Plot local explanations of results
:param results:
:param indexs: a list of frame indices, or None. If None, will take the
second frame.
:return: figures of local explanation plots
"""
# assert results are indeed generated by class
for result in results:
assert cls.get_explanations_key() in result.result_dict
N = len(results)
if indexs is None:
indexs = [1] # default: second frame
figss = []
for n in range(N):
exps = results[n][cls.get_explanations_key()]
asset = results[n].asset
exps2 = LocalExplainer.select_from_exps(exps, indexs)
ys_pred = results[n][cls.get_scores_key()][indexs]
N2 = LocalExplainer.assert_explanations(exps2)
assets2 = [asset for _ in range(N2)]
# LocalExplainer.print_explanations(exps2, assets=assets2, ys=None, ys_pred=ys_pred)
figs = LocalExplainer.plot_explanations(exps2, assets=assets2, ys=None, ys_pred=ys_pred)
figss.append(figs)
return figss
def test_explain_train_test_model(self):
model_class = SklearnRandomForestTrainTestModel
train_dataset_path = config.ROOT + '/python/test/resource/' \
'test_image_dataset_diffdim.py'
train_dataset = import_python_file(train_dataset_path)
train_assets = read_dataset(train_dataset)
_, self.features = run_executors_in_parallel(
MomentNorefFeatureExtractor,
train_assets,
fifo_mode=True,
delete_workdir=True,
parallelize=True,
result_store=None,
optional_dict=None,
optional_dict2=None,
)
xys = model_class.get_xys_from_results(self.features[:7])
model = model_class({'norm_type':'normalize', 'random_state':0}, None)
model.train(xys)
np.random.seed(0)
xs = model_class.get_xs_from_results(self.features[7:])
explainer = LocalExplainer(neighbor_samples=1000)
exps = explainer.explain(model, xs)
self.assertAlmostEqual(exps['feature_weights'][0, 0], -0.12416, places=4)
self.assertAlmostEqual(exps['feature_weights'][1, 0], 0.00076, places=4)
self.assertAlmostEqual(exps['feature_weights'][0, 1], -0.20931, places=4)
self.assertAlmostEqual(exps['feature_weights'][1, 1], -0.01245, places=4)
self.assertAlmostEqual(exps['feature_weights'][0, 2], 0.02322, places=4)
self.assertAlmostEqual(exps['feature_weights'][1, 2], 0.03673, places=4)
self.assertAlmostEqual(exps['features'][0, 0], 107.73501, places=4)
self.assertAlmostEqual(exps['features'][1, 0], 35.81638, places=4)
self.assertAlmostEqual(exps['features'][0, 1], 13691.23881, places=4)
self.assertAlmostEqual(exps['features'][1, 1], 1611.56764, places=4)
self.assertAlmostEqual(exps['features'][0, 2], 2084.40542, places=4)
self.assertAlmostEqual(exps['features'][1, 2], 328.75389, places=4)
self.assertAlmostEqual(exps['features_normalized'][0, 0], -0.65527, places=4)
self.assertAlmostEqual(exps['features_normalized'][1, 0], -3.74922, places=4)
self.assertAlmostEqual(exps['features_normalized'][0, 1], -0.68872, places=4)
self.assertAlmostEqual(exps['features_normalized'][1, 1], -2.79586, places=4)
self.assertAlmostEqual(exps['features_normalized'][0, 2], 0.08524, places=4)
self.assertAlmostEqual(exps['features_normalized'][1, 2], -1.32625, places=4)
self.assertEqual(exps['feature_names'],
['Moment_noref_feature_1st_score',
'Moment_noref_feature_2nd_score',
'Moment_noref_feature_var_score']
)
def test_explain_train_test_model(self):
model_class = SklearnRandomForestTrainTestModel
train_dataset_path = config.ROOT + '/python/test/resource/' \
'test_image_dataset_diffdim.py'
train_dataset = import_python_file(train_dataset_path)
train_assets = read_dataset(train_dataset)
_, self.features = run_executors_in_parallel(
MomentNorefFeatureExtractor,
train_assets,
fifo_mode=True,
delete_workdir=True,
parallelize=True,
result_store=None,
optional_dict=None,
optional_dict2=None,
)
xys = model_class.get_xys_from_results(self.features[:7])
model = model_class({'norm_type':'normalize', 'random_state':0}, None)
model.train(xys)
np.random.seed(0)
xs = model_class.get_xs_from_results(self.features[7:])
explainer = LocalExplainer(neighbor_samples=1000)
exps = explainer.explain(model, xs)
self.assertAlmostEqual(exps['feature_weights'][0, 0], -0.12416, places=4)
self.assertAlmostEqual(exps['feature_weights'][1, 0], 0.00076, places=4)
self.assertAlmostEqual(exps['feature_weights'][0, 1], -0.20931, places=4)
self.assertAlmostEqual(exps['feature_weights'][1, 1], -0.01245, places=4)
self.assertAlmostEqual(exps['feature_weights'][0, 2], 0.02322, places=4)
self.assertAlmostEqual(exps['feature_weights'][1, 2], 0.03673, places=4)
self.assertAlmostEqual(exps['features'][0, 0], 107.73501, places=4)
self.assertAlmostEqual(exps['features'][1, 0], 35.81638, places=4)
self.assertAlmostEqual(exps['features'][0, 1], 13691.23881, places=4)
self.assertAlmostEqual(exps['features'][1, 1], 1611.56764, places=4)
self.assertAlmostEqual(exps['features'][0, 2], 2084.40542, places=4)
self.assertAlmostEqual(exps['features'][1, 2], 328.75389, places=4)
self.assertAlmostEqual(exps['features_normalized'][0, 0], -0.65527, places=4)
self.assertAlmostEqual(exps['features_normalized'][1, 0], -3.74922, places=4)
self.assertAlmostEqual(exps['features_normalized'][0, 1], -0.68872, places=4)
self.assertAlmostEqual(exps['features_normalized'][1, 1], -2.79586, places=4)
self.assertAlmostEqual(exps['features_normalized'][0, 2], 0.08524, places=4)
self.assertAlmostEqual(exps['features_normalized'][1, 2], -1.32625, places=4)
self.assertEqual(exps['feature_names'],
['Moment_noref_feature_1st_score',
'Moment_noref_feature_2nd_score',
'Moment_noref_feature_var_score']
)
def test_explain_vmaf_results(self):
print 'test on running VMAF runner with local explainer...'
ref_path = config.ROOT + "/resource/yuv/src01_hrc00_576x324.yuv"
dis_path = config.ROOT + "/resource/yuv/src01_hrc01_576x324.yuv"
asset = Asset(dataset="test", content_id=0, asset_id=0,
workdir_root=config.ROOT + "/workspace/workdir",
ref_path=ref_path,
dis_path=dis_path,
asset_dict={'width':576, 'height':324})
asset_original = Asset(dataset="test", content_id=0, asset_id=1,
workdir_root=config.ROOT + "/workspace/workdir",
ref_path=ref_path,
dis_path=ref_path,
asset_dict={'width':576, 'height':324})
self.runner = VmafQualityRunnerWithLocalExplainer(
[asset, asset_original],
None, fifo_mode=True,
delete_workdir=True,
result_store=None,
optional_dict2={'explainer': LocalExplainer(neighbor_samples=100)}
)
np.random.seed(0)
self.runner.run()
results = self.runner.results
self.assertAlmostEqual(results[0]['VMAF_score'], 65.4488588759, places=4)
self.assertAlmostEqual(results[1]['VMAF_score'], 99.2259317881, places=4)
expected_feature_names = ['VMAF_feature_adm2_score',
'VMAF_feature_motion_score',
'VMAF_feature_vif_scale0_score',
'VMAF_feature_vif_scale1_score',
'VMAF_feature_vif_scale2_score',
'VMAF_feature_vif_scale3_score']
weights = np.mean(results[0]['VMAF_scores_exps']['feature_weights'], axis=0)
self.assertAlmostEqual(weights[0], 0.75441663, places=4)
self.assertAlmostEqual(weights[1], 0.06816105, places=4)
self.assertAlmostEqual(weights[2], -0.10934421, places=4)
self.assertAlmostEqual(weights[3], 0.22051127, places=4)
self.assertAlmostEqual(weights[4], 0.12517884, places=4)
self.assertAlmostEqual(weights[5], 0.04639162, places=4)
self.assertEqual(results[0]['VMAF_scores_exps']['feature_names'],
expected_feature_names)
weights = np.mean(results[1]['VMAF_scores_exps']['feature_weights'], axis=0)
self.assertAlmostEqual(weights[0], 0.77096087, places=4)
self.assertAlmostEqual(weights[1], 0.01491754, places=4)
self.assertAlmostEqual(weights[2], -0.08025557, places=4)
self.assertAlmostEqual(weights[3], 0.2511188, places=4)
self.assertAlmostEqual(weights[4], 0.14953561, places=4)
self.assertAlmostEqual(weights[5], 0.07960753, places=4)
self.assertEqual(results[1]['VMAF_scores_exps']['feature_names'],
expected_feature_names)
def test_explain_train_test_model(self):
model_class = MomentRandomForestTrainTestModel
xys = model_class.get_xys_from_results(self.features[:7])
del xys['dis_u']
del xys['dis_v']
model = model_class({'norm_type':'normalize', 'random_state':0})
model.train(xys)
np.random.seed(0)
xs = model_class.get_xs_from_results(self.features[7:])
del xs['dis_u']
del xs['dis_v']
explainer = LocalExplainer(neighbor_samples=1000)
exps = explainer.explain(model, xs)
self.assertAlmostEqual(exps['feature_weights'][0, 0], -0.12416, places=4)
self.assertAlmostEqual(exps['feature_weights'][1, 0], 0.00076, places=4)
self.assertAlmostEqual(exps['feature_weights'][0, 1], -0.20931, places=4)
self.assertAlmostEqual(exps['feature_weights'][1, 1], -0.01245, places=4)
self.assertAlmostEqual(exps['feature_weights'][0, 2], 0.02322, places=4)
self.assertAlmostEqual(exps['feature_weights'][1, 2], 0.03673, places=4)
self.assertAlmostEqual(exps['features'][0, 0], 107.73501, places=4)
self.assertAlmostEqual(exps['features'][1, 0], 35.81638, places=4)
self.assertAlmostEqual(exps['features'][0, 1], 13691.23881, places=4)
self.assertAlmostEqual(exps['features'][1, 1], 1611.56764, places=4)
self.assertAlmostEqual(exps['features'][0, 2], 2084.40542, places=4)
self.assertAlmostEqual(exps['features'][1, 2], 328.75389, places=4)
self.assertAlmostEqual(exps['features_normalized'][0, 0], -0.65527, places=4)
self.assertAlmostEqual(exps['features_normalized'][1, 0], -3.74922, places=4)
self.assertAlmostEqual(exps['features_normalized'][0, 1], -0.68872, places=4)
self.assertAlmostEqual(exps['features_normalized'][1, 1], -2.79586, places=4)
self.assertAlmostEqual(exps['features_normalized'][0, 2], 0.08524, places=4)
self.assertAlmostEqual(exps['features_normalized'][1, 2], -1.32625, places=4)
self.assertEqual(exps['feature_names'], ['dis_y'])
def test_explain_train_test_model(self):
model_class = MomentRandomForestTrainTestModel
xys = model_class.get_xys_from_results(self.features[:7])
del xys['dis_u']
del xys['dis_v']
model = model_class({'norm_type':'normalize', 'random_state':0})
model.train(xys)
np.random.seed(0)
xs = model_class.get_xs_from_results(self.features[7:])
del xs['dis_u']
del xs['dis_v']
explainer = LocalExplainer(neighbor_samples=1000)
exps = explainer.explain(model, xs)
self.assertAlmostEqual(exps['feature_weights'][0, 0], -0.12416, places=4)
self.assertAlmostEqual(exps['feature_weights'][1, 0], 0.00076, places=4)
self.assertAlmostEqual(exps['feature_weights'][0, 1], -0.20931, places=4)
self.assertAlmostEqual(exps['feature_weights'][1, 1], -0.01245, places=4)
self.assertAlmostEqual(exps['feature_weights'][0, 2], 0.02322, places=4)
self.assertAlmostEqual(exps['feature_weights'][1, 2], 0.03673, places=4)
self.assertAlmostEqual(exps['features'][0, 0], 107.73501, places=4)
self.assertAlmostEqual(exps['features'][1, 0], 35.81638, places=4)
self.assertAlmostEqual(exps['features'][0, 1], 13691.23881, places=4)
self.assertAlmostEqual(exps['features'][1, 1], 1611.56764, places=4)
self.assertAlmostEqual(exps['features'][0, 2], 2084.40542, places=4)
self.assertAlmostEqual(exps['features'][1, 2], 328.75389, places=4)
self.assertAlmostEqual(exps['features_normalized'][0, 0], -0.65527, places=4)
self.assertAlmostEqual(exps['features_normalized'][1, 0], -3.74922, places=4)
self.assertAlmostEqual(exps['features_normalized'][0, 1], -0.68872, places=4)
self.assertAlmostEqual(exps['features_normalized'][1, 1], -2.79586, places=4)
self.assertAlmostEqual(exps['features_normalized'][0, 2], 0.08524, places=4)
self.assertAlmostEqual(exps['features_normalized'][1, 2], -1.32625, places=4)
self.assertEqual(exps['feature_names'], ['dis_y'])
def _run_on_asset(self, asset):
# Override VmafQualityRunner._run_on_asset(self, asset), by adding
# additional local explanation info.
vmaf_fassembler = self._get_vmaf_feature_assembler_instance(asset)
vmaf_fassembler.run()
feature_result = vmaf_fassembler.results[0]
model = self._load_model(asset)
xs = model.get_per_unit_xs_from_a_result(feature_result)
ys_pred = self.predict_with_model(model, xs)
if self.optional_dict2 is not None and \
'explainer' in self.optional_dict2:
explainer = self.optional_dict2['explainer']
else:
explainer = LocalExplainer()
exps = explainer.explain(model, xs)
result_dict = {}
result_dict.update(feature_result.result_dict) # add feature result
result_dict[self.get_scores_key()] = ys_pred # add quality score
result_dict[self.get_explanations_key()] = exps # add local explanations
return Result(asset, self.executor_id, result_dict)
def show_local_explanations(cls, results, indexs=None):
"""Plot local explanations of results
:param results:
:param indexs: a list of frame indices, or None. If None, will take the
second frame.
:return: figures of local explanation plots
"""
# assert results are indeed generated by class
for result in results:
assert cls.get_explanations_key() in result.result_dict
N = len(results)
if indexs is None:
indexs = [1] # default: second frame
figss = []
for n in range(N):
exps = results[n][cls.get_explanations_key()]
asset = results[n].asset
exps2 = LocalExplainer.select_from_exps(exps, indexs)
ys_pred = results[n][cls.get_scores_key()][indexs]
N2 = LocalExplainer.assert_explanations(exps2)
assets2 = [asset for _ in range(N2)]
# LocalExplainer.print_explanations(exps2, assets=assets2, ys=None, ys_pred=ys_pred)
figs = LocalExplainer.plot_explanations(exps2,
assets=assets2,
ys=None,
ys_pred=ys_pred)
figss.append(figs)
return figss
def _run_on_asset(self, asset):
# Override VmafQualityRunner._run_on_asset(self, asset), by adding
# additional local explanation info.
vmaf_fassembler = self._get_vmaf_feature_assembler_instance(asset)
vmaf_fassembler.run()
feature_result = vmaf_fassembler.results[0]
model = self._load_model(asset)
xs = model.get_per_unit_xs_from_a_result(feature_result)
ys_pred = self.predict_with_model(model, xs)
if self.optional_dict2 is not None and \
'explainer' in self.optional_dict2:
explainer = self.optional_dict2['explainer']
else:
explainer = LocalExplainer()
exps = explainer.explain(model, xs)
result_dict = {}
result_dict.update(feature_result.result_dict) # add feature result
result_dict[self.get_scores_key()] = ys_pred # add quality score
result_dict[
self.get_explanations_key()] = exps # add local explanations
return Result(asset, self.executor_id, result_dict)
def explain_model_on_dataset(model, test_assets_selected_indexs,
test_dataset_filepath):
def print_assets(test_assets):
print '\n'.join(
map(
lambda (i, asset): "Asset {i}: {name}".format(
i=i, name=get_file_name_without_extension(asset.dis_path)),
enumerate(test_assets)))
test_dataset = import_python_file(test_dataset_filepath)
test_assets = read_dataset(test_dataset)
print_assets(test_assets)
print "Assets selected for local explanation: {}".format(
test_assets_selected_indexs)
result_store = FileSystemResultStore()
test_assets = [test_assets[i] for i in test_assets_selected_indexs]
test_fassembler = FeatureAssembler(
feature_dict=model.model_dict['feature_dict'],
feature_option_dict=None,
assets=test_assets,
logger=None,
fifo_mode=True,
delete_workdir=True,
result_store=result_store,
optional_dict=None,
optional_dict2=None,
parallelize=True,
)
test_fassembler.run()
test_feature_results = test_fassembler.results
test_xs = model.get_xs_from_results(test_feature_results)
test_ys = model.get_ys_from_results(test_feature_results)
test_ys_pred = model.predict(test_xs)
explainer = LocalExplainer(neighbor_samples=1000)
test_exps = explainer.explain(model, test_xs)
explainer.print_explanations(test_exps,
assets=test_assets,
ys=test_ys,
ys_pred=test_ys_pred)
explainer.plot_explanations(test_exps,
assets=test_assets,
ys=test_ys,
ys_pred=test_ys_pred)
plt.show()
def explain_model_on_dataset(model, test_assets_selected_indexs, test_dataset_filepath):
def print_assets(test_assets):
print "\n".join(
map(
lambda (i, asset): "Asset {i}: {name}".format(
i=i, name=get_file_name_without_extension(asset.dis_path)
),
enumerate(test_assets),
)
)
test_dataset = import_python_file(test_dataset_filepath)
test_assets = read_dataset(test_dataset)
print_assets(test_assets)
print "Assets selected for local explanation: {}".format(test_assets_selected_indexs)
result_store = FileSystemResultStore()
test_assets = [test_assets[i] for i in test_assets_selected_indexs]
test_fassembler = FeatureAssembler(
feature_dict=model.model_dict["feature_dict"],
feature_option_dict=None,
assets=test_assets,
logger=None,
fifo_mode=True,
delete_workdir=True,
result_store=result_store,
optional_dict=None,
optional_dict2=None,
parallelize=True,
)
test_fassembler.run()
test_feature_results = test_fassembler.results
test_xs = model.get_xs_from_results(test_feature_results)
test_ys = model.get_ys_from_results(test_feature_results)
test_ys_pred = model.predict(test_xs)
explainer = LocalExplainer(neighbor_samples=1000)
test_exps = explainer.explain(model, test_xs)
explainer.print_explanations(test_exps, assets=test_assets, ys=test_ys, ys_pred=test_ys_pred)
explainer.plot_explanations(test_exps, assets=test_assets, ys=test_ys, ys_pred=test_ys_pred)
plt.show()
