def predict_pipeline(args):
test_meta = _load_meta_testing(args)
if args.dev_mode:
test_meta = test_meta.sample(128, replace=False, random_state=1234)
pipeline_name = _parse_neptune_params(args, 'pipeline_name')
Pipeline = registered_pipelines[pipeline_name]
pipeline = Pipeline(
num_classes=_parse_neptune_params(args, 'top_categories') + 1,
epochs=_parse_neptune_params(args, 'epochs'),
workers=args.nb_workers,
models_dir=os.path.join(args.models_dir, 'single_models'),
)
pipeline_filepath = os.path.join(
os.path.join(args.models_dir, 'pipelines'),
'{}_{}'.format(args.name, pipeline_name))
pipeline = pipeline_load(pipeline, pipeline_filepath)
test_filepath = os.path.join(args.raw_data_dir, 'test.bson')
y_test_pred = pipeline.predict(X=test_meta,
img_dataset_filepath=test_filepath)
submission = test_meta[['_id']]
submission['category_id'] = y_test_pred
timestr = datetime.now().strftime("%Y%m%d-%H%M%S")
submission_filepath = os.path.join(
args.submissions_dir,
'{}_{}.csv'.format('{}_{}'.format(args.name, pipeline_name), timestr))
submission.to_csv(submission_filepath, index=None)
def evaluate_pipeline(args):
train_meta, valid_meta = _load_meta_training(args)
if args.sample_validation:
valid_meta = valid_meta.sample(args.sample_validation,
replace=False,
random_state=1234)
if args.dev_mode:
valid_meta = valid_meta.sample(128, replace=False, random_state=1234)
train_filepath = os.path.join(args.raw_data_dir, 'train.bson')
pipeline_name = _parse_neptune_params(args, 'pipeline_name')
Pipeline = registered_pipelines[pipeline_name]
pipeline = Pipeline(
num_classes=_parse_neptune_params(args, 'top_categories') + 1,
epochs=_parse_neptune_params(args, 'epochs'),
workers=args.nb_workers,
models_dir=os.path.join(args.models_dir, 'single_models'),
)
pipeline_filepath = os.path.join(
os.path.join(args.models_dir, 'pipelines'),
'{}_{}'.format(args.name, pipeline_name))
pipeline = pipeline_load(pipeline, pipeline_filepath)
y_pred = pipeline.predict(X=valid_meta,
img_dataset_filepath=train_filepath)
y_true = valid_meta['category_id']
score = accuracy_score(y_true, y_pred)
neptune_post_pipeline_score(score)