def test_pytorch_model_country_as_dense_id_list(self):
net_spec, pytorch_net = train_bandit.build_pytorch_net(
feature_specs=Params.
EXPERIMENT_SPECIFIC_PARAMS_COUNTRY_AS_DENSE_ID_LIST["features"],
product_sets=Params.
EXPERIMENT_SPECIFIC_PARAMS_COUNTRY_AS_DENSE_ID_LIST[
"product_sets"],
float_feature_order=Datasets.
DATA_COUNTRY_DENSE_ID_LIST["final_float_feature_order"],
id_feature_order=Datasets.
DATA_COUNTRY_DENSE_ID_LIST["final_id_feature_order"],
layers=Params.ML_PARAMS["model"]["layers"],
activations=Params.ML_PARAMS["model"]["activations"],
input_dim=train_bandit.num_float_dim(
Datasets.DATA_COUNTRY_DENSE_ID_LIST),
)
skorch_net = train_bandit.fit_custom_pytorch_module_w_skorch(
module=pytorch_net,
X=Datasets.X_COUNTRY_DENSE_ID_LIST["X_train"],
y=Datasets.X_COUNTRY_DENSE_ID_LIST["y_train"],
hyperparams=Params.ML_PARAMS,
)
test_mse = skorch_net.history[-1]["valid_loss"]
# make sure mse is better or close to out of the box GBDT & MLP
# the GBDT doesn't need as much training so make tolerance more forgiving
assert test_mse < self.results_gbdt["mse_test"] * 1.15
assert test_mse < self.results_mlp["mse_test"] * 1.15
def test_same_predictions_country_as_categorical(self):
raw_data = shuffle(Datasets._raw_data)
rand_idx = 0
test_input = raw_data.iloc[rand_idx]
data = preprocessor.preprocess_data(
raw_data,
self.ml_params["data_reader"]["reward_function"],
Params.EXPERIMENT_SPECIFIC_PARAMS_COUNTRY_AS_CATEGORICAL,
shuffle_data=False, # don't shuffle so we can test the same observation
)
_X, _y = preprocessor.data_to_pytorch(data)
X_COUNTRY_CATEG = {
"X_train": {"X_float": _X["X_float"][: Datasets._offset]},
"y_train": _y[: Datasets._offset],
"X_test": {"X_float": _X["X_float"][Datasets._offset :]},
"y_test": _y[Datasets._offset :],
}
net_spec, pytorch_net = train_bandit.build_pytorch_net(
feature_specs=Params.EXPERIMENT_SPECIFIC_PARAMS_COUNTRY_AS_CATEGORICAL[
"features"
],
product_sets=Params.EXPERIMENT_SPECIFIC_PARAMS_COUNTRY_AS_CATEGORICAL[
"product_sets"
],
float_feature_order=Datasets.DATA_COUNTRY_CATEG[
"final_float_feature_order"
],
id_feature_order=Datasets.DATA_COUNTRY_CATEG["final_id_feature_order"],
layers=self.ml_params["model"]["layers"],
activations=self.ml_params["model"]["activations"],
input_dim=train_bandit.num_float_dim(Datasets.DATA_COUNTRY_CATEG),
)
pre_serialized_predictor = BanditPredictor(
experiment_params=Params.EXPERIMENT_SPECIFIC_PARAMS_COUNTRY_AS_CATEGORICAL,
float_feature_order=Datasets.DATA_COUNTRY_CATEG["float_feature_order"],
id_feature_order=Datasets.DATA_COUNTRY_CATEG["id_feature_order"],
id_feature_str_to_int_map=Datasets.DATA_COUNTRY_CATEG[
"id_feature_str_to_int_map"
],
transforms=Datasets.DATA_COUNTRY_CATEG["transforms"],
imputers=Datasets.DATA_COUNTRY_CATEG["imputers"],
net=pytorch_net,
net_spec=net_spec,
)
skorch_net = train_bandit.fit_custom_pytorch_module_w_skorch(
module=pre_serialized_predictor.net,
X=X_COUNTRY_CATEG["X_train"],
y=X_COUNTRY_CATEG["y_train"],
hyperparams=self.ml_params,
)
pre_serialized_predictor.config_to_file(self.tmp_config_path)
pre_serialized_predictor.net_to_file(self.tmp_net_path)
post_serialized_predictor = BanditPredictor.predictor_from_file(
self.tmp_config_path, self.tmp_net_path
)
pre_pred = pre_serialized_predictor.predict(json.loads(test_input.context))
post_pred = post_serialized_predictor.predict(json.loads(test_input.context))
assert np.allclose(pre_pred["scores"], post_pred["scores"], self.tol)
assert pre_pred["ids"] == post_pred["ids"]