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"]
def train(
ml_params: Dict,
experiment_params: Dict,
predictor_save_dir: str = None,
s3_bucket_to_write_to: str = None,
):
logger.info("Initializing data reader...")
data_reader = BigQueryReader(
credential_path=ml_params["data_reader"]["credential_path"],
bq_project=ml_params["data_reader"]["bq_project"],
bq_dataset=ml_params["data_reader"]["bq_dataset"],
decisions_ds_start=ml_params["data_reader"]["decisions_ds_start"],
decisions_ds_end=ml_params["data_reader"]["decisions_ds_end"],
rewards_ds_end=ml_params["data_reader"]["rewards_ds_end"],
reward_function=ml_params["data_reader"]["reward_function"],
experiment_id=experiment_params["experiment_id"],
)
raw_data = data_reader.get_training_data()
if len(raw_data) == 0:
logger.error(f"Got no raws of training data. Training aborted.")
sys.exit()
logger.info(f"Got {len(raw_data)} rows of training data.")
logger.info(raw_data.head())
utils.fancy_print("Kicking off data preprocessing")
# always add decision as a feature to use if not using all features
features_to_use = ml_params["data_reader"].get("features_to_use", ["*"])
if features_to_use != ["*"]:
features_to_use.append(preprocessor.DECISION_FEATURE_NAME)
features_to_use = list(set(features_to_use))
dense_features_to_use = ml_params["data_reader"].get("dense_features_to_use", ["*"])
data = preprocessor.preprocess_data(
raw_data,
experiment_params,
ml_params["reward_type"],
features_to_use,
dense_features_to_use,
)
X, y = preprocessor.data_to_pytorch(data)
model_type = ml_params["model_type"]
model_params = ml_params["model_params"][model_type]
reward_type = ml_params["reward_type"]
feature_importance_params = ml_params.get("feature_importance", {})
if feature_importance_params.get("calc_feature_importance", False):
# calculate feature importances - only works on non id list features at this time
utils.fancy_print("Calculating feature importances")
feature_scores = feature_importance.calculate_feature_importance(
reward_type=reward_type,
feature_names=data["final_float_feature_order"],
X=X,
y=y,
)
feature_importance.display_feature_importances(feature_scores)
# TODO: Make keeping the top "n" features work in predictor. Right now
# using this feature breaks predictor, so don't use it in a final model,
# just use it to experiment in seeing how model performance is.
if feature_importance_params.get("keep_only_top_n", False):
utils.fancy_print("Keeping only top N features")
X, final_float_feature_order = feature_importance.keep_top_n_features(
n=feature_importance_params["n"],
X=X,
feature_order=data["final_float_feature_order"],
feature_scores=feature_scores,
)
data["final_float_feature_order"] = final_float_feature_order
logger.info(f"Keeping top {feature_importance_params['n']} features:")
logger.info(final_float_feature_order)
utils.fancy_print("Starting training")
# build the model
if model_type == "neural_bandit":
model_spec, model = model_constructors.build_pytorch_net(
feature_specs=experiment_params["features"],
product_sets=experiment_params["product_sets"],
float_feature_order=data["final_float_feature_order"],
id_feature_order=data["final_id_feature_order"],
reward_type=reward_type,
layers=model_params["layers"],
activations=model_params["activations"],
dropout_ratio=model_params["dropout_ratio"],
input_dim=num_float_dim(data),
)
logger.info(f"Initialized model: {model}")
elif model_type == "linear_bandit":
assert utils.pset_features_have_dense(experiment_params["features"]), (
"Linear models require that product set features have associated"
"dense representations."
)
model = model_constructors.build_linear_model(
reward_type=reward_type,
penalty=model_params.get("penalty"),
alpha=model_params.get("alpha"),
)
model_spec = None
elif model_type == "gbdt_bandit":
assert utils.pset_features_have_dense(experiment_params["features"]), (
"GBDT models require that product set features have associated"
"dense representations."
)
model = model_constructors.build_gbdt(
reward_type=reward_type,
learning_rate=model_params["learning_rate"],
n_estimators=model_params["n_estimators"],
max_depth=model_params["max_depth"],
)
model_spec = None
elif model_type == "random_forest_bandit":
assert utils.pset_features_have_dense(experiment_params["features"]), (
"Random forest models require that product set features have associated"
"dense representations."
)
model = model_constructors.build_random_forest(
reward_type=reward_type,
n_estimators=model_params["n_estimators"],
max_depth=model_params["max_depth"],
)
model_spec = None
# build the predictor
predictor = BanditPredictor(
experiment_params=experiment_params,
float_feature_order=data["float_feature_order"],
id_feature_order=data["id_feature_order"],
id_feature_str_to_int_map=data["id_feature_str_to_int_map"],
transforms=data["transforms"],
imputers=data["imputers"],
model=model,
model_type=model_type,
reward_type=reward_type,
model_spec=model_spec,
dense_features_to_use=dense_features_to_use,
)
# train the model
if model_type == "neural_bandit":
logger.info(f"Training {model_type} for {model_params['max_epochs']} epochs")
skorch_net = model_trainers.fit_custom_pytorch_module_w_skorch(
reward_type=reward_type,
model=predictor.model,
X=X,
y=y,
hyperparams=model_params,
train_percent=ml_params["train_percent"],
)
elif model_type in ("gbdt_bandit", "random_forest_bandit", "linear_bandit"):
logger.info(f"Training {model_type}")
sklearn_model, _ = model_trainers.fit_sklearn_model(
reward_type=reward_type,
model=model,
X=X,
y=y,
train_percent=ml_params["train_percent"],
)
if predictor_save_dir is not None:
logger.info("Saving predictor artifacts to disk...")
experiment_id = experiment_params.get("experiment_id", "test")
model_name = ml_params.get("model_name", "model")
save_dir = f"{predictor_save_dir}/{experiment_id}"
if not os.path.exists(save_dir):
os.makedirs(save_dir)
predictor_net_path = f"{save_dir}/{model_name}.pt"
predictor_config_path = f"{save_dir}/{model_name}.json"
predictor.config_to_file(predictor_config_path)
predictor.model_to_file(predictor_net_path)
if s3_bucket_to_write_to is not None:
logger.info("Writing predictor artifacts to s3...")
# Assumes aws credentials stored in ~/.aws/credentials that looks like:
# [default]
# aws_access_key_id = YOUR_ACCESS_KEY
# aws_secret_access_key = YOUR_SECRET_KEY
dir_to_zip = save_dir
output_path = save_dir
shutil.make_archive(output_path, "zip", dir_to_zip)
s3_client = boto3.client("s3")
s3_client.upload_file(
Filename=f"{output_path}.zip",
Bucket=s3_bucket_to_write_to,
Key=f"{experiment_id}.zip",
)
class Datasets:
TEST_DIR = os.path.dirname(os.path.abspath(__file__))
TEST_DATASET_DIR = "datasets"
# continuous reward
TEST_DATASET_FILENAME = "height_dataset.csv"
DATASET_PATH = os.path.join(TEST_DIR, TEST_DATASET_DIR,
TEST_DATASET_FILENAME)
_raw_data = pd.read_csv(DATASET_PATH)
_offset = int(len(_raw_data) * Params.ML_PARAMS["train_percent"])
# dataset for country as categorical variable
DATA_COUNTRY_CATEG = preprocessor.preprocess_data(
_raw_data,
Params.ML_PARAMS["data_reader"]["reward_function"],
Params.EXPERIMENT_SPECIFIC_PARAMS_COUNTRY_AS_CATEGORICAL,
)
_X, _y = preprocessor.data_to_pytorch(DATA_COUNTRY_CATEG)
X_COUNTRY_CATEG = {
"X_train": {
"X_float": _X["X_float"][:_offset]
},
"y_train": _y[:_offset],
"X_test": {
"X_float": _X["X_float"][_offset:]
},
"y_test": _y[_offset:],
}
# dataset for country as ID list variable
DATA_COUNTRY_ID_LIST = preprocessor.preprocess_data(
_raw_data,
Params.ML_PARAMS["data_reader"]["reward_function"],
Params.EXPERIMENT_SPECIFIC_PARAMS_COUNTRY_AS_ID_LIST,
)
_X, _y = preprocessor.data_to_pytorch(DATA_COUNTRY_ID_LIST)
X_COUNTRY_ID_LIST = {
"X_train": {
"X_float": _X["X_float"][:_offset],
"X_id_list": _X["X_id_list"][:_offset],
"X_id_list_idxs": _X["X_id_list_idxs"][:_offset],
},
"y_train": _y[:_offset],
"X_test": {
"X_float": _X["X_float"][_offset:],
"X_id_list": _X["X_id_list"][_offset:],
"X_id_list_idxs": _X["X_id_list_idxs"][_offset:],
},
"y_test": _y[_offset:],
}
# dataset for country as dense ID list variable
DATA_COUNTRY_DENSE_ID_LIST = preprocessor.preprocess_data(
_raw_data,
Params.ML_PARAMS["data_reader"]["reward_function"],
Params.EXPERIMENT_SPECIFIC_PARAMS_COUNTRY_AS_DENSE_ID_LIST,
)
_X, _y = preprocessor.data_to_pytorch(DATA_COUNTRY_DENSE_ID_LIST)
X_COUNTRY_DENSE_ID_LIST = {
"X_train": {
"X_float": _X["X_float"][:_offset]
},
"y_train": _y[:_offset],
"X_test": {
"X_float": _X["X_float"][_offset:]
},
"y_test": _y[_offset:],
}
# dataset for country as ID list AND decision as ID list variables
DATA_COUNTRY_AND_DECISION_ID_LIST = preprocessor.preprocess_data(
_raw_data,
Params.ML_PARAMS["data_reader"]["reward_function"],
Params.EXPERIMENT_SPECIFIC_PARAMS_COUNTRY_AND_DECISION_AS_ID_LIST,
)
_X, _y = preprocessor.data_to_pytorch(DATA_COUNTRY_AND_DECISION_ID_LIST)
X_COUNTRY_AND_DECISION_ID_LIST = {
"X_train": {
"X_float": _X["X_float"][:_offset],
"X_id_list": _X["X_id_list"][:_offset],
"X_id_list_idxs": _X["X_id_list_idxs"][:_offset],
},
"y_train": _y[:_offset],
"X_test": {
"X_float": _X["X_float"][_offset:],
"X_id_list": _X["X_id_list"][_offset:],
"X_id_list_idxs": _X["X_id_list_idxs"][_offset:],
},
"y_test": _y[_offset:],
}
# binary reward
TEST_BINARY_REWARD_DATASET_FILENAME = "height_dataset_binary.csv"
BINARY_REWARD_DATASET_PATH = os.path.join(
TEST_DIR, TEST_DATASET_DIR, TEST_BINARY_REWARD_DATASET_FILENAME)
_raw_data_binary_reward = pd.read_csv(BINARY_REWARD_DATASET_PATH)
_offset_binary_reward = int(
len(_raw_data_binary_reward) * Params.ML_PARAMS["train_percent"])
# dataset for country as categorical variable & binary reward
DATA_COUNTRY_CATEG_BINARY_REWARD = preprocessor.preprocess_data(
_raw_data_binary_reward,
Params.REWARD_FUNCTION_BINARY,
Params.EXPERIMENT_SPECIFIC_PARAMS_COUNTRY_AS_CATEGORICAL,
)
_X_binary_reward, _y_binary_reward = preprocessor.data_to_pytorch(
DATA_COUNTRY_CATEG_BINARY_REWARD)
X_COUNTRY_CATEG_BINARY_REWARD = {
"X_train": {
"X_float": _X_binary_reward["X_float"][:_offset_binary_reward]
},
"y_train": _y_binary_reward[:_offset_binary_reward],
"X_test": {
"X_float": _X_binary_reward["X_float"][_offset_binary_reward:]
},
"y_test": _y_binary_reward[_offset_binary_reward:],
}
def train(
ml_params: Dict,
experiment_params: Dict,
model_name: str = None,
predictor_save_dir: str = None,
s3_bucket_to_write_to: str = None,
):
logger.info("Initializing data reader...")
data_reader = BigQueryReader(
credential_path=ml_params["data_reader"]["credential_path"],
bq_project=ml_params["data_reader"]["bq_project"],
bq_dataset=ml_params["data_reader"]["bq_dataset"],
decisions_table=ml_params["data_reader"]["decisions_table"],
rewards_table=ml_params["data_reader"]["rewards_table"],
decisions_ds_start=ml_params["data_reader"]["decisions_ds_start"],
decisions_ds_end=ml_params["data_reader"]["decisions_ds_end"],
rewards_ds_end=ml_params["data_reader"]["rewards_ds_end"],
experiment_id=experiment_params["experiment_id"],
)
raw_data = data_reader.get_training_data()
if len(raw_data) == 0:
logger.error(f"Got no raws of training data. Training aborted.")
sys.exit()
logger.info(f"Got {len(raw_data)} rows of training data.")
logger.info(raw_data.head())
logger.info("Kicking off data preprocessing...")
data = preprocessor.preprocess_data(
raw_data, ml_params["data_reader"]["reward_function"],
experiment_params)
X, y = preprocessor.data_to_pytorch(data)
model_type = ml_params["model_type"]
model_params = ml_params["model_params"][model_type]
reward_type = ml_params["reward_type"]
# build the model
if model_type == "neural_bandit":
model_spec, model = model_constructors.build_pytorch_net(
feature_specs=experiment_params["features"],
product_sets=experiment_params["product_sets"],
float_feature_order=data["final_float_feature_order"],
id_feature_order=data["final_id_feature_order"],
reward_type=reward_type,
layers=model_params["layers"],
activations=model_params["activations"],
dropout_ratio=model_params["dropout_ratio"],
input_dim=num_float_dim(data),
)
logger.info(f"Initialized model: {model}")
elif model_type == "gbdt_bandit":
assert utils.pset_features_have_dense(experiment_params["features"]), (
"GBDT models require that product set features have associated"
"dense reprenstations.")
model = model_constructors.build_gbdt(
reward_type=reward_type,
learning_rate=model_params["learning_rate"],
n_estimators=model_params["n_estimators"],
max_depth=model_params["max_depth"],
)
model_spec = None
elif model_type == "random_forest_bandit":
assert utils.pset_features_have_dense(experiment_params["features"]), (
"Random forest models require that product set features have associated"
"dense reprenstations.")
model = model_constructors.build_random_forest(
reward_type=reward_type,
n_estimators=model_params["n_estimators"],
max_depth=model_params["max_depth"],
)
model_spec = None
# build the predictor
predictor = BanditPredictor(
experiment_params=experiment_params,
float_feature_order=data["float_feature_order"],
id_feature_order=data["id_feature_order"],
id_feature_str_to_int_map=data["id_feature_str_to_int_map"],
transforms=data["transforms"],
imputers=data["imputers"],
model=model,
reward_type=reward_type,
model_spec=model_spec,
)
# train the model
if model_type == "neural_bandit":
logger.info(f"Starting training: {model_params} epochs")
skorch_net = model_trainers.fit_custom_pytorch_module_w_skorch(
reward_type=reward_type,
model=predictor.model,
X=X,
y=y,
hyperparams=model_params,
train_percent=ml_params["train_percent"],
)
elif model_type in ("gbdt_bandit", "random_forest_bandit"):
logger.info(f"Starting training: {model_type}")
sklearn_model, _ = model_trainers.fit_sklearn_model(
reward_type=reward_type,
model=model,
X=X,
y=y,
train_percent=ml_params["train_percent"],
)
if predictor_save_dir is not None:
logger.info("Saving predictor artifacts to disk...")
experiment_id = experiment_params.get("experiment_id", "test")
model_name = experiment_params.get("model_name", "model")
save_dir = f"{predictor_save_dir}/{experiment_id}"
if not os.path.exists(save_dir):
os.makedirs(save_dir)
predictor_net_path = f"{save_dir}/{model_name}.pt"
predictor_config_path = f"{save_dir}/{model_name}.json"
predictor.config_to_file(predictor_config_path)
predictor.model_to_file(predictor_net_path)
if s3_bucket_to_write_to is not None:
logger.info("Writing predictor artifacts to s3...")
# Assumes aws credentials stored in ~/.aws/credentials that looks like:
# [default]
# aws_access_key_id = YOUR_ACCESS_KEY
# aws_secret_access_key = YOUR_SECRET_KEY
dir_to_zip = save_dir
output_path = save_dir
shutil.make_archive(output_path, "zip", dir_to_zip)
s3_client = boto3.client("s3")
s3_client.upload_file(
Filename=f"{output_path}.zip",
Bucket=s3_bucket_to_write_to,
Key=f"{experiment_id}.zip",
)
class Datasets:
TEST_DIR = os.path.dirname(os.path.abspath(__file__))
TEST_DATASET_DIR = "datasets"
TEST_DATASET_FILENAME = "height_dataset.csv"
DATASET_PATH = os.path.join(TEST_DIR, TEST_DATASET_DIR, TEST_DATASET_FILENAME)
_raw_data = pd.read_csv(DATASET_PATH)
_offset = int(len(_raw_data) * Params.SHARED_PARAMS["train_test_split"])
# dataset for country as categorical variable
DATA_COUNTRY_CATEG = preprocessor.preprocess_data(
_raw_data, Params.EXPERIMENT_SPECIFIC_PARAMS_COUNTRY_AS_CATEGORICAL
)
_X, _y = preprocessor.data_to_pytorch(DATA_COUNTRY_CATEG)
X_COUNTRY_CATEG = {
"X_train": {"X_float": _X["X_float"][:_offset]},
"y_train": _y[:_offset],
"X_test": {"X_float": _X["X_float"][_offset:]},
"y_test": _y[_offset:],
}
# dataset for country as ID list variable
DATA_COUNTRY_ID_LIST = preprocessor.preprocess_data(
_raw_data, Params.EXPERIMENT_SPECIFIC_PARAMS_COUNTRY_AS_ID_LIST
)
_X, _y = preprocessor.data_to_pytorch(DATA_COUNTRY_ID_LIST)
X_COUNTRY_ID_LIST = {
"X_train": {
"X_float": _X["X_float"][:_offset],
"X_id_list": _X["X_id_list"][:_offset],
"X_id_list_idxs": _X["X_id_list_idxs"][:_offset],
},
"y_train": _y[:_offset],
"X_test": {
"X_float": _X["X_float"][_offset:],
"X_id_list": _X["X_id_list"][_offset:],
"X_id_list_idxs": _X["X_id_list_idxs"][_offset:],
},
"y_test": _y[_offset:],
}
# dataset for country as dense ID list variable
DATA_COUNTRY_DENSE_ID_LIST = preprocessor.preprocess_data(
_raw_data, Params.EXPERIMENT_SPECIFIC_PARAMS_COUNTRY_AS_DENSE_ID_LIST
)
_X, _y = preprocessor.data_to_pytorch(DATA_COUNTRY_DENSE_ID_LIST)
X_COUNTRY_DENSE_ID_LIST = {
"X_train": {"X_float": _X["X_float"][:_offset]},
"y_train": _y[:_offset],
"X_test": {"X_float": _X["X_float"][_offset:]},
"y_test": _y[_offset:],
}
# dataset for country as ID list AND decision as ID list variables
DATA_COUNTRY_AND_DECISION_ID_LIST = preprocessor.preprocess_data(
_raw_data, Params.EXPERIMENT_SPECIFIC_PARAMS_COUNTRY_AND_DECISION_AS_ID_LIST
)
_X, _y = preprocessor.data_to_pytorch(DATA_COUNTRY_AND_DECISION_ID_LIST)
X_COUNTRY_AND_DECISION_ID_LIST = {
"X_train": {
"X_float": _X["X_float"][:_offset],
"X_id_list": _X["X_id_list"][:_offset],
"X_id_list_idxs": _X["X_id_list_idxs"][:_offset],
},
"y_train": _y[:_offset],
"X_test": {
"X_float": _X["X_float"][_offset:],
"X_id_list": _X["X_id_list"][_offset:],
"X_id_list_idxs": _X["X_id_list_idxs"][_offset:],
},
"y_test": _y[_offset:],
}
def test_same_predictions_country_as_categorical_sklearn_model_binary_reward(
self):
"""
Tests sklearn Linear model + binary reward.
"""
reward_type = "binary"
raw_data = shuffle(Datasets._raw_data_binary_reward)
rand_idx = 0
test_input = raw_data.iloc[rand_idx]
data = preprocessor.preprocess_data(
raw_data,
Params.EXPERIMENT_SPECIFIC_PARAMS_COUNTRY_AS_CATEGORICAL,
reward_type,
shuffle_data=
False, # don't shuffle so we can test the same observation
)
_X, _y = preprocessor.data_to_pytorch(data)
X_COUNTRY_CATEG_BINARY_REWARD = {
"X_train": {
"X_float": _X["X_float"][:Datasets._offset_binary_reward]
},
"y_train": _y[:Datasets._offset_binary_reward],
"X_test": {
"X_float": _X["X_float"][Datasets._offset_binary_reward:]
},
"y_test": _y[Datasets._offset_binary_reward:],
}
model = model_constructors.build_linear_model(reward_type=reward_type)
pre_serialized_predictor = BanditPredictor(
experiment_params=Params.
EXPERIMENT_SPECIFIC_PARAMS_COUNTRY_AS_CATEGORICAL,
float_feature_order=Datasets.
DATA_COUNTRY_CATEG_BINARY_REWARD["float_feature_order"],
id_feature_order=Datasets.
DATA_COUNTRY_CATEG_BINARY_REWARD["id_feature_order"],
id_feature_str_to_int_map=Datasets.
DATA_COUNTRY_CATEG_BINARY_REWARD["id_feature_str_to_int_map"],
transforms=Datasets.DATA_COUNTRY_CATEG_BINARY_REWARD["transforms"],
imputers=Datasets.DATA_COUNTRY_CATEG_BINARY_REWARD["imputers"],
model=model,
model_type="linear_bandit",
reward_type=reward_type,
model_spec=None,
)
skorch_net = model_trainers.fit_sklearn_model(
reward_type=reward_type,
model=model,
X=X_COUNTRY_CATEG_BINARY_REWARD["X_train"],
y=X_COUNTRY_CATEG_BINARY_REWARD["y_train"],
)
pre_serialized_predictor.config_to_file(self.tmp_config_path)
pre_serialized_predictor.model_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"]
def test_same_predictions_country_as_categorical_binary_reward(self):
reward_type = "binary"
raw_data = shuffle(Datasets._raw_data_binary_reward)
rand_idx = 0
test_input = raw_data.iloc[rand_idx]
data = preprocessor.preprocess_data(
raw_data,
Params.EXPERIMENT_SPECIFIC_PARAMS_COUNTRY_AS_CATEGORICAL,
reward_type,
shuffle_data=
False, # don't shuffle so we can test the same observation
)
_X, _y = preprocessor.data_to_pytorch(data)
X_COUNTRY_CATEG_BINARY_REWARD = {
"X_train": {
"X_float": _X["X_float"][:Datasets._offset_binary_reward]
},
"y_train": _y[:Datasets._offset_binary_reward],
"X_test": {
"X_float": _X["X_float"][Datasets._offset_binary_reward:]
},
"y_test": _y[Datasets._offset_binary_reward:],
}
model_spec, pytorch_net = model_constructors.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_BINARY_REWARD["final_float_feature_order"],
id_feature_order=Datasets.
DATA_COUNTRY_CATEG_BINARY_REWARD["final_id_feature_order"],
reward_type=reward_type,
layers=self.model_params["layers"],
activations=self.model_params["activations"],
input_dim=train_bandit.num_float_dim(
Datasets.DATA_COUNTRY_CATEG_BINARY_REWARD),
)
pre_serialized_predictor = BanditPredictor(
experiment_params=Params.
EXPERIMENT_SPECIFIC_PARAMS_COUNTRY_AS_CATEGORICAL,
float_feature_order=Datasets.
DATA_COUNTRY_CATEG_BINARY_REWARD["float_feature_order"],
id_feature_order=Datasets.
DATA_COUNTRY_CATEG_BINARY_REWARD["id_feature_order"],
id_feature_str_to_int_map=Datasets.
DATA_COUNTRY_CATEG_BINARY_REWARD["id_feature_str_to_int_map"],
transforms=Datasets.DATA_COUNTRY_CATEG_BINARY_REWARD["transforms"],
imputers=Datasets.DATA_COUNTRY_CATEG_BINARY_REWARD["imputers"],
model=pytorch_net,
model_type=self.model_type,
reward_type=reward_type,
model_spec=model_spec,
)
skorch_net = model_trainers.fit_custom_pytorch_module_w_skorch(
reward_type=reward_type,
model=pre_serialized_predictor.model,
X=X_COUNTRY_CATEG_BINARY_REWARD["X_train"],
y=X_COUNTRY_CATEG_BINARY_REWARD["y_train"],
hyperparams=self.model_params,
)
pre_serialized_predictor.config_to_file(self.tmp_config_path)
pre_serialized_predictor.model_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"]
# add a test case for missing features in provided context
pre_pred_missing_feature = pre_serialized_predictor.predict({})
post_pred_missing_feature = post_serialized_predictor.predict({})
assert np.allclose(
pre_pred_missing_feature["scores"],
post_pred_missing_feature["scores"],
self.tol,
)
assert pre_pred_missing_feature["ids"] == post_pred_missing_feature[
"ids"]
# add a test case for garbage feature keys provided in context
pre_pred_garbage_feature = pre_serialized_predictor.predict(
{"blah": 42})
post_pred_garbage_feature = post_serialized_predictor.predict(
{"blah": 42})
assert np.allclose(
pre_pred_garbage_feature["scores"],
post_pred_garbage_feature["scores"],
self.tol,
)
assert pre_pred_garbage_feature["ids"] == post_pred_garbage_feature[
"ids"]
def train(
ml_params: Dict,
experiment_params: Dict,
model_name: str = None,
predictor_save_dir: str = None,
s3_bucket_to_write_to: str = None,
):
logger.info("Initializing data reader...")
data_reader = BigQueryReader(
credential_path=ml_params["data_reader"]["credential_path"],
bq_project=ml_params["data_reader"]["bq_project"],
bq_dataset=ml_params["data_reader"]["bq_dataset"],
decisions_table=ml_params["data_reader"]["decisions_table"],
rewards_table=ml_params["data_reader"]["rewards_table"],
decisions_ds_start=ml_params["data_reader"]["decisions_ds_start"],
decisions_ds_end=ml_params["data_reader"]["decisions_ds_end"],
rewards_ds_end=ml_params["data_reader"]["rewards_ds_end"],
experiment_id=experiment_params["experiment_id"],
)
raw_data = data_reader.get_training_data()
if len(raw_data) == 0:
logger.error(f"Got no raws of training data. Training aborted.")
sys.exit()
logger.info(f"Got {len(raw_data)} rows of training data.")
logger.info(raw_data.head())
data = preprocessor.preprocess_data(
raw_data, ml_params["data_reader"]["reward_function"],
experiment_params)
X, y = preprocessor.data_to_pytorch(data)
net_spec, pytorch_net = build_pytorch_net(
feature_specs=experiment_params["features"],
product_sets=experiment_params["product_sets"],
float_feature_order=data["final_float_feature_order"],
id_feature_order=data["final_id_feature_order"],
layers=ml_params["model"]["layers"],
activations=ml_params["model"]["activations"],
dropout_ratio=ml_params["model"]["dropout_ratio"],
input_dim=num_float_dim(data),
)
logger.info(f"Initialized model: {pytorch_net}")
logger.info(f"Starting training: {ml_params['max_epochs']} epochs")
predictor = BanditPredictor(
experiment_params=experiment_params,
float_feature_order=data["float_feature_order"],
id_feature_order=data["id_feature_order"],
id_feature_str_to_int_map=data["id_feature_str_to_int_map"],
transforms=data["transforms"],
imputers=data["imputers"],
net=pytorch_net,
net_spec=net_spec,
)
skorch_net = fit_custom_pytorch_module_w_skorch(module=predictor.net,
X=X,
y=y,
hyperparams=ml_params)
if predictor_save_dir is not None:
logger.info("Saving predictor artifacts to disk...")
experiment_id = experiment_params.get("experiment_id", "test")
model_name = experiment_params.get("model_name", "model")
save_dir = f"{predictor_save_dir}/{experiment_id}"
if not os.path.exists(save_dir):
os.makedirs(save_dir)
predictor_net_path = f"{save_dir}/{model_name}.pt"
predictor_config_path = f"{save_dir}/{model_name}.json"
predictor.config_to_file(predictor_config_path)
predictor.net_to_file(predictor_net_path)
if s3_bucket_to_write_to is not None:
logger.info("Writing predictor artifacts to s3...")
# Assumes aws credentials stored in ~/.aws/credentials that looks like:
# [default]
# aws_access_key_id = YOUR_ACCESS_KEY
# aws_secret_access_key = YOUR_SECRET_KEY
dir_to_zip = save_dir
output_path = save_dir
shutil.make_archive(output_path, "zip", dir_to_zip)
s3_client = boto3.client("s3")
s3_client.upload_file(
Filename=f"{output_path}.zip",
Bucket=s3_bucket_to_write_to,
Key=f"{experiment_id}.zip",
)
