parser.add_argument("parameters_path", help="Parameters output", type=str)
parser.add_argument('-j',
'--jitter',
help="Jitter on/off",
action='store_true')
parser.add_argument("--with_l2_regularization",
action='store_true',
help="Enable L2 regularization")
args = parser.parse_args()
logging.info("Loading dataset...")
dataset_df = pd.read_csv(args.dataset)
smote = SMOTE(random_state=0, k_neighbors=args.k_neighbors)
smote_resample = build_resample_regression_dataset_func(smote,
stepsize=args.stepsize,
jitter=args.jitter)
class ExperimentData(TFExperiment):
def load_dataset(self):
X = dataset_df.drop("year_0", axis=1).values
y = dataset_df["year_0"].values
return (X, y)
def preprocess(self, X, y):
return smote_resample(X, y)
logging.info(
"Running grid search for learning rate and # of neurons %s(num trials = %d, num_epochs = %d)...",
**Evaluation method:** {}-fold cross validation
**Optimization method:** SGD
Parameters:
* num_neurons_p = {} — number of neurons in the hidden layer
* learning_rate_p = {} — learning rate
* momentum = {} — momentum
* l2_reg = {} — L2 regularization
* discretization_stepsize = {} — discretization step size
* k_neighbors = {} — number of neighbors for ADASYN
* jitter = {} — jitter on/off""".format(
args.num_folds, params["num_neurons_p"],
params["learning_rate_p"], params["momentum_p"], params["l2_reg_p"], args.stepsize, args.k_neighbors, args.jitter)
logging.info(experiment_description)
adasyn = ADASYN(random_state=0, n_neighbors=args.k_neighbors)
adasyn_resample = build_resample_regression_dataset_func(adasyn, args.stepsize, jitter=args.jitter)
class Experiment(TFExperiment):
def load_dataset(self):
X = dataset_df.drop("year_0", axis=1).values
y = dataset_df["year_0"].values
return (X, y)
def build_model(self):
model = build_1layer_perceptron(
params["num_neurons_p"], params["learning_rate_p"], params["momentum_p"], params["l2_reg_p"]
)
return model
def preprocess(self, X, y):
return adasyn_resample(X, y)
Using randomized resampling with jitter to balance the dataset.
**Evaluation method:** {}-fold cross validation
**Optimization method:** SGD
Parameters:
* num_neurons_p = {} — number of neurons in the hidden layer
* learning_rate_p = {} — learning rate
* momentum = {} — momentum
* l2_reg = {} — L2 regularization
* discretization_stepsize = {} — discretization step size""".format(
args.num_folds, params["num_neurons_p"], params["learning_rate_p"],
params["momentum_p"], params["l2_reg_p"], args.stepsize)
logging.info(experiment_description)
ros = RandomOverSampler(random_state=0)
random_resample = build_resample_regression_dataset_func(
ros, stepsize=args.stepsize)
class Experiment(TFExperiment):
def load_dataset(self):
X = dataset_df.drop("year_0", axis=1).values
y = dataset_df["year_0"].values
return (X, y)
def build_model(self):
model = build_1layer_perceptron(params["num_neurons_p"],
params["learning_rate_p"],
params["momentum_p"],
params["l2_reg_p"])
return model