class MLPBandit:
def __init__(self, data, lr, final_hidden_dim=50):
self.mlp = MLP(hidden_layer_sizes=(50, 50, final_hidden_dim),
learning_rate_init=lr,
verbose=True)
if type(data) is str:
raw = pd.read_csv(data).sample(frac=1, replace=False)
else:
raw = data
cutoff = int(.8 * len(raw))
labels = raw[GROUND]
bins = [0, 3 * 7, 7 * 7, labels.max()]
bucketed = pd.cut(labels, bins, include_lowest=True, labels=False)
self.train_y = bucketed[:cutoff]
self.test_y = bucketed[cutoff:]
raw.drop(GROUND, inplace=True, axis=1)
self.train_x = raw[:cutoff]
self.test_x = raw[cutoff:]
def train(self):
self.mlp.fit(self.train_x, self.train_y)
def test(self):
return self.mlp.score(self.test_x, self.test_y)
def get_activations(self, X):
hidden_layer_sizes = self.mlp.hidden_layer_sizes
if not hasattr(hidden_layer_sizes, "__iter__"):
hidden_layer_sizes = [hidden_layer_sizes]
hidden_layer_sizes = list(hidden_layer_sizes)
layer_units = [X.shape[1]] + hidden_layer_sizes + \
[self.mlp.n_outputs_]
activations = [X]
for i in range(self.mlp.n_layers_ - 1):
activations.append(np.empty((X.shape[0], layer_units[i + 1])))
self.mlp._forward_pass(activations)
return activations[-2]
