def test_mobilenet():
data = load_train()
r = convert_to_normalized_array(data['band_1']).reshape(-1, 75, 75)
g = convert_to_normalized_array(data['band_2']).reshape(-1, 75, 75)
b = r / (g + 0.01)
X = np.dstack([r, g, b]).reshape(-1, 75, 75, 3)
y = data['is_iceberg']
X_resized = []
for i in range(len(X)):
new_img = resize(X[i], (128, 128))
X_resized.append(new_img)
X_resized = np.concatenate(X_resized).reshape(-1, 128, 128, 3)
X_tr, y_tr = X_resized[:1408], y[:1408]
X_val, y_val = X_resized[1408:], y[1408:]
model = build_model()
model.compile(
loss='binary_crossentropy',
optimizer=optimizers.Adadelta(),
metrics=['binary_accuracy'])
model.fit(
X_tr, y_tr,
batch_size=128,
epochs=50,
validation_data=(X_val, y_val),
callbacks=[TensorBoard(log_dir='/tmp/mobilenet/run-{}'.format(datetime.utcnow().strftime("%Y%m%d%H%M%S")))])
def test_vgg19():
data = load_train()
r = convert_to_normalized_array(data['band_1']).reshape(-1, 75, 75)
g = convert_to_normalized_array(data['band_2']).reshape(-1, 75, 75)
b = r / (g + 0.01)
X = np.dstack([r, g, b]).reshape(-1, 75, 75, 3)
y = data['is_iceberg']
X_tr, y_tr = X[:1408], y[:1408]
X_val, y_val = X[1408:], y[1408:]
model = build_model()
model.compile(loss='binary_crossentropy',
optimizer=optimizers.Adadelta(),
metrics=['binary_accuracy'])
model.fit(X_tr,
y_tr,
batch_size=128,
epochs=50,
validation_data=(X_val, y_val),
callbacks=[
TensorBoard(log_dir='/tmp/vgg19/run-{}'.format(
datetime.utcnow().strftime("%Y%m%d%H%M%S")))
])
def test_knn():
data = load_train()
X = convert_to_normalized_array(data['band_1'])[:, 25:50, 25:50].reshape(
-1, 25 * 25)
y = data['is_iceberg']
model = KNeighborsClassifier(n_neighbors=20)
scores = cross_val_score(model, X, y, scoring='neg_log_loss')
assert np.mean(-scores) == approx(0.48, abs=0.02)