def test_PFN_required(input_dim, Phi_sizes, F_sizes):
n, m = 50, 10
X_train = np.random.rand(n, m, input_dim)
Y_train = np.random.rand(n, 2)
pfn = archs.PFN(input_dim=input_dim,
Phi_sizes=Phi_sizes,
F_sizes=F_sizes,
summary=False)
pfn.fit(X_train, Y_train, epochs=1, batch_size=10)
pfn.inputs, pfn.latent, pfn.weights, pfn.outputs, pfn.Phi, pfn.F
def test_PFN_required(nglobal):
n, m = 50, 10
X_train = [np.random.rand(n, m, 3), np.random.rand(n, nglobal)]
Y_train = np.random.rand(n, 2)
X_val = [np.random.rand(n // 10, m, 3), np.random.rand(n // 10, nglobal)]
Y_val = np.random.rand(n // 10, 2)
pfn = archs.PFN(input_dim=3,
Phi_sizes=[10],
F_sizes=[10],
num_global_features=nglobal,
summary=False)
hist = pfn.fit(X_train,
Y_train,
epochs=1,
batch_size=5,
validation_data=(X_val, Y_val))
pfn._global_feature_tensor
def test_PFN_masking(mask_val):
n, m1, m2 = (50, 10, 20)
input_dim = 3
X_train = np.random.rand(n, m1, input_dim)
y_train = np.random.rand(n, 2)
pfn = archs.PFN(input_dim=input_dim,
Phi_sizes=[10],
F_sizes=[10],
mask_val=mask_val)
pfn.fit(X_train, y_train, epochs=1)
X_test = np.random.rand(n, m2, input_dim)
X_test_mask = np.concatenate((X_test, mask_val * np.ones(
(n, 5, input_dim))),
axis=1)
assert epsilon_diff(pfn.predict(X_test), pfn.predict(X_test_mask), 10**-15)
kf = K.function(inputs=pfn.inputs, outputs=pfn.latent)
pure_mask = kf([0 * X_test + mask_val])[0]
assert epsilon_diff(pure_mask, 0, 10**-15)