def output(m, d):
return {
"matmul":
Matmul.specification(
name="matmul",
num_nodes=m,
num_features=d,
weights_initialization_scheme="he",
weights_optimizer_specification=optimizer.SGD.specification(
lr=0.05, l2=1e-3)),
"loss":
CrossEntropyLogLoss.specification(name="loss", num_nodes=m)
}
def test():
M = 1
D = 2
N = 100
X, T, V = linear_separable(d=D, n=N)
x_min, x_max = X[:, 0].min(), X[:, 0].max()
y_min, y_max = X[:, 1].min(), X[:, 1].max()
sigmoid_classifier_specification = {
_NAME: "softmax_classifier",
_NUM_NODES: M,
_LOG_LEVEL: logging.ERROR,
_COMPOSITE_LAYER_SPEC: {
"matmul01":
Matmul.specification(
name="matmul",
num_nodes=M,
num_features=D,
weights_initialization_scheme="he",
weights_optimizer_specification=SGD.specification(
lr=TYPE_FLOAT(0.2), l2=TYPE_FLOAT(1e-3))),
"loss":
CrossEntropyLogLoss.specification(
name="loss",
num_nodes=M,
loss_function=sigmoid_cross_entropy_log_loss.__qualname__)
}
}
logistic_classifier = SequentialNetwork.build(
specification=sigmoid_classifier_specification, )
for i in range(50):
logistic_classifier.train(X=X, T=T)
prediction = logistic_classifier.predict(
np.array([-1., -1.], dtype=TYPE_FLOAT))
np.isin(prediction, [0, 1])
print(prediction)
def multilayer_network_specification_bn_to_fail(D, M01, M02, M):
sequential_layer_specification_bn_to_fail = {
"matmul01": layer.Matmul.specification(
name="matmul01",
num_nodes=M01,
num_features=D,
weights_initialization_scheme="he",
weights_optimizer_specification=optimiser.SGD.specification(
lr=TYPE_FLOAT(0.05),
l2=TYPE_FLOAT(1e-3)
)
),
"bn01": layer.BatchNormalization.specification(
name="bn01",
num_nodes=M01,
gamma_optimizer_specification=optimiser.SGD.specification(
lr=TYPE_FLOAT(0.05),
l2=TYPE_FLOAT(1e-3)
),
beta_optimizer_specification=optimiser.SGD.specification(
lr=TYPE_FLOAT(0.05),
l2=TYPE_FLOAT(1e-3),
),
momentum=TYPE_FLOAT(0.9)
),
"relu01": layer.ReLU.specification(
name="relu01",
num_nodes=M01,
),
"matmul02": layer.Matmul.specification(
name="matmul01",
num_nodes=M02,
num_features=M01,
weights_initialization_scheme="he",
weights_optimizer_specification=optimiser.SGD.specification(
lr=TYPE_FLOAT(0.05),
l2=TYPE_FLOAT(1e-3)
)
),
"bn02": layer.BatchNormalization.specification(
name="bn02",
num_nodes=M02,
gamma_optimizer_specification=optimiser.SGD.specification(
lr=TYPE_FLOAT(0.05),
l2=TYPE_FLOAT(1e-3)
),
beta_optimizer_specification=optimiser.SGD.specification(
lr=TYPE_FLOAT(0.05),
l2=TYPE_FLOAT(1e-3),
),
momentum=TYPE_FLOAT(0.9)
),
"relu02": layer.ReLU.specification(
name="relu02",
num_nodes=M02,
),
"matmul03": layer.Matmul.specification(
name="matmul03",
num_nodes=M,
num_features=M02,
weights_initialization_scheme="he",
weights_optimizer_specification=optimiser.SGD.specification(
lr=TYPE_FLOAT(0.05),
l2=TYPE_FLOAT(1e-3)
)
),
"bn03": layer.BatchNormalization.specification(
name="bn03",
num_nodes=M,
gamma_optimizer_specification=optimiser.SGD.specification(
lr=TYPE_FLOAT(0.05),
l2=TYPE_FLOAT(1e-3)
),
beta_optimizer_specification=optimiser.SGD.specification(
lr=TYPE_FLOAT(0.05),
l2=TYPE_FLOAT(1e-3),
),
momentum=TYPE_FLOAT(0.9)
),
"loss": CrossEntropyLogLoss.specification(
name="loss001", num_nodes=M
)
}
return {
_NAME: "two_layer_classifier_with_batch_normalization",
_NUM_NODES: M,
_LOG_LEVEL: logging.ERROR,
_COMPOSITE_LAYER_SPEC: sequential_layer_specification_bn_to_fail
}