def entire_network(self, features, params, is_training, reuse_variables):
"""The definition of the entire network.
Sometimes, feature normalization is applied after the main network.
We combine them together (except for the loss layer).
Args:
features: The network input.
params: The parameters.
is_training: True if the network is for training.
reuse_variables: Share variables.
:return: The network output and the endpoints (for other usage).
"""
features, endpoints = self.network(features, params, is_training, reuse_variables)
endpoints["output"] = features
# Add more components (post-processing) after the main network.
if "feature_norm" in params.dict and params.feature_norm:
assert "feature_scaling_factor" in params.dict, "If feature normalization is applied, scaling factor is necessary."
features = l2_scaling(features, params.feature_scaling_factor)
endpoints["output"] = features
return features, endpoints
params.dict["amsoftmax_lambda_min"] = 10
params.dict["amsoftmax_lambda_base"] = 1000
params.dict["amsoftmax_lambda_gamma"] = 1
params.dict["amsoftmax_lambda_power"] = 4
params.dict["arcsoftmax_lambda_min"] = 10
params.dict["arcsoftmax_lambda_base"] = 1000
params.dict["arcsoftmax_lambda_gamma"] = 1
params.dict["arcsoftmax_lambda_power"] = 4
params.dict["feature_norm"] = True
params.dict["feature_scaling_factor"] = 20
from model.common import l2_scaling
outputs, endpoints = tdnn(features, params, is_training=True, reuse_variables=False)
outputs = l2_scaling(outputs, params.feature_scaling_factor)
outputs_norm = tf.norm(outputs, axis=1)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
[outputs_val, outputs_norm_val] = sess.run([outputs, outputs_norm], feed_dict={features: features_val})
assert np.allclose(np.sqrt(np.sum(outputs_val ** 2, axis=1)), params.feature_scaling_factor)
assert np.allclose(outputs_norm_val, params.feature_scaling_factor)
# Test loss functions
# It only works on debug mode, since the loss is asked to output weights for our numpy computation.
from model.loss import asoftmax, additive_margin_softmax, additive_angular_margin_softmax
from model.test_utils import compute_asoftmax, compute_amsoftmax, compute_arcsoftmax
params.dict["global_step"] = 1
print("Asoftmax")
for scaling in [True, False]:
