ml_x_t = train_x_t @ L
ml_val_x_t = val_x_t @ L
ml_test_x = test_x_t @ L
print('Done')
# =============================================================================
num_inputs = ml_x_s.shape[1] # input layer size
# =============================================================================
if 'ml_model' in locals():
del ml_model
if 'ml_model_obj' in locals():
del ml_model_obj
ml_model_obj = my_models(num_inputs, dropout=dropout_pr)
ml_model = ml_model_obj.build_model()
ml_model = ml_model_obj.fit(ml_x_s,
train_y_s,
ml_val_x_s,
val_y_s,
scale=NN_scaling)
error_metric, ml_train_loss, ml_val_loss, ml_test_loss = \
ml_model_obj.evaluate(ml_x_t,train_y_t, ml_val_x_t, val_y_t, ml_test_x,
test_y_t, scale = NN_scaling)
if fine_tuning:
ml_model = ml_model_obj.fit(ml_x_t,
train_y_t,
#coef_s, coef_t = ekmm_emb(emb_y_t, emb_y_s, kern = kmm_kernel, B = B,
# embedder_type = 'no_embedding', n_components = 100)
# =============================================================================
coef_val_s = coef_s[train_y_s.shape[0]:]
coef_s = coef_s[:train_y_s.shape[0]]
# =============================================================================
import numpy.matlib as npm
training_weights = npm.repmat(coef_s, 1, train_y_s.shape[1])
training_val_weights = npm.repmat(coef_val_s, 1, val_y_s.shape[1])
num_weights = training_weights.shape[1]
num_inputs = emb_x_s.shape[1] # input layer size
# =============================================================================
w_model_obj = my_models(num_inputs,
num_weights=num_weights,
model_type='weighted',
dropout=dropout_pr)
model = w_model_obj.build_model()
model = w_model_obj.fit(emb_x_s,
train_y_s,
emb_val_x_s,
val_y_s,
val_w=training_val_weights,
scale=NN_scaling,
training_w=training_weights)
error_sample_bias, train_loss, val_loss, test_loss =\
w_model_obj.evaluate(emb_x_t, train_y_t, emb_val_x_t, val_y_t,
emb_test_x, test_y_t, scale = NN_scaling)
title = 'metric learning plus sample selection bias'
print_stat(title, error_sample_bias, train_loss, val_loss, test_loss)