train_fn = mlp.train_mlp_probe
test_fn = mlp.test_mlp_class
hyperparams = {
'batch_size': [100],
'learning_rate': [0.2],
'lambda_reg': [0.0001],
'num_hidden': [100, 250, 400],
'n_epochs': 1000,
'error_func': 'cross_ent',
'final_layer': ['tanh', 'sigmoid'],
'validate_params': ["batch_size", "learning_rate", "lambda_reg", 'num_hidden']}
# Cross-validate here
best_params, all_params = validation_helpers.validate_grid_search_cheat(train_fn, test_fn,
False, train_samples, train_labels, valid_samples,
valid_labels, hyperparams, num_repeat=2)
# Average results due to non-deterministic nature of the model
f1s = numpy.zeros((1, train_labels.shape[1]))
precisions = numpy.zeros((1, train_labels.shape[1]))
recalls = numpy.zeros((1, train_labels.shape[1]))
num_repeat = 3
print 'All params', all_params
print 'Best params', best_params
for i in range(num_repeat):
model = train_fn(train_labels, train_samples, valid_labels, valid_samples, best_params)
_, _, _, _, f1_c, precision_c, recall_c = test_fn(valid_labels, valid_samples, model)
hyperparams = {
'batch_size': [100],
'learning_rate': [0.2, 0.4],
'lambda_reg': [0.000001, 0.00001],
'num_hidden': [300, 400],
'n_epochs': 50,
'validate_params':
["batch_size", "learning_rate", "lambda_reg", 'num_hidden']
}
# Cross-validate here
best_params, all_params = validation_helpers.validate_grid_search_cheat(
train_fn,
test_fn,
False,
train_samples,
train_labels,
valid_samples,
valid_labels,
hyperparams,
num_repeat=3)
# Average results due to non-deterministic nature of the model
f1s = numpy.zeros((1, train_labels.shape[1]))
precisions = numpy.zeros((1, train_labels.shape[1]))
recalls = numpy.zeros((1, train_labels.shape[1]))
num_repeat = 3
print 'All params', all_params
print 'Best params', best_params
