def test_many_to_many_nested(self): hc.reset() hc.set('x', [1, 2, 3]) hc.set('y', [1, 2, 3]) hc.set('z', [1, 2, 3]) self.assertEqual(hc.count_configs(), 27) self.assertEqual(hc.configs(1, serial=True,offset=0,create_uuid=False), [{'x':1, 'y': 1, 'z': 1}])
def test_only_permutation(self): hc.reset() for i in range(20): hc.set('a'+str(i), [1, 2, 3]) self.assertEqual(hc.count_configs(), 3**20) self.assertEqual(hc.configs(1, serial=True, offset=0)[0]['a1'], 1) self.assertEqual(hc.configs(1, serial=True, offset=0)[0]['a2'], 1)
def test_many_to_many_nested(self): hc.reset() hc.set('x', [1, 2, 3]) hc.set('y', [1, 2, 3]) hc.set('z', [1, 2, 3]) self.assertEqual(hc.count_configs(), 27) self.assertEqual(hc.configs(1, serial=True,offset=0,create_uuid=False), [{'x':1, 'y': 1, 'z': 1}])
def test_only_permutation(self): hc.reset() for i in range(20): hc.set('a' + str(i), [1, 2, 3]) self.assertEqual(hc.count_configs(), 3**20) self.assertEqual(hc.configs(1, serial=True, offset=0)[0]['a1'], 1) self.assertEqual(hc.configs(1, serial=True, offset=0)[0]['a2'], 1)
def test_many_to_many(self): hc.reset() hc.set('x', [1,2,3]) hc.set('y', [1, 2, 3]) self.assertEqual(hc.count_configs(), 9) configs = hc.configs(3, serial=True,create_uuid=False,offset=0) self.assertNotEqual(configs[0], configs[1]) self.assertNotEqual(configs[1], configs[2])
def test_many_to_many(self): hc.reset() hc.set('x', [1, 2, 3]) hc.set('y', [1, 2, 3]) self.assertEqual(hc.count_configs(), 9) configs = hc.configs(3, serial=True, create_uuid=False, offset=0) self.assertNotEqual(configs[0], configs[1]) self.assertNotEqual(configs[1], configs[2])
def test_one_permute(self): hc.reset() hc.set('x', 1) hc.set('y', [1]) self.assertEqual(hc.count_configs(), 1) self.assertEqual( hc.configs(1, serial=True, offset=0, create_uuid=False), [{ 'x': 1, 'y': 1 }])
results = { 'difficulty':float(difficulty), 'ranking':float(ranking), 'g_loss':float(g_loss), 'd_fake':float(d_fake), 'd_real':float(d_real), 'e_loss':float(e_loss) } print("Recording ", results) #hc.io.record(config, results) print("Generating configs with hyper search space of ", hc.count_configs()) j=0 k=0 def get_function(name): if not isinstance(name, str): return name print('name', name); if(name == "function:tensorflow.python.ops.gen_nn_ops.relu"): return tf.nn.relu if(name == "function:tensorflow.python.ops.nn_ops.relu"): return tf.nn.relu if(name == "function:tensorflow.python.ops.gen_nn_ops.relu6"): return tf.nn.relu6 if(name == "function:tensorflow.python.ops.nn_ops.relu6"):
plt.tight_layout() img = "samples/reconstruction.png" plt.savefig(img) hc.io.sample(config, [{"label": "Reconstruction", "image": img}]) hc.permute.set("learning_rate", list(np.linspace(0.0001, 0.003, num=10, dtype=np.float32))) hc.permute.set("n_hidden_recog_1", list(np.linspace(100, 1000, num=10, dtype=np.int32))) hc.permute.set("n_hidden_recog_2", list(np.linspace(100, 1000, num=10, dtype=np.int32))) hc.permute.set("n_hidden_gener_1", list(np.linspace(100, 1000, num=10, dtype=np.int32))) hc.permute.set("n_hidden_gener_2", list(np.linspace(100, 1000, num=10, dtype=np.int32))) hc.set("n_input", 784) # MNIST data input (img shape: 28*28) hc.permute.set("n_z", [1,2,4,8,16,32,64,128]) # dimensionality of latent space hc.set('batch_size', 100) hc.permute.set("transfer_fct", [tf.tanh, tf.nn.elu, tf.nn.relu, tf.nn.relu6, tf.nn.softplus, tf.nn.softsign]); hc.set("epochs", 10) print("hypersearch space", hc.count_configs()) for config in hc.configs(10000): print("Config", config) vae = VariationalAutoencoder(config, learning_rate=config['learning_rate'], transfer_fct=config['transfer_fct']) cost = 1000 for i in range(config['epochs']): cost = train(config, vae, training_epochs=1) sample(config, vae) hc.io.measure(config, {"rank": cost}) visualize(config, vae)
def test_one_permute(self): hc.reset() hc.set('x', 1) hc.set('y', [1]) self.assertEqual(hc.count_configs(), 1) self.assertEqual(hc.configs(1, serial=True,offset=0,create_uuid=False), [{'x':1, 'y': 1}])
def test_store_size(self): hc.reset() hc.set('x', [1,2]) self.assertEqual(hc.count_configs(), 2)
def test_reset(self): hc.reset() self.assertEqual(hc.configs(), []) self.assertEqual(hc.count_configs(), 1)
def test_store_size(self): hc.reset() hc.set('x', [1,2]) self.assertEqual(hc.count_configs(), 2)
def test_reset(self): hc.reset() self.assertEqual(hc.configs(), []) self.assertEqual(hc.count_configs(), 1)
def test_config(sess, config): batch_size = config["batch_size"] n_samples = mnist.test.num_examples total_batch = int(n_samples / batch_size) accuracies = [] costs = [] for i in range(total_batch): x, y = mnist.test.next_batch(batch_size) x=np.reshape(x, [batch_size, X_DIMS[0], X_DIMS[1], 1]) accuracy, cost = test(sess, config, x, y) accuracies.append(accuracy) costs.append(cost) return accuracies, costs print("Searching randomly with %d possible configurations." % hc.count_configs()) for i in range(100): config = hc.random_config() print("Testing configuration", config) sess = tf.Session() graph = create(config) init = tf.initialize_all_variables() sess.run(init) for i in range(10): epoch(sess, config) accuracies, costs = test_config(sess, config) accuracy, cost = np.mean(accuracies), np.mean(costs) results = { 'accuracy':accuracy, 'cost':cost }
ranking = g_loss * (1.0 - difficulty) ranking = e_loss results = { 'difficulty': float(difficulty), 'ranking': float(ranking), 'g_loss': float(g_loss), 'd_fake': float(d_fake), 'd_real': float(d_real), 'e_loss': float(e_loss) } print("Recording ", results) #hc.io.record(config, results) print("Generating configs with hyper search space of ", hc.count_configs()) j = 0 k = 0 def get_function(name): if not isinstance(name, str): return name print('name', name) if (name == "function:tensorflow.python.ops.gen_nn_ops.relu"): return tf.nn.relu if (name == "function:tensorflow.python.ops.nn_ops.relu"): return tf.nn.relu if (name == "function:tensorflow.python.ops.gen_nn_ops.relu6"): return tf.nn.relu6