tree.add_layer(FullyConnected((512, 1, 1, 1))) tree.add_layer(PReLULayer()) tree.add_layer(FullyConnected((512, 1, 1, 1))) tree.add_layer(PReLULayer()) tree.add_layer(FullyConnected()) tree.add_input(X_reward[0]) tree.add_input(X_reward[1]) model = Model() model.set_tree(tree) model.set_loss(Losses.MeanSquared()) model.add_output(y_reward) learning_rate = 0.00002 model.set_optimizer(Optimizers.RMSprop(learning_rate=learning_rate)) model.compile(X_reward, y_reward, initialize_params=True) file = open('reward_model_params_vec.pk', 'rb') params = pickle.load(file) file.close() model.set_params_as_vec(params) model.compile(X_reward, y_reward, initialize_params=False) '''ok.save_model(model, 'okapi_reward_model.pk') model = ok.load_model('okapi_reward_model.pk')''' # model.set_dream_optimizer(Optimizers.RMSprop(learning_rate=0.0001, momentum=0.99)) preds = model.predict_dream([None, np.array([0, 0, 0, 1, 0, 0, 0, 0, 0, 0])], [(28, 28)], max_dream_length=100) # print(model.predict([preds[0], np.asarray([1, 0, 0, 0, 0, 0, 0, 0, 0, 0])])) pyplot.imshow(np.reshape(preds[0], (28, 28)), interpolation='nearest', cmap='Greys') # pyplot.imshow(np.reshape(X_train[[0]], (28, 28)), interpolation='nearest', cmap='Greys') pyplot.show()
num_generations = 100 rm_p = 0.8 init_mut_p = 1e-5 init_mut_std = 1e-3 init_cross_p = 0.7 model = Model() model.add(Convolutional(num_filters, filter_size, filter_size, pad=pad)) model.add(ActivationLayer(Activations.tanh)) model.add(MaxPooling(pool_size, pool_size)) model.add(PReLULayer()) model.add(Dropout(dropout_p)) model.add(FullyConnected(bias_initializer=Initializers.glorot_uniform)) model.add(ActivationLayer(Activations.alt_softmax)) model.compile(X_train, y_train) X_batches, y_batches, num_batches = \ ok.make_batches(X_train, y_train, batch_size) def initialize(population_size): population = [] for i in range(population_size): model.randomize_params(X_train, y_train) individual = {'genome': model.get_params_as_vec()} individual['genome'] = np.append(individual['genome'], [init_mut_std, init_mut_p, init_cross_p]) population.append(individual) return population
tree.add_layer(PReLULayer()) tree.add_layer(FullyConnected((512, 1, 1, 1))) tree.add_layer(PReLULayer()) tree.add_layer(FullyConnected((512, 1, 1, 1))) tree.add_layer(PReLULayer()) tree.add_layer(FullyConnected()) tree.add_layer(ActivationLayer(Activations.tanh)) tree.add_input(X_reward[0]) tree.add_input(X_reward[1]) model = Model() model.set_tree(tree) model.set_loss(Losses.MeanSquared()) model.set_optimizer(Optimizers.RMSprop(learning_rate=0.00005)) model.add_output(y_reward) model.compile(X_reward, y_reward) model.train(X_reward, y_reward, 24) # print(model.predict([X_train[[0]], y_train[0]])) # print(model.predict([X_train[[0]], np.asarray([1, 0, 0, 0, 0, 0, 0, 0, 0, 0])])) # prediction = model.predict_dream([X_test, None], X_reward[1][[0]].shape) X_batches, y_batches, num_batches = ok.make_batches([X_train], y_train, batch_size=10000) for i in range(12): for i, [X_batch, y_batch] in enumerate(zip(X_batches, y_batches)): accuracy, preds = model.get_dream_accuracy([X_batch[0], None], y_batch) preds = preds[0] print('Accuracy: {}%'.format(accuracy)) preds_reward = reward(preds.reshape(preds.shape[0], preds.shape[1]).astype('float32'), y_batch.reshape(preds.shape[0], preds.shape[1]).astype('float32')) X_reward[0] = np.append(X_reward[0], X_batch[0], axis=0) X_reward[1] = np.append(X_reward[1], preds, axis=0)