#dataset, std, mean = load_locomotion(rng) dataset, std, mean = load_terrain(rng) #train_motion_input_dataset = dataset[0][0][300:320] train_motion_input_dataset = dataset[0][0][1500:1520] train_motion_input_dataset = train_motion_input_dataset.swapaxes(0, 1)[:-3] train_motion_input_dataset = train_motion_input_dataset.swapaxes(0, 1) print "motion input dataset shape = ", train_motion_input_dataset.shape M_I = shared(train_motion_input_dataset) BATCH_SIZE = 20 H_SIZE = 128 encoder = HiddenLayer(rng, (128, H_SIZE)) encode_igate = HiddenLayer(rng, (128, H_SIZE)) encode_fgate = HiddenLayer(rng, (128, H_SIZE)) recoder = HiddenLayer(rng, (H_SIZE, H_SIZE)) recode_igate = HiddenLayer(rng, (H_SIZE, H_SIZE)) recode_fgate = HiddenLayer(rng, (H_SIZE, H_SIZE)) activation = ActivationLayer(rng, f='elu') dropout = DropoutLayer(rng, 0.2) encoder_network = Network( Conv1DLayer(rng, (64, 63, 25), (BATCH_SIZE, 63, 240)), Pool1DLayer(rng, (2, ), (BATCH_SIZE, 64, 240)), ActivationLayer(rng, f='elu'), Conv1DLayer(rng, (256, 64, 25), (BATCH_SIZE, 64, 120)),
train_set_x, train_set_y = map(shared, datasets[0]) valid_set_x, valid_set_y = map(shared, datasets[1]) test_set_x, test_set_y = map(shared, datasets[2]) batchsize = 100 train_set_x = train_set_x.reshape((50000, 1, 28, 28)) valid_set_x = valid_set_x.reshape((10000, 1, 28, 28)) test_set_x = test_set_x.reshape((10000, 1, 28, 28)) network = Network(Conv2DLayer(rng, (4, 1, 5, 5), (batchsize, 1, 28, 28)), BatchNormLayer(rng, (batchsize, 4, 28, 28), axes=(0, 2, 3)), ActivationLayer(rng, f='ReLU'), Pool2DLayer(rng, (batchsize, 4, 28, 28)), ReshapeLayer(rng, (batchsize, 4 * 14 * 14)), HiddenLayer(rng, (4 * 14 * 14, 10)), ActivationLayer(rng, f='softmax')) trainer = AdamTrainer(rng=rng, batchsize=batchsize, epochs=15, alpha=0.0001, cost='cross_entropy') trainer.train(network=network, train_input=train_set_x, train_output=train_set_y, valid_input=valid_set_x, valid_output=valid_set_y, test_input=test_set_x, test_output=test_set_y, filename=None)
from nn.AdamTrainer import AdamTrainer from utils import load_data rng = np.random.RandomState(23455) dataset = '../data/mnist/mnist.pkl.gz' datasets = load_data(dataset) shared = lambda d: theano.shared(d) train_set_x, train_set_y = map(shared, datasets[0]) valid_set_x, valid_set_y = map(shared, datasets[1]) test_set_x, test_set_y = map(shared, datasets[2]) network = Network(HiddenLayer(rng, (784, 500)), BatchNormLayer((784, 500)), ActivationLayer(rng, f='ReLU'), HiddenLayer(rng, (500, 10)), BatchNormLayer((500, 10)), ActivationLayer(rng, f='softmax')) trainer = AdamTrainer(rng=rng, batchsize=2, epochs=1, alpha=0.00001, cost='cross_entropy') trainer.train(network=network, train_input=train_set_x, train_output=train_set_y, valid_input=valid_set_x, valid_output=valid_set_y, test_input=test_set_x, test_output=test_set_y,
from nn.BatchNormLayer import BatchNormLayer from nn.HiddenLayer import HiddenLayer from nn.Network import Network from nn.AdamTrainer import AdamTrainer from utils import load_data rng = np.random.RandomState(23455) dataset = '../data/mnist/mnist.pkl.gz' datasets = load_data(dataset) train_set_x, train_set_y = datasets[0] valid_set_x, valid_set_y = datasets[1] test_set_x, test_set_y = datasets[2] network = Network( HiddenLayer(rng, (784, 500)), BatchNormLayer((784, 500)), ActivationLayer(rng, f='ReLU'), HiddenLayer(rng, (500, 10)), BatchNormLayer((500, 10)), ActivationLayer(rng, f='softmax') ) trainer = AdamTrainer(rng=rng, batchsize=100, epochs=1, alpha=0.00001, cost='cross_entropy') trainer.train(network=network, train_input=train_set_x, train_output=train_set_y, valid_input=valid_set_x, valid_output=valid_set_y, test_input=test_set_x, test_output=test_set_y, filename=None)
encoderNetwork = Network( Conv1DLayer(rng, (64, 66, 25), (BATCH_SIZE, 66, 240)), BatchNormLayer(rng, (BATCH_SIZE, 64, 240)), ActivationLayer(rng, f='elu'), Pool1DLayer(rng, (2,), (BATCH_SIZE, 64, 240)), DropoutLayer(rng, 0.25), Conv1DLayer(rng, (128, 64, 25), (BATCH_SIZE, 64, 120)), BatchNormLayer(rng, (BATCH_SIZE, 128, 120)), ActivationLayer(rng, f='elu'), Pool1DLayer(rng, (2,), (BATCH_SIZE, 128, 120)), ReshapeLayer(rng, (BATCH_SIZE, 128*60)), DropoutLayer(rng, 0.25), HiddenLayer(rng, (128*60, FC_SIZE)), BatchNormLayer(rng, (128*60, FC_SIZE)), ActivationLayer(rng, f='elu'), ) variationalNetwork = Network( VariationalLayer(rng), ) decoderNetwork = Network( HiddenLayer(rng, (FC_SIZE/2, 64*30)), BatchNormLayer(rng, (FC_SIZE/2, 64*30)), ActivationLayer(rng, f='elu'), ReshapeLayer(rng, (BATCH_SIZE, 64, 30)), InverseNetwork(Pool1DLayer(rng, (2,), (BATCH_SIZE, 64, 60))),
0, 2, )), ActivationLayer(rng, f='ReLU'), Pool1DLayer(rng, (2, ), (batchsize, 128, 120)), Conv1DLayer(rng, (256, 128, 25), (batchsize, 128, 60)), BatchNormLayer(rng, (batchsize, 256, 60), axes=( 0, 2, )), ActivationLayer(rng, f='ReLU'), Pool1DLayer(rng, (2, ), (batchsize, 256, 60)), # 256*60 = 7680 ReshapeLayer(rng, (batchsize, 7680)), HiddenLayer(rng, (np.prod([256, 30]), 8)), ActivationLayer(rng, f='softmax'), ) # Load the pre-trained conv-layers network.load([ '../models/conv_ae/layer_0.npz', None, None, '../models/conv_ae/layer_1.npz', None, None, '../models/conv_ae/layer_2.npz', None, None, None, None, None ]) trainer = AdamTrainer(rng=rng, batchsize=batchsize, epochs=10, alpha=0.00001, cost='cross_entropy')
from nn.AnimationPlotLines import animation_plot from tools.utils import load_locomotion rng = np.random.RandomState(23455) shared = lambda d: theano.shared(d, borrow=True) dataset, std, mean = load_locomotion(rng) E = shared(dataset[0][0]) BATCH_SIZE = 40 generatorNetwork = Network( DropoutLayer(rng, 0.15), HiddenLayer(rng, (800, 64 * 30)), BatchNormLayer(rng, (800, 64 * 30)), ActivationLayer(rng, f='elu'), ReshapeLayer(rng, (BATCH_SIZE, 64, 30)), InverseNetwork(Pool1DLayer(rng, (2, ), (BATCH_SIZE, 64, 60))), DropoutLayer(rng, 0.15), Conv1DLayer(rng, (64, 64, 25), (BATCH_SIZE, 64, 60)), ActivationLayer(rng, f='elu'), InverseNetwork(Pool1DLayer(rng, (2, ), (BATCH_SIZE, 64, 120))), DropoutLayer(rng, 0.25), Conv1DLayer(rng, (64, 64, 25), (BATCH_SIZE, 64, 120)), ActivationLayer(rng, f='elu'), InverseNetwork(Pool1DLayer(rng, (2, ), (BATCH_SIZE, 64, 240))), DropoutLayer(rng, 0.25), Conv1DLayer(rng, (66, 64, 25), (BATCH_SIZE, 64, 240)), ActivationLayer(rng, f='elu'),