plot_sample_grid(5, decoder_params, (21, 21), gaussian_decoder) plt.savefig('pendulum.png') if __name__ == "__main__": npr.seed(0) trX = load_pendulum(100) encoder_params, decoder_params, fit = \ make_gaussian_fitter(trX, 3, [200], [200]) # 2 also works well, 1 not quite as well fit(1 * 500, 50, 1, adadelta()) plot() fit(25 * 500, 50, 1, adadelta()) plot() fit(50 * 500, 100, 1, rmsprop(1e-4)) plot() fit(50 * 500, 100, 1, rmsprop(1e-5)) plot() fit(50 * 500, 100, 1, rmsprop(1e-6)) plot() params = get_ndarrays(encoder_params), get_ndarrays(decoder_params) with gzip.open('pendulum_params.pkl.gz', 'w') as f: pickle.dump(params, f, protocol=-1) plt.show() # TODO try adam
import logging.config logging.config.fileConfig('logging.conf') from vae.vae import make_gaussian_fitter from vae.optimization import sgd, adagrad, rmsprop, adadelta, adam, \ momentum_sgd, nesterov from vae.util import get_ndarrays from load import load_mice if __name__ == '__main__': logging.info('\n\nStarting experiment!') np.random.seed(0) N = 750000 # 750k is about the memory limit on 3GB GPU trX = load_mice(N) encoder_params, decoder_params, fit = \ make_gaussian_fitter(trX, 20, [200, 200], [200, 200]) fit(1, 50, 1, adadelta()) fit(1, 250, 1, adadelta()) fit(10, 500, 1, rmsprop(1e-4)) fit(25, 500, 1, rmsprop(1e-5)) fit(25, 1000, 1, rmsprop(1e-5)) params = get_ndarrays(encoder_params), get_ndarrays(decoder_params) with gzip.open('params.pkl.gz', 'w') as f: pickle.dump(params, f, protocol=-1)
def plot(): plot_sample_grid(5, decoder_params, (21, 21), gaussian_decoder) plt.savefig('pendulum.png') if __name__ == "__main__": npr.seed(0) trX = load_pendulum(100) encoder_params, decoder_params, fit = \ make_gaussian_fitter(trX, 3, [200], [200]) # 2 also works well, 1 not quite as well fit(1*500, 50, 1, adadelta()) plot() fit(25*500, 50, 1, adadelta()) plot() fit(50*500, 100, 1, rmsprop(1e-4)) plot() fit(50*500, 100, 1, rmsprop(1e-5)) plot() fit(50*500, 100, 1, rmsprop(1e-6)) plot() params = get_ndarrays(encoder_params), get_ndarrays(decoder_params) with gzip.open('pendulum_params.pkl.gz', 'w') as f: pickle.dump(params, f, protocol=-1) plt.show() # TODO try adam
def plot(): plot_sample_grid(10, decoder_params, (30, 30), gaussian_decoder) plt.savefig('mice.png') if __name__ == '__main__': logging.info('\n\nStarting experiment!') np.random.seed(0) N = 750000 # 750k is about the memory limit on 3GB GPU trX = load_mice(N) encoder_params, decoder_params, fit = \ make_gaussian_fitter(trX, 20, [200, 200], [200, 200]) fit(1, 50, 1, adadelta()) plot() fit(1, 250, 1, adadelta()) plot() fit(10, 500, 1, rmsprop(1e-4)) plot() fit(25, 500, 1, rmsprop(1e-5)) plot() fit(25, 1000, 1, rmsprop(1e-5)) params = get_ndarrays(encoder_params), get_ndarrays(decoder_params) with gzip.open('mice_params.pkl.gz', 'w') as f: pickle.dump(params, f, protocol=-1)
import theano from vae.vae import make_binary_fitter, binary_decoder from vae.optimization import adadelta, rmsprop from vae.util import get_ndarrays, floatX from vae.viz import plot_sample_grid from load import load_letters if __name__ == "__main__": npr.seed(0) trX, labels = load_letters('f') encoder_params, decoder_params, fit = make_binary_fitter(trX, 5, [200], [200]) fit(1, 50, 1, adadelta()) fit(3, 250, 1, adadelta()) fit(2000, 50, 1, rmsprop(1e-3)) fit(2000, 250, 10, rmsprop(1e-4)) params = get_ndarrays(encoder_params), get_ndarrays(decoder_params) with gzip.open('letter_params.pkl.gz', 'w') as f: pickle.dump(params, f, protocol=-1) plot_sample_grid(5, decoder_params, (16, 8), binary_decoder) plt.savefig('letters.png') plt.show()
import gzip import theano from vae.vae import make_binary_fitter, binary_decoder from vae.optimization import adadelta, rmsprop from vae.util import get_ndarrays, floatX from vae.viz import plot_sample_grid from load import load_letters if __name__ == "__main__": npr.seed(0) trX, labels = load_letters('f') encoder_params, decoder_params, fit = make_binary_fitter( trX, 5, [200], [200]) fit(1, 50, 1, adadelta()) fit(3, 250, 1, adadelta()) fit(2000, 50, 1, rmsprop(1e-3)) fit(2000, 250, 10, rmsprop(1e-4)) params = get_ndarrays(encoder_params), get_ndarrays(decoder_params) with gzip.open('letter_params.pkl.gz', 'w') as f: pickle.dump(params, f, protocol=-1) plot_sample_grid(5, decoder_params, (16, 8), binary_decoder) plt.savefig('letters.png') plt.show()