train_labels=strain.y,valid_labels=valid.y, learning_rate=0.001, batch_size=configs['batch_size'], epochs=epochs,wait_for=20, epsylon=0.0001, aug=1.01) X = T.matrix() f = theano.function([X],model.fprop(X)) Y_hat = f(strain.X[:10]) Y = strain.y[:10] for y_hat,y in zip(Y_hat,Y): print y_hat[:2], y[:2] y = np.vstack(sgd.iterate([testset.X],[f],configs['batch_size'])[0]) print "results shape" print y.shape out_path = save_path+"/test.csv" submission = [] with open('submissionFileFormat.csv', 'rb') as cvsTemplate: reader = csv.reader(cvsTemplate) for row in reader: submission.append(row) mapping = dict(zip(['left_eye_center_x', 'left_eye_center_y', 'right_eye_center_x',
models = [] h_in, h_out = zip([trtrainset.shape[1]]+configs['hid'],configs['hid'])[i] print h_in,h_out model = cA.cA(numpy_rng=numpy_rng, theano_rng=theano_rng, numvis=h_in, numhid=h_out, activation=T.tanh, vistype="real", contraction=configs['contract'][i]) sgd.train(trtrainset,trvalidset,model, batch_size=configs['batch_size'], wait_for=20, learning_rate=configs['lr'][i], epochs=training_epochs, epsylon=configs['epsylon'][i], aug=1.01) X = T.matrix() encoding = model.hiddens(X) f = theano.function([X],encoding) trtrainset = np.vstack(sgd.iterate([trtrainset],[f],configs['batch_size'])[0]) trvalidset = np.vstack(sgd.iterate([trvalidset],[f],configs['batch_size'])[0]) f = None print "save params" for param in model.params[:2]: print param.name np.save(save_path+"/layer%d%s.npy" % (i,param.name),param.get_value()) np.save(save_path+"/trainset%d.npy" % i,trtrainset) np.save(save_path+"/validset%d.npy" % i,trvalidset)