def main(): # -- top-level parameters of this script dtype = 'float32' # XXX n_examples = 50000 online_batch_size = 1 online_epochs = 2 batch_epochs = 30 lbfgs_m = 20 # -- load and prepare the data set data_view = mnist.views.OfficialVectorClassification(x_dtype=dtype) n_classes = 10 x = data_view.train.x[:n_examples] y = data_view.train.y[:n_examples] y1 = -1 * ones((len(y), n_classes)).astype(dtype) y1[arange(len(y)), y] = 1 # --initialize the SVM model w = zeros((x.shape[1], n_classes), dtype=dtype) b = zeros(n_classes, dtype=dtype) def svm(ww, bb, xx=x, yy=y1): # -- one vs. all linear SVM loss margin = yy * (dot(xx, ww) + bb) hinge = maximum(0, 1 - margin) cost = hinge.mean(axis=0).sum() return cost # -- stage-1 optimization by stochastic gradient descent print 'Starting SGD' n_batches = n_examples / online_batch_size w, b = fmin_sgd( svm, (w, b), streams={ 'xx': x.reshape((n_batches, online_batch_size, x.shape[1])), 'yy': y1.reshape((n_batches, online_batch_size, y1.shape[1])) }, loops=online_epochs, stepsize=0.001, print_interval=10000, ) print 'SGD complete, about to start L-BFGS' show_filters(w.T, (28, 28), ( 2, 5, )) # -- stage-2 optimization by L-BFGS print 'Starting L-BFGS' w, b = fmin_l_bfgs_b(svm, (w, b), maxfun=batch_epochs, iprint=1, m=lbfgs_m) print 'L-BFGS complete' show_filters(w.T, (28, 28), ( 2, 5, ))
def main(): # -- top-level parameters of this script dtype = 'float32' # XXX n_examples = 50000 online_batch_size = 1 online_epochs = 2 batch_epochs = 30 lbfgs_m = 20 n_mlp_hiddens = [200] # -- one entry per hidden layer # -- load and prepare the data set data_view = mnist.views.OfficialVectorClassification(x_dtype=dtype) n_classes = 10 x = data_view.train.x[:n_examples] y = data_view.train.y[:n_examples] y1 = -1 * ones((len(y), n_classes)).astype(dtype) y1[arange(len(y)), y] = 1 # -- allocate the model by running one example through it init_params = {} mlp_svm(init_params, x[:1], y[:1], n_mlp_hiddens, n_classes) if online_epochs: # -- stage-1 optimization by stochastic gradient descent print 'Starting SGD' n_batches = n_examples / online_batch_size stage1_params, = fmin_sgd(mlp_svm, (init_params,), streams={ 'x': x.reshape((n_batches, online_batch_size, x.shape[1])), 'y1': y1.reshape((n_batches, online_batch_size, y1.shape[1]))}, loops=online_epochs, stepsize=0.001, print_interval=10000, ) print 'SGD complete, about to start L-BFGS' show_filters(stage1_params['mlp']['weights'][0].T, (28, 28), (8, 25,)) else: print 'Skipping stage-1 SGD' stage1_params = init_params # -- stage-2 optimization by L-BFGS if batch_epochs: def batch_mlp_svm(p): return mlp_svm(p, x, y1) print 'Starting L-BFGS' stage2_params, = fmin_l_bfgs_b(lambda p: mlp_svm(p, x, y1), args=(stage1_params,), maxfun=batch_epochs, iprint=1, m=lbfgs_m) print 'L-BFGS complete' show_filters(stage2_params['mlp']['weights'][0].T, (28, 28), (8, 25,))
def main(): # -- top-level parameters of this script dtype = "float32" # XXX n_examples = 50000 online_batch_size = 1 online_epochs = 2 batch_epochs = 30 lbfgs_m = 20 # -- load and prepare the data set data_view = mnist.views.OfficialVectorClassification(x_dtype=dtype) n_classes = 10 x = data_view.train.x[:n_examples] y = data_view.train.y[:n_examples] y1 = -1 * ones((len(y), n_classes)).astype(dtype) y1[arange(len(y)), y] = 1 # --initialize the SVM model w = zeros((x.shape[1], n_classes), dtype=dtype) b = zeros(n_classes, dtype=dtype) def svm(ww, bb, xx=x, yy=y1): # -- one vs. all linear SVM loss margin = yy * (dot(xx, ww) + bb) hinge = maximum(0, 1 - margin) cost = hinge.mean(axis=0).sum() return cost # -- stage-1 optimization by stochastic gradient descent print "Starting SGD" n_batches = n_examples / online_batch_size w, b = fmin_sgd( svm, (w, b), streams={ "xx": x.reshape((n_batches, online_batch_size, x.shape[1])), "yy": y1.reshape((n_batches, online_batch_size, y1.shape[1])), }, loops=online_epochs, stepsize=0.001, print_interval=10000, ) print "SGD complete, about to start L-BFGS" show_filters(w.T, (28, 28), (2, 5)) # -- stage-2 optimization by L-BFGS print "Starting L-BFGS" w, b = fmin_l_bfgs_b(svm, (w, b), maxfun=batch_epochs, iprint=1, m=lbfgs_m) print "L-BFGS complete" show_filters(w.T, (28, 28), (2, 5))
def main(): # -- top-level parameters of this script n_hidden1 = n_hidden2 = 25 dtype = "float32" n_examples = 10000 online_batch_size = 1 online_epochs = 3 # -- TIP: partial creates a new function with some parameters filled in # algo = partial(denoising_autoencoder_binary_x, noise_level=0.3) algo = logistic_autoencoder_binary_x batch_epochs = 10 lbfgs_m = 20 n_hidden = n_hidden1 * n_hidden2 rng = np.random.RandomState(123) data_view = mnist.views.OfficialVectorClassification(x_dtype=dtype) x = data_view.train.x[:n_examples] n_examples, n_visible = x.shape x_img_res = 28, 28 # -- uncomment this line to see sample images from the data set # show_filters(x[:100], x_img_res, (10, 10)) # -- create a new model (w, visbias, hidbias) w = rng.uniform( low=-4 * np.sqrt(6.0 / (n_hidden + n_visible)), high=4 * np.sqrt(6.0 / (n_hidden + n_visible)), size=(n_visible, n_hidden), ).astype(dtype) visbias = np.zeros(n_visible).astype(dtype) hidbias = np.zeros(n_hidden).astype(dtype) # show_filters(w.T, x_img_res, (n_hidden1, n_hidden2)) x_stream = x.reshape((n_examples / online_batch_size, online_batch_size, x.shape[1])) def train_criterion(ww, hbias, vbias, x_i=x): cost, hid = algo(x_i, ww, hbias, vbias) l1_cost = abs(ww).sum() * 0.0 # -- raise 0.0 to enforce l1 penalty l2_cost = (ww ** 2).sum() * 0.0 # -- raise 0.0 to enforce l2 penalty return cost.mean() + l1_cost + l2_cost # -- ONLINE TRAINING for epoch in range(online_epochs): t0 = time.time() w, hidbias, visbias = autodiff.fmin_sgd( train_criterion, args=(w, hidbias, visbias), stream=x_stream, # -- fmin_sgd will loop through this once stepsize=0.005, # -- QQ: you should always tune this print_interval=1000, ) print "Online training epoch %i took %f seconds" % (epoch, time.time() - t0) show_filters(w.T, x_img_res, (n_hidden1, n_hidden2)) # -- BATCH TRAINING w, hidbias, visbias = autodiff.fmin_l_bfgs_b( train_criterion, args=(w, hidbias, visbias), # -- scipy.fmin_l_bfgs_b kwargs follow maxfun=batch_epochs, iprint=1, # -- 1 for verbose, 0 for normal, -1 for quiet m=lbfgs_m, # -- how well to approximate the Hessian ) show_filters(w.T, x_img_res, (n_hidden1, n_hidden2))
def main(): # -- top-level parameters of this script n_hidden1 = n_hidden2 = 25 dtype = 'float32' n_examples = 10000 online_batch_size = 1 online_epochs = 3 # -- TIP: partial creates a new function with some parameters filled in # algo = partial(denoising_autoencoder_binary_x, noise_level=0.3) algo = logistic_autoencoder_binary_x batch_epochs = 10 lbfgs_m = 20 n_hidden = n_hidden1 * n_hidden2 rng = np.random.RandomState(123) data_view = mnist.views.OfficialVectorClassification(x_dtype=dtype) x = data_view.train.x[:n_examples] n_examples, n_visible = x.shape x_img_res = 28, 28 # -- uncomment this line to see sample images from the data set # show_filters(x[:100], x_img_res, (10, 10)) # -- create a new model (w, visbias, hidbias) w = rng.uniform(low=-4 * np.sqrt(6. / (n_hidden + n_visible)), high=4 * np.sqrt(6. / (n_hidden + n_visible)), size=(n_visible, n_hidden)).astype(dtype) visbias = np.zeros(n_visible).astype(dtype) hidbias = np.zeros(n_hidden).astype(dtype) # show_filters(w.T, x_img_res, (n_hidden1, n_hidden2)) x_stream = x.reshape( (n_examples / online_batch_size, online_batch_size, x.shape[1])) def train_criterion(ww, hbias, vbias, x_i=x): cost, hid = algo(x_i, ww, hbias, vbias) l1_cost = abs(ww).sum() * 0.0 # -- raise 0.0 to enforce l1 penalty l2_cost = (ww**2).sum() * 0.0 # -- raise 0.0 to enforce l2 penalty return cost.mean() + l1_cost + l2_cost # -- ONLINE TRAINING for epoch in range(online_epochs): t0 = time.time() w, hidbias, visbias = autodiff.fmin_sgd( train_criterion, args=(w, hidbias, visbias), stream=x_stream, # -- fmin_sgd will loop through this once stepsize=0.005, # -- QQ: you should always tune this print_interval=1000, ) print 'Online training epoch %i took %f seconds' % (epoch, time.time() - t0) show_filters(w.T, x_img_res, (n_hidden1, n_hidden2)) # -- BATCH TRAINING w, hidbias, visbias = autodiff.fmin_l_bfgs_b( train_criterion, args=(w, hidbias, visbias), # -- scipy.fmin_l_bfgs_b kwargs follow maxfun=batch_epochs, iprint=1, # -- 1 for verbose, 0 for normal, -1 for quiet m=lbfgs_m, # -- how well to approximate the Hessian ) show_filters(w.T, x_img_res, (n_hidden1, n_hidden2))