def save(self):
#TODO-- save state of dataset and training algorithm so training can be resumed after a crash
if self.save_path is not None:
print 'saving to '+self.save_path
serial.save(self.save_path, self.model)
#replicate the preprocessing described in Kai Yu's paper Improving LCC with Local Tangents
from framework.utils import serial
from framework.datasets import cifar10
from framework.datasets import preprocessing
train = cifar10.CIFAR10(which_set="train",center=True)
pipeline = preprocessing.Pipeline()
pipeline.items.append(preprocessing.GlobalContrastNormalization(subtract_mean=False,std_bias=0.0))
pipeline.items.append(preprocessing.PCA(num_components=512))
test = cifar10.CIFAR10(which_set="test")
train.apply_preprocessor(preprocessor = pipeline, can_fit = True)
test.apply_preprocessor(preprocessor = pipeline, can_fit = False)
serial.save('cifar10_preprocessed_train.pkl',train)
serial.save('cifar10_preprocessed_test.pkl',test)
from framework.utils import serial
from framework.datasets import cifar10
from framework.datasets import preprocessing
train = cifar10.CIFAR10(which_set="train")
pipeline = preprocessing.Pipeline()
pipeline.items.append(preprocessing.ExtractPatches(patch_shape=(8,8),num_patches=200000))
pipeline.items.append(preprocessing.GlobalContrastNormalization())
pipeline.items.append(preprocessing.ZCA())
train.apply_preprocessor(preprocessor = pipeline, can_fit = True)
serial.save('preprocessor.pkl',pipeline)
W_norms = N.square(W).sum(axis=0)
b += beta * (- N.dot(model.vis_mean.get_value(), W) - 0.5 * W_norms )
#
#compensate for beta scaling the responses
W = (W.T * beta).T
print 'making dot products'
dots = N.cast['float32'](N.dot(X,W))
print 'done'
print 'making activations'
acts = N.cast['float32'](N.zeros((X.shape[0],nhid)))
for i in xrange(0,X.shape[0],batch_size):
if i % 1000 == 0:
print i
cur_batch_size = min(batch_size, X.shape[0]-batch_size+1)
acts[i:i+cur_batch_size,:] = model.hid_exp_func(X[i:i+cur_batch_size,:])
print 'saving data'
dots_path = output_path + '_dots.npy'
acts_path = output_path + '_acts.npy'
N.save(dots_path, dots)
N.save(acts_path, acts)
serial.save(output_path+'.pkl',{ 'dots' : dots_path, 'acts' : acts_path, 'b' : b , 'beta' : beta} )
from framework.utils import serial
from framework.datasets import cifar10
from framework.datasets import preprocessing
train = cifar10.CIFAR10(which_set="train")
pipeline = preprocessing.Pipeline()
pipeline.items.append(
preprocessing.ExtractPatches(patch_shape=(8, 8), num_patches=150000))
pipeline.items.append(preprocessing.GlobalContrastNormalization())
pipeline.items.append(preprocessing.ZCA())
test = cifar10.CIFAR10(which_set="test")
train.apply_preprocessor(preprocessor=pipeline, can_fit=True)
test.apply_preprocessor(preprocessor=pipeline, can_fit=False)
train.use_design_loc('cifar10_preprocessed_train_design.npy')
test.use_design_loc('cifar10_preprocessed_test_design.npy')
serial.save('cifar10_preprocessed_train.pkl', train)
serial.save('cifar10_preprocessed_test.pkl', test)
train.apply_preprocessor(
preprocessing.ExtractPatches(patch_shape=(8, 8), num_patches=150000))
orig_patches = train.get_topological_view().copy()
print(orig_patches.min(), orig_patches.max())
#orig_patches -= 0.5
orig_patches -= 127.5
orig_patches /= N.abs(orig_patches).max()
print(orig_patches.min(), orig_patches.max())
train.apply_preprocessor(
preprocessing.GlobalContrastNormalization(std_bias=10.))
processed_patches = train.get_topological_view().copy()
processed_patches /= N.abs(processed_patches).max()
train.apply_preprocessor(preprocessing.ZCA(), can_fit=True)
zca_patches = train.get_topological_view().copy()
zca_patches /= N.abs(zca_patches).max()
print id(zca_patches)
print id(processed_patches)
print N.abs(zca_patches - processed_patches).mean()
concat = N.concatenate((orig_patches, processed_patches, zca_patches), axis=2)
train.set_topological_view(concat)
serial.save('debug.pkl', train)
print 'making dot products'
dots = N.cast['float32'](N.dot(X, W))
print 'done'
print 'making activations'
acts = N.cast['float32'](N.zeros((X.shape[0], nhid)))
for i in xrange(0, X.shape[0], batch_size):
if i % 1000 == 0:
print i
cur_batch_size = min(batch_size, X.shape[0] - batch_size + 1)
acts[i:i + cur_batch_size, :] = model.hid_exp_func(X[i:i +
cur_batch_size, :])
print 'saving data'
dots_path = output_path + '_dots.npy'
acts_path = output_path + '_acts.npy'
N.save(dots_path, dots)
N.save(acts_path, acts)
serial.save(output_path + '.pkl', {
'dots': dots_path,
'acts': acts_path,
'b': b,
'beta': beta
})
orig_patches = train.get_topological_view().copy()
print (orig_patches.min(),orig_patches.max())
#orig_patches -= 0.5
orig_patches -= 127.5
orig_patches /= N.abs(orig_patches).max()
print (orig_patches.min(),orig_patches.max())
train.apply_preprocessor(preprocessing.GlobalContrastNormalization(std_bias=10.))
processed_patches = train.get_topological_view().copy()
processed_patches /= N.abs(processed_patches).max()
train.apply_preprocessor(preprocessing.ZCA(), can_fit = True)
zca_patches = train.get_topological_view().copy()
zca_patches /= N.abs(zca_patches).max()
print id(zca_patches)
print id(processed_patches)
print N.abs(zca_patches-processed_patches).mean()
concat = N.concatenate((orig_patches,processed_patches,zca_patches),axis=2)
train.set_topological_view(concat)
serial.save('debug.pkl',train)
from framework.utils import serial
from framework.datasets import cifar10
from framework.datasets import preprocessing
train = cifar10.CIFAR10(which_set="train")
pipeline = preprocessing.Pipeline()
pipeline.items.append(preprocessing.ExtractPatches(patch_shape=(8,8),num_patches=2000000))
pipeline.items.append(preprocessing.GlobalContrastNormalization())
pipeline.items.append(preprocessing.ZCA())
test = cifar10.CIFAR10(which_set="test")
train.apply_preprocessor(preprocessor = pipeline, can_fit = True)
serial.save('cifar10_preprocessor_2M.pkl',train)