def save_top(data, column, columnuid, layeruid, save_path):
t = column.fwd(data)
top_shape = column.top_shape()
if len(top_shape) == 4:
im = w2i.tile_imgs(t)
# im = Image.fromarray(dp.denormalize(c[0,:]).astype(np.uint8).squeeze(),mode='L')
im.save(save_path + 'col' + str(columnuid) + '_level' + str(layeruid) +
'_top.png')
def save_column_means(dp, columns, immap):
#generate mean image for each column
means = np.zeros([N_COLUMNS] + list(dp.shape()[1:]))
for i, (col, keys) in enumerate(immap['col2key'].iteritems()):
mb = dp.get_mb_by_keys(keys)
x = np.mean(mb[0], axis=0)
means[i, :, :, :] = x
im = w2i.tile_imgs(means)
im.save(IMG_DIR + "mean_imgs.png")
def save_top(dp, columns, immap):
for n, column in columns.iteritems():
mapped_samples, _, _ = get_mapped_batch(dp, n, immap)
if len(mapped_samples) == 0:
return
mapped_batch = np.zeros(dp.shape())
mapped_batch[0:len(mapped_samples), :] = mapped_samples
t = column.fwd(mapped_batch)
top_shape = column.top_shape()
if len(top_shape) == 4:
im = w2i.tile_imgs(t)
# im = Image.fromarray(dp.denormalize(c[0,:]).astype(np.uint8).squeeze(),mode='L')
im.save(IMG_DIR + 'col' + str(n) + '_level' +
str(layer_number + 1) + '_top.png')
elif len(top_shape) == 2:
t = t.reshape([1] + top_shape + [1])
im = w2i.tile_imgs(t)
im.save(IMG_DIR + 'col' + str(n) + '_level' +
str(layer_number + 1) + '_top.png')
else:
print("Top not saved. Dimenstions {}".format(top_shape))
def save_recon(data, column, columnuid, layeruid, save_path):
#Reconstruction of a batch
d, r = column.fwd_back(data)
print(np.max(d))
print(np.max(r))
s = list(d.shape)
s[0] = s[0] * 2
d_r_array = np.empty(s, dtype=d.dtype)
d_r_array[0::2, :, :, :] = d
d_r_array[1::2, :, :, :] = r
im = w2i.tile_imgs(d_r_array)
im.save(save_path + 'col' + str(columnuid) + '_img_recon_level' +
str(layeruid) + '.png')
def save_recon(dp, columns, immap):
for n, column in columns.iteritems():
#Reconstruct mapped examples
mapped_samples, _, _ = get_mapped_batch(dp, n, immap)
if len(mapped_samples) == dp.shape()[0]:
d, r = column.fwd_back(mapped_samples)
s = list(d.shape)
s[0] = s[0] * 2
d_r_array = np.empty(s, dtype=d.dtype)
d_r_array[0::2, :, :, :] = d
d_r_array[1::2, :, :, :] = r
im = w2i.tile_imgs(d_r_array)
im.save(IMG_DIR + 'col' + str(n) + '_mapped_recon_level' +
str(layer_number + 1) + '.png')
def save_injection(column, columnuid, layeruid, save_path):
#Reconstruction of an injected filter
top_shape = column.top_shape()
a = np.zeros(top_shape)
input_shape = column.bottom_shape()
imgs = np.zeros([top_shape[-1]] + list(input_shape[1:]))
for channel in range(top_shape[-1]):
b = a.copy()
if len(top_shape) == 4:
b[0, top_shape[1] / 2, top_shape[2] / 2, channel] = 1
elif len(top_shape) == 2:
b[0, channel] = 1
c = column.inject(b)
imgs[channel, :, :, :] = c[0, :, :, :]
im = w2i.tile_imgs(imgs, normalize=True)
# im = Image.fromarray(dp.denormalize(c[0,:]).astype(np.uint8).squeeze(),mode='L')
im.save(save_path + 'col' + str(columnuid) + '_level' + str(layeruid) +
'_inject.png')
def save_injection(dp, columns, immap):
for n, column in columns.iteritems():
#Reconstruct an injected value
top_shape = column.top_shape()
a = np.zeros(top_shape)
input_shape = dp.shape()
imgs = np.zeros([top_shape[-1]] + list(input_shape[1:]))
for channel in range(top_shape[-1]):
b = a.copy()
if len(top_shape) == 4:
b[0, top_shape[1] / 2, top_shape[2] / 2, channel] = 1
elif len(top_shape) == 2:
b[0, channel] = 1
c = column.inject(b)
imgs[channel, :, :, :] = c[0, :, :, :]
im = w2i.tile_imgs(imgs, normalize=True)
# im = Image.fromarray(dp.denormalize(c[0,:]).astype(np.uint8).squeeze(),mode='L')
im.save(IMG_DIR + 'col' + str(n) + '_level' + str(layer_number + 1) +
'_decode.png')
if __name__ == '__main__':
cifarpath = sys.argv[1]
with open(cifarpath, 'rb') as fi:
datadict = cPickle.load(fi)
data = datadict['data']
data = np.reshape(data,[len(data), 3,32,32])
original_shape = data.shape
print("Original Data Shape {}".format(original_shape))
a,b,dog_data = dog(data,1,0.5)
print("Dog Data Shape {}".format(dog_data.shape))
appended_mat = np.append(
np.append(data[0:2],
a[0:2],
axis=0
),
np.append(b[0:2],
dog_data[0:2],
axis =0
),
axis=0
)
print(appended_mat.shape)
comparison_mat = np.transpose(appended_mat,[0,2,3,1])
comp = w2i.tile_imgs(comparison_mat)
plt.imshow(comp)
plt.show()
newdatadict = {'data':dog_data,
'labels':datadict['labels']}
with open("dog_cifar_data", 'w') as fo:
cPickle.dump(newdatadict, fo)
def save_exemplars(dp, columns, immap):
for i in range(len(columns)):
sample, l, k = get_mapped_batch(dp, i, immap)
if k != None and len(k) > 0:
im = w2i.tile_imgs(sample)
im.save(IMG_DIR + "col" + str(i) + "_exemplars.png")
return np.dot(inputs, ZCAMatrix) #Data whitening
def zca_whitening_sk(inputs):
pca = PCA(whiten=True)
transformed = pca.fit_transform(inputs)
pca.whiten = False
zca = pca.inverse_transform(transformed)
return zca
if __name__ == '__main__':
cifarpath = sys.argv[1]
with open(cifarpath, 'rb') as fi:
datadict = cPickle.load(fi)
data = datadict['data']
original_shape = data.shape
print("Original Data Shape {}".format(original_shape))
data = np.reshape(data, [len(data), -1])
white_data = zca_whitening(data)
comp = w2i.tile_imgs(
np.transpose(
np.reshape(np.append(data[0, :], white_data[0, :], axis=0),
[2, 3, 32, 32])), [0, 2, 3, 1])
plt.imshow(comp)
plt.show()
white_data = np.reshape(white_data, original_shape)
newdatadict = {'data': white_data, 'labels': datadict['labels']}
with open("zca_cifar_data", 'w') as fo:
cPickle.dump(newdatadict, fo)
