def get_weights_as_images(self,
width,
height,
outdir='img/',
n_images=10,
img_type='grey'):
""" Create and save the weights of the hidden units with respect to the
visible units as images.
:param width:
:param height:
:param outdir:
:param n_images:
:param img_type:
:return: self
"""
outdir = self.data_dir + outdir
with tf.Session() as self.tf_session:
self.tf_saver.restore(self.tf_session, self.model_path)
weights = self.W.eval()
perm = np.random.permutation(self.num_hidden)[:n_images]
for p in perm:
w = np.array([i[p] for i in weights])
image_path = outdir + self.model_name + '_{}.png'.format(p)
utils.gen_image(w, width, height, image_path, img_type)
def get_weights_as_images(self,
width,
height,
outdir='img/',
max_images=10,
model_path=None):
""" Save the weights of this autoencoder as images, one image per hidden unit.
Useful to visualize what the autoencoder has learned.
:type width: int
:param width: Width of the images
:type height: int
:param height: Height of the images
:type outdir: string, default 'data/sdae/img'
:param outdir: Output directory for the images. This path is appended to self.data_dir
:type max_images: int, default 10
:param max_images: Number of images to return.
"""
assert max_images <= self.n_components
outdir = self.data_dir + outdir
if not os.path.isdir(outdir):
os.mkdir(outdir)
with tf.Session() as self.sess:
# Restore trained model
if model_path is not None:
self.saver.restore(self.sess, model_path)
else:
self.saver.restore(self.sess,
self.models_dir + self.model_name)
# Extract encoding weights as numpy array
enc_weights = self.Wf_.eval()
# Extract decoding weights as numpy arrays
dec_weights = self.Wg_.eval() if not self.tied_weights else None
perm = np.random.permutation(self.n_components)[:max_images]
for p in perm:
enc_w = np.array([i[p] for i in enc_weights])
image_path = outdir + self.model_name + '-enc_weights_{}.png'.format(
p)
utils.gen_image(enc_w, width, height, image_path)
if not self.tied_weights:
dec_w = np.array([i[p] for i in dec_weights])
image_path = outdir + self.model_name + '-dec_weights_{}.png'.format(
p)
utils.gen_image(dec_w, width, height, image_path)
def output_char(source_name, c, output_dir, post_processing=downsample):
global counter
name = '%s_%d' % (source_name, counter)
counter += 1
img = gen_image([c])
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
detected_lines = extract_lines(img)
if len(detected_lines) == 0:
return
img = detected_lines[0]
detected_chars, _ = extract_chars(img)
if len(detected_chars) == 0:
return
img = detected_chars[0]
img = post_processing(img)
# write the generated image with various quality
for qual in qualities:
out_name = '%s.%d.jpg' % (name, qual)
cv2.imwrite(os.path.join(output_dir, out_name), img,
(cv2.IMWRITE_JPEG_QUALITY, qual))
# write the text, which will be the labels.
with open(os.path.join(output_dir, name + '.txt'), 'w') as text_f:
text_f.write(c + '\n')
def get_weights_as_images(self,
width,
height,
outdir='img/',
n_images=10,
img_type='grey'):
outdir = self.data_dir + outdir
with tf.Session() as self.sess:
self.saver.restore(self.sess, self.models_dir + self.model_name)
weights = self.W.eval()
perm = np.random.permutation(self.nhid)[:n_images]
for p in perm:
w = np.array([i[p] for i in weights])
image_path = outdir + self.model_name + '_{}.png'.format(p)
utils.gen_image(w, width, height, image_path, img_type)
def get_weights_as_images(self, width, height, main_dir='dae/', outdir='img/', n_images=10, model_path=None, img_type='grey'):
''' Saves the weights of this autoencoder as images, one image per hidden unit.
This is useful to visualize what the autoencoder has learned
Parameters
----------
width : width of the images, int
height : height of the images, int
main_dir : path where all your encoding information is placed
outdir : output directory for the images -- this path is appended to self.data_dir, string (default is 'data/sdae/img')
n_images : number of images to return, int (default is 10)
'''
assert n_images <= self.n_components
main_dir = main_dir + '/' if main_dir[-1] != '/' else main_dir
outdir = outdir + '/' if outdir[-1] != '/' else outdir
outdir = main_dir + outdir
if not os.path.isdir(outdir):
os.mkdir(outdir)
with tf.Session() as self.tf_session:
if model_path is not None:
self.tf_saver.restore(self.tf_session, model_path)
else:
self.tf_saver.restore(self.tf_session, self.model_path)
enc_weights = self.W_.eval()
perm = np.random.permutation(enc_weights.shape[1])[:n_images]
for p in perm:
enc_w = np.array([i[p] for i in enc_weights])
image_path = outdir + self.model_name + '-enc_weights_{}.png'.format(p)
utils.gen_image(enc_w, width, height, image_path, img_type)