def classify_from_patchlist(imlist, imdict, pyparams, net, scorelayer = 'score', startlayer = 'conv1_1'):
estlist, scorelist, gtlist = [], [], []
transformer = Transformer(pyparams['im_mean'])
for imname in imlist:
patchlist = []
(point_anns, height_cm) = imdict[os.path.basename(imname)]
# Load image
im = np.asarray(Image.open(imname))
(im, scale) = coral_image_resize(im, pyparams['scaling_method'], pyparams['scaling_factor'], height_cm) #resize.
# Pad the boundaries
im = np.pad(im, ((pyparams['crop_size']*2, pyparams['crop_size']*2),(pyparams['crop_size']*2, pyparams['crop_size']*2), (0, 0)), mode='reflect')
# Extract patches
for (row, col, label) in point_anns:
center_org = np.asarray([row, col])
center = np.round(pyparams['crop_size']*2 + center_org * scale).astype(np.int)
patchlist.append(crop_and_rotate(im, center, pyparams['crop_size'], 0, tile = False))
gtlist.append(label)
# Classify and append
[this_estlist, this_scorelist] = classify_imlist(patchlist, net, transformer, pyparams['batch_size'], scorelayer = scorelayer, startlayer = startlayer)
estlist.extend(this_estlist)
scorelist.extend(this_scorelist)
return (gtlist, estlist, scorelist)
def __call__(self):
t1 = timer()
self.result['data'] = []
self.result['label'] = []
if self._cur + self.params['imgs_per_batch'] >= len(self.imlist):
self._cur = 0
shuffle(self.imlist)
# Grab images names from imlist
imnames = self.imlist[self._cur : self._cur + self.params['imgs_per_batch']]
# Figure out how many patches to grab from each image
patches_per_image = self.chunkify(self.params['batch_size'], self.params['imgs_per_batch'])
# Make nice output string
output_str = [str(npatches) + ' from ' + os.path.basename(imname) + '(id ' + str(itt) + ')' for imname, npatches, itt in zip(imnames, patches_per_image, range(self._cur, self._cur + self.params['imgs_per_batch']))]
# Loop over each image
for imname, npatches in zip(imnames, patches_per_image):
self._cur += 1
# randomly select the rotation angle for each patch
angles = np.random.choice(360, size = npatches, replace = True)
# randomly select whether to flip this particular patch.
flips = np.round(np.random.rand(npatches))*2-1
# get random offsets
rand_offsets = np.round(np.random.rand(npatches, 2) * (self.params['rand_offset'] * 2) - self.params['rand_offset'])
# Randomly permute the patch list for this image. Sampling is done with replacement
# so that if we ask for more patches than is available, it still computes.
(point_anns, height_cm) = self.imdict[os.path.basename(imname)] # read point annotations and image height in centimeters.
point_anns = [point_anns[pp] for pp in np.random.choice(len(point_anns), size = npatches, replace = True)]
# Load image
im = np.asarray(Image.open(imname))
(im, scale) = coral_image_resize(im, self.params['scaling_method'], self.params['scaling_factor'], height_cm) #resize.
# Pad the boundaries
crop_size = self.params['crop_size'] #for convenience, store this value locally
im = np.pad(im, ((crop_size * 2, crop_size * 2),(crop_size * 2, crop_size * 2), (0, 0)), mode='reflect')
for ((row, col, label), angle, flip, rand_offset) in zip(point_anns, angles, flips, rand_offsets):
center_org = np.asarray([row, col])
center = np.round(crop_size * 2 + center_org * scale + rand_offset).astype(np.int)
patch = self.transformer(crop_and_rotate(im, center, crop_size, angle, tile = False))
self.result['data'].append(patch[::flip, :, :])
self.result['label'].append(label)
def classify_from_patchlist(imlist, imdict, pyparams, workdir, scorelayer = 'score', startlayer = 'conv1_1', net_prototxt = 'testnet.prototxt', gpuid = 0, snapshot_prefix = 'snapshot', save = False):
# Preliminaries
caffemodel = find_latest_caffemodel(workdir, snapshot_prefix = snapshot_prefix)
net = load_model(workdir, caffemodel, gpuid = gpuid, net_prototxt = net_prototxt)
transformer = Transformer(pyparams['im_mean'])
estlist, scorelist, gtlist = [], [], []
print "classifying {} images in {} using {}".format(len(imlist), workdir, caffemodel)
for imname in tqdm(imlist):
patchlist = []
(point_anns, height_cm) = imdict[os.path.basename(imname)]
# Load image
im = np.asarray(Image.open(imname))
(im, scale) = coral_image_resize(im, pyparams['scaling_method'], pyparams['scaling_factor'], height_cm) #resize.
# Pad the boundaries
im = np.pad(im, ((pyparams['crop_size']*2, pyparams['crop_size']*2),(pyparams['crop_size']*2, pyparams['crop_size']*2), (0, 0)), mode='reflect')
# Extract patches
for (row, col, label) in point_anns:
center_org = np.asarray([row, col])
center = np.round(pyparams['crop_size']*2 + center_org * scale).astype(np.int)
patchlist.append(crop_and_rotate(im, center, pyparams['crop_size'], 0, tile = False))
gtlist.append(label)
# Classify and append
[this_estlist, this_scorelist] = classify_imlist(patchlist, net, transformer, pyparams['batch_size'], scorelayer = scorelayer, startlayer = startlayer)
estlist.extend(this_estlist)
scorelist.extend(this_scorelist)
if (save):
pickle.dump((gtlist, estlist, scorelist), open(os.path.join(workdir, 'predictions_using_' + caffemodel + '.p'), 'wb'))
return [gtlist, estlist, scorelist]
def __call__(self):
t1 = timer()
self.result['data'] = []
self.result['label'] = []
if self._cur + self.params['imgs_per_batch'] >= len(self.imlist):
self._cur = 0
shuffle(self.imlist)
# Grab images names from imlist
imnames = self.imlist[self._cur:self._cur +
self.params['imgs_per_batch']]
# Figure out how many patches to grab from each image
patches_per_image = self.chunkify(self.params['batch_size'],
self.params['imgs_per_batch'])
# Make nice output string
output_str = [
str(npatches) + ' from ' + os.path.basename(imname) + '(id ' +
str(itt) + ')' for imname, npatches, itt in zip(
imnames, patches_per_image,
range(self._cur, self._cur + self.params['imgs_per_batch']))
]
# Loop over each image
for imname, npatches in zip(imnames, patches_per_image):
self._cur += 1
# randomly select the rotation angle for each patch
angles = np.random.choice(360, size=npatches, replace=True)
# randomly select whether to flip this particular patch.
flips = np.round(np.random.rand(npatches)) * 2 - 1
# get random offsets
rand_offsets = np.round(
np.random.rand(npatches, 2) *
(self.params['rand_offset'] * 2) - self.params['rand_offset'])
# Randomly permute the patch list for this image. Sampling is done with replacement
# so that if we ask for more patches than is available, it still computes.
(point_anns, height_cm) = self.imdict[os.path.basename(
imname
)] # read point annotations and image height in centimeters.
point_anns = [
point_anns[pp] for pp in np.random.choice(
len(point_anns), size=npatches, replace=True)
]
# Load image
im = np.asarray(Image.open(imname))
(im, scale) = coral_image_resize(im, self.params['scaling_method'],
self.params['scaling_factor'],
height_cm) #resize.
# Pad the boundaries
crop_size = self.params[
'crop_size'] #for convenience, store this value locally
im = np.pad(im, ((crop_size * 2, crop_size * 2),
(crop_size * 2, crop_size * 2), (0, 0)),
mode='reflect')
for ((row, col, label), angle, flip,
rand_offset) in zip(point_anns, angles, flips, rand_offsets):
center_org = np.asarray([row, col])
center = np.round(crop_size * 2 + center_org * scale +
rand_offset).astype(np.int)
patch = self.transformer(
crop_and_rotate(im, center, crop_size, angle, tile=False))
self.result['data'].append(patch[::flip, :, :])
self.result['label'].append(label)