def main(model_name, iteration, storage_name):
name = '%s-%s' % (model_name, iteration)
print settings.model(name), settings.pretrained(name)
safe = datastore(settings.storage(storage_name))
safe.super_name = 'features'
safe.sub_name = name
layer_names = ['fc7', 'fc6', 'pool5', 'conv5', 'conv4', 'conv3']
layer_dims = [4096, 4096, 9216, 43264, 64896, 64896]
net = caffe.Classifier(settings.model(name), settings.pretrained(name), mean=np.load(settings.ILSVRC_MEAN), channel_swap=(2, 1, 0), raw_scale=255)
net.set_mode_gpu()
net.set_phase_test()
cub = CUB_200_2011(settings.CUB_ROOT)
dataset_size = sum(1 for _ in cub.get_all_images())
instance = {}
for layer, dim in zip(layer_names, layer_dims):
instance[layer] = np.zeros((dataset_size, dim))
print instance[layer].shape
for i, info in enumerate(cub.get_all_images(cropped=True)):
print info['img_id']
img = caffe.io.load_image(info['img_file'])
net.predict([img], oversample=False)
for layer in layer_names:
instance[layer][i, :] = net.blobs[layer].data[0].flatten()
for layer in layer_names:
safe.save_large_instance(safe.get_instance_path(safe.super_name, safe.sub_name, 'feat_cache_%s' % layer), instance[layer], 4)
def main(out_path, part, random_state, pgs, net_name):
utils.ensure_dir(out_path)
cub = CUB_200_2011(settings.CUB_ROOT)
lfrg = rects.BerkeleyRG(settings.BERKELEY_ANNOTATION_BASE_PATH, cub, part)
RG = rects.RandomForestRG(datastore(settings.storage('rf')),
lfrg,
cub_utils.DeepHelper.get_custom_net(
settings.model(net_name),
settings.pretrained(net_name)),
net_name,
cub,
random_state=random_state,
point_gen_strategy=pgs,
use_seg=True,
pt_n_part=20,
pt_n_bg=100)
RG.setup()
for i, image in enumerate(cub.get_all_images()):
print i
image_path = image['img_file']
img_id = int(image['img_id'])
rel_image_path = image['img_file_rel']
o_image = cv2.imread(image_path)
rect = RG.generate(img_id)
t_img_part = rect.get_rect(o_image)
out_image_path = os.path.join(out_path, rel_image_path)
utils.ensure_dir(os.path.dirname(out_image_path))
cv2.imwrite(out_image_path, t_img_part)
print 'Done'
def compute_estimated_part_data(model_name, shape, IDS, part_names_to_filter, add_noise, noise_std_c, noise_std_d):
net = caffe.Classifier(settings.model(model_name), settings.pretrained(model_name), mean=np.load(settings.ILSVRC_MEAN), channel_swap=(2, 1, 0), raw_scale=255)
net.set_phase_test()
net.set_mode_gpu()
# compute estimated head data
new_Xtest_part = np.zeros(shape)
for i, t_id in enumerate(IDS):
if to_oracle:
t_parts = all_parts_cub.for_image(t_id)
else:
t_parts = estimated_test_parts.for_image(t_id)
t_img_addr = all_image_infos[t_id]
t_img = caffe.io.load_image(t_img_addr)
t_parts_part = t_parts.filter_by_name(part_names_to_filter)
t_img_part = t_parts_part.get_rect(t_img, add_noise=add_noise, noise_std_c=noise_std_c, noise_std_d=noise_std_d)
try:
net.predict([t_img_part], oversample=False)
except Exception, e:
print '------', t_id, '----------'
print part_names_to_filter
print t_img_addr
print '------------'
print t_img.shape
print t_parts
print '------------'
print t_img_part.shape
print t_parts_part
raise e
new_Xtest_part[i, :] = net.blobs[feat_layer].data[0].flatten()
def main(model_name, iteration, storage_name):
name = '%s-%s' % (model_name, iteration)
print settings.model(name), settings.pretrained(name)
safe = datastore(settings.storage(storage_name))
safe.super_name = 'features'
safe.sub_name = name
layer_names = ['fc7', 'fc6', 'pool5', 'conv5', 'conv4', 'conv3']
layer_dims = [4096, 4096, 9216, 43264, 64896, 64896]
net = caffe.Classifier(settings.model(name),
settings.pretrained(name),
mean=np.load(settings.ILSVRC_MEAN),
channel_swap=(2, 1, 0),
raw_scale=255)
net.set_mode_gpu()
net.set_phase_test()
cub = CUB_200_2011(settings.CUB_ROOT)
dataset_size = sum(1 for _ in cub.get_all_images())
instance = {}
for layer, dim in zip(layer_names, layer_dims):
instance[layer] = np.zeros((dataset_size, dim))
print instance[layer].shape
for i, info in enumerate(cub.get_all_images(cropped=True)):
print info['img_id']
img = caffe.io.load_image(info['img_file'])
net.predict([img], oversample=False)
for layer in layer_names:
instance[layer][i, :] = net.blobs[layer].data[0].flatten()
for layer in layer_names:
safe.save_large_instance(
safe.get_instance_path(safe.super_name, safe.sub_name,
'feat_cache_%s' % layer), instance[layer],
4)
def compute_estimated_part_data(model_name, shape, IDS, model_rf):
net = caffe.Classifier(settings.model(model_name),
settings.pretrained(model_name),
mean=np.load(settings.ILSVRC_MEAN),
channel_swap=(2, 1, 0),
raw_scale=255)
net.set_phase_test()
net.set_mode_gpu()
# compute estimated head data
new_Xtest_part = np.zeros(shape)
for i, t_id in enumerate(IDS):
print i
img = caffe.io.load_image(all_image_infos[t_id])
dh.init_with_image(img)
X = dh.features(dense_points)
preds_prob = model_rf.predict_proba(X)
max_prob = np.max(preds_prob[:, 1])
preds_prob = preds_prob[:, 1].reshape((227, 227)).T
preds = preds_prob >= (max_prob / 2)
preds = skimage.morphology.closing(preds,
skimage.morphology.square(10))
preds = skimage.morphology.remove_small_objects(preds,
min_size=10,
connectivity=1)
L, N = skimage.measure.label(preds, return_num=True, background=0)
L_no_bg = L[L != -1].flatten()
vals, counts = scipy.stats.mode(L_no_bg)
part_label = int(vals[0])
indices = np.where(L == part_label)
xmin = indices[0].min()
xmax = indices[0].max()
ymin = indices[1].min()
ymax = indices[1].max()
pmin = Part(-1, '?', -1, xmin, ymin, 1)
pmax = Part(-1, '?', -1, xmax, ymax, 1)
rect_parts = Parts(parts=[pmin, pmax])
rect_parts.denorm_for_size(img.shape[0], img.shape[1], size=227)
rect_info = rect_parts[0].x, rect_parts[1].x, rect_parts[
0].y, rect_parts[1].y
t_img_part = Parts().get_rect(img, rect_info=rect_info)
try:
net.predict([t_img_part], oversample=False)
except Exception:
print '------', t_id, '----------'
new_Xtest_part[i, :] = net.blobs[feat_layer].data[0].flatten()
return new_Xtest_part
def compute_estimated_part_data(model_name, shape, IDS, model_rf):
net = caffe.Classifier(settings.model(model_name), settings.pretrained(model_name), mean=np.load(settings.ILSVRC_MEAN), channel_swap=(2, 1, 0), raw_scale=255)
net.set_phase_test()
net.set_mode_gpu()
# compute estimated head data
new_Xtest_part = np.zeros(shape)
for i, t_id in enumerate(IDS):
print i
img = caffe.io.load_image(all_image_infos[t_id])
dh.init_with_image(img)
X = dh.features(dense_points)
preds_prob = model_rf.predict_proba(X)
max_prob = np.max(preds_prob[:, 1])
preds_prob = preds_prob[:, 1].reshape((227, 227)).T
preds = preds_prob >= (max_prob/2)
preds = skimage.morphology.closing(preds, skimage.morphology.square(10))
preds = skimage.morphology.remove_small_objects(preds, min_size=10, connectivity=1)
L, N = skimage.measure.label(preds, return_num=True, background=0)
L_no_bg = L[L != -1].flatten()
vals, counts = scipy.stats.mode(L_no_bg)
part_label = int(vals[0])
indices = np.where(L == part_label)
xmin = indices[0].min()
xmax = indices[0].max()
ymin = indices[1].min()
ymax = indices[1].max()
pmin = Part(-1, '?', -1, xmin, ymin, 1)
pmax = Part(-1, '?', -1, xmax, ymax, 1)
rect_parts = Parts(parts=[pmin, pmax])
rect_parts.denorm_for_size(img.shape[0], img.shape[1], size=227)
rect_info = rect_parts[0].x, rect_parts[1].x, rect_parts[0].y, rect_parts[1].y
t_img_part = Parts().get_rect(img, rect_info=rect_info)
try:
net.predict([t_img_part], oversample=False)
except Exception:
print '------', t_id, '----------'
new_Xtest_part[i, :] = net.blobs[feat_layer].data[0].flatten()
return new_Xtest_part
def compute_estimated_part_data(model_name, shape, IDS,
part_names_to_filter, add_noise,
noise_std_c, noise_std_d):
net = caffe.Classifier(settings.model(model_name),
settings.pretrained(model_name),
mean=np.load(settings.ILSVRC_MEAN),
channel_swap=(2, 1, 0),
raw_scale=255)
net.set_phase_test()
net.set_mode_gpu()
# compute estimated head data
new_Xtest_part = np.zeros(shape)
for i, t_id in enumerate(IDS):
if to_oracle:
t_parts = all_parts_cub.for_image(t_id)
else:
t_parts = estimated_test_parts.for_image(t_id)
t_img_addr = all_image_infos[t_id]
t_img = caffe.io.load_image(t_img_addr)
t_parts_part = t_parts.filter_by_name(part_names_to_filter)
t_img_part = t_parts_part.get_rect(t_img,
add_noise=add_noise,
noise_std_c=noise_std_c,
noise_std_d=noise_std_d)
try:
net.predict([t_img_part], oversample=False)
except Exception, e:
print '------', t_id, '----------'
print part_names_to_filter
print t_img_addr
print '------------'
print t_img.shape
print t_parts
print '------------'
print t_img_part.shape
print t_parts_part
raise e
new_Xtest_part[i, :] = net.blobs[feat_layer].data[0].flatten()
def main(sname, iteration, cropped, full, flipped, force, dataset,
storage_name):
new_name = '%s-%d' % (sname, iteration)
if dataset == 'segmented':
cub = CUB_200_2011_Segmented(settings.CUB_ROOT, full=full)
elif dataset == 'part-head':
cub = CUB_200_2011_Parts_Head(settings.CUB_ROOT, full=full)
elif dataset == 'part-body':
cub = CUB_200_2011_Parts_Body(settings.CUB_ROOT, full=full)
elif dataset == 'part-head-rf-new':
cub = CUB_200_2011(settings.CUB_ROOT, 'images_head_rf_new')
elif dataset == 'part-body-rf-new':
cub = CUB_200_2011(settings.CUB_ROOT, 'images_body_rf_new')
else:
cub = CUB_200_2011(settings.CUB_ROOT,
images_folder_name=dataset,
full=full)
if not storage_name:
ft_storage = datastore(settings.storage(new_name))
else:
ft_storage = datastore(settings.storage(storage_name))
ft_extractor = CNN_Features_CAFFE_REFERENCE(
ft_storage,
model_file=settings.model(new_name),
pretrained_file=settings.pretrained(new_name),
full=full,
crop_index=0)
number_of_images_in_dataset = sum(1 for _ in cub.get_all_images())
bar = pyprind.ProgBar(number_of_images_in_dataset, width=80)
for t, des in ft_extractor.extract_all(cub.get_all_images(),
flip=flipped,
crop=cropped,
bbox=cub.get_bbox(),
force=force):
bar.update()
print 'DONE'
def main(sname, iteration, cropped, full, flipped, force, dataset, storage_name):
new_name = '%s-%d' % (sname, iteration)
if dataset == 'segmented':
cub = CUB_200_2011_Segmented(settings.CUB_ROOT, full=full)
elif dataset == 'part-head':
cub = CUB_200_2011_Parts_Head(settings.CUB_ROOT, full=full)
elif dataset == 'part-body':
cub = CUB_200_2011_Parts_Body(settings.CUB_ROOT, full=full)
elif dataset == 'part-head-rf-new':
cub = CUB_200_2011(settings.CUB_ROOT, 'images_head_rf_new')
elif dataset == 'part-body-rf-new':
cub = CUB_200_2011(settings.CUB_ROOT, 'images_body_rf_new')
else:
cub = CUB_200_2011(settings.CUB_ROOT, images_folder_name=dataset, full=full)
if not storage_name:
ft_storage = datastore(settings.storage(new_name))
else:
ft_storage = datastore(settings.storage(storage_name))
ft_extractor = CNN_Features_CAFFE_REFERENCE(ft_storage, model_file=settings.model(new_name), pretrained_file=settings.pretrained(new_name), full=full, crop_index=0)
number_of_images_in_dataset = sum(1 for _ in cub.get_all_images())
bar = pyprind.ProgBar(number_of_images_in_dataset, width=80)
for t, des in ft_extractor.extract_all(cub.get_all_images(), flip=flipped, crop=cropped, bbox=cub.get_bbox(), force=force):
bar.update()
print 'DONE'
def main(out_path, part, random_state, pgs, net_name):
utils.ensure_dir(out_path)
cub = CUB_200_2011(settings.CUB_ROOT)
lfrg = rects.BerkeleyRG(settings.BERKELEY_ANNOTATION_BASE_PATH, cub, part)
RG = rects.RandomForestRG(datastore(settings.storage('rf')), lfrg, cub_utils.DeepHelper.get_custom_net(settings.model(net_name), settings.pretrained(net_name)), net_name, cub, random_state=random_state, point_gen_strategy=pgs, use_seg=True, pt_n_part=20, pt_n_bg=100)
RG.setup()
for i, image in enumerate(cub.get_all_images()):
print i
image_path = image['img_file']
img_id = int(image['img_id'])
rel_image_path = image['img_file_rel']
o_image = cv2.imread(image_path)
rect = RG.generate(img_id)
t_img_part = rect.get_rect(o_image)
out_image_path = os.path.join(out_path, rel_image_path)
utils.ensure_dir(os.path.dirname(out_image_path))
cv2.imwrite(out_image_path, t_img_part)
print 'Done'