def _add_proposals_from_file(
self, roidb, proposal_file, min_proposal_size, top_k, crowd_thresh
):
"""Add proposals from a proposals file to an roidb."""
logger.info('Loading proposals from: {}'.format(proposal_file))
proposals = load_object(proposal_file)
id_field = 'indexes' if 'indexes' in proposals else 'ids' # compat fix
_sort_proposals(proposals, id_field)
box_list = []
for i, entry in enumerate(roidb):
if i % 2500 == 0:
logger.info(' {:d}/{:d}'.format(i + 1, len(roidb)))
boxes = proposals['boxes'][i]
# Sanity check that these boxes are for the correct image id
assert entry['id'] == proposals[id_field][i]
# Remove duplicate boxes and very small boxes and then take top k
boxes = box_utils.clip_boxes_to_image(
boxes, entry['height'], entry['width']
)
keep = box_utils.unique_boxes(boxes)
boxes = boxes[keep, :]
keep = box_utils.filter_small_boxes(boxes, min_proposal_size)
boxes = boxes[keep, :]
if top_k > 0:
boxes = boxes[:top_k, :]
box_list.append(boxes)
_merge_proposal_boxes_into_roidb(roidb, box_list)
if crowd_thresh > 0:
_filter_crowd_proposals(roidb, crowd_thresh)
def _add_proposals_from_file(self, roidb, proposal_file, min_proposal_size,
top_k, crowd_thresh):
"""Add proposals from a proposals file to an roidb."""
logger.info('Loading proposals from: {}'.format(proposal_file))
proposals = load_object(proposal_file)
id_field = 'indexes' if 'indexes' in proposals else 'ids' # compat fix
_sort_proposals(proposals, id_field)
box_list = []
for i, entry in enumerate(roidb):
if i % 2500 == 0:
logger.info(' {:d}/{:d}'.format(i + 1, len(roidb)))
boxes = proposals['boxes'][i]
# Sanity check that these boxes are for the correct image id
assert entry['id'] == proposals[id_field][i]
# Remove duplicate boxes and very small boxes and then take top k
boxes = box_utils.clip_boxes_to_image(boxes, entry['height'],
entry['width'])
keep = box_utils.unique_boxes(boxes)
boxes = boxes[keep, :]
keep = box_utils.filter_small_boxes(boxes, min_proposal_size)
boxes = boxes[keep, :]
if top_k > 0:
boxes = boxes[:top_k, :]
box_list.append(boxes)
_merge_proposal_boxes_into_roidb(roidb, box_list)
if crowd_thresh > 0:
_filter_crowd_proposals(roidb, crowd_thresh)
def accuracy(dataset, detections_pkl):
# Load predictions and ground truths
ds = JsonDataset(dataset)
roidb = ds.get_roidb(gt=True)
dets = load_object(detections_pkl)
all_boxes = dets['all_boxes']
def id_or_index(ix, val):
if len(val) == 0:
return val
else:
return val[ix]
trues = 0.
# Iterate through all images
for ix, entry in enumerate(roidb):
cls_boxes_i = [
id_or_index(ix, cls_k_boxes) for cls_k_boxes in all_boxes
]
true_boxes = entry['boxes']
if (true_boxes.shape[0] == 0) == (len(cls_boxes_i[3]) == 0):
trues += 1
# Finally, calculate accuracy by dividing the sum of true predictions by total samples
acc = trues/len(roidb)
print("Accuracy: " + str(acc))
return acc
def do_reval(dataset_name, output_dir, args):
dataset = JsonDataset(dataset_name)
dets = load_object(os.path.join(output_dir, 'detections.pkl'))
# Override config with the one saved in the detections file
if args.cfg_file is not None:
core_config.merge_cfg_from_cfg(core_config.load_cfg(dets['cfg']))
else:
core_config._merge_a_into_b(core_config.load_cfg(dets['cfg']), cfg)
# re-filter on score threshold:
dets['all_boxes'] = \
[
[
im[im[:,4] > cfg.TEST.SCORE_THRESH,:] if len(im) != 0 else []
for im in cls
]
for cls in dets['all_boxes']
]
results = task_evaluation.evaluate_all(dataset,
dets['all_boxes'],
dets['all_segms'],
dets['all_keyps'],
output_dir,
use_matlab=args.matlab_eval)
task_evaluation.log_copy_paste_friendly_results(results)
def process_in_parallel(tag, total_range_size, binary, output_dir, opts=''):
"""Run the specified binary cfg.NUM_GPUS times in parallel, each time as a
subprocess that uses one GPU. The binary must accept the command line
arguments `--range {start} {end}` that specify a data processing range.
"""
# Snapshot the current cfg state in order to pass to the inference
# subprocesses
cfg_file = os.path.join(output_dir, '{}_range_config.yaml'.format(tag))
with open(cfg_file, 'w') as f:
envu.yaml_dump(cfg, stream=f)
subprocess_env = os.environ.copy()
processes = []
subinds = np.array_split(range(total_range_size), cfg.NUM_GPUS)
# Determine GPUs to use
cuda_visible_devices = os.environ.get('CUDA_VISIBLE_DEVICES')
if cuda_visible_devices:
gpu_inds = map(int, cuda_visible_devices.split(','))
assert -1 not in gpu_inds, \
'Hiding GPU indices using the \'-1\' index is not supported'
else:
gpu_inds = reversed(range(cfg.NUM_GPUS))
# Run the binary in cfg.NUM_GPUS subprocesses
for i, gpu_ind in enumerate(gpu_inds):
start = subinds[i][0]
end = subinds[i][-1] + 1
subprocess_env['CUDA_VISIBLE_DEVICES'] = str(gpu_ind)
cmd = '{binary} --range {start} {end} --cfg {cfg_file} NUM_GPUS 1 {opts}'
cmd = cmd.format(binary=shlex_quote(binary),
start=int(start),
end=int(end),
cfg_file=shlex_quote(cfg_file),
opts=' '.join([shlex_quote(opt) for opt in opts]))
logger.info('{} range command {}: {}'.format(tag, i, cmd))
if i == 0:
subprocess_stdout = subprocess.PIPE
else:
filename = os.path.join(
output_dir, '%s_range_%s_%s.stdout' % (tag, start, end))
subprocess_stdout = open(filename, 'w') # NOQA (close below)
p = subprocess.Popen(cmd,
shell=True,
env=subprocess_env,
stdout=subprocess_stdout,
stderr=subprocess.STDOUT,
bufsize=1)
processes.append((i, p, start, end, subprocess_stdout))
print('Sleep 15s for gpu_id ', gpu_ind)
time.sleep(15)
# Log output from inference processes and collate their results
outputs = []
for i, p, start, end, subprocess_stdout in processes:
log_subprocess_output(i, p, output_dir, tag, start, end)
if i > 0:
subprocess_stdout.close()
range_file = os.path.join(output_dir,
'%s_range_%s_%s.pkl' % (tag, start, end))
range_data = load_object(range_file)
outputs.append(range_data)
return outputs
def rrpn_loader(rpn_pkl_file):
"""
Open, index and return a pickled proposals dictionary
"""
pkl = load_object(rpn_pkl_file)
proposals = {}
for boxes, scores, id in zip(pkl['boxes'], pkl['scores'], pkl['ids']):
proposals[id] = {'boxes': boxes, 'scores': scores}
return proposals
def vis(dataset, detections_pkl, thresh, output_dir, limit=0):
ds = JsonDataset(dataset)
roidb = ds.get_roidb()
dets = load_object(detections_pkl)
assert all(k in dets for k in ['all_boxes', 'all_segms', 'all_keyps']), \
'Expected detections pkl file in the format used by test_engine.py'
all_boxes = dets['all_boxes']
all_segms = dets['all_segms']
all_keyps = dets['all_keyps']
def id_or_index(ix, val):
if len(val) == 0:
return val
else:
return val[ix]
for ix, entry in enumerate(roidb):
if limit > 0 and ix >= limit:
break
if ix % 10 == 0:
print('{:d}/{:d}'.format(ix + 1, len(roidb)))
im = cv2.imread(entry['image'])
im_name = os.path.splitext(os.path.basename(entry['image']))[0]
cls_boxes_i = [
id_or_index(ix, cls_k_boxes) for cls_k_boxes in all_boxes
]
cls_segms_i = [
id_or_index(ix, cls_k_segms) for cls_k_segms in all_segms
]
cls_keyps_i = [
id_or_index(ix, cls_k_keyps) for cls_k_keyps in all_keyps
]
vis_utils.vis_one_image(
im[:, :, ::-1],
'{:d}_{:s}'.format(ix, im_name),
os.path.join(output_dir, 'vis'),
cls_boxes_i,
segms=cls_segms_i,
keypoints=cls_keyps_i,
thresh=thresh,
box_alpha=0.8,
dataset=ds,
show_class=True,
ext = "jpg"
)
def do_reval(dataset_name, output_dir, args):
dataset = JsonDataset(dataset_name)
dets = load_object(os.path.join(output_dir, 'detections.pkl'))
# Override config with the one saved in the detections file
if args.cfg_file is not None:
core_config.merge_cfg_from_cfg(core_config.load_cfg(dets['cfg']))
else:
core_config._merge_a_into_b(core_config.load_cfg(dets['cfg']), cfg)
results = task_evaluation.evaluate_all(dataset,
dets['all_boxes'],
dets['all_segms'],
dets['all_keyps'],
output_dir,
use_matlab=args.matlab_eval)
task_evaluation.log_copy_paste_friendly_results(results)
def convert_mask_net(args, mask_net):
# Initialization net
init_net = caffe2_pb2.NetDef()
net = caffe2_pb2.NetDef()
blobs = io_utils.load_object(args.wts)['blobs']
externals = set(
c2_utils.UnscopeName(inp) for inp in mask_net.external_input)
for name in set(blobs.keys()).intersection(externals):
blob = blobs[name]
add_custom_op(init_net,
'GivenTensorFill', [], [name],
values=blob.flatten(),
shape=blob.shape)
# Pre-process the ROIs
add_custom_op(net, 'BBoxToRoi', [
Constants.nms_outputs[1], Constants.im_info, Constants.nms_outputs[3]
], [Constants.mask_rois])
# Group the ROIs based on their FPN level
if cfg.FPN.MULTILEVEL_ROIS:
outputs = [Constants.mask_rois + Constants.idx_restore_suffix]
for level in range(cfg.FPN.ROI_MIN_LEVEL, cfg.FPN.ROI_MAX_LEVEL + 1):
outputs.append(Constants.mask_rois +
Constants.fpn_level_suffix(level))
add_custom_op(net,
'MultiLevelRoi', [Constants.mask_rois],
outputs,
min_level=cfg.FPN.ROI_MIN_LEVEL,
canon_scale=cfg.FPN.ROI_CANONICAL_SCALE,
canon_level=cfg.FPN.ROI_CANONICAL_LEVEL)
# Generate the masks
net.op.extend(mask_net.op)
# Post-process the masks
add_custom_op(net,
'SegmentMask',
Constants.nms_outputs[1:-1] +
[Constants.mask_pred, Constants.im_info],
[Constants.main_output, Constants.im_info],
thresh_bin=cfg.MRCNN.THRESH_BINARIZE)
net.name = args.net_name + '_mask'
init_net.name = args.net_name + '_mask_init'
save_model(net, init_net, args.mask_dir)
def do_reval(dataset_name, output_dir, args):
dataset = JsonDataset(dataset_name)
dets = load_object(os.path.join(output_dir, 'detections.pkl'))
# Override config with the one saved in the detections file
if args.cfg_file is not None:
core_config.merge_cfg_from_cfg(core_config.load_cfg(dets['cfg']))
else:
core_config._merge_a_into_b(core_config.load_cfg(dets['cfg']), cfg)
results = task_evaluation.evaluate_all(
dataset,
dets['all_boxes'],
dets['all_segms'],
dets['all_keyps'],
output_dir,
use_matlab=args.matlab_eval
)
task_evaluation.log_copy_paste_friendly_results(results)
def _do_body_uv_eval(json_dataset, res_file, output_dir):
ann_type = 'uv'
imgIds = json_dataset.COCO.getImgIds()
imgIds.sort()
res = load_object(res_file)
coco_dt = json_dataset.COCO.loadRes(res)
# Non-standard params used by the modified COCO API version
# from the DensePose fork
# global normalization factor used in per-instance evaluation
test_sigma = 0.255
coco_eval = denseposeCOCOeval(json_dataset.COCO, coco_dt, ann_type,
test_sigma)
coco_eval.params.imgIds = imgIds
coco_eval.evaluate()
coco_eval.accumulate()
eval_file = os.path.join(output_dir, 'body_uv_results.pkl')
save_object(coco_eval, eval_file)
logger.info('Wrote pickle eval results to: {}'.format(eval_file))
coco_eval.summarize()
return coco_eval
def convert_mask_net(args, mask_net):
# Initialization net
init_net = caffe2_pb2.NetDef()
net = caffe2_pb2.NetDef()
blobs = io_utils.load_object(args.wts)['blobs']
externals = set(c2_utils.UnscopeName(inp) for inp in mask_net.external_input)
for name in set(blobs.keys()).intersection(externals):
blob = blobs[name]
add_custom_op(init_net, 'GivenTensorFill', [], [name],
values = blob.flatten(),
shape = blob.shape)
# Pre-process the ROIs
add_custom_op(net, 'BBoxToRoi',
[Constants.nms_outputs[1], Constants.im_info, Constants.nms_outputs[3]],
[Constants.mask_rois])
# Group the ROIs based on their FPN level
if cfg.FPN.MULTILEVEL_ROIS:
outputs = [Constants.mask_rois + Constants.idx_restore_suffix]
for level in range(cfg.FPN.ROI_MIN_LEVEL, cfg.FPN.ROI_MAX_LEVEL + 1):
outputs.append(Constants.mask_rois + Constants.fpn_level_suffix(level))
add_custom_op(net, 'MultiLevelRoi', [Constants.mask_rois], outputs,
min_level = cfg.FPN.ROI_MIN_LEVEL,
canon_scale = cfg.FPN.ROI_CANONICAL_SCALE,
canon_level = cfg.FPN.ROI_CANONICAL_LEVEL)
# Generate the masks
net.op.extend(mask_net.op)
# Post-process the masks
add_custom_op(net, 'SegmentMask',
Constants.nms_outputs[1:-1] + [Constants.mask_pred, Constants.im_info],
[Constants.main_output, Constants.im_info],
thresh_bin = cfg.MRCNN.THRESH_BINARIZE)
net.name = args.net_name + '_mask'
init_net.name = args.net_name + '_mask_init'
save_model(net, init_net, args.mask_dir)
def load_and_convert_coco_model(args):
model_dict = load_object(args.coco_model_file_name)
remove_momentum(model_dict)
convert_coco_blobs_to_cityscape_blobs(model_dict)
return model_dict
parser.add_argument('--out_dir', dest='out_dir',
help='Test outputs directory',
default='../data/test')
args = parser.parse_args()
return args
##------------------------------------------------------------------------------
if __name__ == '__main__':
args = parse_args()
fig = plt.figure(figsize=(16, 9))
# Load the dataset and proposals
coco = COCO_PLUS(args.ann_file)
proposals = load_object(args.proposals)
num_imgs = len(coco.imgs)
assert len(proposals['ids']) == num_imgs, \
"Number of proposals do not match the number of images in the dataset"
step = 1
for i in tqdm(range(0, num_imgs, step)):
img_id = proposals['ids'][i]
points = coco.imgToPc[img_id]['points']
img_path = os.path.join(args.imgs_dir, coco.imgs[img_id]["file_name"])
img = np.array(plt.imread(img_path))
boxes = proposals['boxes'][i]
## Plot proposals
img = draw_xyxy_bbox(img, list(boxes), lineWidth=2)
img = draw_points(img, points, color=(0,0,255), radius=5, thickness=-1, format='RGB')
def visualize_ranking(dataset, detections_pkl, opts):
# Load predictions and ground truths
ds = JsonDataset(dataset)
roidb = ds.get_roidb(gt=True)
dets = load_object(detections_pkl)
all_boxes = dets['all_boxes']
def id_or_index(ix, val):
if len(val) == 0:
return val
else:
return val[ix]
# Load coordinates
with open(opts.coord_file) as json_file:
coord_data = json.load(json_file)
# Iterate through all images and note false positive and negatives, as well as entry scores
false_positives = []
false_negatives = []
scores = []
for ix, entry in enumerate(roidb):
cls_boxes_i = [
id_or_index(ix, cls_k_boxes) for cls_k_boxes in all_boxes
]
preds = np.array(cls_boxes_i[opts.class_id])
entry['preds'] = preds
true_boxes = entry['boxes']
if preds.shape[0] > 0 and np.max(preds[:, -1]) > opts.threshold:
box_scores = preds[:, -1]
box_scores = box_scores[np.where(box_scores > opts.threshold)]
score = np.sum(
box_scores[np.argsort(box_scores)[-opts.box_count:]])
scores.append([entry, score])
if true_boxes.shape[0] == 0:
false_positives.append(entry)
else:
if true_boxes.shape[0] > 0:
false_negatives = add_negative(false_negatives, entry,
coord_data, opts.min_distance)
# Find top rated entries
scores = np.array(scores)
scores = scores[np.argsort(scores[:, 1])[::-1]]
for entry in scores[:, 0]:
entry['coords'] = coord_data[os.path.split(entry['image'])[-1]]
# Filter by proximity
for i in range(scores.shape[0]):
if scores[i][1] > 0:
current_entry = scores[i][0]
for j in range(i + 1, scores.shape[0]):
second_entry = scores[j][0]
dist = distance(
(current_entry['coords'][0], current_entry['coords'][1]),
(second_entry['coords'][0],
second_entry['coords'][1])).km * 1000
if dist < opts.min_distance:
scores[j][1] = 0
scores = scores[np.where(scores[:, 1] > 0)]
top_entries = scores[np.argsort(scores[:, 1])[-opts.image_count:][::-1]]
# Choose random negative samples
false_samples = np.append(false_negatives, false_positives)
np.random.shuffle(false_samples)
# Visualize positive and negative samples
rows_cols = (opts.image_count,
2) if opts.angle == 'ver' else (2, opts.image_count)
plt_shape = (6., opts.image_count *
2.5) if opts.angle == 'ver' else (opts.image_count * 2.5, 6.)
fig = plt.figure(1, plt_shape)
grid = ImageGrid(
fig,
111,
nrows_ncols=rows_cols,
axes_pad=0.03,
label_mode='L',
)
# Show top ranked images
for i, result in enumerate(top_entries):
entry = result[0]
score = result[1]
grid_idx = i
if opts.angle == 'ver':
grid_idx = i * 2
# Load image and add bounding boxes
im = cv2.imread(entry['image'])
preds = entry['preds']
true_boxes = entry['boxes']
for bbox in true_boxes:
im = vis_bbox(
im, (bbox[0], bbox[1], bbox[2] - bbox[0], bbox[3] - bbox[1]),
_GT_COLOR, opts.box_thickness)
count = 0
for bbox in preds:
if bbox[-1] > opts.threshold:
count += 1
print(
os.path.split(entry['image'])[-1] + ': ' + str(bbox[0:4]))
im = vis_bbox(
im,
(bbox[0], bbox[1], bbox[2] - bbox[0], bbox[3] - bbox[1]),
_PRED_COLOR, opts.box_thickness)
if count >= opts.box_count:
break
# Adjust grid setting
im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
show_img(grid, im, grid_idx)
t = grid[grid_idx].text(12,
42,
"Score: " + str(round(score, 3)),
fontsize=8,
bbox=dict(boxstyle='square',
fc='white',
ec='none',
alpha=0.6))
if i == 0:
if opts.angle == 'ver':
grid[grid_idx].set_title("Top\nPredictions", size=18)
else:
grid[grid_idx].set_ylabel("Top Predictions", fontsize=13)
# Show random negative samples (false positive, false negative)
for i, entry in enumerate(false_samples):
if i >= opts.image_count:
break
grid_idx = opts.image_count + i
if opts.angle == 'ver':
grid_idx = 2 * i + 1
# Load image and add bounding boxes
im = cv2.imread(entry['image'])
preds = entry['preds']
true_boxes = entry['boxes']
for bbox in true_boxes:
im = vis_bbox(
im, (bbox[0], bbox[1], bbox[2] - bbox[0], bbox[3] - bbox[1]),
_GT_COLOR, opts.box_thickness)
for bbox in preds:
if bbox[-1] > opts.threshold:
im = vis_bbox(
im,
(bbox[0], bbox[1], bbox[2] - bbox[0], bbox[3] - bbox[1]),
_PRED_COLOR, opts.box_thickness)
# Adjust grid setting
im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
grid[grid_idx].imshow(im)
grid[grid_idx].grid(False)
grid[grid_idx].set_xticks([])
grid[grid_idx].set_yticks([])
if i == 0:
if opts.angle == 'ver':
grid[grid_idx].set_title("Errors", size=18)
else:
grid[grid_idx].set_ylabel("Errors", fontsize=13)
plt.axis('off')
plt.subplots_adjust(hspace=1)
plt.savefig("ranking.png", dpi=300, bbox_inches='tight')
# def plot_dist(dist,dist2=None):
# fig,ax = plt.subplots()
# plt_bars(dist, coco_classes, dist2, ax=ax)
# plt.show()
if __name__ == '__main__':
# workspace.GlobalInit(['caffe2', '--caffe2_log_level=0'])
logger = setup_logging(__name__)
args = parse_args()
logger.info('Called with args:')
logger.info(args)
assert args.class_weights is not None, 'class_weights file required'
print(str(args.class_weights[0]))
coco_wts = load_object(args.class_weights[0])
print(coco_wts.shape)
coco_dist = get_dist(coco_wts)
print("coco size: ", coco_wts.sum())
assert args.voc_class_weights is not None, 'voc dist files needed'
voc_files = args.voc_class_weights
if args.do_val_set:
voc_files = [
'collecting/test/voc_2007_test/generalized_rcnn/class_weights.pkl',
'collecting/test/voc_2012_val/generalized_rcnn/class_weights.pkl'
]
voc_wts = np.concatenate([load_object(vocfile) for vocfile in voc_files],
axis=0)
def _add_proposals_from_file(self, roidb, proposal_file, min_proposal_size,
top_k, crowd_thresh):
"""Add proposals from a proposals file to an roidb."""
logger.info('Loading proposals from: {}'.format(proposal_file))
proposals = load_object(proposal_file)
id_field = 'indexes' if 'indexes' in proposals else 'ids' # compat fix
# _remove_proposals_not_in_roidb(proposals, roidb, id_field)
_sort_proposals(proposals, id_field)
box_list = []
score_list = []
total_roi = 0
up_1024 = 0
up_2048 = 0
up_3072 = 0
up_4096 = 0
for i, entry in enumerate(roidb):
if i % 2500 == 0:
logger.info(' {:d}/{:d}'.format(i + 1, len(roidb)))
boxes = proposals['boxes'][i]
scores = proposals['scores'][i]
# Sanity check that these boxes are for the correct image id
assert entry['id'] == proposals[id_field][i]
# Remove duplicate boxes and very small boxes and then take top k
# boxes = box_utils.clip_boxes_to_image(
# boxes, entry['height'], entry['width']
# )
assert (boxes[:, 0] >= 0).all()
assert (boxes[:, 1] >= 0).all()
assert (boxes[:, 2] >= boxes[:, 0]).all()
assert (boxes[:, 3] >= boxes[:, 1]).all()
assert (boxes[:, 2] < entry['width']).all(), entry['image']
assert (boxes[:, 3] < entry['height']).all(), entry['image']
keep = box_utils.unique_boxes(boxes)
boxes = boxes[keep, :]
scores = scores[keep]
keep = box_utils.filter_small_boxes(boxes, min_proposal_size)
boxes = boxes[keep, :]
scores = scores[keep]
# sort by confidence
sorted_ind = np.argsort(-scores.flatten())
boxes = boxes[sorted_ind, :]
scores = scores[sorted_ind, :]
if top_k > 0:
boxes = boxes[:top_k, :]
scores = scores[:top_k]
total_roi += boxes.shape[0]
if boxes.shape[0] > 1024:
up_1024 += 1
if boxes.shape[0] > 2048:
up_2048 += 1
if boxes.shape[0] > 3072:
up_3072 += 1
if boxes.shape[0] > 4096:
up_4096 += 1
box_list.append(boxes)
score_list.append(scores)
print('total_roi: ', total_roi, ' ave roi: ',
total_roi / len(box_list))
print('up_1024: ', up_1024)
print('up_2048: ', up_2048)
print('up_3072: ', up_3072)
print('up_4096: ', up_4096)
_merge_proposal_boxes_into_roidb(roidb, box_list, score_list)
if crowd_thresh > 0:
_filter_crowd_proposals(roidb, crowd_thresh)
def _add_pseudo_gt_annotations(self, roidb, gt_roidb):
"""
Return the database of pseudo ground-truth regions of interest from detect result.
This function loads/saves from/to a cache file to speed up future calls.
"""
# gt_roidb = copy.deepcopy(roidb)
# for entry in roidb:
# self._add_gt_annotations(entry)
assert 'train' in self.name or 'val' in self.name, 'Only trainval dataset has pseudo gt.'
# detection.pkl is 0-based indices
# the VOC result file is 1-based indices
cache_files = cfg.PSEUDO_PATH
if isinstance(cache_files, basestring):
cache_files = (cache_files, )
all_dets = None
for cache_file in cache_files:
if self.name not in cache_file:
continue
assert os.path.exists(cache_file), cache_file
# with open(cache_file, 'rb') as fid:
# res = cPickle.load(fid)
res = load_object(cache_file)
print('{} pseudo gt roidb loaded from {}'.format(
self.name, cache_file))
if all_dets is None:
all_dets = res['all_boxes']
else:
for i in range(len(all_dets)):
all_dets[i].extend(res['all_boxes'][i])
assert len(all_dets[1]) == len(roidb), len(all_dets[1])
if len(all_dets) == self.num_classes:
cls_offset = 0
elif len(all_dets) + 1 == self.num_classes:
cls_offset = -1
else:
raise Exception('Unknown mode.')
threshold = 1.0
for im_i, entry in enumerate(roidb):
if im_i % 1000 == 0:
print('{:d} / {:d}'.format(im_i + 1, len(roidb)))
num_valid_objs = 0
if len(gt_roidb[im_i]['gt_classes']) == 0:
print(gt_roidb[im_i])
if len(gt_roidb[im_i]['is_crowd']) == sum(
gt_roidb[im_i]['is_crowd']):
print(gt_roidb[im_i])
# when cfg.WSL = False, background class is in.
# detection.pkl only has 20 classes
# fast_rcnn need 21 classes
for cls in range(1, self.num_classes):
# TODO(YH): we need threshold the pseudo label
# filter the pseudo label
# self._gt_class has 21 classes
# if self._gt_classes[ix][cls] == 0:
if cls not in gt_roidb[im_i]['gt_classes']:
continue
dets = all_dets[cls + cls_offset][im_i]
if dets.shape[0] <= 0:
continue
# TODO(YH): keep only one box
# if dets.shape[0] > 0:
# num_valid_objs += 1
max_score = 0
num_valid_objs_cls = 0
for i in range(dets.shape[0]):
det = dets[i]
score = det[4]
if score > max_score:
max_score = score
if score < threshold:
continue
num_valid_objs += 1
num_valid_objs_cls += 1
if num_valid_objs_cls == 0:
if max_score > 0:
num_valid_objs += 1
boxes = np.zeros((num_valid_objs, 4), dtype=entry['boxes'].dtype)
# obn_scores = np.zeros((num_valid_objs, 1), dtype=entry['obn_scores'].dtype)
gt_classes = np.zeros((num_valid_objs),
dtype=entry['gt_classes'].dtype)
gt_overlaps = np.zeros((num_valid_objs, self.num_classes),
dtype=entry['gt_overlaps'].dtype)
seg_areas = np.zeros((num_valid_objs),
dtype=entry['seg_areas'].dtype)
is_crowd = np.zeros((num_valid_objs),
dtype=entry['is_crowd'].dtype)
box_to_gt_ind_map = np.zeros(
(num_valid_objs), dtype=entry['box_to_gt_ind_map'].dtype)
obj_i = 0
valid_segms = []
for cls in range(1, self.num_classes):
# filter the pseudo label
# self._gt_class has 21 classes
# if self._gt_classes[ix][cls] == 0:
if cls not in gt_roidb[im_i]['gt_classes']:
continue
dets = all_dets[cls + cls_offset][im_i]
if dets.shape[0] <= 0:
continue
max_score = 0
max_score_bb = []
num_valid_objs_cls = 0
for i in range(dets.shape[0]):
det = dets[i]
x1 = det[0]
y1 = det[1]
x2 = det[2]
y2 = det[3]
score = det[4]
if score > max_score:
max_score = score
max_score_bb = [x1, y1, x2, y2]
if score < threshold:
continue
assert x1 >= 0
assert y1 >= 0
assert x2 >= x1
assert y2 >= y1
assert x2 < entry['width']
assert y2 < entry['height']
boxes[obj_i, :] = [x1, y1, x2, y2]
gt_classes[obj_i] = cls
seg_areas[obj_i] = (x2 - x1 + 1) * (y2 - y1 + 1)
is_crowd[obj_i] = 0
box_to_gt_ind_map[obj_i] = obj_i
gt_overlaps[obj_i, cls] = 1.0
valid_segms.append([])
obj_i += 1
num_valid_objs_cls += 1
if num_valid_objs_cls == 0:
x1, y1, x2, y2 = max_score_bb[:]
assert x1 >= 0
assert y1 >= 0
assert x2 >= x1
assert y2 >= y1
assert x2 < entry['width']
assert y2 < entry['height']
boxes[obj_i, :] = [x1, y1, x2, y2]
gt_classes[obj_i] = cls
seg_areas[obj_i] = (x2 - x1 + 1) * (y2 - y1 + 1)
is_crowd[obj_i] = 0
box_to_gt_ind_map[obj_i] = obj_i
gt_overlaps[obj_i, cls] = 1.0
valid_segms.append([])
obj_i += 1
assert obj_i == num_valid_objs
# Show Pseudo GT boxes
if True:
# if False:
import cv2
im = cv2.imread(entry['image'])
for obj_i in range(num_valid_objs):
cv2.rectangle(im, (boxes[obj_i][0], boxes[obj_i][1]),
(boxes[obj_i][2], boxes[obj_i][3]),
(255, 0, 0), 5)
save_dir = os.path.join(cfg.OUTPUT_DIR, 'pgt')
if not os.path.exists(save_dir):
os.makedirs(save_dir)
save_path = os.path.join(save_dir, str(im_i) + '.png')
# print(save_path)
cv2.imwrite(save_path, im)
# cv2.imshow('im', im)
# cv2.waitKey()
entry['boxes'] = np.append(entry['boxes'], boxes, axis=0)
# entry['obn_scores'] = np.append(entry['obn_scores'], obn_scores, axis=0)
entry['segms'].extend(valid_segms)
# To match the original implementation:
# entry['boxes'] = np.append(
# entry['boxes'], boxes.astype(np.int).astype(np.float), axis=0)
entry['gt_classes'] = np.append(entry['gt_classes'], gt_classes)
entry['seg_areas'] = np.append(entry['seg_areas'], seg_areas)
entry['gt_overlaps'] = np.append(entry['gt_overlaps'].toarray(),
gt_overlaps,
axis=0)
entry['gt_overlaps'] = scipy.sparse.csr_matrix(
entry['gt_overlaps'])
entry['is_crowd'] = np.append(entry['is_crowd'], is_crowd)
entry['box_to_gt_ind_map'] = np.append(entry['box_to_gt_ind_map'],
box_to_gt_ind_map)
help='Proposals file',
default='../output/temp_proposals_ss.pkl')
args = parser.parse_args()
args.image_file = os.path.abspath(args.image_file)
args.proposals_file = os.path.abspath(args.proposals_file)
return args
if __name__ == '__main__':
args = parse_args()
fig = plt.figure(figsize=(16, 6))
# Load the proposals
proposals = load_object(args.proposals_file)
# print(len(proposals1['indexes']))
# print(proposals2['ids'])
# np.set_printoptions(threshold=np.nan)
# for i in range(20):
# print(proposals1['boxes'][i][0:10,:])
# print('---------------------------')
# print(proposals2['boxes'][i][0:10,:])
# input('something')
# print('Nucoco proposal keys:')
# for key, value in proposals1.items() :
# print(key, value)
#
def test_net(weights_file,
dataset_name,
proposal_file,
output_dir,
ind_range=None,
gpu_id=0,
subset_pointer=None):
"""Run inference on all images in a dataset or over an index range of images
in a dataset using a single GPU.
"""
assert not cfg.MODEL.RPN_ONLY, \
'Use rpn_generate to generate proposals from RPN-only models'
# determine file name
if ind_range is not None:
det_name = 'detection_range_%s_%s.pkl' % tuple(ind_range)
else:
det_name = 'detections.pkl'
det_file = os.path.join(output_dir, det_name)
# load results if already present
if os.path.exists(det_file):
res = load_object(det_file)
all_boxes, all_segms, all_keyps = res['all_boxes'], res[
'all_segms'], res['all_keyps']
else:
roidb, dataset, start_ind, end_ind, total_num_images = get_roidb_and_dataset(
dataset_name, proposal_file, ind_range)
if subset_pointer is not None:
voc_subset = subset_pointer.subset
this_sub = voc_subset[:len(roidb)]
# subset_pointer.subset = voc_subset[len(roidb):]
# filter roidb:
roidb = [roi for taking, roi in zip(this_sub, roidb) if taking]
total_num_images = len(roidb)
end_ind = total_num_images
model = initialize_model_from_cfg(weights_file, gpu_id=gpu_id)
num_images = len(roidb)
num_classes = cfg.MODEL.NUM_CLASSES
all_boxes, all_segms, all_keyps = empty_results(
num_classes, num_images)
if cfg.TEST.COLLECT_ALL:
all_feats = []
all_class_weights = np.empty(shape=(num_images, num_classes),
dtype=np.float32)
timers = defaultdict(Timer)
for i, entry in enumerate(roidb):
if cfg.TEST.PRECOMPUTED_PROPOSALS:
# The roidb may contain ground-truth rois (for example, if the roidb
# comes from the training or val split). We only want to evaluate
# detection on the *non*-ground-truth rois. We select only the rois
# that have the gt_classes field set to 0, which means there's no
# ground truth.
box_proposals = entry['boxes'][entry['gt_classes'] == 0]
if len(box_proposals) == 0:
continue
else:
# Faster R-CNN type models generate proposals on-the-fly with an
# in-network RPN; 1-stage models don't require proposals.
box_proposals = None
im = cv2.imread(entry['image'])
with c2_utils.NamedCudaScope(gpu_id):
cls_boxes_i, cls_segms_i, cls_keyps_i, sum_softmax, topk_feats = im_detect_all(
model,
im,
box_proposals,
timers,
return_feats=cfg.TEST.COLLECT_ALL)
# print('nfeats:', topk_feats.shape[0])
# print(topk_feats)
extend_results(i, all_boxes, cls_boxes_i)
if cls_segms_i is not None:
extend_results(i, all_segms, cls_segms_i)
if cls_keyps_i is not None:
extend_results(i, all_keyps, cls_keyps_i)
if cfg.TEST.COLLECT_ALL:
all_class_weights[i] = sum_softmax
all_feats.append(
topk_feats
) # will accumulate about 9 Gb of feats on COCO train set (118K imgs)
if i % 10 == 0: # Reduce log file size
ave_total_time = np.sum(
[t.average_time for t in timers.values()])
eta_seconds = ave_total_time * (num_images - i - 1)
eta = str(datetime.timedelta(seconds=int(eta_seconds)))
det_time = (timers['im_detect_bbox'].average_time +
timers['im_detect_mask'].average_time +
timers['im_detect_keypoints'].average_time)
misc_time = (timers['misc_bbox'].average_time +
timers['misc_mask'].average_time +
timers['misc_keypoints'].average_time)
logger.info(('im_detect: range [{:d}, {:d}] of {:d}: '
'{:d}/{:d} {:.3f}s + {:.3f}s (eta: {})').format(
start_ind + 1, end_ind, total_num_images,
start_ind + i + 1, start_ind + num_images,
det_time, misc_time, eta))
if cfg.VIS:
im_name = os.path.splitext(os.path.basename(entry['image']))[0]
vis_utils.vis_one_image(im[:, :, ::-1],
'{:d}_{:s}'.format(i, im_name),
os.path.join(output_dir, 'vis'),
cls_boxes_i,
segms=cls_segms_i,
keypoints=cls_keyps_i,
thresh=cfg.VIS_TH,
box_alpha=0.8,
dataset=dataset,
show_class=True)
cfg_yaml = envu.yaml_dump(cfg)
save_object(
dict(all_boxes=all_boxes,
all_segms=all_segms,
all_keyps=all_keyps,
cfg=cfg_yaml), det_file)
logger.info('Wrote detections to: {}'.format(
os.path.abspath(det_file)))
if cfg.TEST.COLLECT_ALL:
save_object(all_class_weights,
os.path.join(output_dir, 'class_weights.pkl'))
save_object(all_feats,
os.path.join(output_dir, 'feature_vectors.pkl'))
logger.info(
'Wrote class weights and feature vectors to output folder')
return all_boxes, all_segms, all_keyps
fc6s = []
fc7s = []
max_classes = []
gt_classes = []
gt_classes_cnt = np.zeros(N, dtype=np.int32)
cnt = 0
cnt2 = 0
for root, dirs, files in os.walk(dir_in, topdown=False):
for name in files:
if name.endswith('pkl'):
pass
else:
continue
print(cnt, os.path.join(root, name))
path = os.path.join(root, name)
pkl = load_object(path)
roi_feat = pkl['roi_feats']
if has_fc:
fc6 = pkl['fc6']
fc7 = pkl['fc7']
gt_class = pkl['gt_classes']
print(cnt, cnt2, 'roi_feat: ', roi_feat.shape)
t = []
for i in range(roi_feat.shape[0]):
if gt_class[i] in t:
# continue
pass
def voc_eval_corloc(detpath,
annopath,
imagesetfile,
classname,
cachedir,
ovthresh=0.5,
use_07_metric=False):
# assumes detections are in detpath.format(classname)
# assumes annotations are in annopath.format(imagename)
# assumes imagesetfile is a text file with each line an image name
# cachedir caches the annotations in a pickle file
# first load gt
if not os.path.isdir(cachedir):
os.mkdir(cachedir)
imageset = os.path.splitext(os.path.basename(imagesetfile))[0]
cachefile = os.path.join(cachedir, imageset + '_annots.pkl')
# read list of images
with open(imagesetfile, 'r') as f:
lines = f.readlines()
imagenames = [x.strip() for x in lines]
if not os.path.isfile(cachefile):
# load annots
recs = {}
for i, imagename in enumerate(imagenames):
recs[imagename] = parse_rec(annopath.format(imagename))
if i % 100 == 0:
logger.info(
'Reading annotation for {:d}/{:d}'.format(
i + 1, len(imagenames)))
# save
logger.info('Saving cached annotations to {:s}'.format(cachefile))
save_object(recs, cachefile)
else:
# load
recs = load_object(cachefile)
# extract gt objects for this class
class_recs = {}
npos = 0
npos_im = 0
for imagename in imagenames:
R = [obj for obj in recs[imagename] if obj['name'] == classname]
bbox = np.array([x['bbox'] for x in R])
difficult = np.array([x['difficult'] for x in R]).astype(np.bool)
det = [False] * len(R)
npos = npos + sum(~difficult)
class_recs[imagename] = {'bbox': bbox,
'difficult': difficult,
'det': det}
if len(R) > 0:
npos_im += min(1, sum(~difficult))
# read dets
detfile = detpath.format(classname)
with open(detfile, 'r') as f:
lines = f.readlines()
if len(lines) == 0:
return 0.0
splitlines = [x.strip().split(' ') for x in lines]
image_ids = [x[0] for x in splitlines]
confidence = np.array([float(x[1]) for x in splitlines])
BB = np.array([[float(z) for z in x[2:]] for x in splitlines])
# sort by confidence
sorted_ind = np.argsort(-confidence)
sorted_scores = np.sort(-confidence)
BB = BB[sorted_ind, :]
image_ids = [image_ids[x] for x in sorted_ind]
nd = len(image_ids)
T = []
F = []
too_min = 0
for d in range(nd):
if image_ids[d] in T or image_ids[d] in F:
continue
R = class_recs[image_ids[d]]
all_difficult = True
for difficult in R['difficult']:
if not difficult:
all_difficult = False
if all_difficult:
continue
bb = BB[d, :].astype(float)
ovmax = -np.inf
BBGT = R['bbox'].astype(float)
if BBGT.size > 0:
# compute overlaps
# intersection
ixmin = np.maximum(BBGT[:, 0], bb[0])
iymin = np.maximum(BBGT[:, 1], bb[1])
ixmax = np.minimum(BBGT[:, 2], bb[2])
iymax = np.minimum(BBGT[:, 3], bb[3])
iw = np.maximum(ixmax - ixmin + 1., 0.)
ih = np.maximum(iymax - iymin + 1., 0.)
inters = iw * ih
# union
uni = ((bb[2] - bb[0] + 1.) * (bb[3] - bb[1] + 1.) +
(BBGT[:, 2] - BBGT[:, 0] + 1.) *
(BBGT[:, 3] - BBGT[:, 1] + 1.) - inters)
overlaps = inters / uni
ovmax = np.max(overlaps)
jmax = np.argmax(overlaps)
if ovmax > ovthresh:
T.append(image_ids[d])
else:
F.append(image_ids[d])
uni = (bb[2] - bb[0] + 1.) * (bb[3] - bb[1] + 1.)
overlaps = inters / uni
ovmax = np.max(overlaps)
if ovmax > ovthresh:
too_min += 1
if len(F) == 0:
too_min_rate = 0.0
else:
too_min_rate = 1.0 * too_min / len(F)
print('npos_im: ', npos_im, '#F: ', len(F), ' too_min: ', too_min, ' rate: ', too_min_rate)
return 1.0 * len(T) / npos_im, too_min_rate
import numpy as np
import os
import sys
from detectron.utils.io import save_object
from detectron.utils.io import load_object
if __name__ == '__main__':
original_weights_file = sys.argv[1]
file_out = sys.argv[2]
out_blobs = {}
used_blobs = set()
original_src_blobs = load_object(original_weights_file)
print('====================================')
print('get params in original weights')
for blob_name in sorted(original_src_blobs.keys()):
if 'bn_s' in blob_name:
pass
else:
continue
bn_name = blob_name.rsplit('_', 1)[0]
conv_name = blob_name.rsplit('_', 2)[0]
if 'res_conv1_bn_s' in blob_name:
conv_name = 'conv1'
bn_s_name = blob_name
def grid_search():
dataset_name, proposal_file = get_inference_dataset(0, is_parent=False)
roidb, dataset, start_ind, end_ind, total_num_images = get_roidb_and_dataset(
dataset_name, proposal_file, None
)
num_images = len(roidb)
num_classes = cfg.MODEL.NUM_CLASSES
subinds = np.array_split(range(num_images), cfg.NUM_GPUS)
tag = 'detection'
output_dir = get_output_dir(cfg.TEST.DATASETS, training=False)
det_file = os.path.join(output_dir, 'detections.pkl')
outputs = load_object(det_file)
print(len(outputs))
all_dets_cache = outputs['all_boxes']
print(len(all_dets_cache))
all_boxes_cache = []
all_scores_cache = []
for i, entry in enumerate(roidb):
print(i)
max_det = all_dets_cache[1][i].shape[0]
print(max_det, num_classes)
boxes = np.zeros((max_det, 4), dtype=np.float32)
scores = np.zeros((max_det, num_classes), dtype=np.float32)
boxes[:] = -1
scores[:] = -1
for j in range(num_classes):
if len(all_dets_cache[j]) > 0:
pass
else:
continue
scores[:, j] = all_dets_cache[j][i][:, 4]
boxes[:, 0:4] = all_dets_cache[1][i][:, :4]
boxes = np.tile(boxes, (1, scores.shape[1]))
print(scores.shape, boxes.shape)
all_boxes_cache.append(boxes)
all_scores_cache.append(scores)
timers = defaultdict(Timer)
resultss = []
nmses = [1.0, 0.9, 0.8, 0.7, 0.6, 0.5, 0.4, 0.3, 0.2, 0.1, 0.0]
threshs = [1e-10, 1e-9, 1e-8, 1e-7, 1e-6, 1e-5, 1e-4, 1e-3, 1e-2, 1e-1]
max_per_images = [10000, 1000, 100, 10, 1]
for nms in nmses:
for thresh in threshs:
for max_per_image in max_per_images:
print("----------------------------------------------------")
print('NUM: ', nms, ' Thresh: ', thresh, ' MAX_PER_IM: ', max_per_image)
cfg.immutable(False)
cfg.TEST.NMS = nms
cfg.TEST.SCORE_THRESH = thresh
cfg.TEST.DETECTIONS_PER_IM = max_per_image
cfg.immutable(True)
all_boxes, all_segms, all_keyps = empty_results(num_classes, num_images)
for i, entry in enumerate(roidb):
# print(i)
timers['im_detect_bbox'].tic()
scores = all_scores_cache[i]
boxes = all_boxes_cache[i]
# print(scores.shape, boxes.shape)
timers['im_detect_bbox'].toc()
timers['misc_bbox'].tic()
scores, boxes, cls_boxes_i = box_results_with_nms_and_limit(scores, boxes)
timers['misc_bbox'].toc()
extend_results(i, all_boxes, cls_boxes_i)
results = task_evaluation.evaluate_all(
dataset, all_boxes, all_segms, all_keyps, output_dir
)
print(results)
print(resultss)
f = open('grid_search.csv', 'wb')
wr = csv.writer(f, dialect='excel')
wr.writerows(resultss)
def process_in_parallel(
tag, total_range_size, binary, output_dir, opts=''
):
"""Run the specified binary cfg.NUM_GPUS times in parallel, each time as a
subprocess that uses one GPU. The binary must accept the command line
arguments `--range {start} {end}` that specify a data processing range.
"""
# Snapshot the current cfg state in order to pass to the inference
# subprocesses
cfg_file = os.path.join(output_dir, '{}_range_config.yaml'.format(tag))
with open(cfg_file, 'w') as f:
envu.yaml_dump(cfg, stream=f)
subprocess_env = os.environ.copy()
processes = []
subinds = np.array_split(range(total_range_size), cfg.NUM_GPUS)
# Determine GPUs to use
cuda_visible_devices = os.environ.get('CUDA_VISIBLE_DEVICES')
if cuda_visible_devices:
gpu_inds = map(int, cuda_visible_devices.split(','))
assert -1 not in gpu_inds, \
'Hiding GPU indices using the \'-1\' index is not supported'
else:
gpu_inds = range(cfg.NUM_GPUS)
# Run the binary in cfg.NUM_GPUS subprocesses
for i, gpu_ind in enumerate(gpu_inds):
start = subinds[i][0]
end = subinds[i][-1] + 1
subprocess_env['CUDA_VISIBLE_DEVICES'] = str(gpu_ind)
cmd = '{binary} --range {start} {end} --cfg {cfg_file} NUM_GPUS 1 {opts}'
cmd = cmd.format(
binary=shlex_quote(binary),
start=int(start),
end=int(end),
cfg_file=shlex_quote(cfg_file),
opts=' '.join([shlex_quote(opt) for opt in opts])
)
logger.info('{} range command {}: {}'.format(tag, i, cmd))
if i == 0:
subprocess_stdout = subprocess.PIPE
else:
filename = os.path.join(
output_dir, '%s_range_%s_%s.stdout' % (tag, start, end)
)
subprocess_stdout = open(filename, 'w') # NOQA (close below)
p = subprocess.Popen(
cmd,
shell=True,
env=subprocess_env,
stdout=subprocess_stdout,
stderr=subprocess.STDOUT,
bufsize=1
)
processes.append((i, p, start, end, subprocess_stdout))
# Log output from inference processes and collate their results
outputs = []
for i, p, start, end, subprocess_stdout in processes:
log_subprocess_output(i, p, output_dir, tag, start, end)
if i > 0:
subprocess_stdout.close()
range_file = os.path.join(
output_dir, '%s_range_%s_%s.pkl' % (tag, start, end)
)
range_data = load_object(range_file)
outputs.append(range_data)
return outputs
def complete_stats(dataset, detections_pkl, threshold, box_idx, long_output=True):
# Load predictions and ground truths
ds = JsonDataset(dataset)
roidb = ds.get_roidb(gt=True)
dets = load_object(detections_pkl)
all_boxes = dets['all_boxes']
def id_or_index(ix, val):
if len(val) == 0:
return val
else:
return val[ix]
true_positives = 0
true_negatives = 0
false_positives = 0
false_negatives = 0
total = len(roidb)
# Iterate through all images
for ix, entry in enumerate(roidb):
cls_boxes_i = [
id_or_index(ix, cls_k_boxes) for cls_k_boxes in all_boxes
]
preds = np.array(cls_boxes_i[box_idx])
true_boxes = entry['boxes']
# Check if the images resulted in a true/false positive/negative
if (true_boxes.shape[0] == 0):
if preds.shape[0] > 0 and np.max(preds[:,4]) > threshold:
false_positives += 1
else:
true_negatives += 1
else:
if preds.shape[0] > 0 and np.max(preds[:,4]) > threshold:
true_positives += 1
else:
false_negatives += 1
# Calculate the statistics
prec = float('nan')
if true_positives+false_positives > 0:
prec = true_positives/float(true_positives+false_positives)
elif false_negatives == 0:
prec = 1.
rec = float('nan')
if true_positives+false_negatives > 0:
rec = true_positives/float(true_positives+false_negatives)
elif false_positives == 0:
rec = 1.
acc = float(true_positives+true_negatives)/total
fm = 0
if prec > 0 or rec > 0:
fm = 2.0*prec*rec/(prec+rec)
# Re-enable printing
enablePrint()
# Print results
if (long_output):
print("True positives: {}\tFalse positives: {}".format(true_positives, false_positives))
print("True negatives: {}\tFalse negatives: {}".format(true_negatives, false_negatives))
print("Total: {}".format(total))
print("Precision: " + str(prec*100))
print("Recall: " + str(rec*100))
print("F-measure: " + str(fm*100))
print("Accuracy: " + str(acc*100))
else:
print("{};{};{};{};".format(acc, prec, rec, fm))
return acc
import os
from time import strftime
from detectron.datasets.json_dataset_wsl import JsonDataset
import matplotlib as mpl
import matplotlib.pyplot as plt
from detectron.utils.io import save_object
from detectron.utils.io import load_object
if __name__ == '__main__':
net_1 = sys.argv[1]
net_2 = sys.argv[2]
pkl_1 = load_object(net_1)
pkl_2 = load_object(net_2)
if 'blobs' in pkl_1.keys():
pkl_1 = pkl_1['blobs']
if 'blobs' in pkl_2.keys():
pkl_2 = pkl_2['blobs']
print(pkl_1.keys())
print(pkl_2.keys())
for key in sorted(pkl_1.keys()):
if 'momentum' in key:
continue
if key.endswith('_b'):
def load_and_convert_PubLayNet_model(args):
model_dict = load_object(args.PubLayNet_model_file_name)
remove_momentum(model_dict)
convert_PubLayNet_blobs_to_target_blobs(model_dict)
return model_dict
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals
import numpy as np
import os
import sys
from collections import OrderedDict
from six.moves import cPickle as pickle
from detectron.utils.io import save_object
from detectron.utils.io import load_object
if __name__ == '__main__':
in_path = sys.argv[1]
out_path = sys.argv[2]
pkl_data = load_object(in_path)
pkl_data = pkl_data['blobs']
keys = pkl_data.keys()
for k in list(keys):
if 'momentum' in k:
print('delete ', k)
t = pkl_data.pop(k, None)
save_object(pkl_data, out_path)
help='Proposals file 2',
default='../output/proposals/nucoco_sw_f/eb/proposals_nucoco_train.pkl'
)
args = parser.parse_args()
return args
if __name__ == '__main__':
args = parse_args()
fig = plt.figure(figsize=(16, 6))
coco = COCO_PLUS(args.ann_file, args.imgs_dir)
# Load the proposals
proposals1 = load_object(args.proposals_file1)
proposals2 = load_object(args.proposals_file2)
num_imgs = len(proposals1['ids'])
assert len(proposals1['ids']) == len(proposals2['ids']), \
"Number of images are different in proposal files"
# print(len(proposals1['indexes']))
# print(proposals2['ids'])
# np.set_printoptions(threshold=np.nan)
# for i in range(20):
# print(proposals1['boxes'][i][0:10,:])
# print('---------------------------')
# print(proposals2['boxes'][i][0:10,:])
# input('something')
def initialize_gpu_from_weights_file(model, weights_file, gpu_id=0):
"""Initialize a network with ops on a specific GPU.
If you use CUDA_VISIBLE_DEVICES to target specific GPUs, Caffe2 will
automatically map logical GPU ids (starting from 0) to the physical GPUs
specified in CUDA_VISIBLE_DEVICES.
"""
logger.info('Loading weights from: {}'.format(weights_file))
ws_blobs = workspace.Blobs()
src_blobs = load_object(weights_file)
if 'cfg' in src_blobs:
saved_cfg = load_cfg(src_blobs['cfg'])
configure_bbox_reg_weights(model, saved_cfg)
if 'blobs' in src_blobs:
# Backwards compat--dictionary used to be only blobs, now they are
# stored under the 'blobs' key
src_blobs = src_blobs['blobs']
# Initialize weights on GPU gpu_id only
unscoped_param_names = OrderedDict() # Print these out in model order
for blob in model.params:
unscoped_param_names[c2_utils.UnscopeName(str(blob))] = True
with c2_utils.NamedCudaScope(gpu_id):
for unscoped_param_name in unscoped_param_names.keys():
if (unscoped_param_name.find(']_') >= 0
and unscoped_param_name not in src_blobs):
# Special case for sharing initialization from a pretrained
# model:
# If a blob named '_[xyz]_foo' is in model.params and not in
# the initialization blob dictionary, then load source blob
# 'foo' into destination blob '_[xyz]_foo'
src_name = unscoped_param_name[unscoped_param_name.find(']_') +
2:]
else:
src_name = unscoped_param_name
if src_name not in src_blobs:
logger.info('{:s} not found'.format(src_name))
continue
dst_name = core.ScopedName(unscoped_param_name)
has_momentum = src_name + '_momentum' in src_blobs
has_momentum_str = ' [+ momentum]' if has_momentum else ''
logger.info(
'{:s}{:} loaded from weights file into {:s}: {}'.format(
src_name, has_momentum_str, dst_name,
src_blobs[src_name].shape))
if dst_name in ws_blobs:
# If the blob is already in the workspace, make sure that it
# matches the shape of the loaded blob
ws_blob = workspace.FetchBlob(dst_name)
assert ws_blob.shape == src_blobs[src_name].shape, \
('Workspace blob {} with shape {} does not match '
'weights file shape {}').format(
src_name,
ws_blob.shape,
src_blobs[src_name].shape)
workspace.FeedBlob(
dst_name, src_blobs[src_name].astype(np.float32, copy=False))
if has_momentum:
workspace.FeedBlob(
dst_name + '_momentum',
src_blobs[src_name + '_momentum'].astype(np.float32,
copy=False))
# We preserve blobs that are in the weights file but not used by the current
# model. We load these into CPU memory under the '__preserve__/' namescope.
# These blobs will be stored when saving a model to a weights file. This
# feature allows for alternating optimization of Faster R-CNN in which blobs
# unused by one step can still be preserved forward and used to initialize
# another step.
for src_name in src_blobs.keys():
if (src_name not in unscoped_param_names
and not src_name.endswith('_momentum')
and src_blobs[src_name] is not None):
with c2_utils.CpuScope():
workspace.FeedBlob('__preserve__/{:s}'.format(src_name),
src_blobs[src_name])
logger.info(
'{:s} preserved in workspace (unused)'.format(src_name))
def load_and_convert_coco_model(args):
model_dict = load_object(args.coco_model_file_name)
remove_momentum(model_dict)
convert_coco_blobs_to_cityscape_blobs(model_dict)
return model_dict
def voc_eval(detpath,
annopath,
imagesetfile,
classname,
cachedir,
ovthresh=0.5,
use_07_metric=False,
subset_pointer=None):
"""rec, prec, ap = voc_eval(detpath,
annopath,
imagesetfile,
classname,
[ovthresh],
[use_07_metric])
Top level function that does the PASCAL VOC evaluation.
detpath: Path to detections
detpath.format(classname) should produce the detection results file.
annopath: Path to annotations
annopath.format(imagename) should be the xml annotations file.
imagesetfile: Text file containing the list of images, one image per line.
classname: Category name (duh)
cachedir: Directory for caching the annotations
[ovthresh]: Overlap threshold (default = 0.5)
[use_07_metric]: Whether to use VOC07's 11 point AP computation
(default False)
"""
# assumes detections are in detpath.format(classname)
# assumes annotations are in annopath.format(imagename)
# assumes imagesetfile is a text file with each line an image name
# cachedir caches the annotations in a pickle file
this_sub = None
if subset_pointer is not None:
this_sub = subset_pointer.subset
# first load gt
if not os.path.isdir(cachedir):
os.mkdir(cachedir)
imageset = os.path.splitext(os.path.basename(imagesetfile))[0]
cachefile = os.path.join(cachedir, imageset + '_annots.pkl')
# read list of images
with open(imagesetfile, 'r') as f:
lines = f.readlines()
imagenames = [x.strip() for x in lines]
if subset_pointer is not None:
imagenames = [n for n,taken in zip(imagenames,this_sub) if taken]
if not os.path.isfile(cachefile):
# load annots
recs = {}
for i, imagename in enumerate(imagenames):
recs[imagename] = parse_rec(annopath.format(imagename))
if i % 100 == 0:
logger.info(
'Reading annotation for {:d}/{:d}'.format(
i + 1, len(imagenames)))
# save
logger.info('Saving cached annotations to {:s}'.format(cachefile))
save_object(recs, cachefile)
else:
recs = load_object(cachefile)
# extract gt objects for this class
class_recs = {}
npos = 0
for imagename in imagenames:
R = [obj for obj in recs[imagename] if obj['name'] == classname]
bbox = np.array([x['bbox'] for x in R])
difficult = np.array([x['difficult'] for x in R]).astype(np.bool)
det = [False] * len(R)
npos = npos + sum(~difficult)
class_recs[imagename] = {'bbox': bbox,
'difficult': difficult,
'det': det}
# read dets
detfile = detpath.format(classname)
with open(detfile, 'r') as f:
lines = f.readlines()
splitlines = [x.strip().split(' ') for x in lines]
image_ids = [x[0] for x in splitlines]
confidence = np.array([float(x[1]) for x in splitlines])
BB = np.array([[float(z) for z in x[2:]] for x in splitlines])
# sort by confidence
sorted_ind = np.argsort(-confidence)
BB = BB[sorted_ind, :]
image_ids = [image_ids[x] for x in sorted_ind]
# go down dets and mark TPs and FPs
nd = len(image_ids)
tp = np.zeros(nd)
fp = np.zeros(nd)
for d in range(nd):
R = class_recs[image_ids[d]]
bb = BB[d, :].astype(float)
ovmax = -np.inf
BBGT = R['bbox'].astype(float)
if BBGT.size > 0:
# compute overlaps
# intersection
ixmin = np.maximum(BBGT[:, 0], bb[0])
iymin = np.maximum(BBGT[:, 1], bb[1])
ixmax = np.minimum(BBGT[:, 2], bb[2])
iymax = np.minimum(BBGT[:, 3], bb[3])
iw = np.maximum(ixmax - ixmin + 1., 0.)
ih = np.maximum(iymax - iymin + 1., 0.)
inters = iw * ih
# union
uni = ((bb[2] - bb[0] + 1.) * (bb[3] - bb[1] + 1.) +
(BBGT[:, 2] - BBGT[:, 0] + 1.) *
(BBGT[:, 3] - BBGT[:, 1] + 1.) - inters)
overlaps = inters / uni
ovmax = np.max(overlaps)
jmax = np.argmax(overlaps)
if ovmax > ovthresh:
if not R['difficult'][jmax]:
if not R['det'][jmax]:
tp[d] = 1.
R['det'][jmax] = 1
else:
fp[d] = 1.
else:
fp[d] = 1.
# compute precision recall
fp = np.cumsum(fp)
tp = np.cumsum(tp)
rec = tp / float(npos)
# avoid divide by zero in case the first detection matches a difficult
# ground truth
prec = tp / np.maximum(tp + fp, np.finfo(np.float64).eps)
ap = voc_ap(rec, prec, use_07_metric)
return rec, prec, ap, npos
def initialize_gpu_from_weights_file(model, weights_file, gup_id=0):
logger.info('Loading weights from: {}'.format(weights_file))
ws_blobs = workspace.Blobs()
src_blobs = load_object(weights_file)
if 'cfg' in src_blobs:
saved_cfg = load_cfg(src_blobs['cfg'])
configure_bbox_reg_weights(model, saved_cfg)
if 'blobs' in src_blobs:
src_blobs = src_blobs['blobs']
unscoped_para_names = OrderedDict()
for blob in model.params:
unscoped_para_names[c2_utils.UnscopeName(str(blob))] = True
with c2_utils.NamedCudaScope(gup_id):
for unscoped_para_name in unscoped_para_names.keys():
if (unscoped_para_name.find(']_') >= 0
and unscoped_para_name not in src_blobs):
# Special case for sharing initialization from a pretrained
# model:
# If a blob named '_[xyz]_foo' is in model.params and not in
# the initialization blob dictionary, then load source blob
# 'foo' into destination blob '_[xyz]_foo'
src_name = unscoped_para_name[unscoped_para_name.find((']_') +
2)]
else:
src_name = unscoped_para_name
if src_name not in src_blobs:
logger.info('{:s} not found.'.format(src_name))
continue
dst_name = core.ScopedName(unscoped_para_name)
has_momentum = src_name + '_momentum' in src_blobs
has_momentum_str = ' [+ momentum] ' if has_momentum else ''
logger.info(
'{:s}{:} loaded form weights file into {:s}: {}'.format(
src_name, has_momentum_str, dst_name,
src_blobs[src_name].shape))
if dst_name in ws_blobs:
# if the blob is already in the workspace, make sure that it matches
# the shape of the loaded blob
ws_blob = workspace.FetchBlob(dst_name)
assert ws_blob.shape == src_blobs[src_name].shape, \
('Workspace blob {} with shape {} does not match '
'weights file shape {}').format(
src_name,
ws_blob.shape,
src_blobs[src_name].shape)
workspace.FeedBlob(
dst_name, src_blobs[src_name].astype(np.float32, copy=False))
if has_momentum:
workspace.FeedBlob(
dst_name + '_momentum',
src_blobs[src_name + '_momentum'].astype(np.float32,
copy=False))
for src_name in src_blobs.keys():
if (src_name not in unscoped_para_names
and not src_name.endswith('_momentum')
and src_blobs[src_name] is not None):
with c2_utils.CpuScope():
workspace.FeedBlob('__presever__/{:s}'.format(src_name),
src_blobs[src_name])
logger.info(
'{:s} preserved in workspace (unused)'.format(src_name))
dest='input_model',
help='Trained model weights',
default='../output/models/R_50_C4_2x_original/model_final.pkl')
parser.add_argument(
'--output_model',
dest='output_model',
help='Ouput model weights',
default='../output/models/R_50_C4_2x_original_RW/model_final.pkl')
args = parser.parse_args()
args.input_model = os.path.abspath(args.input_model)
args.output_model = os.path.abspath(args.output_model)
return args
if __name__ == '__main__':
args = parse_args()
out_dir = os.path.dirname(args.output_model)
os.makedirs(out_dir, exist_ok=True)
wts = load_object(args.input_model)
for blob in list(wts['blobs'].keys()):
if blob.startswith('cls_score_') or blob.startswith('bbox_pred_'):
del wts['blobs'][blob]
save_object(wts, args.output_model)
#with open(args.output_model, 'wb') as f:
# pkl.dump(wts, f)
def voc_eval(detpath,
annopath,
imagesetfile,
classname,
cachedir,
ovthresh=0.5,
use_07_metric=False):
"""rec, prec, ap = voc_eval(detpath,
annopath,
imagesetfile,
classname,
[ovthresh],
[use_07_metric])
Top level function that does the PASCAL VOC evaluation.
detpath: Path to detections
detpath.format(classname) should produce the detection results file.
annopath: Path to annotations
annopath.format(imagename) should be the xml annotations file.
imagesetfile: Text file containing the list of images, one image per line.
classname: Category name (duh)
cachedir: Directory for caching the annotations
[ovthresh]: Overlap threshold (default = 0.5)
[use_07_metric]: Whether to use VOC07's 11 point AP computation
(default False)
"""
# assumes detections are in detpath.format(classname)
# assumes annotations are in annopath.format(imagename)
# assumes imagesetfile is a text file with each line an image name
# cachedir caches the annotations in a pickle file
# first load gt
if not os.path.isdir(cachedir):
os.mkdir(cachedir)
imageset = os.path.splitext(os.path.basename(imagesetfile))[0]
cachefile = os.path.join(cachedir, imageset + '_annots.pkl')
# read list of images
with open(imagesetfile, 'r') as f:
lines = f.readlines()
imagenames = [x.strip() for x in lines]
if not os.path.isfile(cachefile):
# load annots
recs = {}
for i, imagename in enumerate(imagenames):
recs[imagename] = parse_rec(annopath.format(imagename))
if i % 100 == 0:
logger.info(
'Reading annotation for {:d}/{:d}'.format(
i + 1, len(imagenames)))
# save
logger.info('Saving cached annotations to {:s}'.format(cachefile))
save_object(recs, cachefile)
else:
recs = load_object(cachefile)
# extract gt objects for this class
class_recs = {}
npos = 0
for imagename in imagenames:
R = [obj for obj in recs[imagename] if obj['name'] == classname]
bbox = np.array([x['bbox'] for x in R])
difficult = np.array([x['difficult'] for x in R]).astype(np.bool)
det = [False] * len(R)
npos = npos + sum(~difficult)
class_recs[imagename] = {'bbox': bbox,
'difficult': difficult,
'det': det}
# read dets
detfile = detpath.format(classname)
with open(detfile, 'r') as f:
lines = f.readlines()
splitlines = [x.strip().split(' ') for x in lines]
image_ids = [x[0] for x in splitlines]
confidence = np.array([float(x[1]) for x in splitlines])
BB = np.array([[float(z) for z in x[2:]] for x in splitlines])
# sort by confidence
sorted_ind = np.argsort(-confidence)
BB = BB[sorted_ind, :]
image_ids = [image_ids[x] for x in sorted_ind]
# go down dets and mark TPs and FPs
nd = len(image_ids)
tp = np.zeros(nd)
fp = np.zeros(nd)
for d in range(nd):
R = class_recs[image_ids[d]]
bb = BB[d, :].astype(float)
ovmax = -np.inf
BBGT = R['bbox'].astype(float)
if BBGT.size > 0:
# compute overlaps
# intersection
ixmin = np.maximum(BBGT[:, 0], bb[0])
iymin = np.maximum(BBGT[:, 1], bb[1])
ixmax = np.minimum(BBGT[:, 2], bb[2])
iymax = np.minimum(BBGT[:, 3], bb[3])
iw = np.maximum(ixmax - ixmin + 1., 0.)
ih = np.maximum(iymax - iymin + 1., 0.)
inters = iw * ih
# union
uni = ((bb[2] - bb[0] + 1.) * (bb[3] - bb[1] + 1.) +
(BBGT[:, 2] - BBGT[:, 0] + 1.) *
(BBGT[:, 3] - BBGT[:, 1] + 1.) - inters)
overlaps = inters / uni
ovmax = np.max(overlaps)
jmax = np.argmax(overlaps)
if ovmax > ovthresh:
if not R['difficult'][jmax]:
if not R['det'][jmax]:
tp[d] = 1.
R['det'][jmax] = 1
else:
fp[d] = 1.
else:
fp[d] = 1.
# compute precision recall
fp = np.cumsum(fp)
tp = np.cumsum(tp)
rec = tp / float(npos)
# avoid divide by zero in case the first detection matches a difficult
# ground truth
prec = tp / np.maximum(tp + fp, np.finfo(np.float64).eps)
ap = voc_ap(rec, prec, use_07_metric)
return rec, prec, ap
