def demo_net(sym, class_names, args):
# print config
print('called with args\n{}'.format(pprint.pformat(vars(args))))
# setup context
if args.gpu:
ctx = mx.gpu(int(args.gpu))
else:
ctx = mx.cpu(0)
# load single test
im_tensor, im_info, im_orig = load_test(args.image, short=args.img_short_side, max_size=args.img_long_side,
mean=args.img_pixel_means, std=args.img_pixel_stds)
# generate data batch
data_batch = generate_batch(im_tensor, im_info)
# load params
arg_params, aux_params = load_param(args.params, ctx=ctx)
# produce shape max possible
data_names = ['data', 'im_info']
label_names = None
data_shapes = [('data', (1, 3, args.img_long_side, args.img_long_side)), ('im_info', (1, 3))]
label_shapes = None
# check shapes
check_shape(sym, data_shapes, arg_params, aux_params)
# create and bind module
mod = Module(sym, data_names, label_names, context=ctx)
mod.bind(data_shapes, label_shapes, for_training=False)
mod.init_params(arg_params=arg_params, aux_params=aux_params)
# forward
forward_starts = time.time()
mod.forward(data_batch)
rois, scores, bbox_deltas = mod.get_outputs()
rois.wait_to_read()
rois = rois[:, 1:]
scores = scores[0]
bbox_deltas = bbox_deltas[0]
forward_costs = time.time() - forward_starts
print("forward costs %.4f" % (forward_costs))
im_info = im_info[0]
# decode detection
det = im_detect(rois, scores, bbox_deltas, im_info,
bbox_stds=args.rcnn_bbox_stds, nms_thresh=args.rcnn_nms_thresh,
conf_thresh=args.rcnn_conf_thresh)
# print out
for [cls, conf, x1, y1, x2, y2] in det:
if cls > 0 and conf > args.vis_thresh:
print(class_names[int(cls)], conf, [x1, y1, x2, y2])
# if vis
if args.vis:
vis_detection(im_orig, det, class_names, thresh=args.vis_thresh, prefix=args.image)
def test_net(sym, imdb, args):
# print config
logger.info('called with args\n{}'.format(pprint.pformat(vars(args))))
# setup context
ctx = mx.gpu(args.gpu)
# load testing data
test_data = TestLoader(imdb.roidb, batch_size=1, short=args.img_short_side, max_size=args.img_long_side,
mean=args.img_pixel_means, std=args.img_pixel_stds)
# load params
arg_params, aux_params = load_param(args.params, ctx=ctx)
# produce shape max possible
data_names = ['data', 'im_info']
label_names = None
data_shapes = [('data', (1, 3, args.img_long_side, args.img_long_side)), ('im_info', (1, 3))]
label_shapes = None
# check shapes
check_shape(sym, data_shapes, arg_params, aux_params)
# create and bind module
mod = Module(sym, data_names, label_names, context=ctx)
mod.bind(data_shapes, label_shapes, for_training=False)
mod.init_params(arg_params=arg_params, aux_params=aux_params)
# all detections are collected into:
# all_boxes[cls][image] = N x 5 array of detections in
# (x1, y1, x2, y2, score)
all_boxes = [[[] for _ in range(imdb.num_images)]
for _ in range(imdb.num_classes)]
# start detection
with tqdm(total=imdb.num_images) as pbar:
for i, data_batch in enumerate(test_data):
# forward
im_info = data_batch.data[1][0]
mod.forward(data_batch)
rois, scores, bbox_deltas = mod.get_outputs()
rois = rois[:, 1:]
scores = scores[0]
bbox_deltas = bbox_deltas[0]
det = im_detect(rois, scores, bbox_deltas, im_info,
bbox_stds=args.rcnn_bbox_stds, nms_thresh=args.rcnn_nms_thresh,
conf_thresh=args.rcnn_conf_thresh)
for j in range(1, imdb.num_classes):
indexes = np.where(det[:, 0] == j)[0]
all_boxes[j][i] = np.concatenate((det[:, -4:], det[:, [1]]), axis=-1)[indexes, :]
pbar.update(data_batch.data[0].shape[0])
# evaluate model
imdb.evaluate_detections(all_boxes)
def demo_net(sym, class_names, args):
# print config
print('called with args\n{}'.format(pprint.pformat(vars(args))))
# setup context
if args.gpu:
ctx = mx.gpu(int(args.gpu))
else:
ctx = mx.cpu(0)
# load single test
im_tensor, im_info, im_orig = load_test(args.image, short=args.img_short_side, max_size=args.img_long_side,
mean=args.img_pixel_means, std=args.img_pixel_stds)
# generate data batch
data_batch = generate_batch(im_tensor, im_info)
# load params
arg_params, aux_params = load_param(args.params, ctx=ctx)
# produce shape max possible
data_names = ['data', 'im_info']
label_names = None
data_shapes = [('data', (1, 3, args.img_long_side, args.img_long_side)), ('im_info', (1, 3))]
label_shapes = None
# check shapes
check_shape(sym, data_shapes, arg_params, aux_params)
# create and bind module
mod = Module(sym, data_names, label_names, context=ctx)
mod.bind(data_shapes, label_shapes, for_training=False)
mod.init_params(arg_params=arg_params, aux_params=aux_params)
# forward
mod.forward(data_batch)
rois, scores, bbox_deltas = mod.get_outputs()
rois = rois[:, 1:]
scores = scores[0]
bbox_deltas = bbox_deltas[0]
im_info = im_info[0]
# decode detection
det = im_detect(rois, scores, bbox_deltas, im_info,
bbox_stds=args.rcnn_bbox_stds, nms_thresh=args.rcnn_nms_thresh,
conf_thresh=args.rcnn_conf_thresh)
# print out
for [cls, conf, x1, y1, x2, y2] in det:
if cls > 0 and conf > args.vis_thresh:
print(class_names[int(cls)], conf, [x1, y1, x2, y2])
# if vis
if args.vis:
vis_detection(im_orig, det, class_names, thresh=args.vis_thresh)
def train_net(sym, roidb, args):
# print config
logger.info('called with args\n{}'.format(pprint.pformat(vars(args))))
# setup multi-gpu
ctx = [mx.gpu(int(i)) for i in args.gpus.split(',')]
batch_size = args.rcnn_batch_size * len(ctx)
# load training data
feat_sym = sym.get_internals()['rpn_cls_score_output']
ag = AnchorGenerator(feat_stride=args.rpn_feat_stride,
anchor_scales=args.rpn_anchor_scales,
anchor_ratios=args.rpn_anchor_ratios)
asp = AnchorSampler(allowed_border=args.rpn_allowed_border,
batch_rois=args.rpn_batch_rois,
fg_fraction=args.rpn_fg_fraction,
fg_overlap=args.rpn_fg_overlap,
bg_overlap=args.rpn_bg_overlap)
train_data = AnchorLoader(roidb,
batch_size,
args.img_short_side,
args.img_long_side,
args.img_pixel_means,
args.img_pixel_stds,
feat_sym,
ag,
asp,
shuffle=True)
# produce shape max possible
_, out_shape, _ = feat_sym.infer_shape(data=(1, 3, args.img_long_side,
args.img_long_side))
feat_height, feat_width = out_shape[0][-2:]
rpn_num_anchors = len(args.rpn_anchor_scales) * len(args.rpn_anchor_ratios)
data_names = ['data', 'im_info', 'gt_boxes']
label_names = ['label', 'bbox_target', 'bbox_weight']
data_shapes = [('data', (batch_size, 3,
args.img_long_side, args.img_long_side)),
('im_info', (batch_size, 3)),
('gt_boxes', (batch_size, 100, 5))]
label_shapes = [('label', (batch_size, 1, rpn_num_anchors * feat_height,
feat_width)),
('bbox_target', (batch_size, 4 * rpn_num_anchors,
feat_height, feat_width)),
('bbox_weight', (batch_size, 4 * rpn_num_anchors,
feat_height, feat_width))]
# print shapes
data_shape_dict, out_shape_dict = infer_data_shape(
sym, data_shapes + label_shapes)
logger.info('max input shape\n%s' % pprint.pformat(data_shape_dict))
logger.info('max output shape\n%s' % pprint.pformat(out_shape_dict))
# load and initialize params
if args.resume:
arg_params, aux_params = load_param(args.resume)
else:
arg_params, aux_params = load_param(args.pretrained)
arg_params, aux_params = initialize_frcnn(sym, data_shapes, arg_params,
aux_params)
# check parameter shapes
check_shape(sym, data_shapes + label_shapes, arg_params, aux_params)
# check fixed params
fixed_param_names = get_fixed_params(sym, args.net_fixed_params)
logger.info('locking params\n%s' % pprint.pformat(fixed_param_names))
# metric
rpn_eval_metric = RPNAccMetric()
rpn_cls_metric = RPNLogLossMetric()
rpn_bbox_metric = RPNL1LossMetric()
eval_metric = RCNNAccMetric()
cls_metric = RCNNLogLossMetric()
bbox_metric = RCNNL1LossMetric()
eval_metrics = mx.metric.CompositeEvalMetric()
for child_metric in [
rpn_eval_metric, rpn_cls_metric, rpn_bbox_metric, eval_metric,
cls_metric, bbox_metric
]:
eval_metrics.add(child_metric)
# callback
batch_end_callback = mx.callback.Speedometer(batch_size,
frequent=args.log_interval,
auto_reset=False)
epoch_end_callback = mx.callback.do_checkpoint(args.save_prefix)
# learning schedule
base_lr = args.lr
lr_factor = 0.1
lr_epoch = [int(epoch) for epoch in args.lr_decay_epoch.split(',')]
lr_epoch_diff = [
epoch - args.start_epoch for epoch in lr_epoch
if epoch > args.start_epoch
]
lr = base_lr * (lr_factor**(len(lr_epoch) - len(lr_epoch_diff)))
lr_iters = [
int(epoch * len(roidb) / batch_size) for epoch in lr_epoch_diff
]
logger.info('lr %f lr_epoch_diff %s lr_iters %s' %
(lr, lr_epoch_diff, lr_iters))
lr_scheduler = mx.lr_scheduler.MultiFactorScheduler(lr_iters, lr_factor)
# optimizer
optimizer_params = {
'momentum': 0.9,
'wd': 0.0005,
'learning_rate': lr,
'lr_scheduler': lr_scheduler,
'rescale_grad': (1.0 / batch_size),
'clip_gradient': 5
}
# train
mod = Module(sym,
data_names=data_names,
label_names=label_names,
logger=logger,
context=ctx,
work_load_list=None,
fixed_param_names=fixed_param_names)
mod.fit(train_data,
eval_metric=eval_metrics,
epoch_end_callback=epoch_end_callback,
batch_end_callback=batch_end_callback,
kvstore='device',
optimizer='sgd',
optimizer_params=optimizer_params,
arg_params=arg_params,
aux_params=aux_params,
begin_epoch=args.start_epoch,
num_epoch=args.epochs)
def train_net(sym, roidb, args):
# print config
logger.info('called with args\n{}'.format(pprint.pformat(vars(args))))
# setup multi-gpu
ctx = [mx.gpu(int(i)) for i in args.gpus.split(',')]
batch_size = args.rcnn_batch_size * len(ctx)
# load training data
feat_sym = sym.get_internals()['rpn_cls_score_output']
ag = AnchorGenerator(feat_stride=args.rpn_feat_stride,
anchor_scales=args.rpn_anchor_scales, anchor_ratios=args.rpn_anchor_ratios)
asp = AnchorSampler(allowed_border=args.rpn_allowed_border, batch_rois=args.rpn_batch_rois,
fg_fraction=args.rpn_fg_fraction, fg_overlap=args.rpn_fg_overlap,
bg_overlap=args.rpn_bg_overlap)
train_data = AnchorLoader(roidb, batch_size, args.img_short_side, args.img_long_side,
args.img_pixel_means, args.img_pixel_stds, feat_sym, ag, asp, shuffle=True)
# produce shape max possible
_, out_shape, _ = feat_sym.infer_shape(data=(1, 3, args.img_long_side, args.img_long_side))
feat_height, feat_width = out_shape[0][-2:]
rpn_num_anchors = len(args.rpn_anchor_scales) * len(args.rpn_anchor_ratios)
data_names = ['data', 'im_info', 'gt_boxes']
label_names = ['label', 'bbox_target', 'bbox_weight']
data_shapes = [('data', (batch_size, 3, args.img_long_side, args.img_long_side)),
('im_info', (batch_size, 3)),
('gt_boxes', (batch_size, 100, 5))]
label_shapes = [('label', (batch_size, 1, rpn_num_anchors * feat_height, feat_width)),
('bbox_target', (batch_size, 4 * rpn_num_anchors, feat_height, feat_width)),
('bbox_weight', (batch_size, 4 * rpn_num_anchors, feat_height, feat_width))]
# print shapes
data_shape_dict, out_shape_dict = infer_data_shape(sym, data_shapes + label_shapes)
logger.info('max input shape\n%s' % pprint.pformat(data_shape_dict))
logger.info('max output shape\n%s' % pprint.pformat(out_shape_dict))
# load and initialize params
if args.resume:
arg_params, aux_params = load_param(args.resume)
else:
arg_params, aux_params = load_param(args.pretrained)
arg_params, aux_params = initialize_frcnn(sym, data_shapes, arg_params, aux_params)
# check parameter shapes
check_shape(sym, data_shapes + label_shapes, arg_params, aux_params)
# check fixed params
fixed_param_names = get_fixed_params(sym, args.net_fixed_params)
logger.info('locking params\n%s' % pprint.pformat(fixed_param_names))
# metric
rpn_eval_metric = RPNAccMetric()
rpn_cls_metric = RPNLogLossMetric()
rpn_bbox_metric = RPNL1LossMetric()
eval_metric = RCNNAccMetric()
cls_metric = RCNNLogLossMetric()
bbox_metric = RCNNL1LossMetric()
eval_metrics = mx.metric.CompositeEvalMetric()
for child_metric in [rpn_eval_metric, rpn_cls_metric, rpn_bbox_metric, eval_metric, cls_metric, bbox_metric]:
eval_metrics.add(child_metric)
# callback
batch_end_callback = mx.callback.Speedometer(batch_size, frequent=args.log_interval, auto_reset=False)
epoch_end_callback = mx.callback.do_checkpoint(args.save_prefix)
# learning schedule
base_lr = args.lr
lr_factor = 0.1
lr_epoch = [int(epoch) for epoch in args.lr_decay_epoch.split(',')]
lr_epoch_diff = [epoch - args.start_epoch for epoch in lr_epoch if epoch > args.start_epoch]
lr = base_lr * (lr_factor ** (len(lr_epoch) - len(lr_epoch_diff)))
lr_iters = [int(epoch * len(roidb) / batch_size) for epoch in lr_epoch_diff]
logger.info('lr %f lr_epoch_diff %s lr_iters %s' % (lr, lr_epoch_diff, lr_iters))
lr_scheduler = mx.lr_scheduler.MultiFactorScheduler(lr_iters, lr_factor)
# optimizer
optimizer_params = {'momentum': 0.9,
'wd': 0.0005,
'learning_rate': lr,
'lr_scheduler': lr_scheduler,
'rescale_grad': (1.0 / batch_size),
'clip_gradient': 5}
# train
mod = Module(sym, data_names=data_names, label_names=label_names,
logger=logger, context=ctx, work_load_list=None,
fixed_param_names=fixed_param_names)
mod.fit(train_data, eval_metric=eval_metrics, epoch_end_callback=epoch_end_callback,
batch_end_callback=batch_end_callback, kvstore='device',
optimizer='sgd', optimizer_params=optimizer_params,
arg_params=arg_params, aux_params=aux_params, begin_epoch=args.start_epoch, num_epoch=args.epochs)
def test_net(sym, imdb, args):
# print config
logger.info('called with args\n{}'.format(pprint.pformat(vars(args))))
# setup context
ctx = mx.gpu(args.gpu)
# load testing data
test_data = TestLoader(imdb.roidb,
batch_size=1,
short=args.img_short_side,
max_size=args.img_long_side,
mean=args.img_pixel_means,
std=args.img_pixel_stds)
# load params
arg_params, aux_params = load_param(args.params, ctx=ctx)
# produce shape max possible
data_names = ['data', 'im_info']
label_names = None
data_shapes = [('data', (1, 3, args.img_long_side, args.img_long_side)),
('im_info', (1, 3))]
label_shapes = None
# check shapes
check_shape(sym, data_shapes, arg_params, aux_params)
# create and bind module
mod = Module(sym, data_names, label_names, context=ctx)
mod.bind(data_shapes, label_shapes, for_training=False)
mod.init_params(arg_params=arg_params, aux_params=aux_params)
# all detections are collected into:
# all_boxes[cls][image] = N x 5 array of detections in
# (x1, y1, x2, y2, score)
all_boxes = [[[] for _ in range(imdb.num_images)]
for _ in range(imdb.num_classes)]
# start detection
with tqdm(total=imdb.num_images) as pbar:
for i, data_batch in enumerate(test_data):
# forward
im_info = data_batch.data[1][0]
mod.forward(data_batch)
rois, scores, bbox_deltas = mod.get_outputs()
rois = rois[:, 1:]
scores = scores[0]
bbox_deltas = bbox_deltas[0]
det = im_detect(rois,
scores,
bbox_deltas,
im_info,
bbox_stds=args.rcnn_bbox_stds,
nms_thresh=args.rcnn_nms_thresh,
conf_thresh=args.rcnn_conf_thresh)
for j in range(1, imdb.num_classes):
indexes = np.where(det[:, 0] == j)[0]
all_boxes[j][i] = np.concatenate((det[:, -4:], det[:, [1]]),
axis=-1)[indexes, :]
pbar.update(data_batch.data[0].shape[0])
# evaluate model
imdb.evaluate_detections(all_boxes)
def demo_net(sym, class_names, args, result_path):
# print config
print('called with args\n{}'.format(pprint.pformat(vars(args))))
# setup context
if args.gpu:
ctx = mx.gpu(int(args.gpu))
else:
ctx = mx.cpu(0)
# load single test
im_tensor, im_info, im_orig = load_test(args.image,
short=args.img_short_side,
max_size=args.img_long_side,
mean=args.img_pixel_means,
std=args.img_pixel_stds)
# generate data batch
data_batch = generate_batch(im_tensor, im_info)
# load params
arg_params, aux_params = load_param(args.params, ctx=ctx)
# produce shape max possible
data_names = ['data', 'im_info']
label_names = None
data_shapes = [('data', (1, 3, args.img_long_side, args.img_long_side)),
('im_info', (1, 3))]
label_shapes = None
# check shapes
check_shape(sym, data_shapes, arg_params, aux_params)
# create and bind module
mod = Module(sym, data_names, label_names, context=ctx)
mod.bind(data_shapes, label_shapes, for_training=False)
mod.init_params(arg_params=arg_params, aux_params=aux_params)
# forward
forward_starts = time.time()
mod.forward(data_batch)
rois, scores, bbox_deltas = mod.get_outputs()
rois.wait_to_read()
rois = rois[:, 1:]
scores = scores[0]
bbox_deltas = bbox_deltas[0]
forward_costs = time.time() - forward_starts
print("forward costs %.4f" % (forward_costs))
im_info = im_info[0]
# decode detection
det = im_detect(rois,
scores,
bbox_deltas,
im_info,
bbox_stds=args.rcnn_bbox_stds,
nms_thresh=args.rcnn_nms_thresh,
conf_thresh=args.rcnn_conf_thresh)
fieldnames = ['name', 'coordinate']
if result_path.exists():
csvfile = result_path.open("a")
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
else:
csvfile = result_path.open("w+")
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
writer.writeheader()
img_name = Path(args.image).name
bbox_str = ''
for [cls, conf, x1, y1, x2, y2] in det:
if cls > 0 and conf > args.vis_thresh:
print(class_names[int(cls)], conf, [x1, y1, x2, y2])
bbox_str += "%d_%d_%d_%d;" % (int(x1), int(y1), int(x2 - x1),
int(y2 - y1))
writer.writerow({'name': img_name, 'coordinate': bbox_str[:-1]})
csvfile.close()
print("detect image %s" % img_name)
# if vis
if args.vis:
vis_detection(im_orig,
det,
class_names,
thresh=args.vis_thresh,
prefix=args.image)