def get_model(opt, pretrained=None, trn=True, weights_FIMs=None, alpha=1.):
'''Getting model and initializing.
Args:
pretrained, None or path to pretrained model weights
trn, True for training and False for evaluating'''
# Model structure
model = Darknet(opt.model_config_path, opt.img_size, weights_FIMs, alpha)
print(model)
# Initialize
model.apply(weights_init_normal)
# Pretrained or not
coco_weights = True if pretrained == 'weights/yolov3.weights' else False
try:
model.load_weights(pretrained, use_coco=coco_weights)
except TypeError:
pass
# Cuda or not
if opt.cuda:
model = model.cuda()
cudnn.benchmark = True
# Mode = train or eval
if trn:
model.train()
else:
model.eval()
return model
def eval_flowchart(init_style, para_part, reg, alpha, ablation_type):
'''Main body for evaluation'''
args = all_args()
# Storage path 'eval/'
os.makedirs(args.save_folder, exist_ok=True)
# Dataset
dataset = get_data(args)
# Load net
net = Darknet(args.model_config_path, img_size=args.img_size)
# Visdom
viz, epoch_aps = init_viz(args, init_style, para_part, reg, alpha, dataset)
# Evaluate
ckpt_path, ckpts = get_ckpts(args, init_style, para_part, reg, alpha,
ablation_type)
mAP_max = 0
for ckpt_idx, ckpt in enumerate(ckpts):
# sample one for hyperparameter adjustment
if ckpt_idx < 120:
continue
# Make output dir
dir_name = '_'.join([
ablation_type, args.arc, args.dataset, args.set_type, init_style,
para_part, reg,
str(alpha), ckpt,
str(ckpt_idx)
])
output_dir = get_output_dir(args.save_folder, dir_name)
# Load weight
args.weight_path = os.path.join(ckpt_path, ckpt)
# assert os.path.isfile(args.weight_path)
try:
net.load_weights(args.weight_path)
except FileNotFoundError as err:
print(err)
# Cuda or not
if args.cuda:
net = net.cuda()
cudnn.benchmark = True
net.eval()
print('Finished loading model!')
# Evaluation, use_07_eval False
aps, mAP = test_net(output_dir,
net,
args.cuda,
dataset,
args.score_thres,
args.nms_thres,
use_07_eval=False,
iou_thres=args.iou_thres)
# If not greater than before, delete
if mAP_max >= mAP:
rmtree(output_dir)
else:
mAP_max = mAP
# Visdom
update_vis(viz, epoch_aps, ckpt_idx + 1, *aps, mAP)
def main():
img_size = 512 # 必须是32的整数倍 [416, 512, 608]
cfg = "cfg/yolov3-spp.cfg"
weights = "weights/yolov3-spp-ultralytics-{}.pt".format(img_size)
img_path = "test.jpg"
input_size = (img_size, img_size)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = Darknet(cfg, img_size)
model.load_state_dict(torch.load(weights, map_location=device)["model"])
model.to(device)
model.eval()
# init
img = torch.zeros((1, 3, img_size, img_size), device=device)
model(img)
img_o = cv2.imread(img_path) # BGR
assert img_o is not None, "Image Not Found " + img_path
img = img_utils.letterbox(img_o, new_shape=input_size, auto=True, color=(0, 0, 0))[0]
# Convert
img = img[:, :, ::-1].transpose(2, 0, 1) # BGR to RGB, to 3x416x416
img = np.ascontiguousarray(img)
img = torch.from_numpy(img).to(device).float()
img /= 255.0 # scale (0, 255) to (0, 1)
img = img.unsqueeze(0) # add batch dimension
t1 = torch_utils.time_synchronized()
pred = model(img)[0] # only get inference result
t2 = torch_utils.time_synchronized()
print(t2 - t1)
pred = utils.non_max_suppression(pred, conf_thres=0.3, iou_thres=0.6, multi_label=True)[0]
t3 = time.time()
print(t3 - t2)
# process detections
pred[:, :4] = utils.scale_coords(img.shape[2:], pred[:, :4], img_o.shape).round()
print(pred.shape)
bboxes = pred[:, :4].detach().cpu().numpy()
scores = pred[:, 4].detach().cpu().numpy()
classes = pred[:, 5].detach().cpu().numpy().astype(np.int) + 1
category_index = dict([(i + 1, str(i + 1)) for i in range(90)])
img_o = draw_box(img_o[:, :, ::-1], bboxes, classes, scores, category_index)
plt.imshow(img_o)
plt.show()
img_o.save("test_result.jpg")
def setup_detector(opt):
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
weights_path = os.path.join(
opt.weights_path,
"weights_RADAR.pth" if opt.radar else "weights_LIDAR.pth")
# Set up model
model = Darknet(opt.model_def, img_size=cnf.BEV_WIDTH).to(device)
# Load checkpoint weights
model.load_state_dict(torch.load(weights_path, map_location=device))
# Eval mode
model.eval()
return model
def Load_Yolo(device):
#Load Darknet
yolo_model_def = os.path.join(yolo_path, 'config/yolov3-tiny.cfg')
yolo_img_size = 416
yolo_weights_path = os.path.join(yolo_path, 'weights/yolov3-tiny.weights')
model = Darknet(yolo_model_def, img_size=yolo_img_size).to(device)
if yolo_weights_path.endswith(".weights"):
# Load darknet weights
model.load_darknet_weights(yolo_weights_path)
else:
# Load checkpoint weights
model.load_state_dict(torch.load(yolo_weights_path))
model.eval() # Set in evaluation mode
return model
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--config_file", default="runs/config.json")
parser.add_argument("--output_dir", default='output')
parser.add_argument("--model_checkpoint")
args = parser.parse_args()
with open(args.config_file) as config_buffer:
config = json.loads(config_buffer.read())
val_loader_dict = dict()
for i, dataset_config in enumerate(config['val']["datasets"]):
val_dataset = VOCDetection(
img_dir=dataset_config["image_folder"],
annotation_dir=dataset_config["annot_folder"],
cache_dir=dataset_config["cache_dir"],
split_file=dataset_config['split_file'],
img_size=config['model']['input_size'],
filter_labels=config['model']['labels'],
multiscale=False,
augment=False)
val_dataset.name = dataset_config.get('name')
val_loader = DataLoader(val_dataset,
batch_size=config["val"]["batch_size"],
collate_fn=val_dataset.collate_fn,
shuffle=True)
dataset_name = val_dataset.name if val_dataset.name else f"Dataset #{i}"
val_loader_dict[dataset_name] = val_loader
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = Darknet(config["model"]["config"]).to(device)
model.load_state_dict(torch.load(args.model_checkpoint))
model.eval()
result_dict = evaluate(model, val_loader_dict, config["val"])
for name, results in result_dict.items():
output_str = f"{name} evaluation results:\n" \
f"precision-{results['precision']},\n" \
f"recall-{results['recall']},\n" \
f"AP-{results['AP']},\n" \
f"F1-{results['F1']},\n" \
f"ap_class-{results['AP_class']}"
print(output_str)
class Detector(object):
def __init__(self, device, model_def, load_path, reg_threshold,
cls_threshold, nms_threshold, image_size):
self.image_size = image_size
self.model = Darknet(model_def, img_size=self.image_size).to(device)
# TODO
# change device to GPU
self.model.load_state_dict(torch.load(load_path, map_location='cpu'))
self.model.eval()
self.reg_threshold = reg_threshold
self.cls_threshold = cls_threshold
self.nms_threshold = nms_threshold
self.device = device
@torch.no_grad()
def __call__(self, image):
original_size = image.shape[:2]
tensor = torch.from_numpy(image).to(self.device).permute(2, 0, 1)
tensor = tensor.contiguous().float().div_(255)
_, h, w = tensor.shape
dim_diff = np.abs(h - w)
pad1, pad2 = dim_diff // 2, dim_diff - dim_diff // 2
pad = (0, 0, pad1, pad2) if h <= w else (pad1, pad2, 0, 0)
tensor = f.pad(tensor, pad, "constant", value=0)
tensor = f.interpolate(tensor.unsqueeze(0),
size=self.image_size,
mode="nearest").squeeze_(0)
result = self.model(tensor.unsqueeze_(0))
detection = non_max_suppression(result, self.reg_threshold,
self.nms_threshold)[0]
if detection is not None:
detection = detection[detection[:, -2] > self.cls_threshold]
detection = rescale_boxes(detection, self.image_size,
original_size)
else:
print("detection result is None")
return detection
def channels_select(prune_cfg, data, origin_model, aux_util, device,
data_loader, select_layer, pruned_rate):
with open(progress_result, 'a') as f:
f.write(('\n' + '%10s' * 9 + '\n') %
('Stage', 'Change', 'MSELoss', 'AuxLoss', 'Total', 'P', 'R',
'[email protected]', 'F1'))
logger.info(('%10s' * 6) %
('Stage', 'Channels', 'Batch', 'MSELoss', 'AuxLoss', 'Total'))
batch_size = data_loader.batch_size
img_size = data_loader.dataset.img_size
accumulate = 64 // batch_size
hook_util = HookUtils()
handles = []
n_iter = math.floor(500 / batch_size)
pruning_model = Darknet(prune_cfg,
img_size=(img_size, img_size)).to(device)
chkpt = torch.load(progress_chkpt, map_location=device)
pruning_model.load_state_dict(chkpt['model'], strict=True)
aux_in_layer = aux_util.conv_layer_dict[select_layer]
aux_model = aux_util.creat_aux_model(aux_in_layer)
aux_model.to(device)
aux_model.load_state_dict(chkpt['aux_in{}'.format(aux_in_layer)],
strict=True)
aux_loss_scalar = max(0.01, pow((int(aux_in_layer) + 1) / 75, 2))
del chkpt
solve_sub_problem_optimizer = optim.SGD(
pruning_model.module_list[int(aux_in_layer)].MaskConv2d.parameters(),
lr=hyp['lr0'],
momentum=hyp['momentum'])
for name, child in origin_model.module_list.named_children():
if name == aux_in_layer:
handles.append(
child.register_forward_hook(hook_util.hook_origin_output))
if name == select_layer:
handles.append(
child.register_forward_hook(hook_util.hook_origin_output))
for name, child in pruning_model.module_list.named_children():
if name == aux_in_layer:
handles.append(
child.register_forward_hook(hook_util.hook_prune_output))
if name == select_layer:
handles.append(
child.register_forward_hook(hook_util.hook_prune_output))
if device.type != 'cpu' and torch.cuda.device_count() > 1:
origin_model = torch.nn.parallel.DistributedDataParallel(
origin_model, find_unused_parameters=True)
origin_model.yolo_layers = origin_model.module.yolo_layers
pruning_model = torch.nn.parallel.DistributedDataParallel(
pruning_model, find_unused_parameters=True)
pruning_model.yolo_layers = pruning_model.module.yolo_layers
retain_channels_num = math.floor(
aux_util.layer_info[select_layer]["in_channels"] * (1 - pruned_rate))
pruning_model.nc = 80
pruning_model.hyp = hyp
pruning_model.arc = 'default'
pruning_model.eval()
aux_model.eval()
MSE = nn.MSELoss(reduction='mean')
mloss = torch.zeros(3).to(device)
for i_k in range(retain_channels_num):
data_iter = iter(data_loader)
pbar = tqdm(range(n_iter), total=n_iter)
print(('\n' + '%10s' * 6) %
('Stage', 'gpu_mem', 'channels', 'MSELoss', 'AuxLoss', 'Total'))
for i in pbar:
imgs, targets, _, _ = data_iter.next()
if len(targets) == 0:
continue
imgs = imgs.to(device).float(
) / 255.0 # uint8 to float32, 0 - 255 to 0.0 - 1.0
targets = targets.to(device)
with torch.no_grad():
_ = origin_model(imgs)
_, pruning_pred = pruning_model(imgs)
pruning_loss, _ = compute_loss(pruning_pred, targets,
pruning_model)
hook_util.cat_to_gpu0()
mse_loss = torch.zeros(1, device=device)
aux_pred = aux_model(hook_util.prune_features['gpu0'][1], targets)
aux_loss = compute_loss_for_DCP(aux_pred, targets)
mse_loss += MSE(hook_util.prune_features['gpu0'][0],
hook_util.origin_features['gpu0'][0])
loss = hyp['joint_loss'] * mse_loss + aux_loss + 0 * pruning_loss
loss.backward()
mem = torch.cuda.memory_cached() / 1E9 if torch.cuda.is_available(
) else 0
s = ('%10s' * 3 + '%10.3g' * 3) % (
'Prune ' + select_layer, '%.3gG' % mem, '%g/%g' %
(i_k, retain_channels_num), hyp['joint_loss'] * mse_loss,
aux_loss, loss)
pbar.set_description(s)
# if (i + 1) % 10 == 0:
# logger.info(('%10s' * 3 + '%10.3g' * 3) %
# ('Prune' + select_layer, str(i_k), '%g/%g' % (i, n_iter), hyp['joint_loss'] * mse_loss,
# aux_loss, loss))
hook_util.clean_hook_out()
grad = pruning_model.module.module_list[int(
select_layer)].MaskConv2d.weight.grad.detach()**2
grad = grad.sum((2, 3)).sqrt().sum(0)
if i_k == 0:
pruning_model.module.module_list[int(
select_layer)].MaskConv2d.selected_channels_mask[:] = 1e-5
if select_layer in aux_util.sync_guide.keys():
sync_layer = aux_util.sync_guide[select_layer]
pruning_model.module.module_list[int(
sync_layer)].MaskConv2d.selected_channels_mask[(
-1 * aux_util.layer_info[select_layer]["in_channels"]
):] = 1e-5
selected_channels_mask = pruning_model.module.module_list[int(
select_layer)].MaskConv2d.selected_channels_mask
_, indices = torch.topk(grad * (1 - selected_channels_mask), 1)
pruning_model.module.module_list[int(
select_layer)].MaskConv2d.selected_channels_mask[indices] = 1
if select_layer in aux_util.sync_guide.keys():
pruning_model.module.module_list[int(
sync_layer)].MaskConv2d.selected_channels_mask[-(
aux_util.layer_info[select_layer]["in_channels"] -
indices)] = 1
pruning_model.zero_grad()
pbar = tqdm(range(n_iter), total=n_iter)
print(('\n' + '%10s' * 6) %
('Stage', 'gpu_mem', 'channels', 'MSELoss', 'AuxLoss', 'Total'))
for i in pbar:
imgs, targets, _, _ = data_iter.next()
if len(targets) == 0:
continue
imgs = imgs.to(device).float(
) / 255.0 # uint8 to float32, 0 - 255 to 0.0 - 1.0
targets = targets.to(device)
with torch.no_grad():
_ = origin_model(imgs)
_, pruning_pred = pruning_model(imgs)
pruning_loss, _ = compute_loss(pruning_pred, targets,
pruning_model)
hook_util.cat_to_gpu0()
mse_loss = torch.zeros(1, device=device)
aux_pred = aux_model(hook_util.prune_features['gpu0'][1], targets)
aux_loss = compute_loss_for_DCP(aux_pred, targets)
mse_loss += MSE(hook_util.prune_features['gpu0'][0],
hook_util.origin_features['gpu0'][0])
loss = hyp[
'joint_loss'] * mse_loss + aux_loss_scalar * aux_loss + 0 * pruning_loss
loss.backward()
if i % accumulate == 0:
solve_sub_problem_optimizer.step()
solve_sub_problem_optimizer.zero_grad()
mem = torch.cuda.memory_cached() / 1E9 if torch.cuda.is_available(
) else 0
mloss = (mloss * i +
torch.cat([hyp['joint_loss'] * mse_loss, aux_loss, loss
]).detach()) / (i + 1)
s = ('%10s' * 3 + '%10.3g' * 3) % (
'SubProm ' + select_layer, '%.3gG' % mem, '%g/%g' %
(i_k, retain_channels_num), *mloss)
pbar.set_description(s)
if (i + 1) % n_iter == 0:
logger.info(('%10s' * 3 + '%10.3g' * 3) %
('SubPro' + select_layer, str(i_k), '%g/%g' %
(i, n_iter), *mloss))
hook_util.clean_hook_out()
for handle in handles:
handle.remove()
greedy_indices = pruning_model.module.module_list[int(
select_layer)].MaskConv2d.selected_channels_mask < 1
pruning_model.module.module_list[int(
select_layer)].MaskConv2d.selected_channels_mask[greedy_indices] = 0
res, _ = test.test(prune_cfg,
data,
batch_size=batch_size * 2,
img_size=416,
model=pruning_model,
conf_thres=0.1,
iou_thres=0.5,
save_json=False,
dataloader=None)
chkpt = torch.load(progress_chkpt, map_location=device)
chkpt['current_layer'] = aux_util.next_prune_layer(select_layer)
chkpt['epoch'] = -1
chkpt['model'] = pruning_model.module.state_dict() if type(
pruning_model
) is nn.parallel.DistributedDataParallel else pruning_model.state_dict()
chkpt['optimizer'] = None
torch.save(chkpt, progress_chkpt)
torch.save(chkpt, last)
del chkpt
with open(progress_result, 'a') as f:
f.write(('%10s' * 2 + '%10.3g' * 7) %
('Pruning ' + select_layer,
str(aux_util.layer_info[select_layer]['in_channels']) + '->' +
str(retain_channels_num), *mloss, *res[:4]) + '\n')
torch.cuda.empty_cache()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
os.makedirs("output", exist_ok=True)
# Set up model
model = Darknet(opt.model_def, img_size=opt.img_size).to(device)
if opt.weights_path.endswith(".weights"):
# Load darknet weights
model.load_darknet_weights(opt.weights_path)
else:
# Load checkpoint weights
model.load_state_dict(torch.load(opt.weights_path)["model_state_dict"])
model.eval() # Set in evaluation mode
dataloader = DataLoader(
ImageFolder(opt.image_folder, img_size=opt.img_size),
batch_size=opt.batch_size,
shuffle=False,
num_workers=opt.n_cpu,
)
classes = load_classes(opt.class_path) # Extracts class labels from file
Tensor = torch.cuda.FloatTensor if torch.cuda.is_available(
) else torch.FloatTensor
imgs = [] # Stores image paths
img_detections = [] # Stores detections for each image index
class YOLOV4(object):
if CWD == THIS_DIR:
_defaults = {
"weights": "weights/yolov4.weights",
"config": "cfg/yolov4.cfg",
"classes_path": 'cfg/coco.names',
"thresh": 0.5,
"nms_thresh": 0.4,
"model_image_size": (608, 608),
"max_batch_size": 4,
"half": True
}
else:
_defaults = {
"weights": "yolov4_pytorch/weights/yolov4.weights",
"config": "yolov4_pytorch/cfg/yolov4.cfg",
"classes_path": 'yolov4_pytorch/cfg/coco.names',
"thresh": 0.5,
"nms_thresh": 0.4,
"model_image_size": (608, 608),
"max_batch_size": 4,
"half": True
}
def __init__(self, bgr=True, gpu_device=0, **kwargs):
self.__dict__.update(self._defaults) # set up default values
# for portability between keras-yolo3/yolo.py and this
if 'model_path' in kwargs:
kwargs['weights'] = kwargs['model_path']
if 'score' in kwargs:
kwargs['thresh'] = kwargs['score']
self.__dict__.update(kwargs) # update with user overrides
self.class_names = self._get_class()
self.model = Darknet(self.config)
self.model.load_darknet_weights(self.weights)
self.device = gpu_device
self.model.cuda(self.device)
self.model.eval()
self.bgr = bgr
if self.half:
self.model.half()
# warm up
self._detect([np.zeros((10, 10, 3), dtype=np.uint8)])
print('Warmed up!')
def _get_class(self):
with open(self.classes_path) as f:
class_names = f.readlines()
class_names = [c.strip() for c in class_names]
return class_names
def _detect(self, list_of_imgs):
inputs = []
for img in list_of_imgs:
if self.bgr:
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
# print('bgr: {}'.format(img.shape))
# print('size: {}'.format(self.model_image_size))
image = cv2.resize(img, self.model_image_size)
# print('image: {}'.format(image.shape))
inputs.append(np.expand_dims(np.array(image), axis=0))
images = np.concatenate(inputs, 0)
# print('images: {}'.format(images.shape))
images = torch.from_numpy(images.transpose(0, 3, 1,
2)).float().div(255.0)
images = images.cuda()
images = torch.autograd.Variable(images)
if self.half:
images = images.half()
batches = []
for i in range(0, len(images), self.max_batch_size):
these_imgs = images[i:i + self.max_batch_size]
batches.append(these_imgs)
feature_list = None
with torch.no_grad():
for batch in batches:
img = batch.cuda(self.device)
features = self.model(img)
if feature_list is None:
feature_list = features
else:
feature_list = torch.cat((feature_list, features))
# feature_list: (batch, height * width * num_anchors, 5 + num_classes)
return feature_list
def detect_get_box_in(self,
images,
box_format='ltrb',
classes=None,
buffer_ratio=0.0):
'''
Params
------
- images : ndarray-like or list of ndarray-like
- box_format : string of characters representing format order, where l = left, t = top, r = right, b = bottom, w = width and h = height
- classes : list of string, classes to focus on
- buffer : float, proportion of buffer around the width and height of the bounding box
Returns
-------
if one ndarray given, this returns a list (boxes in one image) of tuple (box_infos, score, predicted_class),
else if a list of ndarray given, this return a list (batch) containing the former as the elements,
where,
- box_infos : list of floats in the given box format
- score : float, confidence level of prediction
- predicted_class : string
'''
single = False
if isinstance(images, list):
if len(images) <= 0:
return None
else:
assert all(isinstance(im, np.ndarray) for im in images)
elif isinstance(images, np.ndarray):
images = [images]
single = True
res = self._detect(images)
frame_shapes = [image.shape for image in images]
all_dets = self._postprocess(res,
shapes=frame_shapes,
box_format=box_format,
classes=classes,
buffer_ratio=buffer_ratio)
if single:
return all_dets[0]
else:
return all_dets
def get_detections_dict(self, frames, classes=None, buffer_ratio=0.0):
'''
Params: frames, list of ndarray-like
Returns: detections, list of dict, whose key: label, confidence, t, l, w, h
'''
if frames is None or len(frames) == 0:
return None
all_dets = self.detect_get_box_in(frames,
box_format='tlbrwh',
classes=classes,
buffer_ratio=buffer_ratio)
all_detections = []
for dets in all_dets:
detections = []
for tlbrwh, confidence, label in dets:
top, left, bot, right, width, height = tlbrwh
detections.append({
'label': label,
'confidence': confidence,
't': top,
'l': left,
'b': bot,
'r': right,
'w': width,
'h': height
})
all_detections.append(detections)
return all_detections
def _nms(self, predictions):
predictions[..., :4] = self.xywh2p1p2(predictions[..., :4])
outputs = [None for _ in range(len(predictions))]
for i, image_pred in enumerate(predictions):
image_pred = image_pred[image_pred[:, 4] >= self.thresh]
# If none anchor are remaining => process next image
if not image_pred.size(0):
continue
# Object confidence times class confidence (n, ) * (n, )
score = image_pred[:, 4] * image_pred[:, 5:].max(1)[0]
class_confs, class_preds = image_pred[:, 5:].max(1, keepdim=True)
detections = torch.cat(
(image_pred[:, :5], class_confs.type(
predictions.dtype), class_preds.type(predictions.dtype)),
dim=1)
keep = batched_nms(image_pred[:, :4].float(), score,
class_preds[:, 0], self.nms_thresh)
outputs[i] = detections[keep]
return outputs
@staticmethod
def xywh2p1p2(x):
y = x.new(x.shape)
y[..., 0] = x[..., 0] - x[..., 2] / 2.
y[..., 1] = x[..., 1] - x[..., 3] / 2.
y[..., 2] = x[..., 0] + x[..., 2] / 2.
y[..., 3] = x[..., 1] + x[..., 3] / 2.
return y
@staticmethod
def p1p2Toxywh(x):
y = x.new(x.shape)
y[..., 0] = x[..., 0]
y[..., 1] = x[..., 1]
y[..., 2] = x[..., 2] - x[..., 0]
y[..., 3] = x[..., 3] - x[..., 1]
return y
def _postprocess(self,
outputs,
shapes,
box_format='ltrb',
classes=None,
buffer_ratio=0.0):
outputs = self._nms(outputs)
detections = []
for i, frame_bbs in enumerate(outputs):
im_height, im_width, _ = shapes[i]
if frame_bbs is None:
detections.append([])
continue
frame_bbs = self._resize_boxes(frame_bbs, self.model_image_size,
(im_height, im_width))
frame_dets = []
for box in frame_bbs:
pred_box = self.p1p2Toxywh(box[:4]).data.cpu().numpy()
# box = box.data.cpu().numpy()
cls_conf = box[4].item()
cls_id = box[-1]
cls_name = self.class_names[int(cls_id)]
if classes is not None and cls_name not in classes:
continue
left, top, w, h = pred_box
right = left + w
bottom = top + h
width = right - left + 1
height = bottom - top + 1
width_buffer = width * buffer_ratio
height_buffer = height * buffer_ratio
top = max(0.0, top - 0.5 * height_buffer)
left = max(0.0, left - 0.5 * width_buffer)
bottom = min(im_height - 1.0, bottom + 0.5 * height_buffer)
right = min(im_width - 1.0, right + 0.5 * width_buffer)
box_infos = []
for c in box_format:
if c == 't':
box_infos.append(int(round(top)))
elif c == 'l':
box_infos.append(int(round(left)))
elif c == 'b':
box_infos.append(int(round(bottom)))
elif c == 'r':
box_infos.append(int(round(right)))
elif c == 'w':
box_infos.append(int(round(width + width_buffer)))
elif c == 'h':
box_infos.append(int(round(height + height_buffer)))
else:
assert False, 'box_format given in detect unrecognised!'
assert len(box_infos) > 0, 'box infos is blank'
detection = (box_infos, cls_conf, cls_name)
frame_dets.append(detection)
detections.append(frame_dets)
return detections
@staticmethod
def _resize_boxes(boxes, current_dim, original_shape):
h_ratio = original_shape[0] / current_dim[0]
w_ratio = original_shape[1] / current_dim[1]
boxes[..., 0] *= w_ratio
boxes[..., 1] *= h_ratio
boxes[..., 2] *= w_ratio
boxes[..., 3] *= h_ratio
return boxes
def main():
img_size = 512 # 必须是32的整数倍 [416, 512, 608]
cfg = "/home/mist/yolov3_spp/cfg/yolov3-spp.cfg" # 改成生成的.cfg文件
weights = "/home/mist/yolov3_spp/weights/yolov3spp-29.pt".format(
img_size) # 改成自己训练好的权重文件
json_path = "/home/mist/yolov3_spp/data/pascal_voc_classes.json" # json标签文件
img_path = "test.jpg"
assert os.path.exists(cfg), "cfg file {} dose not exist.".format(cfg)
assert os.path.exists(weights), "weights file {} dose not exist.".format(
weights)
assert os.path.exists(json_path), "json file {} dose not exist.".format(
json_path)
assert os.path.exists(img_path), "image file {} dose not exist.".format(
img_path)
json_file = open(json_path, 'r')
class_dict = json.load(json_file)
category_index = {v: k for k, v in class_dict.items()}
input_size = (img_size, img_size)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = Darknet(cfg, img_size)
model.load_state_dict(torch.load(weights, map_location=device)["model"])
model.to(device)
model.eval()
with torch.no_grad():
# init
img = torch.zeros((1, 3, img_size, img_size), device=device)
model(img)
img_o = cv2.imread(img_path) # BGR
assert img_o is not None, "Image Not Found " + img_path
img = img_utils.letterbox(img_o,
new_shape=input_size,
auto=True,
color=(0, 0, 0))[0]
# Convert
img = img[:, :, ::-1].transpose(2, 0, 1) # BGR to RGB, to 3x416x416
img = np.ascontiguousarray(img)
img = torch.from_numpy(img).to(device).float()
img /= 255.0 # scale (0, 255) to (0, 1)
img = img.unsqueeze(0) # add batch dimension
t1 = torch_utils.time_synchronized()
pred = model(img)[0] # only get inference result
t2 = torch_utils.time_synchronized()
print(t2 - t1)
pred = utils.non_max_suppression(pred,
conf_thres=0.1,
iou_thres=0.6,
multi_label=True)[0]
t3 = time.time()
print(t3 - t2)
if pred is None:
print("No target detected.")
exit(0)
# process detections
pred[:, :4] = utils.scale_coords(img.shape[2:], pred[:, :4],
img_o.shape).round()
print(pred.shape)
bboxes = pred[:, :4].detach().cpu().numpy()
scores = pred[:, 4].detach().cpu().numpy()
classes = pred[:, 5].detach().cpu().numpy().astype(np.int) + 1
img_o = draw_box(img_o[:, :, ::-1], bboxes, classes, scores,
category_index)
plt.imshow(img_o)
plt.show()
img_o.save("test_result.jpg")
model3 = Darknet(os.path.join(
BASE_DIR, "yolo_v3/config/yolov3-custom.cfg")).to(device)
model3.load_state_dict(
torch.load(os.path.join(models_path, "yolo_v3_4_25.pt"),
map_location=device))
dataset = MyTestDataset(split='stage1_train',
transforms=get_test_transforms(rescale_size=(416,
416)))
test_loader = DataLoader(dataset,
batch_size=1,
num_workers=0,
shuffle=False)
model.eval()
model2.eval()
model3.eval()
for i, (image, targets) in enumerate(test_loader):
image = image[0].to(device=device)
name = targets["name"][0]
start_time = time.time()
with torch.no_grad():
outputs = model(image)
outputs2 = model2(image)
outputs3 = model3(image)
outputs = non_max_suppression(outputs, conf_thres=0.5)
outputs2 = non_max_suppression(outputs2, conf_thres=0.5)
outputs3 = non_max_suppression(outputs3, conf_thres=0.5)
elapsed_time = time.time() - start_time
def mask_catch(input, output):
parser = argparse.ArgumentParser()
parser.add_argument(u"--input_file_path",
type=unicode,
default=input,
help=u"path to images directory")
parser.add_argument(u"--output_path",
type=unicode,
default=output,
help=u"output image directory")
parser.add_argument(u"--model_def",
type=unicode,
default=u"data/yolov3_mask.cfg",
help=u"path to model definition file")
parser.add_argument(u"--weights_path",
type=unicode,
default=u"checkpoints/yolov3_ckpt_499.pth",
help=u"path to weights file")
parser.add_argument(u"--class_path",
type=unicode,
default=u"data/mask_dataset.names",
help=u"path to class label file")
parser.add_argument(u"--conf_thres",
type=float,
default=0.8,
help=u"object confidence threshold")
parser.add_argument(u"--nms_thres",
type=float,
default=0.3,
help=u"iou thresshold for non-maximum suppression")
parser.add_argument(u"--frame_size",
type=int,
default=416,
help=u"size of each image dimension")
opt = parser.parse_args()
# Output directory
os.makedirs(opt.output_path, exist_ok=True)
# checking for GPU
device = torch.device(u"cuda" if torch.cuda.is_available() else u"cpu")
# Set up model
model = Darknet(opt.model_def, img_size=opt.frame_size).to(device)
# loading weights
if opt.weights_path.endswith(u".weights"):
model.load_darknet_weights(opt.weights_path) # Load weights
else:
model.load_state_dict(torch.load(opt.weights_path)) # Load checkpoints
# Set in evaluation mode
model.eval()
# Extracts class labels from file
classes = load_classes(opt.class_path)
# ckecking for GPU for Tensor
Tensor = torch.cuda.FloatTensor if torch.cuda.is_available(
) else torch.FloatTensor
print u"\nPerforming object detection:"
# for text in output
t_size = cv2.getTextSize(u" ", cv2.FONT_HERSHEY_PLAIN, 1, 1)[0]
for imagename in os.listdir(opt.input_file_path):
print u"\n" + imagename + u"_______"
image_path = os.path.join(opt.input_file_path, imagename)
print image_path
# frame extraction
org_img = cv2.imread(image_path)
# Original image width and height
i_height, i_width = org_img.shape[:2]
# resizing => [BGR -> RGB] => [[0...255] -> [0...1]] => [[3, 416, 416] -> [416, 416, 3]]
# => [[416, 416, 3] => [416, 416, 3, 1]] => [np_array -> tensor] => [tensor -> variable]
# resizing to [416 x 416]
# Create a black image
x = y = i_height if i_height > i_width else i_width
# Black image
img = np.zeros((x, y, 3), np.uint8)
# Putting original image into black image
start_new_i_height = int((y - i_height) / 2)
start_new_i_width = int((x - i_width) / 2)
img[start_new_i_height:(start_new_i_height + i_height),
start_new_i_width:(start_new_i_width + i_width)] = org_img
#resizing to [416x 416]
img = cv2.resize(img, (opt.frame_size, opt.frame_size))
# [BGR -> RGB]
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
# [[0...255] -> [0...1]]
img = np.asarray(img) / 255
# [[3, 416, 416] -> [416, 416, 3]]
img = np.transpose(img, [2, 0, 1])
# [[416, 416, 3] => [416, 416, 3, 1]]
img = np.expand_dims(img, axis=0)
# [np_array -> tensor]
img = torch.Tensor(img)
# plt.imshow(img[0].permute(1, 2, 0))
# plt.show()
# [tensor -> variable]
img = Variable(img.type(Tensor))
# Get detections
with torch.no_grad():
detections = model(img)
detections = non_max_suppression_output(detections, opt.conf_thres,
opt.nms_thres)
# print(detections)
# For accommodate results in original frame
mul_constant = x / opt.frame_size
#We should set a variable for the number of nomask people. i is the variable
i = 0
# For each detection in detections
for detection in detections:
if detection is not None:
print u"{0} Detection found".format(len(detection))
for x1, y1, x2, y2, conf, cls_conf, cls_pred in detection:
# Accommodate bounding box in original frame
x1 = int(x1 * mul_constant - start_new_i_width)
y1 = int(y1 * mul_constant - start_new_i_height)
x2 = int(x2 * mul_constant - start_new_i_width)
y2 = int(y2 * mul_constant - start_new_i_height)
# Bounding box making and setting Bounding box title
if (int(cls_pred) == 0):
# WITH_MASK
cv2.rectangle(org_img, (x1, y1), (x2, y2), (0, 255, 0),
2)
else:
#WITHOUT_MASK
i += 1
cv2.rectangle(org_img, (x1, y1), (x2, y2), (0, 0, 255),
2)
cv2.putText(org_img,
classes[int(cls_pred)] + u": %.2f" % conf,
(x1, y1 + t_size[1] + 4),
cv2.FONT_HERSHEY_PLAIN, 1, [225, 255, 255], 2)
u"""------------Ready to save!-----------------"""
import time
now = time.strftime(u"%Y-%m-%d-%H_%M_%S", time.localtime(time.time()))
#num is the number of people
num = len(detection)
#na=now + '-' + 'NUM:%d'%num +'-'+ 'Nom:%d'%i+'-'+'.jpg'
u"""------------txt_save-----------------"""
u"""------------image_save-----------------"""
na = u'result.jpg'
out_filepath = os.path.join(opt.output_path, na)
cv2.imwrite(out_filepath,
org_img) #org_img is final result with frames
#naa = now + '-' + 'NUM:%d' % num + '-' + 'Nom:%d' % i
#ssh_scp_put('172.21.39.222',22,'tensor','tensor',out_filepath,'/home/tensor/eden/%s.jpg'%naa)
#upload_img(na)
#os.remove(out_filepath)
signal = 1 #we first set signal only 1
if i == 0:
signal = 0
print u"Signal is ", signal
print u"Finish to save!!!"
msg = now + u'-' + u'NUM:%d' % num + u'-' + u'Nomask:%d' % i + u'-'
nam = u'info.txt'
full_path = os.path.join(opt.output_path, nam)
print u"----------------"
file = open(full_path, u'w')
file.write(msg)
cv2.destroyAllWindows()
return signal
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--config_file", default="config/runs/config.json")
parser.add_argument("--output_dir", default='output')
args = parser.parse_args()
with open(args.config_file) as config_buffer:
config = json.loads(config_buffer.read())
exp_name = get_experiment_name(config)
print(f"Experiment name: {exp_name}")
out_dir = os.path.join(args.output_dir, exp_name)
if os.path.exists(out_dir):
print("experiment dir already exists! Removing...")
shutil.rmtree(out_dir)
os.makedirs(out_dir)
log_dir = f"{out_dir}/logs"
checkpoint_dir = f"{out_dir}/checkpoints"
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
tb_logger = SummaryWriter(log_dir)
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
handlers=[
logging.FileHandler(f"{out_dir}/log.log"),
logging.StreamHandler(sys.stdout)
],
level=logging.INFO)
logger = logging.getLogger(__name__)
logging.info("New session")
seed = config["train"]["seed"]
if seed > 0:
np.random.seed(seed)
torch.manual_seed(seed)
###############################
# Prepare data loaders
###############################
print("Loading datasets...")
if config['val']['validate']:
train_loader, val_concat_loader, val_loader_dict = prepare_dataloaders(
config)
else:
train_loader = prepare_dataloaders(config)
print("Loaded!")
if config["train"]["debug"]:
image_batch, target = next(iter(train_loader))
draw_image_batch_with_targets(image_batch[:4], target, cols=2)
if config['val']['validate']:
val_image_batch, val_target = next(iter(val_concat_loader))
draw_image_batch_with_targets(val_image_batch[:4],
val_target,
cols=2)
###############################
# Construct the model
###############################
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = Darknet(config["model"]["config"]).to(device)
model.apply(weights_init_normal)
print("Model initialized!")
if config["train"]["freeze_feature_extractor"]:
model.freeze_feature_extractor()
print(f"Trainable params: {get_trainable_params_num(model):,}")
# If specified we start from checkpoint
if config["model"]["pretrained_weights"]:
if config["model"]["pretrained_weights"].endswith(".pth"):
model.load_state_dict(
torch.load(config["model"]["pretrained_weights"]))
else:
model.load_darknet_weights(config["model"]["pretrained_weights"])
print("Pretrained weights loaded!")
optimizer = torch.optim.Adam(model.parameters(),
lr=config["train"]["learning_rate"])
###############################
# Training
###############################
batches_done = 0
grad_accumulations = config["train"]["gradient_accumulations"]
save_every = config["train"]["save_every"]
if config["val"]["validate"]:
val_iterator = iter(val_concat_loader)
for epoch in range(config["train"]["nb_epochs"]):
effective_loss = 0
loss_history = torch.zeros(len(train_loader))
logger.info(f"Epoch {epoch} started!")
bar = tqdm(train_loader)
for i, (image_batch, bboxes) in enumerate(bar):
model.train()
image_batch = image_batch.to(device)
bboxes = bboxes.to(device)
loss, outputs = model(image_batch, bboxes)
effective_loss += loss.item()
loss_history[i] = loss.item()
loss.backward()
if i % grad_accumulations == 0:
# Accumulates gradient before each step
optimizer.step()
if config["train"]["gradient_clipping"]:
torch.nn.utils.clip_grad_norm_(model.parameters(), 5)
grad_norm = get_grad_norm(model)
optimizer.zero_grad()
if config["val"]["validate"]:
model.eval()
try:
val_image_batch, val_bboxes = next(val_iterator)
except StopIteration:
val_iterator = iter(val_concat_loader)
val_image_batch, val_bboxes = next(val_iterator)
val_image_batch = val_image_batch.to(device)
val_bboxes = val_bboxes.to(device)
with torch.no_grad():
val_loss, val_outputs = model(val_image_batch,
val_bboxes)
tb_logger.add_scalar("loss/validation", val_loss,
batches_done)
bar.set_description(
f"Loss: {effective_loss / grad_accumulations:.6f}")
batches_done += 1
# Tensorboard logging
for metric_name in metrics:
metric_dict = {}
for j, yolo_layer in enumerate(model.yolo_layers):
metric_dict[f"yolo_{j}"] = yolo_layer.metrics[
metric_name]
if metric_name == 'loss':
metric_dict["overall"] = loss.item()
tb_logger.add_scalars(metric_name, metric_dict,
batches_done)
tb_logger.add_scalar("grad_norm", grad_norm, batches_done)
tb_logger.add_scalar("loss/effective_loss", effective_loss,
batches_done)
effective_loss = 0
# save model
if save_every > 0 and batches_done % save_every == 0:
torch.save(model.state_dict(),
f"{checkpoint_dir}/yolov3_{batches_done}.pth")
epoch_loss = loss_history.mean()
print(f"Epoch loss: {epoch_loss}")
tb_logger.add_scalar("epoch_loss", epoch_loss, epoch)
if config["val"]["validate"]:
result_dict = evaluate(model, val_loader_dict, config["val"])
for name, results in result_dict.items():
output_str = f"{name} evaluation results:\n" \
f"precision-{results['precision']},\n" \
f"recall-{results['recall']},\n" \
f"AP-{results['AP']},\n" \
f"F1-{results['F1']},\n" \
f"ap_class-{results['AP_class']}"
logging.info(output_str)
print(output_str)
tb_logger.add_scalar(f"val_precision/{name}",
results['precision'], epoch)
tb_logger.add_scalar(f"val_recall/{name}", results['recall'],
epoch)
tb_logger.add_scalar(f"val_F1/{name}", results['F1'], epoch)
tb_logger.add_scalar(f"val_AP/{name}", results['AP'], epoch)
# save model
torch.save(model.state_dict(),
f"{checkpoint_dir}/yolov3_epoch_{epoch}.pth")
opt = parser.parse_args()
print(opt)
os.makedirs("output", exist_ok=True)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = Darknet(opt.model_def, img_size=opt.img_size).to(device)
if opt.weights_path.endswith(".weights"):
"""Load darknet weights"""
model.load_darknet_weights(opt.weights_path)
else:
"""Load checkpoint weights"""
model.load_state_dict(torch.load(opt.weights_path))
model.eval()
dataloader = DataLoader(
ImageFolder(opt.image_folder, img_size=opt.img_size),
batch_size=opt.batch_size,
shuffle=False,
num_workers=opt.n_cpu,
)
classes = load_classes(opt.class_path) # Extracts class labels from file
imgs = [] # Stores image paths
img_detections = [] # Stores detections for each image index
print("\nPerforming object detection:")
prev_time = time.time()
def detect_shoes(img, conf_thres=0.1, nms_thres=0.4, box_extension=0):
'''Detect shoes in an image.
Given an image, detect where the shoes are and output the bounding box coordinates,
class confidence scores and confidence score.
Input:
- img: image data from Image.open(img_path).
- conf_thres: confidence score threshold. Float.
- nms_thres: threshold for non maximum suppression.
Output:
- croppend images?
- bounding box coordinates
- confidence scores
'''
model_def = 'config/yolov3-openimages.cfg'
weights_path = 'config/yolov3-openimages.weights'
class_path = 'config/oidv6.names'
conf_thres = 0.1
nms_thres = 0.4
batch_size = 1
n_cpu = 0
img_size = 416
# Extract image as PyTorch tensor
img_original = transforms.ToTensor()(img)
img_shape_original = img_original.shape.permute(1,2,0)
# Pad to square resolution
img, _ = pad_to_square(img, 0)
# Resize
img = resize(img, img_size)
img = img.unsqueeze_(0)
# Set up device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# Set up model
model = Darknet(model_def, img_size=img_size).to(device)
model.load_darknet_weights(weights_path)
model.eval() # Set in evaluation mode
classes = load_classes(class_path) # Extracts class labels from file
Tensor = torch.cuda.FloatTensor if torch.cuda.is_available() else torch.FloatTensor
input_imgs = Variable(img.type(Tensor))
# Get detections
with torch.no_grad():
detections = model(input_imgs) # (B, A, )
detections = non_max_suppression_for_footwear(detections, conf_thres, nms_thres)[0]
if detections is not None:
detections = rescale_boxes(detections, img_size, img_shape_original[:2])
cropped_imgs = []
bbox_coords = []
for x1, y1, x2, y2, conf, cls_conf, cls_pred in detections:
x1 = round(x1)
y1 = round(y1)
x2 = round(x2)
y2 = round(y2)
cropped_imgs.append(img_original[:, x1:x2, y1:y2])
bbox_coords.append([x1,y1,x2,y2])
return cropped_imgs, bbox_coords
else:
return None, None
# Bounding-box colors
cmap = plt.get_cmap("tab20b")
colors = [cmap(i) for i in np.linspace(0, 1, 20)]
# # Create plot
# img = np.array(Image.open(img_path))
# plt.figure()
# fig, ax = plt.subplots(1)
# ax.imshow(img)
# # Draw bounding boxes and labels of detections
# if detections is not None:
# # Rescale boxes to original image
# detections = rescale_boxes(detections, img_size, img.shape[:2])
# unique_labels = detections[:, -1].cpu().unique()
# n_cls_preds = len(unique_labels)
# bbox_colors = random.sample(colors, n_cls_preds)
# for x1, y1, x2, y2, conf, cls_conf, cls_pred in detections:
# print("\t+ Label: %s, Conf: %.5f" % (classes[int(cls_pred)], cls_conf.item()))
# box_w = x2 - x1
# box_h = y2 - y1
# color = bbox_colors[int(np.where(unique_labels == int(cls_pred))[0])]
# # Create a Rectangle patch
# bbox = patches.Rectangle((x1, y1), box_w, box_h, linewidth=2, edgecolor=color, facecolor="none")
# # Add the bbox to the plot
# ax.add_patch(bbox)
# # Add label
# plt.text(
# x1,
# y1,
# s=classes[int(cls_pred)]+', %.2f'%conf.item(),
# color="white",
# verticalalignment="top",
# bbox={"color": color, "pad": 0},
# )
# # Save generated image with detections
# plt.axis("off")
def main():
# load model
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
model = Darknet(cfg.MODEL, img_size=cfg.SIZE).to(device)
model.load_darknet_weights(cfg.WEIGHTS)
model.eval()
# coco classes
classes = load_classes(cfg.CLASSES)
# animals and person
app_classes = [0, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23]
# tensor type
Tensor = torch.cuda.FloatTensor if torch.cuda.is_available() else torch.FloatTensor
# create video capture
cap = cv2.VideoCapture('udp://127.0.0.1:5000', cv2.CAP_FFMPEG)
if not cap.isOpened():
print('VideoCapture not opened')
exit(-1)
# preprocess pipeline
t = transforms.Compose([
transforms.Resize((cfg.SIZE, cfg.SIZE)),
transforms.ToTensor()
])
# tracker
tracker = Sort()
# bbox colors
colors=[
(255,0,0),
(0,255,0),
(0,0,255),
(255,0,255),
(128,0,0),
(0,128,0),
(0,0,128),
(128,0,128),
(128,128,0),
(0,128,128)
]
# process stream
while True:
# read frame
ret, frame = cap.read()
# frame = cv2.flip(cv2.flip(frame, 0), 1)
orig = frame
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
img = Image.fromarray(frame)
# process image
img = t(img).unsqueeze(0).type(Tensor)
with torch.no_grad():
detections = model(img)
detections = non_max_suppression(detections, cfg.CONF, cfg.NMS)
detections = detections[0]
if detections is not None:
# track objects
tracked_objects = tracker.update(detections.cpu())
det = rescale_boxes(tracked_objects, cfg.SIZE, frame.shape[:2])
for x1, y1, x2, y2, obj_id, cls_pred in det:
# ignore not necessary classes
if int(cls_pred) not in app_classes:
continue
# draw bbox
color = colors[int(obj_id) % len(colors)]
cls = classes[int(cls_pred)]
x1, x2, y1, y2 = int(x1), int(x2), int(y1), int(y2)
cv2.rectangle(orig, (x1, y1), (x2, y2), color, 2)
cv2.putText(
orig,
cls + '-' + str(int(obj_id)),
(x1, y1 - 10),
cv2.FONT_HERSHEY_SIMPLEX,
1,
(255, 255, 255),
3
)
cv2.imshow('YoloV3', orig)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
cap.release()
cv2.destroyAllWindows()
def greedy_channel_select(origin_model, prune_cfg, origin_weights,
select_layer, device, aux_util, data_loader,
pruned_rate):
init_state_dict = mask_converted(prune_cfg, origin_weights, target=None)
prune_model = Darknet(prune_cfg).to(device)
prune_model.load_state_dict(init_state_dict, strict=True)
del init_state_dict
solve_sub_problem_optimizer = optim.SGD(
prune_model.module_list[int(select_layer)].MaskConv2d.parameters(),
lr=hyp['lr0'],
momentum=hyp['momentum'])
hook_util = HookUtils()
handles = []
info = aux_util.layer_info[int(select_layer)]
in_channels = info['in_channels']
remove_k = math.floor(in_channels * pruned_rate)
k = in_channels - remove_k
for name, child in origin_model.module_list.named_children():
if name == select_layer:
handles.append(
child.BatchNorm2d.register_forward_hook(
hook_util.hook_origin_input))
aux_idx = aux_util.conv_layer_dict[select_layer]
hook_layer_aux = aux_util.down_sample_layer[aux_idx]
for name, child in prune_model.module_list.named_children():
if name == select_layer:
handles.append(
child.BatchNorm2d.register_forward_hook(
hook_util.hook_prune_input))
elif name == hook_layer_aux:
handles.append(
child.register_forward_hook(hook_util.hook_prune_input))
aux_net = aux_util.creat_aux_list(416,
device,
conv_layer_name=select_layer)
chkpt_aux = torch.load(aux_weight, map_location=device)
aux_net.load_state_dict(chkpt_aux['aux{}'.format(aux_idx)])
del chkpt_aux
if device.type != 'cpu' and torch.cuda.device_count() > 1:
prune_model = torch.nn.parallel.DistributedDataParallel(
prune_model, find_unused_parameters=True)
prune_model.yolo_layers = prune_model.module.yolo_layers
aux_net = torch.nn.parallel.DistributedDataParallel(
aux_net, find_unused_parameters=True)
nb = len(data_loader)
prune_model.nc = 80
prune_model.hyp = hyp
prune_model.arc = 'default'
prune_model.eval()
aux_net.eval()
MSE = nn.MSELoss(reduction='mean')
greedy = torch.zeros(k)
for i_k in range(k):
pbar = tqdm(enumerate(data_loader), total=nb)
print(('\n' + '%10s' * 8) % ('Stage', 'gpu_mem', 'iter', 'MSELoss',
'PdLoss', 'AuxLoss', 'Total', 'targets'))
for i, (imgs, targets, _, _) in pbar:
if len(targets) == 0:
continue
imgs = imgs.to(device).float(
) / 255.0 # uint8 to float32, 0 - 255 to 0.0 - 1.0
targets = targets.to(device)
with torch.no_grad():
_ = origin_model(imgs)
_, pruning_pred = prune_model(imgs)
pruning_loss, _ = compute_loss(pruning_pred, targets, prune_model)
hook_util.cat_to_gpu0('prune')
aux_pred = aux_net(hook_util.prune_features['gpu0'][1])
aux_loss, _ = AuxNetUtils.compute_loss_for_aux(
aux_pred, aux_net, targets)
mse_loss = torch.zeros(1).to(device)
mse_loss += MSE(hook_util.prune_features['gpu0'][0],
hook_util.origin_features['gpu0'][0])
loss = hyp['joint_loss'] * mse_loss + pruning_loss + aux_loss
loss.backward()
mem = torch.cuda.memory_cached() / 1E9 if torch.cuda.is_available(
) else 0
s = ('%10s' * 3 + '%10.3g' * 5) % (
'Pruning ' + select_layer, '%.3gG' % mem, '%g/%g' %
(i_k, k), mse_loss, pruning_loss, aux_loss, loss, len(targets))
pbar.set_description(s)
hook_util.clean_hook_out('origin')
hook_util.clean_hook_out('prune')
grad = prune_model.module.module_list[int(
select_layer)].MaskConv2d.weight.grad.detach().clone()**2
grad = grad.sum((2, 3)).sqrt().sum(0)
if i_k == 0:
prune_model.module.module_list[int(
select_layer)].MaskConv2d.selected_channels_mask[:] = 1e-5
_, non_greedy_indices = torch.topk(grad, k)
logger.info('non greedy layer{}: selected==>{}'.format(
select_layer, str(non_greedy_indices)))
selected_channels_mask = prune_model.module.module_list[int(
select_layer)].MaskConv2d.selected_channels_mask
_, indices = torch.topk(grad * (1 - selected_channels_mask), 1)
prune_model.module.module_list[int(
select_layer)].MaskConv2d.selected_channels_mask[indices] = 1
greedy[i_k] = indices
logger.info('greedy layer{} iter{}: indices==>{}'.format(
select_layer, str(i_k), str(indices)))
prune_model.zero_grad()
pbar = tqdm(enumerate(data_loader), total=nb)
mloss = torch.zeros(4).to(device)
print(('\n' + '%10s' * 8) % ('Stage', 'gpu_mem', 'iter', 'MSELoss',
'PdLoss', 'AuxLoss', 'Total', 'targets'))
for i, (imgs, targets, _, _) in pbar:
if len(targets) == 0:
continue
imgs = imgs.to(device).float(
) / 255.0 # uint8 to float32, 0 - 255 to 0.0 - 1.0
targets = targets.to(device)
with torch.no_grad():
_ = origin_model(imgs)
_, pruning_pred = prune_model(imgs)
pruning_loss, _ = compute_loss(pruning_pred, targets, prune_model)
hook_util.cat_to_gpu0('prune')
aux_pred = aux_net(hook_util.prune_features['gpu0'][1])
aux_loss, _ = AuxNetUtils.compute_loss_for_aux(
aux_pred, aux_net, targets)
mse_loss = torch.zeros(1).to(device)
mse_loss += MSE(hook_util.prune_features['gpu0'][0],
hook_util.origin_features['gpu0'][0])
loss = hyp['joint_loss'] * mse_loss + pruning_loss + aux_loss
loss.backward()
solve_sub_problem_optimizer.step()
solve_sub_problem_optimizer.zero_grad()
mem = torch.cuda.memory_cached() / 1E9 if torch.cuda.is_available(
) else 0
mloss = (mloss * i + torch.cat(
[mse_loss, pruning_loss, aux_loss, loss]).detach()) / (i + 1)
s = ('%10s' * 3 + '%10.3g' * 5) % ('SubProm ' + select_layer,
'%.3gG' % mem, '%g/%g' %
(i_k, k), *mloss, len(targets))
pbar.set_description(s)
hook_util.clean_hook_out('origin')
hook_util.clean_hook_out('prune')
for handle in handles:
handle.remove()
logger.info(
("greedy layer{}: selected==>{}".format(select_layer, str(greedy))))
class BBDetection():
def __init__(self):
print(os.getcwd())
self.model_cfg = "./src/akhenaten_dv/scripts/Perception/BBoxDetection/model_cfg/yolo_baseline_tiny.cfg"
self.weights_path = './src/akhenaten_dv/scripts/Perception/BBoxDetection/7.weights'
self.conf_thres = 0.8
self.nms_thres = 0.25
self.vanilla_anchor = False
self.xy_loss = 2
self.wh_loss = 1.6
self.no_object_loss = 25
self.object_loss = 0.1
cuda = torch.cuda.is_available()
self.device = torch.device('cuda:0' if cuda else 'cpu')
random.seed(0)
torch.manual_seed(0)
if cuda:
torch.cuda.manual_seed(0)
torch.cuda.manual_seed_all(0)
torch.backends.cudnn.benchmark = True
torch.cuda.empty_cache()
self.model = Darknet(config_path=self.model_cfg,
xy_loss=self.xy_loss,
wh_loss=self.wh_loss,
no_object_loss=self.no_object_loss,
object_loss=self.object_loss,
vanilla_anchor=self.vanilla_anchor)
# Load weights
self.model.load_weights(self.weights_path,
self.model.get_start_weight_dim())
self.model.to(self.device, non_blocking=True)
def detect(self, cv_img):
cv_img = cv2.cvtColor(cv_img, cv2.COLOR_BGR2RGB)
img = img_pil.fromarray(cv_img)
w, h = img.size
new_width, new_height = self.model.img_size()
pad_h, pad_w, ratio = calculate_padding(h, w, new_height, new_width)
img = torchvision.transforms.functional.pad(img,
padding=(pad_w, pad_h,
pad_w, pad_h),
fill=(127, 127, 127),
padding_mode="constant")
img = torchvision.transforms.functional.resize(img,
(new_height, new_width))
bw = self.model.get_bw()
if bw:
img = torchvision.transforms.functional.to_grayscale(
img, num_output_channels=1)
img = torchvision.transforms.functional.to_tensor(img)
img = img.unsqueeze(0)
with torch.no_grad():
self.model.eval()
img = img.to(self.device, non_blocking=True)
# output,first_layer,second_layer,third_layer = model(img)
output = self.model(img)
for detections in output:
detections = detections[detections[:, 4] > self.conf_thres]
box_corner = torch.zeros((detections.shape[0], 4),
device=detections.device)
xy = detections[:, 0:2]
wh = detections[:, 2:4] / 2
box_corner[:, 0:2] = xy - wh
box_corner[:, 2:4] = xy + wh
probabilities = detections[:, 4]
nms_indices = nms(box_corner, probabilities, self.nms_thres)
main_box_corner = box_corner[nms_indices]
if nms_indices.shape[0] == 0:
continue
bboxes = []
for i in range(len(main_box_corner)):
x0 = main_box_corner[i, 0].to('cpu').item() / ratio - pad_w
y0 = main_box_corner[i, 1].to('cpu').item() / ratio - pad_h
x1 = main_box_corner[i, 2].to('cpu').item() / ratio - pad_w
y1 = main_box_corner[i, 3].to('cpu').item() / ratio - pad_h
bboxes.append([x0, y0, x1, y1])
return bboxes