random seed, hyp for random perspective

data/hyp.tir_od.tiny_aug_gamma_anchor_3.yaml

@@ -0,0 +1,36 @@
+lr0: 0.001 #0.001  # initial learning rate (SGD=1E-2, Adam=1E-3)
+lrf: 0.01  # final OneCycleLR learning rate (lr0 * lrf)
+momentum: 0.937  # SGD momentum/Adam beta1
+weight_decay: 0.005  # optimizer weight decay 5e-4  It resolve mAP of overfitting test
+warmup_epochs: 3.0  # warmup epochs (fractions ok)
+warmup_momentum: 0.8  # warmup initial momentum
+warmup_bias_lr: 0.001 #0.001  # warmup initial bias lr
+box: 0.05  # box loss gain
+cls: 0.5  # cls loss gain
+cls_pw: 1.0  # cls BCELoss positive_weight
+obj: 1.0  # obj loss gain (scale with pixels)
+obj_pw: 1.0  # obj BCELoss positive_weight
+iou_t: 0.60  # like the default in the code was 0.2 IoU training threshold
+anchor_t: 4.0  # anchor-multiple threshold
+anchors: 3  # anchors per output layer (0 to ignore)  @@HK was 3
+fl_gamma: 1.5 #1.5  # focal loss gamma (efficientDet default gamma=1.5)
+hsv_h: 0.0  # image HSV-Hue augmentation (fraction)
+hsv_s: 0.0  # image HSV-Saturation augmentation (fraction)
+hsv_v: 0.0  # image HSV-Value augmentation (fraction)
+degrees: 0 # image rotation (+/- deg)
+translate: 0  # image translation (+/- fraction)
+scale: 0  # image scale (+/- gain)
+shear: 0.0  # image shear (+/- deg)
+perspective: 0.0  # image perspective (+/- fraction), range 0-0.001
+flipud: 0.3  # image flip up-down (probability)
+fliplr: 0.5  # image flip left-right (probability)
+mosaic: 0.5  # image mosaic (probability)
+mixup: 0.15  # image mixup (probability)
+copy_paste: 0.0  # image copy paste (probability)
+paste_in: 0.0  # image copy paste (probability), use 0 for faster training : cutout
+loss_ota: 0 #1 # use ComputeLossOTA, use 0 for faster training
+inversion: 0.5 #opposite temperature
+img_percentile_removal: 0.3
+beta : 0.3
+random_perspective : 0
+gamma : 80   # percent

data/hyp.tir_od.tiny_aug_gamma_lr_0p005.yaml

@@ -0,0 +1,36 @@
+lr0: 0.005 #0.001  # initial learning rate (SGD=1E-2, Adam=1E-3)
+lrf: 0.01  # final OneCycleLR learning rate (lr0 * lrf)
+momentum: 0.937  # SGD momentum/Adam beta1
+weight_decay: 0.005  # optimizer weight decay 5e-4  It resolve mAP of overfitting test
+warmup_epochs: 3.0  # warmup epochs (fractions ok)
+warmup_momentum: 0.8  # warmup initial momentum
+warmup_bias_lr: 0.001 #0.001  # warmup initial bias lr
+box: 0.05  # box loss gain
+cls: 0.5  # cls loss gain
+cls_pw: 1.0  # cls BCELoss positive_weight
+obj: 1.0  # obj loss gain (scale with pixels)
+obj_pw: 1.0  # obj BCELoss positive_weight
+iou_t: 0.60  # like the default in the code was 0.2 IoU training threshold
+anchor_t: 4.0  # anchor-multiple threshold
+anchors: 2  # anchors per output layer (0 to ignore)  @@HK was 3
+fl_gamma: 1.5 #1.5  # focal loss gamma (efficientDet default gamma=1.5)
+hsv_h: 0.0  # image HSV-Hue augmentation (fraction)
+hsv_s: 0.0  # image HSV-Saturation augmentation (fraction)
+hsv_v: 0.0  # image HSV-Value augmentation (fraction)
+degrees: 0 # image rotation (+/- deg)
+translate: 0  # image translation (+/- fraction)
+scale: 0  # image scale (+/- gain)
+shear: 0.0  # image shear (+/- deg)
+perspective: 0.0  # image perspective (+/- fraction), range 0-0.001
+flipud: 0.3  # image flip up-down (probability)
+fliplr: 0.5  # image flip left-right (probability)
+mosaic: 0.5  # image mosaic (probability)
+mixup: 0.15  # image mixup (probability)
+copy_paste: 0.0  # image copy paste (probability)
+paste_in: 0.0  # image copy paste (probability), use 0 for faster training : cutout
+loss_ota: 0 #1 # use ComputeLossOTA, use 0 for faster training
+inversion: 0.5 #opposite temperature
+img_percentile_removal: 0.3
+beta : 0.3
+random_perspective : 0
+gamma : 80   # percent

data/hyp.tir_od.tiny_aug_gamma_rnd_perspective.yaml

@@ -0,0 +1,36 @@
+lr0: 0.001 #0.001  # initial learning rate (SGD=1E-2, Adam=1E-3)
+lrf: 0.01  # final OneCycleLR learning rate (lr0 * lrf)
+momentum: 0.937  # SGD momentum/Adam beta1
+weight_decay: 0.005  # optimizer weight decay 5e-4  It resolve mAP of overfitting test
+warmup_epochs: 3.0  # warmup epochs (fractions ok)
+warmup_momentum: 0.8  # warmup initial momentum
+warmup_bias_lr: 0.001 #0.001  # warmup initial bias lr
+box: 0.05  # box loss gain
+cls: 0.5  # cls loss gain
+cls_pw: 1.0  # cls BCELoss positive_weight
+obj: 1.0  # obj loss gain (scale with pixels)
+obj_pw: 1.0  # obj BCELoss positive_weight
+iou_t: 0.60  # like the default in the code was 0.2 IoU training threshold
+anchor_t: 4.0  # anchor-multiple threshold
+anchors: 2  # anchors per output layer (0 to ignore)  @@HK was 3
+fl_gamma: 1.5 #1.5  # focal loss gamma (efficientDet default gamma=1.5)
+hsv_h: 0.0  # image HSV-Hue augmentation (fraction)
+hsv_s: 0.0  # image HSV-Saturation augmentation (fraction)
+hsv_v: 0.0  # image HSV-Value augmentation (fraction)
+degrees: 0.0 # image rotation (+/- deg)
+translate: 0.2  # image translation (+/- fraction)
+scale: 0.9  # image scale (+/- gain)
+shear: 0.0  # image shear (+/- deg)
+perspective: 0.001  # image perspective (+/- fraction), range 0-0.001
+flipud: 0.3  # image flip up-down (probability)
+fliplr: 0.5  # image flip left-right (probability)
+mosaic: 0.5  # image mosaic (probability)
+mixup: 0.15  # image mixup (probability)
+copy_paste: 0.0  # image copy paste (probability)
+paste_in: 0.0  # image copy paste (probability), use 0 for faster training : cutout
+loss_ota: 0 #1 # use ComputeLossOTA, use 0 for faster training
+inversion: 0.5 #opposite temperature
+img_percentile_removal: 0.3
+beta : 0.3
+random_perspective : 0
+gamma : 80   # percent

data/hyp.tir_od.tiny_aug_paste_in.yaml

@@ -0,0 +1,37 @@
+lr0: 0.005 #0.001  # initial learning rate (SGD=1E-2, Adam=1E-3)
+lrf: 0.01  # final OneCycleLR learning rate (lr0 * lrf)
+momentum: 0.937  # SGD momentum/Adam beta1
+weight_decay: 0.005  # optimizer weight decay 5e-4  It resolve mAP of overfitting test
+warmup_epochs: 3.0  # warmup epochs (fractions ok)
+warmup_momentum: 0.8  # warmup initial momentum
+warmup_bias_lr: 0.001 #0.001  # warmup initial bias lr
+box: 0.05  # box loss gain
+cls: 0.5  # cls loss gain
+cls_pw: 1.0  # cls BCELoss positive_weight
+obj: 1.0  # obj loss gain (scale with pixels)
+obj_pw: 1.0  # obj BCELoss positive_weight
+iou_t: 0.20  # IoU training threshold
+anchor_t: 4.0  # anchor-multiple threshold
+anchors: 2  # anchors per output layer (0 to ignore)  @@HK was 3
+fl_gamma: 1.5 #1.5  # focal loss gamma (efficientDet default gamma=1.5)
+hsv_h: 0.0  # image HSV-Hue augmentation (fraction)
+hsv_s: 0.0  # image HSV-Saturation augmentation (fraction)
+hsv_v: 0.0  # image HSV-Value augmentation (fraction)
+degrees: 0 # image rotation (+/- deg)
+translate: 0  # image translation (+/- fraction)
+scale: 0  # image scale (+/- gain)
+shear: 0.0  # image shear (+/- deg)
+perspective: 0.0  # image perspective (+/- fraction), range 0-0.001
+flipud: 0.3  # image flip up-down (probability)
+fliplr: 0.5  # image flip left-right (probability)
+mosaic: 0.5  # image mosaic (probability)
+mixup: 0.15  # image mixup (probability)
+copy_paste: 0.0  # image copy paste (probability)
+paste_in: 0.3  # image copy paste (probability), use 0 for faster training : cutout
+loss_ota: 0 #1 # use ComputeLossOTA, use 0 for faster training
+inversion: 0.5 #opposite temperature
+img_percentile_removal: 0.3
+beta : 0.3
+random_perspective : 0
+gamma : 80   # percent
+gamma_liklihood: 0.25

test.py

@@ -448,4 +448,6 @@ test based on RGB coco model
 
 
 --weights /mnt/Data/hanoch/runs/train/yolov7575/weights/best.pt --device 0 --batch-size 16 --data data/tir_od_test_set.yaml --img-size 640 --conf 0.001 --verbose --norm-type single_image_percentile_0_1 --input-channels 1 --project test --task test --iou-thres 0.6
+--weights /home/hanoch/projects/tir_od/runs/gpu02/yolov74/weights --device 0 --batch-size 16 --data data/tir_od_test_set.yaml --img-size 640 --conf 0.001 --verbose --norm-type single_image_percentile_0_1 --input-channels 1 --project test --task test --iou-thres 0.6
+
 """

train.py

@@ -52,12 +52,16 @@ logger = logging.getLogger(__name__)
 clear_ml = True
 
 from clearml import Task, Logger
-task = Task.init(
+
-        project_name="TIR_OD",
+if clear_ml:  # clearml support
-        task_name="train yolov7 with dummy test"
+
-    )
+    task = Task.init(
-# Task.execute_remotely() will invoke the job immidiately over the remote and not DeV
+            project_name="TIR_OD",
-task.set_base_docker(docker_image="nvcr.io/nvidia/pytorch:24.09-py3", docker_arguments="--shm-size 8G")
+            task_name="train yolov7 with dummy test"
+        )
+    # Task.execute_remotely() will invoke the job immidiately over the remote and not DeV
+    task.set_base_docker(docker_image="nvcr.io/nvidia/pytorch:24.09-py3", docker_arguments="--shm-size 8G")
+
 gradient_clip_value = 100.0
 opt_gradient_clipping = True
 
@@ -121,9 +125,6 @@ def train(hyp, opt, device, tb_writer=None):
     best = wdir / 'best.pt'
     results_file = save_dir / 'results.txt'
 
-    # Save run settings
-    with open(save_dir / 'hyp.yaml', 'w') as f:
-        yaml.dump(hyp, f, sort_keys=False)
     with open(save_dir / 'opt.yaml', 'w') as f:
         yaml.dump(vars(opt), f, sort_keys=False)
 
@@ -132,7 +133,17 @@ def train(hyp, opt, device, tb_writer=None):
     # Configure
     plots = not opt.evolve  # create plots
     cuda = device.type != 'cpu'
-    init_seeds(2 + rank)
+    if opt.predefined_seed:
+        hyp['seed'] = 2 + rank
+        init_seeds(2 + rank)
+    else:
+        rand_seed = int(time.time())
+        hyp['seed'] = rand_seed
+        init_seeds(rand_seed)
+
+    # Save run settings
+    with open(save_dir / 'hyp.yaml', 'w') as f:
+        yaml.dump(hyp, f, sort_keys=False)
 
     if clear_ml: #clearml support
         config_file = task.connect_configuration(opt.data)
@@ -374,7 +385,9 @@ def train(hyp, opt, device, tb_writer=None):
                 check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)
             if opt.amp or 1:
                 model.half().float()  # pre-reduce anchor precision TODO HK Why ? >???!!!!
-
+    if 1:
+        print("opt.local_rank", opt.local_rank)
+        print("opt.local_rank", opt.local_rank)
     # DDP mode
     if cuda and rank != -1:
         model = DDP(model, device_ids=[opt.local_rank], output_device=opt.local_rank,
@@ -694,7 +707,7 @@ if __name__ == '__main__':
     parser.add_argument('--single-cls', action='store_true', help='train multi-class data as single-class')
     parser.add_argument('--adam', action='store_true', help='use torch.optim.Adam() optimizer')
     parser.add_argument('--sync-bn', action='store_true', help='use SyncBatchNorm, only available in DDP mode')
-    parser.add_argument('--local_rank', type=int, default=-1, help='DDP parameter, do not modify')
+    parser.add_argument('--local-rank', type=int, default=-1, help='DDP parameter, do not modify') #Changed in version 2.0.0: The launcher will passes the --local-rank=<rank> argument to your script. From PyTorch 2.0.0 onwards, the dashed --local-rank is preferred over the previously used underscored --local_rank.
     parser.add_argument('--workers', type=int, default=8, help='maximum number of dataloader workers')
     parser.add_argument('--project', default='runs/train', help='save to project/name')
     parser.add_argument('--entity', default=None, help='W&B entity')
@@ -725,6 +738,9 @@ if __name__ == '__main__':
 
     parser.add_argument('--amp', action='store_true', help='Remove torch AMP')
 
+    parser.add_argument('--predefined-seed', action='store_true', help='predefined_seed only set it to constant otherwise add args that load the random one ')
+
+
 
 
     opt = parser.parse_args()

utils/datasets.py

@@ -912,11 +912,6 @@ def load_mosaic(self, index):
     for i, index in enumerate(indices):
         # Load image
         img, _, (h, w) = load_image(self, index)
-        # if 1:
-        #     img = np.repeat(img[:, :, np.newaxis], 3, axis=2)
-        # if img.ndim <3 : #TIR =>unsqueeze ro RGB 1-channel
-        #     tir_signal = True
-        #     img = img[:, :, np.newaxis] #np.repeat(img[:, :, np.newaxis], 3, axis=2) #img[:, :, np.newaxis]
         # place img in img4
         if i == 0:  # top left
             if self.is_tir_signal:
@@ -1436,7 +1431,7 @@ class Albumentations_gamma_contrast:
 
         self.transform = A.Compose([
             # A.CLAHE(p=0.01),
-            A.RandomBrightnessContrast(brightness_limit=0.2, contrast_limit=0.2, p=alb_prob),
+            A.RandomBrightnessContrast(brightness_limit=0.2, contrast_limit=0.2, p=alb_prob), #Contrast adjustment: x' = clip((x - mean) * (1 + a) + mean) ; x'' = clip(x' * (1 + β))
             A.RandomGamma(gamma_limit=gamma_limit, p=alb_prob)])
             # A.Blur(p=0.01),
             # A.MedianBlur(p=0.01),