Fixed 'meta' and 'hyp' may out of order when using evolve

[imyhxy]

Hi, there.
Recently I have try to use the evolve feature of yolov5, and found out that the keys of meta and hyp maybe miss align in train.py. The reason is that, firstly, dictionary in python won't gurantee order. Secondly, the anchors hyper-parameter is possible add to the hyp dictionary in runtime (when it not be defined in hyp.yaml). And after that, all hyper-parameters after anchors will miss align between meta and hyp.

So, I just use the hyp keys to get both gain, which will maintain the order.

We can validate this by set all gain of meta to 0 except one that after anchors key. Use translate key as an example (because translate and scale are initialized with non-zero value, if a hyper-parameter initialize with 0 and it won't get mutation), it will evolve on the scale key:

python train.py --evolve 5 --epochs 2

        # Hyperparameter evolution metadata (mutation scale 0-1, lower_limit, upper_limit)
        meta = {'lr0': (0, 1e-5, 1e-1),  # initial learning rate (SGD=1E-2, Adam=1E-3)
                'lrf': (0, 0.01, 1.0),  # final OneCycleLR learning rate (lr0 * lrf)
                'momentum': (0, 0.6, 0.98),  # SGD momentum/Adam beta1
                'weight_decay': (0, 0.0, 0.001),  # optimizer weight decay
                'warmup_epochs': (0, 0.0, 5.0),  # warmup epochs (fractions ok)
                'warmup_momentum': (0, 0.0, 0.95),  # warmup initial momentum
                'warmup_bias_lr': (0, 0.0, 0.2),  # warmup initial bias lr
                'box': (0, 0.02, 0.2),  # box loss gain
                'cls': (0, 0.2, 4.0),  # cls loss gain
                'cls_pw': (0, 0.5, 2.0),  # cls BCELoss positive_weight
                'obj': (0, 0.2, 4.0),  # obj loss gain (scale with pixels)
                'obj_pw': (0, 0.5, 2.0),  # obj BCELoss positive_weight
                'iou_t': (0, 0.1, 0.7),  # IoU training threshold
                'anchor_t': (0, 2.0, 8.0),  # anchor-multiple threshold
                'anchors': (0, 2.0, 10.0),  # anchors per output grid (0 to ignore)
                'fl_gamma': (0, 0.0, 2.0),  # focal loss gamma (efficientDet default gamma=1.5)
                'hsv_h': (0, 0.0, 0.1),  # image HSV-Hue augmentation (fraction)
                'hsv_s': (0, 0.0, 0.9),  # image HSV-Saturation augmentation (fraction)
                'hsv_v': (0, 0.0, 0.9),  # image HSV-Value augmentation (fraction)
                'degrees': (0, 0.0, 45.0),  # image rotation (+/- deg)
                'translate': (1, 0.0, 0.9),  # image translation (+/- fraction)
                'scale': (0, 0.0, 0.9),  # image scale (+/- gain)
                'shear': (0, 0.0, 10.0),  # image shear (+/- deg)
                'perspective': (0, 0.0, 0.001),  # image perspective (+/- fraction), range 0-0.001
                'flipud': (0, 0.0, 1.0),  # image flip up-down (probability)
                'fliplr': (0, 0.0, 1.0),  # image flip left-right (probability)
                'mosaic': (0, 0.0, 1.0),  # image mixup (probability)
                'mixup': (0, 0.0, 1.0),  # image mixup (probability)
                'copy_paste': (0, 0.0, 1.0)}  # segment copy-paste (probability)

   metrics/precision,      metrics/recall,     metrics/mAP_0.5,metrics/mAP_0.5:0.95,        val/box_loss,        val/obj_loss,        val/cls_loss,                 lr0,                 lrf,            momentum,        weight_decay,       warmup_epochs,     warmup_momentum,      warmup_bias_lr,                 box,                 cls,              cls_pw,                 obj,              obj_pw,               iou_t,            anchor_t,            fl_gamma,               hsv_h,               hsv_s,               hsv_v,             degrees,           translate,               scale,               shear,         perspective,              flipud,              fliplr,              mosaic,               mixup,          copy_paste,             anchors
             0.29013,             0.38144,             0.27229,              0.1401,             0.08883,            0.043384,            0.016148,                0.01,                 0.2,               0.937,              0.0005,                   3,                 0.8,                 0.1,                0.05,                 0.5,                   1,                   1,                   1,                 0.2,                   4,                   0,               0.015,                 0.7,                 0.4,                   0,                 0.1,                 0.5,                   0,                   0,                   0,                 0.5,                   1,                   0,                   0,                   3
             0.28905,             0.38505,              0.2748,             0.14504,            0.088602,            0.043445,            0.016112,                0.01,                 0.2,               0.937,              0.0005,                   3,                 0.8,                 0.1,                0.05,                 0.5,                   1,                   1,                   1,                 0.2,                   4,                   0,               0.015,                 0.7,                 0.4,                   0,                 0.1,             0.62915,                   0,                   0,                   0,                 0.5,                   1,                   0,                   0,                   3
             0.28587,             0.38832,             0.27283,              0.1416,            0.088814,            0.043403,            0.016127,                0.01,                 0.2,               0.937,              0.0005,                   3,                 0.8,                 0.1,                0.05,                 0.5,                   1,                   1,                   1,                 0.2,                   4,                   0,               0.015,                 0.7,                 0.4,                   0,                 0.1,             0.59574,                   0,                   0,                   0,                 0.5,                   1,                   0,                   0,                   3
             0.29403,             0.36923,             0.27276,             0.14222,            0.089021,             0.04337,            0.016128,                0.01,                 0.2,               0.937,              0.0005,                   3,                 0.8,                 0.1,                0.05,                 0.5,                   1,                   1,                   1,                 0.2,                   4,                   0,               0.015,                 0.7,                 0.4,                   0,                 0.1,              0.5751,                   0,                   0,                   0,                 0.5,                   1,                   0,                   0,                   3
             0.28571,             0.38769,             0.27213,             0.14169,            0.088797,            0.043437,            0.016124,                0.01,                 0.2,               0.937,              0.0005,                   3,                 0.8,                 0.1,                0.05,                 0.5,                   1,                   1,                   1,                 0.2,                   4,                   0,               0.015,                 0.7,                 0.4,                   0,                 0.1,             0.61197,                   0,                   0,                   0,                 0.5,                   1,                   0,                   0,                   3

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improving the .gitignore and genetic algorithm in YOLOv5.

📊 Key Changes

• Modified the .gitignore to explicitly ignore the wandb/ directory.
• Updated the genetic algorithm within train.py to ensure hyp keys align with meta values during mutation.

🎯 Purpose & Impact

• Ensuring that the Weights & Biases (wandb) folder, used for experiment tracking, is properly excluded from version control. This helps prevent accidental commits of potentially large and unneeded wandb directories.
• Refining the genetic algorithm that fine-tunes hyperparameters to prevent errors and ensure only valid mutations are considered. This should lead to more efficient and potentially more effective model training.

The changes will make collaborating on YOLOv5 codebase cleaner and may lead to improved performance of models trained with YOLOv5 thanks to the more robust hyperparameter tuning process. 🧬✨

[glenn-jocher]

@imyhxy I reviewed the code here, but I don't see any problem in master. Even if meta and hyp are out of order, the k key indexes hyp directly, so order between the two is not relevant to the code here. hyp[k] will always be matched with meta[k]. If meta lacks a key that hyp has, then simply no limits and significant digits will be applied there, but there is no error I see:

yolov5/train.py

Lines 581 to 586 in fad57c2

	# Constrain to limits
	for k, v in meta.items():
	hyp[k] = max(hyp[k], v[1]) # lower limit
	hyp[k] = min(hyp[k], v[2]) # upper limit
	hyp[k] = round(hyp[k], 5) # significant digits

[glenn-jocher]

@imyhxy took a second look, yes you are right there is a latent bug!

For some reason I was focusing on the other section of code above, not sure why. But yes in the initial gain assignment there is a danger of misalignment, if not now then possibly in the future.

Your fix looks good I think, it enforces alignment again based off the key directly.

[glenn-jocher]

@imyhxy PR is merged. Thank you for your contributions to YOLOv5 🚀 and Vision AI ⭐