Added support for gradual unfreezing while training (transfer-learning)

[senhorinfinito]

Unfreezing layers during training is one the strategy used in ULMFIT to fine-tune model. Based on that, I have updated code to unfreeze layers during training.

Compatibility Changes

• Added --grad_unfreeze , to able the gradual unfreezing.
• --unfreezing_layers , layer ids, (it is basically decided on number of layers has already frozen)
• --unfreeze_epochs, Epoch number, where you want to unfreeze the layer.

Usage

To use gradual unfreeze, if only backbone is frozen & trained for 50 epochs.
python train.py --data coco.yaml --cfg yolov5n.yaml --weights '' --batch-size 128 --grad_unfreeze --unfreezing_layers 9 --unfreezing_layers 8 --unfreezing_layers 7 --unfreezing_layers 6 --unfreeze_epochs 15 20 25 30 35

You can find issue here #9677

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Implemented Gradual Unfreezing in YOLOv5 training 🚀.

📊 Key Changes

• Added the gradual_unfreeze function to allow selective unfreezing of model layers during training.
• Updated train.py to support new command-line arguments for gradual unfreezing.
• Modified training loop to perform gradual unfreezing based on specified layers and epochs.

🎯 Purpose & Impact

• Purpose: To improve model fine-tuning by gradually unfreezing layers during training, allowing for more controlled learning.
• Impact: Users can now fine-tune models more effectively, potentially leading to better performance and quicker convergence. This could be especially useful when adapting pre-trained models to new datasets or tasks. 🆕✨

[WangQvQ]

在训练期间解冻层是 ULMFIT 中用于微调模型的策略之一。基于此，我更新了代码以在训练期间解冻层。

兼容性更改

• 添加，以便能够逐渐解冻。--grad_unfreeze
• --unfreezing_layers，层 ID，（基本上决定了层数已经冻结了）
• --unfreeze_epochs、纪元编号，用于解冻图层的位置。

用法

使用渐进解冻，如果只有骨干被冻结和训练50个epoch。 python train.py --data coco.yaml --cfg yolov5n.yaml --weights '' --batch-size 128 --grad_unfreeze --unfreezing_layers 9 --unfreezing_layers 8 --unfreezing_layers 7 --unfreezing_layers 6 --unfreeze_epochs 15 20 25 30 35

您可以在此处找到问题 #9677

Thank you for your contribution. I modified the file according to your way, but there was an error:
AssertionError: Length of unfreezeing layers [[9], [8], [7], [6]] not matched with [[9], [8], [7], [6]]

[anantsakhare]

在训练期间解冻层是 ULMFIT 中用于微调模型的策略之一。基于此，我更新了代码以在训练期间解冻层。

兼容性更改

• 添加，以便能够逐渐解冻。--grad_unfreeze
• --unfreezing_layers，层 ID，（基本上决定了层数已经冻结了）
• --unfreeze_epochs、纪元编号，用于解冻图层的位置。

用法

使用渐进解冻，如果只有骨干被冻结和训练50个epoch。 python train.py --data coco.yaml --cfg yolov5n.yaml --weights '' --batch-size 128 --grad_unfreeze --unfreezing_layers 9 --unfreezing_layers 8 --unfreezing_layers 7 --unfreezing_layers 6 --unfreeze_epochs 15 20 25 30 35
您可以在此处找到问题 #9677

Thank you for your contribution. I modified the file according to your way, but there was an error: AssertionError: Length of unfreezeing layers [[9], [8], [7], [6]] not matched with [[9], [8], [7], [6]]

Please use an equal number of epoch & equal number of layers if you want to unfreeze 4 layers, use 4 different epoch values.
@WangQvQ Thanks for showing interest in implementations.
eg.
python train.py --data coco.yaml --cfg yolov5n.yaml --weights '' --batch-size 128 --grad_unfreeze --unfreezing_layers 9 --unfreezing_layers 8 --unfreezing_layers 7 --unfreezing_layers 6 --unfreeze_epochs 15 20 25 30