ultralytics / yolov5

YOLOv5 🚀 in PyTorch > ONNX > CoreML > TFLite
https://docs.ultralytics.com
GNU Affero General Public License v3.0
50.6k stars 16.32k forks source link

Transfer Learning with Frozen Layers #1314

Open glenn-jocher opened 3 years ago

glenn-jocher commented 3 years ago

📚 This guide explains how to freeze YOLOv5 🚀 layers when transfer learning. Transfer learning is a useful way to quickly retrain a model on new data without having to retrain the entire network. Instead, part of the initial weights are frozen in place, and the rest of the weights are used to compute loss and are updated by the optimizer. This requires less resources than normal training and allows for faster training times, though it may also results in reductions to final trained accuracy. UPDATED 28 March 2023.

Before You Start

Clone repo and install requirements.txt in a Python>=3.7.0 environment, including PyTorch>=1.7. Models and datasets download automatically from the latest YOLOv5 release.

git clone https://github.com/ultralytics/yolov5  # clone
cd yolov5
pip install -r requirements.txt  # install

Freeze Backbone

All layers that match the freeze list in train.py will be frozen by setting their gradients to zero before training starts. https://github.com/ultralytics/yolov5/blob/771ac6c53ded79c408ed8bd99f7604b7077b7d77/train.py#L119-L126

To see a list of module names:

for k, v in model.named_parameters():
    print(k)

# Output
model.0.conv.conv.weight
model.0.conv.bn.weight
model.0.conv.bn.bias
model.1.conv.weight
model.1.bn.weight
model.1.bn.bias
model.2.cv1.conv.weight
model.2.cv1.bn.weight
...
model.23.m.0.cv2.bn.weight
model.23.m.0.cv2.bn.bias
model.24.m.0.weight
model.24.m.0.bias
model.24.m.1.weight
model.24.m.1.bias
model.24.m.2.weight
model.24.m.2.bias

Looking at the model architecture we can see that the model backbone is layers 0-9: https://github.com/ultralytics/yolov5/blob/58f8ba771e3712b525ca93a1ee66bc2b2df2092f/models/yolov5s.yaml#L12-L48

so we can define the freeze list to contain all modules with 'model.0.' - 'model.9.' in their names:

python train.py --freeze 10

Freeze All Layers

To freeze the full model except for the final output convolution layers in Detect(), we set freeze list to contain all modules with 'model.0.' - 'model.23.' in their names:

python train.py --freeze 24

Results

We train YOLOv5m on VOC on both of the above scenarios, along with a default model (no freezing), starting from the official COCO pretrained --weights yolov5m.pt:

$ train.py --batch 48 --weights yolov5m.pt --data voc.yaml --epochs 50 --cache --img 512 --hyp hyp.finetune.yaml

Accuracy Comparison

The results show that freezing speeds up training, but reduces final accuracy slightly.

Screenshot 2020-11-06 at 18 08 13

GPU Utilization Comparison

Interestingly, the more modules are frozen the less GPU memory is required to train, and the lower GPU utilization. This indicates that larger models, or models trained at larger --image-size may benefit from freezing in order to train faster.

Environments

YOLOv5 may be run in any of the following up-to-date verified environments (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled):

Status

YOLOv5 CI

If this badge is green, all YOLOv5 GitHub Actions Continuous Integration (CI) tests are currently passing. CI tests verify correct operation of YOLOv5 training, validation, inference, export and benchmarks on MacOS, Windows, and Ubuntu every 24 hours and on every commit.

omobayode1 commented 3 years ago

I noticed that there's an argument in yolov3 train.py code "--freeze-layer"?

Please, what does it do?

It states that it freezes all non-output layer?

Please can you provide more clarification about this?

Thank you.

Omobayode

mphillips-valleyit commented 3 years ago

@glenn-jocher Another dimension to this is generalization. I assume your results are shown for a test dataset. But for generalization to new datasets, freezing might also help prevent overfitting to ttraining data (and therefore improve robustness/generalization).

glenn-jocher commented 3 years ago

@mphillips-valleyit interesting point, though hard to quantify beyond existing val/test metrics.

mphillips-valleyit commented 3 years ago

It could be done with separate datasets--models pretrained on COCO, measure their generalization (freezing vs. non-freezing fine-tuning) to OpenImages for common categories. If I'm able to post results on this at some point, I will.

nanhui69 commented 3 years ago

@mphillips-valleyit how is your custom data training result by transfer lerning ? could you attach you train log here?

ramonhollands commented 3 years ago

@glenn-jocher Might be interesting to do a final step by unfreezing and training the complete netwerk again with differentiated learning rate. So complete training process would be (default method in fast.ai):

  1. Freeze the backbone
  2. (optional reset the head weights)
  3. Train the head for a while
  4. Unfreeze the complete network
  5. Train the complete network with lower learning rate for backbone
glenn-jocher commented 3 years ago

@ramonhollands that's an interesting idea, though I'm sure the devil is in the details, such as the epochs you take these actions at, the LRs used, dataset and model etc. I don't have time to investigate further, but you should be able to reproduce the above tutorial and apply the extra steps you propose to quantify differences. If you do please share your results with us.

One point to mention is that classification and detection may not share a common set of optimal training steps, so what works for fast.ai may not correlate perfectly to detection architectures like YOLO. Would be very interested to see experimental results.

ramonhollands commented 3 years ago

Ill take that challenge the coming weeks. Trying to wrap your amazing work in the fast.ai framework to be able to use best of both worlds, including the fastai learning rate finder and discriminate learning rates etc. The method should work for detection architectures as well (https://www.youtube.com/watch?v=0frKXR-2PBY). Ill keep you updated.

aritzLizoain commented 3 years ago

I trained a model with an online dataset containing 5 categories, and now I'm trying to fine-tune it with my own images, which contain the same 5 categories plus an additional one. My images are similar to the ones from the online dataset, so I thought that transfer learning would work. However, this is what I obtain while fine-tuning:

Class = all, Images = 3, Targets = 0, P = 0, R = 0, mAP@.5 = 0, mAP@.5:.95 = 0

When I visualize the labels everything looks correct, so I don't understand is why Targets is 0. I also modified the dataset configuration yaml file adding the new category.

The fine-tuning works when I remove my additional category and fine-tune with the same 5 categories.

Does anybody know what am I doing wrong here?

Thanks in advance!

yushuinanrong commented 3 years ago

@aritzLizoain I have done something similar. In my case, I created a dataset that borrows certain classes from COCO and OpenImages. Then I fine tuned a pretrained yolov5 (trained on COCO) model on my custom dataset. The performance of the fine-tuned model isn't good.

glenn-jocher commented 3 years ago

@ramonhollands LR finder sounds very cool, but be careful because sometimes LRs that work well for training can cause instabilities without a warmup to ramp the LR from 0 to it's initial value.

@aritzLizoain no targets found during testing means no labels are found for your images. Follow the Custom training tutorial to create a custom dataset: https://docs.ultralytics.com/yolov5/tutorials/train_custom_data

ilyalasy commented 3 years ago

While using Transfer Learning (both with layers freeze and without), it happens that model "forgets" data it was trained on (metrics on original data are getting worse). So, I think problem might be in too large learning rate. Can you please give a little bit more details on what hyperparameters should be changed when finetuning the model? (maybe change lr0 to the last lr that was during original training and removing warmup epochs, or is it a wrong approach?)

tusharnitharwal commented 3 years ago

@glenn-jocher Might be interesting to do a final step by unfreezing and training the complete netwerk again with differentiated learning rate. So complete training process would be (default method in fast.ai):

  1. Freeze the backbone
  2. (optional reset the head weights)
  3. Train the head for a while
  4. Unfreeze the complete network
  5. Train the complete network with lower learning rate for backbone

How do you set a different learning rate for the backbone?

ramonhollands commented 3 years ago

You have to split the backbone and head parameters and add additional param groups for both with different 'lr' argument (https://pytorch.org/docs/stable/optim.html). I wrote some initial code which Ill post later today.

ramonhollands commented 3 years ago

See https://github.com/ramonhollands/different_learning_rates/blob/master/train.py

I have added two parameters to experiment with:

I started some experiments which where encouraging but did not have enough time to finish up yet.

laisimiao commented 3 years ago

@glenn-jocher I am curious about result pics in "Accuracy Comparison", why can the mAP of exp9_freeze_all increase as training progresses? Now that all params are frozed, they won't be optimized and performance should be a flat line?

glenn-jocher commented 3 years ago

@laisimiao exp9_freeze_all freezes all layer except output layer, which has an active gradient.

pragyan430 commented 3 years ago

Hey @glenn-jocher , How do I add replace the FC output layer and add a new one??

GiorgioSgl commented 3 years ago

I don't understand why the time to get train on the model with all frozen layers, it's more than the backbone frozen layer. It would not be the inverse??

turgut090 commented 3 years ago

Hi @glenn-jocher . I trained the model with icevision(a wrapper around this repo). However, I could not find out a proper way to export the model with export.py to TorchScript model. https://github.com/ultralytics/yolov5/blob/master/models/export.py Saving the weights of yolo5 throws the following:

torch.save(model.state_dict(), "models/yolo.pth")
python export.py --weights ~/Documents/imageai/models/yolo.pt
Namespace(weights='/home/turgut/Documents/imageai/models/yolo.pt', img_size=[640, 640], batch_size=1, device='cpu', include=['torchscript', 'onnx', 'coreml'], half=False, inplace=False, train=False, optimize=False, dynamic=False, simplify=False, opset_version=12)
YOLOv5 🚀 v5.0-180-ge8c5237 torch 1.7.1+cu110 CPU

Traceback (most recent call last):
  File "/home/turgut/Documents/yolov5/models/export.py", line 165, in <module>
    export(**vars(opt))
  File "/home/turgut/Documents/yolov5/models/export.py", line 47, in export
    model = attempt_load(weights, map_location=device)  # load FP32 model
  File "/home/turgut/Documents/yolov5/models/experimental.py", line 120, in attempt_load
    model.append(ckpt['ema' if ckpt.get('ema') else 'model'].float().fuse().eval())  # FP32 model
KeyError: 'model'

If I save the model, then I get:

torch.save(model, "model.pth")
Namespace(weights='/home/turgut/Documents/imageai/model.pth', img_size=[640, 640], batch_size=1, device='cpu', include=['torchscript', 'onnx', 'coreml'], half=False, inplace=False, train=False, optimize=False, dynamic=False, simplify=False, opset_version=12)
YOLOv5 🚀 v5.0-180-ge8c5237 torch 1.7.1+cu110 CPU

Traceback (most recent call last):
  File "/home/turgut/Documents/yolov5/models/export.py", line 165, in <module>
    export(**vars(opt))
  File "/home/turgut/Documents/yolov5/models/export.py", line 47, in export
    model = attempt_load(weights, map_location=device)  # load FP32 model
  File "/home/turgut/Documents/yolov5/models/experimental.py", line 119, in attempt_load
    ckpt = torch.load(attempt_download(w), map_location=map_location)  # load
  File "/home/turgut/.local/share/r-miniconda/envs/r-reticulate/lib/python3.9/site-packages/torch/serialization.py", line 594, in load
    return _load(opened_zipfile, map_location, pickle_module, **pickle_load_args)
  File "/home/turgut/.local/share/r-miniconda/envs/r-reticulate/lib/python3.9/site-packages/torch/serialization.py", line 853, in _load
    result = unpickler.load()
  File "/home/turgut/.local/share/r-miniconda/envs/r-reticulate/lib/python3.9/site-packages/torch/nn/modules/module.py", line 778, in __getattr__
    raise ModuleAttributeError("'{}' object has no attribute '{}'".format(
torch.nn.modules.module.ModuleAttributeError: 'Model' object has no attribute 'param_groups_fn'

Is there a way to sort this out?

glenn-jocher commented 3 years ago

@turgut090 see TorchScript, ONNX, CoreML Export tutorial:

YOLOv5 Tutorials

turgut090 commented 3 years ago

@glenn-jocher Thanks for your reply. Does this mean that I have to train a model only with this specific script? Here, I see: https://docs.ultralytics.com/yolov5/tutorials/model_export https://github.com/ultralytics/yolov5/blob/master/train.py in order to have checkpoints?

https://github.com/ultralytics/yolov5/blob/bb79e13d521c54b20b06555fe79cdff055f28721/train.py#L401-L408

Other options are yolov5m.pt, yolov5l.pt, and yolov5x.pt, or your own checkpoint from 
training a custom dataset runs/exp0/weights/best.pt.
glenn-jocher commented 3 years ago

@turgut090 all tutorials in this repo operate correctly with models trained in this repo.

Naturally we can't speak to 3rd party implementations, nor do we provide support for them at this time.

JihwanEom commented 3 years ago

@glenn-jocher

Hi, thank you for your awesome works!

I have a question about the freezing backbone layers in detail.

# Freeze 
 freeze = [f'model.{x}.' for x in range(freeze)]  # layers to freeze 
 for k, v in model.named_parameters(): 
     v.requires_grad = True  # train all layers 
     if any(x in k for x in freeze): 
         print(f'freezing {k}') 
         v.requires_grad = False ](`url`)

If you freeze like above, as my best knowledge, the running vars and means of batch normalization layer will be trained cause the module is not changed to eval mode like module.eval().

How do you think about this issue? I wonder it was intended.

Best, Jihwan Eom

glenn-jocher commented 3 years ago

@JihwanEom the batchnorm learnable parameters gamma and beta are frozen by the above code, stats continue computing as normal while in train mode. See https://pytorch.org/docs/stable/generated/torch.nn.BatchNorm2d.html

JihwanEom commented 3 years ago

@glenn-jocher

Thank you for kind explanation.

Then, the general freezing means only detaching gradients on parameters? Or include off tracking running means and vars?

glenn-jocher commented 3 years ago

@JihwanEom yes that's correct. The above freezing code only affects parameters and does not halt batchnorm stats updates (mean and var).

RamanHrynkevich commented 2 years ago

Hello! Thank you for your article. You can tell us in more detail how to add a new class to an already trained model.

That is, I trained the model to identify 5 objects, now I need the model to detect 6 classes, i.e. start finding another new one. How to do it correctly? I tried a frieze of 10 layers and 24, I tried to specify all 6 classes and only one new one in yaml for training, but the result was a model that finds the new class poorly and forgets the old ones.

glenn-jocher commented 2 years ago

@Partisanus you should train a model on all classes you want it to be able to detect on. See https://en.wikipedia.org/wiki/Catastrophic_interference

You can freeze or not freeze layers also, but this is a separate topic.

dhruvildave22 commented 2 years ago

I am using pre-trained(on 80 coco classes) yolov5s.pt model and training it to detect 3 more classes. I used this data.yml file when i started training.

train: /content/yolov5/datasets/animal-dataset/train/images
val: /content/yolov5/datasets/animal-dataset/valid/images

nc: 83
names: ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light',
        'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow',
        'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee',
        'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard',
        'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple',
        'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch',
        'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone',
        'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear',
        'hair drier', 'toothbrush', 'lion', 'frog', 'tiger']

according to above code i want to add lion, frog and tiger with index 80, 81 and 82 respectively. and as described by @glenn-jocher, i added --freeze 10. !python train.py --img 640 --batch 16 --epochs 5 --data data.yaml --weights yolov5s.pt --cache --freeze 10

but the problem i am facing is that now it only detects last 3 classes only not other 80 classes. I tried this also in data.yaml: names: ['lion', 'frog', 'tiger'] with nc: 83 but nothings seems to work.

Please can someone help and provide more clarification about this?

Thanks, Dhruvil Dave

glenn-jocher commented 2 years ago

@dhruvildave22 see https://community.ultralytics.com/t/how-to-combine-weights-to-detect-from-multiple-datasets

dhruvildave22 commented 2 years ago

@glenn-jocher I will try this also.

train: # train images (relative to 'path')
  - images/data_a
  - images/data_b

but i have one question here, I am using this coco dataset from roboflow https://public.roboflow.com/object-detection/microsoft-coco-subset. it already gives all the train and valid values in images-labels format. This dataset has at-least 118k train images and my custom dataset has only 1900 train images. My only question here: will this going to affect the prediction rate for my custom dataset since the number of images in my dataset is much lower than the coco dataset which i will be using?

glenn-jocher commented 2 years ago

@dhruvildave22 yes, just as you say one dominant dataset will receive more attention than a smaller dataset if you train both at the same time.

We have a parameter called --image-weights in train.py which will sample images based on their performance, so poorly performing images should be then overly represented and the problem you describe may be mitigated. This is highly experimental, but you can try it out and see if it helps your situation.

https://github.com/ultralytics/yolov5/blob/71621df87589faea19ba4c4098bb68e73201f30c/train.py#L487

Juanjojr9 commented 2 years ago

@glenn-jocher Might be interesting to do a final step by unfreezing and training the complete netwerk again with differentiated learning rate. So complete training process would be (default method in fast.ai):

  1. Freeze the backbone
  2. (optional reset the head weights)
  3. Train the head for a while
  4. Unfreeze the complete network
  5. Train the complete network with lower learning rate for backbone

Hi @ramonhollands @glenn-jocher . I would like to ask you about this experiment. First of all, what is the benefit of freezing only the backbone, training the network and then retraining the whole network? I've been searching but I can't find anything related to this. I would like to apply this technique, but I don't know what is the benefit of training it that way, instead of training the net only once with the backbone frozen. On the other hand, what are the disadvantages?

Based on your evidence, is it a correct way to apply transfer learning in this way? what results you obtained?

I hope you can help me. Best regards

jjjonathan14 commented 2 years ago

Hi I have two YOLOV5L models with face detection and hand detection seperately. Is it possible to merge one YOLOv5L model able to detect both face and hand?

glenn-jocher commented 2 years ago

@jjjonathan14 the correct workflow is to train a single model on multiple datasets to create a single face, hand model: https://community.ultralytics.com/t/how-to-combine-weights-to-detect-from-multiple-datasets

bryanbocao commented 2 years ago

@glenn-jocher Might be interesting to do a final step by unfreezing and training the complete netwerk again with differentiated learning rate. So complete training process would be (default method in fast.ai):

  1. Freeze the backbone
  2. (optional reset the head weights)
  3. Train the head for a while
  4. Unfreeze the complete network
  5. Train the complete network with lower learning rate for backbone

Hi @ramonhollands @glenn-jocher . I would like to ask you about this experiment. First of all, what is the benefit of freezing only the backbone, training the network and then retraining the whole network? I've been searching but I can't find anything related to this. I would like to apply this technique, but I don't know what is the benefit of training it that way, instead of training the net only once with the backbone frozen. On the other hand, what are the disadvantages?

Based on your evidence, is it a correct way to apply transfer learning in this way? what results you obtained?

I hope you can help me. Best regards

@Juanjojr9 The backbone mainly acts as a general visual feature extractor that "may" generalize across different datasets. So starting from a well-trained backbone can make your network have good initial parameters that (1) may generalize well and (2) can fasten training. Headers are mainly trained for specific tasks (such as regressing bounding boxes and classification on specific dataset. You might want to train header weights specific to your custom dataset.

jjjonathan14 commented 2 years ago

@jjjonathan14 the correct workflow is to train a single model on multiple datasets to create a single face, hand model: https://community.ultralytics.com/t/how-to-combine-weights-to-detect-from-multiple-datasets

thanks for the anwser. Is it possible to do meta learning with YOLOv5 models?

glenn-jocher commented 2 years ago

@jjjonathan14 what's meta learning?

shizhanhao commented 2 years ago

Dear author, could you save the parameters of each layer of yolov5 into a Pt file, and then use multiple Pt files to form a new model for reasoning? What knowledge do I need to learn if I want to achieve this effect? thank you

glenn-jocher commented 2 years ago

@shizhanhao I don't understand your question exactly but it doesn't sound like a feasible plan to me.

shizhanhao commented 2 years ago

@shizhanhao I don't understand your question exactly but it doesn't sound like a feasible plan to me. thank you!

DavidLandup0 commented 2 years ago

@jjjonathan14 what's meta learning?

Meta learning refers to a pretty broad area of "learning about learning" and can refer to multiple subfields. Usually, it refers to algorithms that learn about other algorithms. A classic example would be training 10 networks on the same task and data, and then having a "meta learner" learn to which of those networks to "listen to" for the final result. In essence - ensemble learning.

ghoshaw commented 2 years ago

I think that only set "requires_grad = False " is not enough for BN layer, because BN has another two params that do not need to calc grad, but they will change as the training. Is that ok?

glenn-jocher commented 2 years ago

@ghoshaw these are not parameters subject to weight decay.

feimadada commented 1 year ago

image Hi, I want to know what hardware and pytorch version do you use in this experiment?

simplecoderx commented 1 year ago

@glenn-jocher sir, when im doing my transfer learning in yolov5, the electricity cuts off and the training stops, my question is, can i still continue the training? can i use --resume? how to continue interrupted transfer learning?

simplecoderx commented 1 year ago

@glenn-jocher sir, when im doing my transfer learning in yolov5, the electricity cuts off and the training stops, my question is, can i still continue the training? can i use --resume? how to continue interrupted transfer learning?

Mr-ind1fferent commented 1 year ago

Ill take that challenge the coming weeks. Trying to wrap your amazing work in the fast.ai framework to be able to use best of both worlds, including the fastai learning rate finder and discriminate learning rates etc. The method should work for detection architectures as well (https://www.youtube.com/watch?v=0frKXR-2PBY). Ill keep you updated.

so have you figure out how to unfreeze backbone after epoch?

KhaKimThuy commented 1 year ago

I just want to detect person and motorcycle classes, how can I reduce the number of parameters?

image