Detr: Need suggestions to improve custom training using detectron2 wrapper

[theoutsider8060]

Hi, thank you for this great repo on DETR. I was trying to fine-tune DETR on my custom dataset using the detectron2 wrapper and I need some suggestions regarding that.

My dataset consists of around 14k training images and 5k validation images with a maximum of 30 objects per image. There is only a single class.

These are the changes I made to the default configuration:

1. I changed the num_classes to 1.
2. I only have 1 gpu with about 12 gb memory. So, I changed my batch size from 64 to 4 and also divided the default learning rate 1e-4 by 16.
3. My dataset is in standard COCO format (for each object, XY(topleft)_WH). I did not make any change to this annotation and feed it directly to the network.
4. In the default configuration file inside d2/configs, the pretrained weights were set to imagenet weights. Instead, I used pretrained weights on COCO using DETR from 'https://dl.fbaipublicfiles.com/detr/detr-r50-e632da11.pth' to initialize my custom DETR network having one class. In addition, I used the converted.py file to convert the pretrained weights into the format of the detectron2 wrapper. After doing this, all the parameters except for class_embed.weight and class_embed.bias were set and those two were skipped as shown below.

WARNING [07/31 15:27:46 fvcore.common.checkpoint]: Skip loading parameter 'detr.class_embed.weight' to the model due to incompatible shapes: (81, 256) in the checkpoint but (2, 256) in the model! You might want to double check if this is expected.
WARNING [07/31 15:27:46 fvcore.common.checkpoint]: Skip loading parameter 'detr.class_embed.bias' to the model due to incompatible shapes: (81,) in the checkpoint but (2,) in the model! You might want to double check if this is expected.
detr.class_embed.{weight, bias}
criterion.empty_weight

1. I trained the network for 70 epochs and I have attached the total training loss curve and the validation maP curve below.

Now, it seems my maP is stuck around 35-38 and not improving much. And, the training curve has plateaued as well. What can I do to make my model perform better? These are some questions I have off the top of my head:

1. I tried changing num_queries to 60 and trained for 50 epochs and my maP only became 6. Should I pursue this further? I do not have much computational resource to train for 300+ epochs.

2. Change the eos_coef? Will that help?

3. I am currently using the default augmentation. Should I add some extra augmentation to better generalize my dataset?

4. Is there something specific I need to change because I have only a single class? The loss values are still pretty high around 20-30 after 70 epoch.

Any help would be greatly appreciated. Thanks in advance.

[alcinos]

Hi @theoutsider8060
Thank you for your interest in DETR

I suggest you to have a look at these issues: #9, #124 and #125 to see if you can gain some insights.

Regarding your specific case, some remarks/questions:

• It seems that you are not dropping the learning rate. Try doing that (cf the Default Detectron2 config, you probably need to adjust the iteration point where it steps). Also, your scaling factor sounds too aggressive, it comes down to about 6e-6 for the transformer, which means that the backbone is going to be basically frozen (it will use learning rate 7e-7). Try starting from higher and dividing by a factor of 2 or 10 regularly.
• If you're finetuning, I don't recommend changing the number of queries, because that requires you to retrain the whole decoder
• How good/bad is 35 AP on your dataset? How much do you get with Faster-RCNN for eg?
• Data augmentation is dependent of what your data looks like. You should know best if some form of augmentations would make sense in your case
• The loss is the sum of the loss of all layers, so it's expected to be a bit high
• If you can't use more gpus in parallel, you should perhaps consider gradient accumulation.

Best of luck.

[theoutsider8060]

Thanks for the reply @alcinos . I will definitely look more into the issues that you mentioned.

With regards to results using other models, at the moment, I am getting AP of 40.2 using faster R-CNN with an inference time of 91 ms. By using DETR, my inference time per image has dropped to around 70 ms but the AP has also dropped as I mentioned.

Just a few queries before I make the change:

1. Since I will be going back to the default learning rate of 1e-4, can you provide any suggestion on exactly when I should drop it further and by how much?
2. I am not sure if I have got this wrong. But since I have only a single class, shouldn't my class_error drop down to zero as soon as I start training? While I was fine tuning, I noticed my class_error started at around 7 and now has steadily declined towards 0. Is this how it is supposed to be?

[alcinos]

1. Hard to tell, monitor training and drop by 2 or 10 when you feel it plateaus. On coco, for a quick run, one could do 150 epochs with drop at 100 (but this is from scratch). 1e-4 might be a bit high considering you have some unstability already. But try, and work from there.
2. you technically have two classes: the class you are trying to detect and the "no-object" class, this is why you don't get 0.

[Antoine-ls]

@theoutsider8060 Hello, I am now facing the same case, as I am using detectron2 wrapper, how do you change the class_nums to 1 as I did not see it in the yaml files ?

@alcinos I have a question about images transformation, in my case, all the images are having the same shape (1080,1440), do I still need to reshape it? And how could I change input part in yaml files like MIN_SIZE_TRAIN: (480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800), it means that images need to be reshaped into a square shape but not rectangular?

Thanks in advance !

[DanielFennhagen]

Hi @Antoine-ls, the second part of your question has been answered by @fmassa in https://github.com/facebookresearch/detr/issues/245#issuecomment-699491315

You should resize the input to have a minimum size of 800 (and ideally a max size of 1333), the model doesn't do the resizing inside it. Check our colab notebook for further details on how to perform the input preprocessing https://colab.research.google.com/github/facebookresearch/detr/blob/colab/notebooks/detr_attention.ipynb

[ruodingt]

@Antoine-ls

1. For changing class_nums

https://github.com/facebookresearch/detr/blob/4e1a9281bc5621dcd65f3438631de25e255c4269/d2/detr/config.py#L11

1. Re. image transformation.

Image will still be rectangular as MIN_SIZE_TRAIN is used along with MAX_SIZE_TRAIN (default to 1333)
You can this how these params are used in the function:

https://github.com/facebookresearch/detr/blob/4e1a9281bc5621dcd65f3438631de25e255c4269/d2/detr/dataset_mapper.py#L15

https://github.com/facebookresearch/detr/blob/4e1a9281bc5621dcd65f3438631de25e255c4269/d2/detr/dataset_mapper.py#L36

class ResizeShortestEdge(Augmentation):
    """
    Scale the shorter edge to the given size, with a limit of `max_size` on the longer edge.
    If `max_size` is reached, then downscale so that the longer edge does not exceed max_size.
    """

Hi @theoutsider8060 Any luck with your DETR training? I am in similar situation as you did.
Did you manage to get a similar performance compared with mask rcnn?

Regards,
Ruoding