Training on top of MsCOCO+ Custom class

[abhigoku10]

@dbolya @fanyix @chongzhou96 @zllrunning Thanks for the wonderful code . I am having few queries
1.How much time did it too to train the network for Mscoco dataset from scratch
2. What are the types of augmentation you have applied
3. What is final number of images used as a training set
4. Can i freeze the layers and add additional classes to MsCOCO(from 80 to 90) and re-train it, if so whats is the procedure to do it

[dbolya]

I'm going to need more info than that, bruv.

[dbolya]

1. 4 Days (on 1 P100, should take ~5 on something like a Titan Xp, depending on your CPU).
2. Copying from data/config.py:

    # SSD data augmentation parameters
    # Randomize hue, vibrance, etc.
    'augment_photometric_distort': True,
    # Have a chance to scale down the image and pad (to emulate smaller detections)
    'augment_expand': True,
    # Potentialy sample a random crop from the image and put it in a random place
    'augment_random_sample_crop': True,
    # Mirror the image with a probability of 1/2
    'augment_random_mirror': True,

See utils/augmentations.py for implementation details.
3. ~118k (COCO's 2017 train split). I can fetch the exact number if you want.
4. Yup, and it's pretty easy too. Just follow the custom dataset procedure in the ReadMe and follow #36 to fine tune from the existing weights.

[abhigoku10]

thanks for the quick response , i have a query on point 5 for which i shall elaborate .
The yolact_base_54_800000.pth was trained on MS coco for 80 classes , i want to perform custom training which is of 20 classes along with 80 classes of MSCOCo , since training from scratch takes time i want to freeze the max conv layers so that the network will not forget the previous information of 80 classes along with it learn the next 20 classes. SO to perform this what is the procedure

[dbolya]

You don't need to freeze the previous weights, just load them in place of the pretrained imagenet weights. So essentially you start training from where I left off training when I uploaded those weights. That's what #36 does. Also now that you're fine-tuning, you don't need to train for a full 800k iterations so you can change max_iter to like 200k then change lr to start lower (let's say 1e-4). Thus it'll take 1/4 the time (these parameters aren't optimal so it might converge before then).

[abhigoku10]

Thanks for the response shall try and respond back to you

[abhigoku10]

Hi, I tried training the model as suggested, the training is executed without any errors but when I test the model it has learnt only the new 10 class but has forgotten or not predicting the other 80 classes from MSCoco. Can you please let me knw if I ahev to do anything else

[dbolya]

Did you add your new data to COCO and train with the combined training set, or did you just train with your new data? If the latter, then it's going to forget everything it learned about the old data (and there's no easy way to freeze some part so it doesn't forget atm so you have to train with both).

[abhigoku10]

Thanks for response, I did the later part so I am correcting it and going to re-train again with both mscoco data and my new data

[abhigoku10]

@dbolya thanks for the response now i am getting the expected results , but have few additional queries
1.Since the MScoco have huge data , for the model to learn the new custom data . what is the minimum number of dataset required
2. will there be any change in the performance of the base model when new data is added and re-trained , if so what should be done to obtain good accuracy
Thanks in advance

[dbolya]

1. Not sure. Perhaps around 100-1k examples per class? Maybe you can just sample 1/10th of the images randomly, or bring the number of examples in each of the original classes in line with your new classes.

2. Also not sure, since I haven't tested training on new classes. I'd expect a small drop in performance for the old classes. However, if the new classes are not that similar to the old ones, adding them probably wouldn't have much effect.

[annael99]

@abhigoku10 could i see the code where you train on the 90/100 classes? Also, i'd love to hear about the results you got (performance-wise and speed-wise) if it possible.
Thanks in advance