Hi Samuel,

Thanks for your fantastic work, even if TrivialAugment was a bit worse than all the other RL-based approaches, it should still be preferrable, given its very low cost and straightforward implementation, I think.

I am trying to understand your implementation in auglib.py in order to modify it to suit my purposes, and I am seeing that, in the standard search spaces, both fixed and wide, the geometric transforms inside ALL_TRANSFORMS (this is, rotate, translates, shears) all have a 50% chance of being applied. This is not true for the intensity-based transforms, though. I think (might be wrong) that this is not mentioned in the paper, e.g. in fig. 1 or the algorithm listing in page 4. So, may I ask if you used the 50% chance of switching off geometric transforms in your experiments?

Thanks!

Adrian

Hello!Hi, the following code was used directly in my model: transforms.Compose([transforms.TrivialAugmentWide(),transforms.RandomResizedCrop(224),transforms.RandomHorizontalFlip(), transforms. ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])]) after I found that both validation loss and training loss increased on mobilev2 and resnet50 , am I using the wrong method？The training set is the flower dataset

Is there an easy way to make a DataLoader, or for example ImageFolder, or to use something like transforms = [ToTensor(), TrivialAugment()] ? Because, a quick look let me to believe that TrivialAugment(img) is needed for the latter and this is not accepted as a standard torch Transform.

Have bug, described bellow:
File "../trivialaugment/TrivialAugment/train.py", line 353, in spawn_process assert worldsize == C.get()['gpus'], f"Did not specify the number of GPUs in Config with which it was started: {worldsize} vs {C.get()['gpus']}"

It happens due to incorrect unpacking "args.config". need to use "args.config[0]" instead "args.config" (line: 343-344). After this it work right :)

Hi, could you please tell me how did you train your model? Because I found in train.py that "aug_lib.blend_images" does not calculate loss during the training process, I don't know how the enhanced data adjusts the model. I would be very grateful if you could reply me.

Hi, firstly thank you for sharing this finding, it is really interesting to find that after going one big round, the simple solutions might work the best after all.

Saw that you all did point out at the end of the paper that TA does not extend well for Object Detection task. Is it possible to share more details regarding this, and what sort of tuning you all did?

Hello. Thanks for the work and code. Do not understand why do we need to have 30 bins in the end, why not just to sample augmentation strength uniformly in the specific intervals? Do we need a granularity? What is the intuinition behind this?

Hi, I'm currently trying to find the perfect trivial augmentation policy for my problem. When I apply the auto_contrast or sharpness at 2 strength levels and i display a few augmentations, i can't tell if there happens anything compared to the original image. I've already tried to rise the minVal but nothing changed. Here's a minimal working example

from PIL import Image
import aug_lib

orig_img = Image.open('pizza.jpeg')
aug_lib.set_augmentation_space(augmentation_space='wide_standard_custom', 
                                           num_strengths=2, 
                                           custom_augmentation_space_augs=["auto_contrast" ]) # or sharpness                                                                   
augmenter = aug_lib.TrivialAugment()
for _ in range(4):
    display(augmenter(orig_img))

Thanks in advance for your commitment!

Kind regards, Johannes