Integrate TResNet to Object Detection

[thuyngch]

Hi authors, Thanks for your great work of extremely fast TResNet.

TResNet is demonstrated to be excellent in Image Classification. I am curious about its robustness in Object Detection. Currently, there are several frameworks pushing research on Object Detection, like mmdetection and detectron.

So, do you have any plan to integrate your great work with these frameworks? Also, I see that you are working on incorporating with Rwrightman in PyTorch-Image-Model in order to add TResNet into that framework. I appreciate your work and hope to see the result in Object Detection.

Thanks,

[mrT23]

Hi Thuy Ng

In my workplace, we have been using TResNet for quite some time for single-label and multi-label classification tasks, with ultra-sota scores.

recently, we used tresnet_m as a replacement for resnet50 in object detection, and we did saw improvement in our Coco score. We are also using TResNet as a basis for text-to-image searching tasks.

I admit that i am not going to open merge request for TResNet in open-source Object Detection frameworks. it's a tedious job, and i would love if someone from the community, who tried and experienced himself the advantages of TResNet models, would do it.

In terms of further promoting TResNet, i am now participating in two Kaggle competitions. i expect to be in the top-3 places in both of them. once the results will be published, it will give a further boost (i hope) to TResNet. the only reason i am so optimistic about my position in the competitions (there are hundreds of participants), is that i have better backbone models then the other competitors

Tal

[thuyngch]

Hi Tal,

I am very impressed by TResNet results on transfer tasks. Can you share more about the improvement in transfer tasks, compared with baseline (ResNet50)? Or do you plan to update more transfer results in your paper?

Thanks, Thuy

[mrT23]

Hi Thuy

regarding TResNet Vs ResNet50 - tresnet_m outperform resnet50 on all downstream datasets. for fine-grain classification datasets, the margin was quite large, due to the usage of AA layer in TResNet, which is a major force in fine-grain classification.

i do not plan on adding more specific details about the way we trained the downstream (transfer) datasets. it contains some inner tricks, and these things are almost never discussed in detail in academic articles (see for example transfer learning section in efficientNet of Gpipe papers - hardly 4 lines).

However, i can refer you to a public talk i gave two weeks ago, that contain more details about the "special sauce" that makes deep learning actually work: https://drive.google.com/file/d/1xPfa3XpqTZTeCrpuJtosqEGNDbmD5M9Q/view

I can also refer you to some more insights about TResNet and its features: https://miil-public-eu.oss-eu-central-1.aliyuncs.com/model-zoo/tresnet/TResnet_discussion.pdf

Tal

[thuyngch]

Thank you very much!

[mrT23]

@thuyngch just to give you proper numbers for object detection: on of my teammates now reported in his weekly that he replaced resnet50 with tresnet_m, and got a bump in is COCO mAP score from 42.8 to 45.2, which is a quite significant bump.

Tal

[thuyngch]

Hi Tal,

That sounds great! Is it the RetinaNet/FCOS or other architecture?

[mrT23]

FCOS

On Thu, Apr 30, 2020 at 5:09 PM Thuy Ng notifications@github.com wrote:

Hi Tal,

That sounds great! Is it the RetinaNet/FCOS or other architecture?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/mrT23/TResNet/issues/14#issuecomment-621876272, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFBXQDFPCVXWPS3EGPXNQYTRPGA73ANCNFSM4MQQV7MQ .