Tweetovi

	Mingxing Tan proslijedio/la je tweet
	Jeff Dean @JeffDean	9. sij
	What I did over my winter break! It gives me great pleasure to share this summary of some of our work in 2019, on behalf of all my colleagues at @GoogleAI & @GoogleHealth. ai.googleblog.com/2020/01/google…
	Prikaz sažetka ·

	Mingxing Tan @tanmingxing	4. pro
	This video explains AdvProp. Thanks @CShorten30 ! twitter.com/CShorten30/sta…
	Prikaži podatke ·

	Mingxing Tan @tanmingxing	2. pro
	Odgovor korisniku/ci @wightmanr @A_K_Nain i 4 ostali My first principle is to reduce model size and computation. However, I agree hardware have different demands: e.g., depthwise conv is fast on CPU, but slow on GPU. In future, we may apply NAS to design GPU-specific models. If you have good ideas on that, please let me know :)
	Prikaz razgovora ·

	Mingxing Tan @tanmingxing	2. pro
	Odgovor korisniku/ci @wightmanr @A_K_Nain i 4 ostali You might want to compare EfficientNet-B0 with ResNet101 , since they have similar ImageNet acc (77.3%). EfficientNets are more friendly to mobile CPU or accelerators (~10x faster than ResNet-50 on EdgeTPU: bit.ly/34FxQk6). More optimizations needed for GPUs.
	Prikaz razgovora ·

	Mingxing Tan @tanmingxing	29. stu
	Odgovor korisniku/ci @assthiam19 @quocleix @ruomingpang Yes, you can. They are tflite compatible.
	Prikaz razgovora ·

	Mingxing Tan @tanmingxing	26. stu
	Odgovor korisniku/ci @klimov_k_v @quocleix Hi Kirill, good question! we use separate scale/offset and stats.
	Prikaz razgovora ·

	Mingxing Tan proslijedio/la je tweet
	Ross Wightman @wightmanr	23. stu
	Odgovor korisniku/ci @tanmingxing @quocleix The TLDR of the paper; use adversarial examples as training data augmentation, maintain separate BatchNorm for normal vs adversarial examples. Neat. As usual I've ported & tested #PyTorch weights github.com/rwightman/gen-…
	Prikaz razgovora ·

	Mingxing Tan @tanmingxing	25. stu
	Can adversarial examples improve image recognition? Check out our recent work: AdvProp, achieving ImageNet top-1 accuracy 85.5% (no extra data) with adversarial examples! Arxiv: arxiv.org/abs/1911.09665 Checkpoints: git.io/JeopW pic.twitter.com/bAu054LGt2
	Prikaz fotografije ·

	Mingxing Tan @tanmingxing	23. stu
	Odgovor korisniku/ci @jeremyphoward @quocleix @ruomingpang Thanks!
	Prikaz razgovora ·

	Mingxing Tan @tanmingxing	22. stu
	Excited to share our work on efficient neural architectures for object detection! New state-of-the-art accuracy (51 mAP on COCO for single-model single-scale), with an order-of-magnitude better efficiency! Collaborated with @quocleix and @ruomingpang. twitter.com/quocleix/statu…
	Prikaži podatke ·

	Mingxing Tan proslijedio/la je tweet
	Quoc Le @quocleix	12. stu
	Odgovor korisniku/ci @quocleix Full comparison against state-of-the-art on ImageNet. Noisy Student is our method. Noisy Student + EfficientNet is 11% better than your favorite ResNet-50 😉 pic.twitter.com/BhwgJvSOYK
	Prikaz razgovora ·

	Mingxing Tan proslijedio/la je tweet
	Quoc Le @quocleix	12. stu
	Want to improve accuracy and robustness of your model? Use unlabeled data! Our new work uses self-training on unlabeled data to achieve 87.4% top-1 on ImageNet, 1% better than SOTA. Huge gains are seen on harder benchmarks (ImageNet-A, C and P). Link: arxiv.org/abs/1911.04252 pic.twitter.com/0umSnX7wui
	Prikaz fotografije ·

	Mingxing Tan @tanmingxing	4. stu
	Odgovor korisniku/ci @jeremyphoward @wightmanr @quocleix Good to know! How much improvement did you get? IIRC, by changing the last stride=2 to stride=1 (final output shape becomes 14x14) for Efficientnet-B0, you would get around 1% accuracy gain, with the cost of more computations in the last few layers.
	Prikaz razgovora ·

	Mingxing Tan @tanmingxing	4. stu
	Odgovor korisniku/ci @jeremyphoward @wightmanr @quocleix Excellent question! It is more like a tradition. However, later stage layers need a lot of params, so it is better to not add many extra layers in later stage if #params is a concern. EfficientNets use the same tradition, mostly to keep the scaling simple and to minimize params
	Prikaz razgovora ·

	Mingxing Tan proslijedio/la je tweet
	Jeff Dean @JeffDean	26. lis
	Great to see this collaboration between Google researchers & engineers launch, with major improvement to search quality! The work brings together many things we've been working on over the last few years: Transformers, BERT, @TensorFlow, TPU pods, ... blog.google/products/searc…
	Prikaz sažetka ·

	Mingxing Tan @tanmingxing	18. lis
	AutoML for video neural architecture design. Results are quite promising! twitter.com/GoogleAI/statu…
	Prikaži podatke ·

	Mingxing Tan @tanmingxing	19. ruj
	Odgovor korisniku/ci @ajlavin @quocleix Hi Andrew, Good point! I will update the paper to make it explicit (also replied on Github). FYI, the purpose of group conv is to reduce the FLOPS in pointwise convs, so we have more FLOPS budge for bigger kernel sizes (since the total FLOPS is constrained).
	Prikaz razgovora ·

	Mingxing Tan @tanmingxing	7. kol
	Odgovor korisniku/ci @karanchahal96 @quocleix Hi Karanbir, AutoML has evolved: recent algorithms (such as DARTS arxiv.org/abs/1806.09055) can finish a search in a couple of GPU hours.
	Prikaz razgovora ·

	Mingxing Tan @tanmingxing	6. kol
	Introducing EfficientNet-EdgeTPU: customized for mobile accelerators, with higher accuracy and 10x faster inference speed. blog post: ai.googleblog.com/2019/08/effici… Code and pertained models: github.com/tensorflow/tpu… twitter.com/GoogleAI/statu… pic.twitter.com/Vbj6aRHQMi
	Prikaz fotografije ·

	Mingxing Tan @tanmingxing	1. kol
	Odgovor korisniku/ci @_wkong @quocleix Depends on how many TPUs are used. Usually takes a couple of days.
	Prikaz razgovora ·