Stable Code Technical Report
Paper
•
2404.01226
•
Published
•
1
Flax Community
non-profit
AI & ML interests
JAX, Flax, TPU, 🤗
Recent Activity
flax-community's activity
ncoop57
authored
2
papers
5 days ago
Post
1130
My latest project is the outcome of the last 2+ years working with TPUs from the amazing TPU Research Cloud (TRC) program and training Encoder-only LMs with the TensorFlow Model Garden library.
👉 Link: https://github.com/stefan-it/model-garden-lms
An overview of some features:
- Cheatsheet for setting-up a TPU VM Pod (with all necessary dependencies) to pretrain LMs with TF Model Garden
- Conversion scripts that convert TF Model Garden weights to Hugging Face Transformers-compatible models
- Supported architectures include BERT, BERT with Token Dropping and TEAMS
I also released BERT-based models pretrained on the great Hugging Face FineWeb and FineWeb-Edu datasets (10BT subset). With more to come!
👉 Model Hub Link: https://huggingface.co/model-garden-lms
If you find these resources useful, please give them a like!
Made from Bavarian Oberland with ❤️ and 🥨.
👉 Link: https://github.com/stefan-it/model-garden-lms
An overview of some features:
- Cheatsheet for setting-up a TPU VM Pod (with all necessary dependencies) to pretrain LMs with TF Model Garden
- Conversion scripts that convert TF Model Garden weights to Hugging Face Transformers-compatible models
- Supported architectures include BERT, BERT with Token Dropping and TEAMS
I also released BERT-based models pretrained on the great Hugging Face FineWeb and FineWeb-Edu datasets (10BT subset). With more to come!
👉 Model Hub Link: https://huggingface.co/model-garden-lms
If you find these resources useful, please give them a like!
Made from Bavarian Oberland with ❤️ and 🥨.
paws
authored
a
paper
21 days ago
Post
1310
There's a new
New optimizers include:
* AdafactorBigVision -
* ADOPT -
* MARS -
* LaProp -
* Cautious Optimizers - a modification to all of the above, prefix with
I shared some caution comparisons in this model repo: rwightman/timm-optim-caution
For details, references, see the code: https://github.com/huggingface/pytorch-image-models/tree/main/timm/optim
timm release, v 1.0.12, with a focus on optimizers. The optimizer factory has been refactored, there's now a timm.optim.list_optimizers() and new way to register optimizers and their attributes. As always you can use an timm optimizer like a torch one, just replace torch.optim with timm.optimNew optimizers include:
* AdafactorBigVision -
adfactorbv* ADOPT -
adopt / adoptw (decoupled decay)* MARS -
mars* LaProp -
laprop* Cautious Optimizers - a modification to all of the above, prefix with
c as well as cadamw, cnadamw, csgdw, clamb, crmsproptf I shared some caution comparisons in this model repo: rwightman/timm-optim-caution
For details, references, see the code: https://github.com/huggingface/pytorch-image-models/tree/main/timm/optim
- 3 replies
Post
1279
I'm currently on a push to expand the scope of image based datasets on the Hub. There's certainly a lot already, but for anyone who's looked closely, there's not a whole lot of standardization. I am to fix that, datasets under the https://huggingface.co/timm and https://huggingface.co/pixparse orgs will serve as canonical examples for various task / modality combinations and be useable without fuss in libraries like
I just uploaded the first multi-label dataset that I'll support with
Next up object detection & segmentation! I've got an annotation spec sorted out, a lot of datasets ready to rip, and yeah that means
timm, OpenCLIP, and hopefully more.I just uploaded the first multi-label dataset that I'll support with
timm scripts soon:
timm/plant-pathology-2021 Next up object detection & segmentation! I've got an annotation spec sorted out, a lot of datasets ready to rip, and yeah that means
timm support for object detection, eventually segmentation, is finally under development :OPost
1056
Want to validate some hparams or figure out what
I had this sitting on my drive and forgot where I pulled it together from. It's 100 classes of imagenet, 50k train and 10k val images (from ImageNet-1k train set), and 5k test images (from ImageNet-1k val set). 7.4GB instead of > 100GB for the full ImageNet-1k. This ver is not reduced resolution like some other 'mini' versions. Super easy to use with timm train/val scripts, checkout the dataset card.
I often check fine-tuning with even smaller datasets like:
* timm/resisc45
* timm/oxford-iiit-pet
But those are a bit small to train any modest size model w/o starting from pretrained weights.
timm model to use before commiting to download or training with a large dataset? Try mini-imagenet:
timm/mini-imagenet I had this sitting on my drive and forgot where I pulled it together from. It's 100 classes of imagenet, 50k train and 10k val images (from ImageNet-1k train set), and 5k test images (from ImageNet-1k val set). 7.4GB instead of > 100GB for the full ImageNet-1k. This ver is not reduced resolution like some other 'mini' versions. Super easy to use with timm train/val scripts, checkout the dataset card.
I often check fine-tuning with even smaller datasets like:
* timm/resisc45
* timm/oxford-iiit-pet
But those are a bit small to train any modest size model w/o starting from pretrained weights.
Post
1580
New MobileNetV4 weights were uploaded a few days ago -- more ImageNet-12k training at 384x384 for the speedy 'Conv Medium' models.
There are 3 weight variants here for those who like to tinker. On my hold-out eval they are ordered as below, not that different, but the Adopt 180 epochs closer to AdamW 250 than to AdamW 180.
* AdamW for 250 epochs - timm/mobilenetv4_conv_medium.e250_r384_in12k
* Adopt for 180 epochs - timm/mobilenetv4_conv_medium.e180_ad_r384_in12k
* AdamW for 180 epochs - timm/mobilenetv4_conv_medium.e180_r384_in12k
This was by request as a user reported impressive results using the 'Conv Large' ImagNet-12k pretrains as object detection backbones. ImageNet-1k fine-tunes are pending, the weights do behave differently with the 180 vs 250 epochs and the Adopt vs AdamW optimizer.
There are 3 weight variants here for those who like to tinker. On my hold-out eval they are ordered as below, not that different, but the Adopt 180 epochs closer to AdamW 250 than to AdamW 180.
* AdamW for 250 epochs - timm/mobilenetv4_conv_medium.e250_r384_in12k
* Adopt for 180 epochs - timm/mobilenetv4_conv_medium.e180_ad_r384_in12k
* AdamW for 180 epochs - timm/mobilenetv4_conv_medium.e180_r384_in12k
This was by request as a user reported impressive results using the 'Conv Large' ImagNet-12k pretrains as object detection backbones. ImageNet-1k fine-tunes are pending, the weights do behave differently with the 180 vs 250 epochs and the Adopt vs AdamW optimizer.
infinitylogesh
authored
a
paper
about 1 month ago
w11wo
authored
a
paper
about 2 months ago
Post
647
A new
Featured in the release are:
* The MambaOut model, a cheeky arch inspired by SSM but without the SSM part, a ConvNeXt with gating.
* Several timm trained MambaOut variations with arch tweaks and ImageNet-12k pretrain to verify scaling, supplement ported weights.
* The smallest MobileNetV4, a 0.5x width scaled Conv-Small.
* Two impressive MobileNetV3 Large models outperforming all previous, using MNV4 Small recipe.
* 'Zepto,' a new compact ConvNeXt variant even smaller than the previous Atto, 2.2M params, RMSNorm, and solid results for its size.
* Newly ported SigLIP SO400M/16 ViT multi-lingual weights, the largest i18n weights, prevous was B/16.
* Two ImageNet-1k fine-tuned SigLIP SO400M models at 378x378
* InternViT 300M weight port. A really solid ViT encoder distilled from OpenGVLab 6B VL model encoder.
* An assortment of very small, sub 1M param pretrained test models to improve library unit tests and serve low-resource applications.
timm release (1.0.11) is out now. A also wrote an article on one of the included models: https://huggingface.co/blog/rwightman/mambaoutFeatured in the release are:
* The MambaOut model, a cheeky arch inspired by SSM but without the SSM part, a ConvNeXt with gating.
* Several timm trained MambaOut variations with arch tweaks and ImageNet-12k pretrain to verify scaling, supplement ported weights.
* The smallest MobileNetV4, a 0.5x width scaled Conv-Small.
* Two impressive MobileNetV3 Large models outperforming all previous, using MNV4 Small recipe.
* 'Zepto,' a new compact ConvNeXt variant even smaller than the previous Atto, 2.2M params, RMSNorm, and solid results for its size.
* Newly ported SigLIP SO400M/16 ViT multi-lingual weights, the largest i18n weights, prevous was B/16.
* Two ImageNet-1k fine-tuned SigLIP SO400M models at 378x378
* InternViT 300M weight port. A really solid ViT encoder distilled from OpenGVLab 6B VL model encoder.
* An assortment of very small, sub 1M param pretrained test models to improve library unit tests and serve low-resource applications.
gkuwanto
authored
a
paper
2 months ago
esantus
authored
a
paper
2 months ago
avramandrei
authored
a
paper
2 months ago
Post
2509
A 'small' MobileNet-V4 update, I just pushed weights for the smallest model I've trained in the series, a 0.5 width multiplier version of the MobileNet-V4 Conv Small.
Now you may look at this and say hey, why is this impressive? 64.8% top-1 and 2.2M params? MobileNetV3-Small 0.75, and MobileNet-V2 0.5 are both fewer params (at ~2M) and over 65% top-1, what gives? Well this is where MobileNet-V4 differs from the previous versions of the model family, it trades off (gives up) a little parameter efficiency for some computational efficiency.
So, let's look at the speed. On a 4090 w/ torchcompile
* 98K img/sec - timm/mobilenetv4_conv_small_050.e3000_r224_in1k
* 58K img/sec - timm/mobilenetv3_small_075.lamb_in1k
* 37K img/sec - timm/mobilenetv2_050.lamb_in1k
And there you go, if you have a need for speed, MNV4 is the better option.
Now you may look at this and say hey, why is this impressive? 64.8% top-1 and 2.2M params? MobileNetV3-Small 0.75, and MobileNet-V2 0.5 are both fewer params (at ~2M) and over 65% top-1, what gives? Well this is where MobileNet-V4 differs from the previous versions of the model family, it trades off (gives up) a little parameter efficiency for some computational efficiency.
So, let's look at the speed. On a 4090 w/ torchcompile
* 98K img/sec - timm/mobilenetv4_conv_small_050.e3000_r224_in1k
* 58K img/sec - timm/mobilenetv3_small_075.lamb_in1k
* 37K img/sec - timm/mobilenetv2_050.lamb_in1k
And there you go, if you have a need for speed, MNV4 is the better option.
Post
1288
The
Also worth pointing out, there are three rather newish 'test' models that you'll see at the top of any samples/sec comparison:
* test_vit ( timm/test_vit.r160_in1k)
* test_efficientnet ( timm/test_efficientnet.r160_in1k)
* test_byobnet ( timm/test_byobnet.r160_in1k, a mix of resnet, darknet, effnet/regnet like blocks)
They are < 0.5M params, insanely fast and originally intended for unit testing w/ real weights. They have awful ImageNet top-1, it's rare to have anyone bother to train a model this small on ImageNet (the classifier is roughly 30-70% of the param count!). However, they are FAST on very limited hadware and you can fine-tune them well on small data. Could be the model you're looking for?
timm leaderboard
timm/leaderboard has been updated with the ability to select different hardware benchmark sets: RTX4090, RTX3090, two different CPUs along with some NCHW / NHWC layout and torch.compile (dynamo) variations. Also worth pointing out, there are three rather newish 'test' models that you'll see at the top of any samples/sec comparison:
* test_vit ( timm/test_vit.r160_in1k)
* test_efficientnet ( timm/test_efficientnet.r160_in1k)
* test_byobnet ( timm/test_byobnet.r160_in1k, a mix of resnet, darknet, effnet/regnet like blocks)
They are < 0.5M params, insanely fast and originally intended for unit testing w/ real weights. They have awful ImageNet top-1, it's rare to have anyone bother to train a model this small on ImageNet (the classifier is roughly 30-70% of the param count!). However, they are FAST on very limited hadware and you can fine-tune them well on small data. Could be the model you're looking for?
Post
2060
The latest timm validation & test set results are now viewable by a leaderboard space:
timm/leaderboard
As of yesterday, I updated all of the results for ImageNet , ImageNet-ReaL, ImageNet-V2, ImageNet-R, ImageNet-A, and Sketch sets. The csv files can be found in the GH repo https://github.com/huggingface/pytorch-image-models/tree/main/results
Unfortunately the latest benchmark csv files are not yet up to date, there are some gaps in dataset results vs throughput/flop numbers impact the plots.
h/t to @MohamedRashad for making the first timm leaderboard.
As of yesterday, I updated all of the results for ImageNet , ImageNet-ReaL, ImageNet-V2, ImageNet-R, ImageNet-A, and Sketch sets. The csv files can be found in the GH repo https://github.com/huggingface/pytorch-image-models/tree/main/results
Unfortunately the latest benchmark csv files are not yet up to date, there are some gaps in dataset results vs throughput/flop numbers impact the plots.
h/t to @MohamedRashad for making the first timm leaderboard.
- 1 reply
Post
1989
I can't resist an opportunity to update an old baseline. Read a new article on my latest look at improving MobileNet-V1 and EfficientNet-B0 baselines.
https://huggingface.co/blog/rwightman/mobilenet-baselines
timm/mobilenetv1_100.ra4_e3600_r224_in1k
timm/efficientnet_b0.ra4_e3600_r224_in1k
https://huggingface.co/blog/rwightman/mobilenet-baselines
timm/mobilenetv1_100.ra4_e3600_r224_in1k
timm/efficientnet_b0.ra4_e3600_r224_in1k
LouisCastricato
authored
a
paper
5 months ago
saattrupdan
authored
a
paper
5 months ago