This thread is aimed to highlight new released models on TensorFlow Hub
9 Likes
New BERT model available:
MobileBERT is a compact task-agnostic BERT for Resource limited devices.
This version is an upgrade on the previous version, compatible with Preprocess model and also provides trainable Masked Language Model!
2 Likes
Mobilenet V3 is now available!
There are 8 new models, 4 [large, small]x[1.0, 0.75] for classification and their respective 4 feature vectors
They have also their TFLite and TF.js versions available
You can find links to all of them on this collection
2 Likes
This is indeed great. Would appreciae a TFLite variant as well. I did try to make it work (an earlier version) but the resulting model was spurious.
Here’s where I have logged all the progress: GitHub - sayakpaul/BERT-for-Mobile: Compares the DistilBERT and MobileBERT architectures for mobile deployments..
My idea was to author a blog post on “Optimizing different BERTs for Mobile” or something like that but didn’t happen because of lack of proper comparison metrics 
3 Likes
Next-Generation Pose Detection with MoveNet
The model has two versions: Lighting and Thunder and both have the regular SavedModel version and also the TensorFlow.js version ready to use.
They are quite fast and with great accuracy. You can try them out on this guide:
[ MoveNet: Ultra fast and accurate pose detection model]
(MoveNet: Ultra fast and accurate pose detection model. | TensorFlow Hub)
3 Likes
Recently the FRILL models were published on TFHub!
These are small, fast and great for on-device human non-semantic speech sound classification like detecting when someone is speaking with or without a mask and related tasks!
if you want even more information there’s also this nice blog post explaining more details on how the model is trained: FRILL: On-Device Speech Representations using TensorFlow-Lite
1 Like
The new EfficientNet V2 models are all available on TensorFlow Hub!
- Collection: Google | efficientnet_v2 | Kaggle
- Inference only tutorial: Image Classification with TensorFlow Hub
- Transfer Learning Tutorial: Retraining an Image Classifier | TensorFlow Hub
go check it out and let us know what you build with them!
2 Likes
MLP-Mixer [1] models were added: Sayak | mlp-mixer | Kaggle.
Vision Transformer [2] models were added: Sayak | vision_transformer | Kaggle.
A total of 28 models was added with Colab Notebooks on Transfer Learning and Inference.
In this new year, I hope the TF-Hub team involves itself a bit more deeply with contributors who tackle this kind of model contributions all by themselves. There are a lot of scopes where the team could be really helpful:
- Review the model conversion code. In this case, the above-mentioned models were converted from JAX.
- Help test the model on validation benchmarks to ensure performance does not drop.
- Review the transfer learning and model inference code.
- Help reach the original model authors if needed.
Apologies if it’s too much to ask 
References
[1] [2105.01601] MLP-Mixer: An all-MLP Architecture for Vision
[2] [2010.11929] An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
2 Likes
loss_reduction=tf.compat.v2.keras.losses.Reduction.SUM_OVER_BATCH_SIZE):
using tensorflow-hub gives keras is not found error
When i import keras i get
~\anaconda3\envs\fatma\lib\site-packages\tensorflow\python\eager\monitoring.py in __init__(self, metric_name, metric_methods, label_length, *args)
133 self._metric_name, len(self._metric_methods)))
134
--> 135 self._metric = self._metric_methods[self._label_length].create(*args)
136
137 def __del__(self):
AlreadyExistsError: Another metric with the same name already exists.
for which model? can you post a code snippet?