search

google-coral / tutorials

Colab/Jupyter tutorials about training TensorFlow models for Edge TPU, and other tutorials
Apache License 2.0
181 stars 39 forks source link

SSD MobileNet V2 TD2.x Tutorial request #16

Closed ulrichMarco closed 3 years ago

ulrichMarco commented 3 years ago

Description

On the official coral object detection models website, there is listed a SSD MobileNet V2 model which is generated with tf 2.x. Is it possible to get an tutorial how to retrain this model? Its because i have an Nvidia RTX 3080 which only supports Cuda 11+, and TF1.15,which is used in the tutorials only supports till Cuda 10.x .

Issue Type

Documentation Feature Request

Operating System

Linux, Ubuntu

Coral Device

USB Accelerator, M.2 Accelerator with dual Edge TPU

Other Devices

No response

Programming Language

C++, Python 3.6

Relevant Log Output

No response

hjonnala commented 3 years ago

@ulrichMarco please check tutorials repo for MobileNet v2 tutorial with tf2.x

ulrichMarco commented 3 years ago

I did, there are only TF 1 tutorials for object detection. The one you linked is an image classification one.

Naveen-Dodda commented 3 years ago

Hi,

I can see your problem to use RTX 3080. I tired to build a TF2.x tutorial but have due few roadblocks, we have postpone the plan. But you can use following recommendations and try it your self it's fairly easy process to do.

  1. Main reason to drop TF2.x for tutorial is it supports less number of architectures suitable for edgetpu compared to TF1.x. you can take a look at supported backbones here https://github.com/tensorflow/models/blob/master/research/object_detection/builders/model_builder.py#L100

  2. You can use tutorial here https://github.com/tensorflow/models/blob/master/research/object_detection/colab_tutorials/eager_few_shot_od_training_tf2_colab.ipynb to train a model.

  3. in TF2.x you have to use Post Training Quantization to export a int8_tflite model.

  4. Finally compile it for edgetpu

Just two different steps to try it with TF2.x

Let us know if you have any questions. Happy Training

google-coral-bot[bot] commented 3 years ago

Have a few minutes? We'd love your feedback about the Coral developer experience! Take our 5-minute survey.

Are you satisfied with the resolution of your issue? Yes No