1回答
- 新しい順
- 投票が多い順
- コメントが多い順
1
Hello Lorenzo,
In general, you can use Ground Truth augmented manifest to train a TensorFlow model with SageMaker using Pipe Mode. Augmented manifests can only support Pipe input mode.
So the specific example you asked here is not going to work using augmented manifest, because it's script mode. If you just meant to use the estimator, you can modify the code to something like this.
Please have a read on and follow this for more details.
Thanks Jann! What I would like to do is to train an object detector model using images annotated with Ground Truth. The model could be YOLO V5 or similar. I would like to have access to Tensorboard for metrics visualization. Do you have code examples to do that? Thanks!
Hello Lorenzo, Yes, you can follow the document below to set up the Tensorboard. Please note this will create an app in your domain so cost will occur.
"Amazon SageMaker with TensorBoard runs the TensorBoard application on an ml.r5.large instance and incurs charges after the SageMaker free tier or the free trial period of the feature. For more information, see Amazon SageMaker Pricing."
https://docs.aws.amazon.com/sagemaker/latest/dg/tensorboard-on-sagemaker.html
Thanks Jann. The problem seems to be that the Jumpstart models (e.g. in s3://jumpstart-cache-prod-us-east-1/source-directory-tarballs/tensorflow/transfer_learning/od1/v1.1.0/sourcedir.tar.gz) don't know how to open those annotation files it seems. I'm seeing the error:
The jumpstart code is still looking for the annotation.json file, and it probably wants to use that to load the annotations
In general, would the models in Jumpstart be able to load the images and annotations listed in the augmented manifest file produced by GroundTruth?
Unfortunately, it is not supported for Jumpstart models as they are expecting a different type of format. You may need to, which you probably have done, write a custom script to convert the format.