object_detection_tutorial.ipynb pretrained model "resnet50 imagenet"

[YonghangW]

thanks your work!
In the file called “object_detection_tutorial.ipynb”，exp_config = exp_factory.get_exp_config('retinanet_resnetfpn_coco') use the init_checkpoint called 'gs://cloud-tpu-checkpoints/vision-2.0/resnet50_imagenet/ckpt-28080', the weight file can be obtained through other channels if I don't download it through the goole cloud storage.

[Ayush-Gupta100]

Hi @YonghangW,

The checkpoint path
gs://cloud-tpu-checkpoints/vision-2.0/resnet50_imagenet/ckpt-28080
refers to a Google Cloud Storage (GCS) bucket, so the recommended way to access it is via gsutil.

You can download it locally using:

gsutil cp -r gs://cloud-tpu-checkpoints/vision-2.0/resnet50_imagenet .

If you prefer not to use GCS, there isn’t an official HTTP mirror for this exact checkpoint. In that case, you have a couple of options:

• Use a different pretrained ResNet50 (e.g., from Keras or TF Hub), though it may not be directly compatible with the RetinaNet config.

• Skip the checkpoint by setting:

  exp_config.task.model.backbone.init_checkpoint = None
  

  and train from scratch or provide your own checkpoint.

It might be helpful to clarify this in the tutorial (e.g., mentioning gsutil or alternative approaches), since it can be confusing for new users.

Hope this helps!

[YonghangW]

@Ayush-Gupta100 Thanks for your reply.I wonder if I can download the checkpoint through "https://storage.googleapis.com/cloud-tpu-checkpoints/vision-2.0/resnet50_imagenet/ckpt-28080.index" and "https://storage.googleapis.com/cloud-tpu-checkpoints/vision-2.0/resnet50_imagenet/ckpt-28080.data-00000-of-00001" because I used the downloaded checkpoint in the retinanet object detection and trained successfully.