Repository navigation
object_detection_tutorial.ipynb pretrained model "resnet50 imagenet" #13634
Description
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.
Reacted by Santosh Kumar
Activity
Ayush-Gupta100 commentedon May 1, 2026
Hi @YonghangW,
The checkpoint path
gs://cloud-tpu-checkpoints/vision-2.0/resnet50_imagenet/ckpt-28080refers to a Google Cloud Storage (GCS) bucket, so the recommended way to access it is via
gsutil.You can download it locally using:
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:
and train from scratch or provide your own checkpoint.
It might be helpful to clarify this in the tutorial (e.g., mentioning
gsutilor alternative approaches), since it can be confusing for new users.Hope this helps!
YonghangW commentedon May 2, 2026
@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.