tensorflow / tpu

Reference models and tools for Cloud TPUs.

Home Page:https://cloud.google.com/tpu/

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Missing checkpoint files for ResNet-RS models.

sebastian-sz opened this issue · comments

It appears that checkpoint files are missing from provided tar archives.
On the models page we can see 11 checkpoint variants.

If we download any checkpoint after (including) ResNet-RS-152 256x256 we will get an error:

!mkdir checkpoint_dir
!curl https://storage.googleapis.com/cloud-tpu-checkpoints/resnet-rs/resnet-rs-152-i256.tar.gz | tar xz -C checkpoint_dir

import tensorflow as tf
variables = tf.train.list_variables("checkpoint_dir")

>>>ValueError: Couldn't find 'checkpoint' file or checkpoints in given directory checkpoint_dir

If we inspect the directory we can see, that the checkpoint file is missing.

Ok, so I figured out the checkpoint file has always the same structure:

model_checkpoint_path: "model.ckpt"
all_model_checkpoint_paths: "model.ckpt"

So you can just copy this file from other checkpoints or create it and paste the above lines.
After that the loading should work.