Checkpoints does not seem to change (eager tflite example)?

TensorFlow General Discussion
, , ,
1

Looking at the example here Few Shot Object Detection for TensorFlow Lite the checkpoint files that are generated seems to be the same? At least their file size appear to be equal:

drwxr-xr-x 3 root root     4096 Sep  5 10:34 ..
-rw-r--r-- 1 root root 10476721 Sep  5 10:38 ckpt-2.data-00000-of-00001
-rw-r--r-- 1 root root    26198 Sep  5 10:38 ckpt-2.index
drwxr-xr-x 2 root root     4096 Sep  5 10:38 .ipynb_checkpoints
-rw-r--r-- 1 root root 10476721 Sep  5 10:39 ckpt-3.data-00000-of-00001
-rw-r--r-- 1 root root    26198 Sep  5 10:39 ckpt-3.index
-rw-r--r-- 1 root root 10476721 Sep  5 10:40 ckpt-4.data-00000-of-00001
-rw-r--r-- 1 root root    26198 Sep  5 10:40 ckpt-4.index
-rw-r--r-- 1 root root 10476721 Sep  5 10:40 ckpt-5.data-00000-of-00001
-rw-r--r-- 1 root root    26198 Sep  5 10:40 ckpt-5.index
-rw-r--r-- 1 root root 10476721 Sep  5 10:41 ckpt-6.data-00000-of-00001
-rw-r--r-- 1 root root    26198 Sep  5 10:41 ckpt-6.index
-rw-r--r-- 1 root root      519 Sep  5 10:41 checkpoint

The only thing I have changed is that I have put the saving of checkpoint for each 100 step like so and set the number of checkpoints to keep (max_to_keep) to 5:

  if idx % 100 == 0:
    print('batch ' + str(idx) + ' of ' + str(num_batches)
    + ', loss=' +  str(total_loss.numpy()), flush=True)
    ckpt_manager.save()

Shouldn’t the file size of the checkpoints differ since the weight values changes?

5

To answer this.
Even thought the size of the checkpoint does not change, the values of the variables in the checkpoint file in fact does change.