File system scheme '[local]' not implemented in Google Colab TPU

For loading file from local file when using TPU - read them as normal python file.read() (not tf.io). In your case:

def load_image(image_path):
    with open(image_path, "rb") as local_file: # <= change here
      img = local_file.read()
    img = tf.image.decode_jpeg(img, channels=3)
    img = tf.image.resize(img, (299, 299))
    img = tf.keras.applications.inception_v3.preprocess_input(img)
    return img, image_path
load_image('/content/train2017/000000000009.jpg')

Cloud TPUs can only access data in GCS as only the GCS file system is registered. Please see: https://cloud.google.com/tpu/docs/troubleshooting#cannot_use_local_filesystem for more details.

Though for checkpointing starting with TF 2.3 release you should be able to use the experimental_io_device='/job:localhost' option (https://www.tensorflow.org/api_docs/python/tf/train/CheckpointOptions) to store/load your checkpoints to and from your Colab runtime. Even with that API though you'll need to load data from GCS.

Example:

checkpoint = tf.train.Checkpoint(model=model)
local_device_option = tf.train.CheckpointOptions(experimental_io_device="/job:localhost")
checkpoint.write(checkpoint_path, options=local_device_option)