UnimplementedError: File system scheme '[local]' not implemented on TPU distribute strategy

[h4ck4l1]

UnimplementedError: File system scheme '[local]' not implemented on TPU distribute strategy

[aaarrti]

Afaik, colab TPUs are using TPU-node architecture, which means they are running on a different host, than the VM, which serves colab runtime, so TPU can't read the local file-system of colab VM (seems logical). One workaround you can do is:

• create google cloud project
• Add this one top of your notebook

!gcloud auth login
!gcloud config set project [PROJECT-id]

from google.colab import auth
auth.authenticate_user()

• Create GCS bucket

resource "google_storage_bucket" "data_bucket" {
  name                        = var.bucket-name
  location                    = var.region
  storage_class               = "STANDARD"
  uniform_bucket_level_access = true
  force_destroy               = true
}

• Then you initialize TPU cluster following way

tf.distribute.cluster_resolver.TPUClusterResolver(tpu='grpc://' + os.environ['COLAB_TPU_ADDR'])

• At this point, you can try to load dataset from your bucket, and you will get a similar error

{
  "code": 403,
  "message": "service-<id>@cloud-tpu.iam.gserviceaccount.com does not have storage.objects.list access to the Google Cloud Storage bucket. Permission 'storage.objects.list' denied on resource (or it may not exist).",
}

• Then you go, and add IAM role bindings for TPUs service account

resource "google_storage_bucket_iam_member" "TPU_ROLE_1" {
  bucket     = var.bucket-name
  role       = "roles/storage.objectViewer"
  member     = var.tpu_sa
  depends_on = [google_storage_bucket.data_bucket]
}

resource "google_storage_bucket_iam_member" "TPU_ROLE_2" {
  bucket     = var.bucket-name
  role       = "roles/storage.legacyBucketReader"
  member     = var.tpu_sa
  depends_on = [google_storage_bucket.data_bucket]
}

resource "google_storage_bucket_iam_member" "TPU_ROLE_3" {
  bucket     = var.bucket-name
  role       = "roles/storage.objectCreator"
  member     = var.tpu_sa
  depends_on = [google_storage_bucket.data_bucket]
}

resource "google_storage_bucket_iam_member" "TPU_ROLE_4" {
  bucket     = var.bucket-name
  role       = "roles/storage.legacyBucketWriter"
  member     = var.tpu_sa
  depends_on = [google_storage_bucket.data_bucket]
}

where var.tpu_sa is "serviceAccount:service-<id>@cloud-tpu.iam.gserviceaccount.com"

• Then finally, your colab TPU should be able to access GCS bucket.

Alternatively, you could use Kaggle Notebooks, since they provide TPU VMs, so you can read the local file-system when running those.

To sum up, this is not a keras-tuner bug 🙈 , but rather an environment misconfiguration.

[h4ck4l1]

Thank you very much for clarifying. I didnt say this was a bug, initially i wrote it on discussion forum as i couldn't figure out why this was happening and i thought i was doing some mistake. but seeing the replies in discussion were way too slow i had to raise issue.
Again thanks for clarifying, i thought i could find other routes than creating a bucket. yea i will try and get back if i get any error