Closed sirtris closed 2 years ago
I was trying to run it on colab and I got a different error 13 frames
/usr/local/lib/python3.7/dist-packages/flash/image/embedding/vissl/hooks.py in on_start(self, task) 50 # get around vissl distributed training by setting MockTask flags 51 num_nodes = lightning_module.trainer.num_nodes ---> 52 accelerators_ids = accelerator_connector(lightning_module.trainer).parallel_device_ids 53 accelerator_per_node = len(accelerators_ids) if accelerators_ids is not None else 1 54 task.world_size = num_nodes * accelerator_per_node
AttributeError: 'AcceleratorConnector' object has no attribute 'parallel_device_ids'
Hi @sirtris @Daniel-R-Armstrong Thanks for reporting these! They should both be fixed on latest master. We will have a patch release later today that includes the fixes. I'll report back here once the release is out for you to try :smiley:
Thanks!
Our release is out! @sirtris @Daniel-R-Armstrong This should now be working for you if you install flash with:
pip install 'lightning-flash[image]==0.7.2'
Also note that we've updated our docs to warn that multi-gpu ssl training is not currently supported: https://lightning-flash.readthedocs.io/en/stable/reference/image_embedder.html
Hope that helps :smiley:
Awesome
Thanks I gave it a quick run with Trainer(..., max_steps=16)
and it seems to work.
🐛 ImageEmbedder example not working
I tried to run the
flash_examples/image_embedder.py
script but it crashes. It gives me the error message:AssertionError: Incorrect embedding shape: torch.Size([16, 8192]) but expected Nx128
To Reproduce
run the example script
Code sample
Expected behavior
Run without error.
Environment
THX