Butch Landingin

Running batch transforms (e.g. torch.nn.functional.grid_sample) is slower on TPU vs CPU

17

## 🐛 Bug Executing the batch transforms which use the torch.nn.functional.grid_sample function seems to run slower on a single TPU core vs the CPU. ## To Reproduce We encountered this...

[CLI] fastai WandbCallback triggers error when running on multi TPU cores

3

**Description** Using the fastai WandbCallback on multiple TPU cores triggers an error: `AssertionError: can only test a child process` **Wandb features** fastai WandbCallback (`fastai.callback.wandb.WandbCallback`) The error is in the wandb...

When running an xla fit method, it triggers an error: ``` Exception in device=TPU:0: Input type (torch.FloatTensor) and weight type (XLAFloatType) should be the same ``` Note: downgrading to 2.3.0...

Setting the `concat_pool = True` in cnn_learner (which is the default) creates a model that doesn't run on multiple tpu cores. Setting it to `concat_pool = False` works fine however....

See this [notebook for performance benchmarks](https://github.com/butchland/fastai_xla_extensions/blob/master/explore_nbs/test_tpu_pets_slow.ipynb) In comparison, a torch based dataloader is much faster -- for comparison see [this sample notebook](https://github.com/butchland/fastai_xla_extensions/blob/master/samples/torch_dataloader_multicore_pets_tpu.ipynb)

batch transforms for vision are slow

4

Confirming that batch transforms are slow Same [notebook without batch tfms ](https://github.com/butchland/fastai_xla_extensions/blob/master/archive_nbs/Explore_fastai_xla_extensions_Pets_TPU_batch_tfms_comparison_demo.ipynb) - each epoch runs at 1:34 to 2:25 mins For exact [same notebook with batch tfms](https://github.com/butchland/fastai_xla_extensions/blob/master/archive_nbs/Explore_fastai_xla_extensions_Pets_TPU_batch_tfms.ipynb)

see the [notebook for the patch](https://github.com/butchland/fastai_xla_extensions/blob/master/archive_nbs/Explore_10_nlp_TPU_Comparison_debug.ipynb)

Language model trainer is very slow or not finishing fit_one_cycle for 1 epoch

2

see [notebook for sample](https://github.com/butchland/fastai_xla_extensions/blob/master/explore_nbs/Explore_10_nlp_TPU_Comparison_debug.ipynb)

Fix `%%model` and `%model` ipython magic error for seed model with csv column name with periods

3

* Add missing tests for `%model` and `%%model` line magic * add tests for full and view kind of models * add tests for seed model (happy path) * Add...