[Bugfix] recipe_template cli

[rahul-tuli]

A recent bug was found in recipe_template where invoking the cli with sparseml.recipe quantization true --target vnni would error out

the reason was that from the cli side the mask_type wasn't settable to it defaulted to unstructured which is wrong.

This PR fixes that by setting the correct mask type from the cli type, and also adds tests to confirm correct working.

before this PR:

(sparseml3.8) 🥃 sparseml sparseml.recipe_template --quantization true --target vnni --lr constant
Traceback (most recent call last):
  File "/home/rahul/virtual_environments/sparseml3.8/bin/sparseml.recipe_template", line 8, in <module>
    sys.exit(main())
  File "/home/rahul/virtual_environments/sparseml3.8/lib/python3.8/site-packages/click/core.py", line 1128, in __call__
    return self.main(*args, **kwargs)
  File "/home/rahul/virtual_environments/sparseml3.8/lib/python3.8/site-packages/click/core.py", line 1053, in main
    rv = self.invoke(ctx)
  File "/home/rahul/virtual_environments/sparseml3.8/lib/python3.8/site-packages/click/core.py", line 1395, in invoke
    return ctx.invoke(self.callback, **ctx.params)
  File "/home/rahul/virtual_environments/sparseml3.8/lib/python3.8/site-packages/click/core.py", line 754, in invoke
    return __callback(*args, **kwargs)
  File "/home/rahul/projects/sparseml/src/sparseml/pytorch/recipe_template/cli.py", line 108, in main
    template = recipe_template(**kwargs)
  File "/home/rahul/projects/sparseml/src/sparseml/pytorch/recipe_template/main.py", line 94, in recipe_template
    mask_type: str = _validate_mask_type(mask_type=mask_type, target=target)
  File "/home/rahul/projects/sparseml/src/sparseml/pytorch/recipe_template/main.py", line 151, in _validate_mask_type
    raise ValueError(
ValueError: The specified mask type unstructured and target vnni are incompatible, try overriding mask_type to block4

after this PR:

$sparseml.recipe_template --quantization true --target vnni --lr constant
2022-11-17 16:58:16 sparseml.pytorch.recipe_template.main INFO     Recipe written to file recipe.md
Template:
---
lr_func: constant
init_lr: 0.001
final_lr: 0.0
num_epochs: 20.0
num_qat_epochs: 5.0
num_qat_finetuning_epochs: 2.5
quantization_submodules: null

training_modifiers:
  - !EpochRangeModifier
    start_epoch: 0.0
    end_epoch: eval(num_epochs)

  - !LearningRateFunctionModifier
    start_epoch: 0.0
    end_epoch: eval(num_epochs)
    lr_func: eval(lr_func)
    init_lr: eval(init_lr)
    final_lr: eval(final_lr)

pruning_modifiers:
  - !ConstantPruningModifier
    start_epoch: 0.0
    params: __ALL_PRUNABLE__

quantization_modifiers:
  - !QuantizationModifier
    start_epoch: 0.0
    submodules: eval(quantization_submodules)
    disable_quantization_observer_epoch: eval(num_qat_epochs - num_qat_finetuning_epochs)
    freeze_bn_stats_epoch: eval(num_qat_epochs - num_qat_finetuning_epochs)
    tensorrt: False
    quantize_linear_activations: False
    quantize_conv_activations: False
    quantize_embedding_activations: False
    exclude_module_types:
    - LayerNorm
    - Tanh
---
This recipe defines the hyperparams necessary to apply specified sparsification
techniques.

Users are encouraged to experiment with the training length and initial learning
rate to either expedite training or to produce a more accurate model.

This can be done by either editing the recipe or supplying the `--recipe_args`
argument to the training commands.

For example, the following appended to the training commands will change the number
of epochs and the initial learning rate:

```bash
--recipe_args '{"num_epochs":8,"init_lr":0.0001}'

The following template can be used to apply this recipe in a one-shot manner:

from sparseml.pytorch.optim import ScheduledModifierManager

model = ... # instantiated model
recipe = ... # local path to this recipe

manager = ScheduledModifierManager.from_yaml(recipe)
manager.apply(model)

[corey-nm]

easy diff!