Distillation support for torchvision script (#1310)

* Add support for `self` distillation and `disable`

* Pull out model creation into a method

* Add support to distill with another model

* Add modifier loss update before backward pass

* bugfix, set loss

* Update src/sparseml/pytorch/torchvision/train.py

Co-authored-by: corey-nm <[email protected]>

Co-authored-by: corey-nm <[email protected]>

Author: rahul-tuli

src/sparseml/pytorch/torchvision/train.py

@@ -23,7 +23,7 @@
 import warnings
 from functools import update_wrapper
 from types import SimpleNamespace
-from typing import Callable
+from typing import Callable, Optional
 
 import torch
 import torch.utils.data
@@ -63,6 +63,7 @@ def train_one_epoch(
     epoch: int,
     args,
     log_metrics_fn: Callable[[str, utils.MetricLogger, int, int], None],
+    manager=None,
     model_ema=None,
     scaler=None,
 ) -> utils.MetricLogger:
@@ -91,13 +92,24 @@ def train_one_epoch(
         start_time = time.time()
         image, target = image.to(device), target.to(device)
         with torch.cuda.amp.autocast(enabled=scaler is not None):
-            output = model(image)
+            outputs = output = model(image)
             if isinstance(output, tuple):
                 # NOTE: sparseml models return two things (logits & probs)
                 output = output[0]
             loss = criterion(output, target)
 
         if steps_accumulated % accum_steps == 0:
+            if manager is not None:
+                loss = manager.loss_update(
+                    loss=loss,
+                    module=model,
+                    optimizer=optimizer,
+                    epoch=epoch,
+                    steps_per_epoch=len(data_loader) / accum_steps,
+                    student_outputs=outputs,
+                    student_inputs=image,
+                )
+
             # first: do training to consume gradients
             if scaler is not None:
                 scaler.scale(loss).backward()
@@ -348,27 +360,28 @@ def collate_fn(batch):
     )
 
     _LOGGER.info("Creating model")
-    if args.arch_key in ModelRegistry.available_keys():
-        with torch_distributed_zero_first(args.rank if args.distributed else None):
-            model = ModelRegistry.create(
-                key=args.arch_key,
-                pretrained=args.pretrained,
-                pretrained_path=args.checkpoint_path,
-                pretrained_dataset=args.pretrained_dataset,
-                num_classes=num_classes,
-            )
-    elif args.arch_key in torchvision.models.__dict__:
-        # fall back to torchvision
-        model = torchvision.models.__dict__[args.arch_key](
-            pretrained=args.pretrained, num_classes=num_classes
-        )
-        if args.checkpoint_path is not None:
-            load_model(args.checkpoint_path, model, strict=True)
-    else:
-        raise ValueError(
-            f"Unable to find {args.arch_key} in ModelRegistry or in torchvision.models"
+    local_rank = args.rank if args.distributed else None
+    model = _create_model(
+        arch_key=args.arch_key,
+        local_rank=local_rank,
+        pretrained=args.pretrained,
+        checkpoint_path=args.checkpoint_path,
+        pretrained_dataset=args.pretrained_dataset,
+        device=device,
+        num_classes=num_classes,
+    )
+
+    if args.distill_teacher not in ["self", "disable", None]:
+        _LOGGER.info("Instantiating teacher")
+        args.distill_teacher = _create_model(
+            arch_key=args.teacher_arch_key,
+            local_rank=local_rank,
+            pretrained=True,  # teacher is always pretrained
+            pretrained_dataset=args.pretrained_teacher_dataset,
+            checkpoint_path=args.distill_teacher,
+            device=device,
+            num_classes=num_classes,
         )
-    model.to(device)
 
     if args.distributed and args.sync_bn:
         model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
@@ -542,7 +555,12 @@ def log_metrics(tag: str, metrics: utils.MetricLogger, epoch: int, epoch_step: i
             )
 
     if manager is not None:
-        manager.initialize(model, epoch=args.start_epoch, loggers=logger)
+        manager.initialize(
+            model,
+            epoch=args.start_epoch,
+            loggers=logger,
+            distillation_teacher=args.distill_teacher,
+        )
         optimizer = manager.modify(
             model, optimizer, steps_per_epoch=steps_per_epoch, epoch=args.start_epoch
         )
@@ -578,6 +596,7 @@ def log_metrics(tag: str, metrics: utils.MetricLogger, epoch: int, epoch_step: i
             epoch,
             args,
             log_metrics,
+            manager=manager,
             model_ema=model_ema,
             scaler=scaler,
         )
@@ -650,6 +669,39 @@ def log_metrics(tag: str, metrics: utils.MetricLogger, epoch: int, epoch_step: i
     _LOGGER.info(f"Training time {total_time_str}")
 
 
+def _create_model(
+    arch_key: Optional[str] = None,
+    local_rank=None,
+    pretrained: Optional[bool] = False,
+    checkpoint_path: Optional[str] = None,
+    pretrained_dataset: Optional[str] = None,
+    device=None,
+    num_classes=None,
+):
+    if arch_key in ModelRegistry.available_keys():
+        with torch_distributed_zero_first(local_rank):
+            model = ModelRegistry.create(
+                key=arch_key,
+                pretrained=pretrained,
+                pretrained_path=checkpoint_path,
+                pretrained_dataset=pretrained_dataset,
+                num_classes=num_classes,
+            )
+    elif arch_key in torchvision.models.__dict__:
+        # fall back to torchvision
+        model = torchvision.models.__dict__[arch_key](
+            pretrained=pretrained, num_classes=num_classes
+        )
+        if checkpoint_path is not None:
+            load_model(checkpoint_path, model, strict=True)
+    else:
+        raise ValueError(
+            f"Unable to find {arch_key} in ModelRegistry or in torchvision.models"
+        )
+    model.to(device)
+    return model
+
+
 def _get_lr_scheduler(args, optimizer, checkpoint=None, manager=None):
     lr_scheduler = None
 
@@ -1026,6 +1078,34 @@ def new_func(*args, **kwargs):
     help="Save the best validation result after the given "
     "epoch completes until the end of training",
 )
+@click.option(
+    "--distill-teacher",
+    default=None,
+    type=str,
+    help="Teacher model for distillation (a trained image classification model)"
+    " can be set to 'self' for self-distillation and 'disable' to switch-off"
+    " distillation, additionally can also take in a SparseZoo stub",
+)
+@click.option(
+    "--pretrained-teacher-dataset",
+    default=None,
+    type=str,
+    help=(
+        "The dataset to load pretrained weights for the teacher"
+        "Load the default dataset for the architecture if set to None. "
+        "examples:`imagenet`, `cifar10`, etc..."
+    ),
+)
+@click.option(
+    "--teacher-arch-key",
+    default=None,
+    type=str,
+    help=(
+        "The architecture key for teacher image classification model; "
+        "example: `resnet50`, `mobilenet`. "
+        "Note: Will be read from the checkpoint if not specified"
+    ),
+)
 @click.pass_context
 def cli(ctx, **kwargs):
     """