Fix densenet pipeline: update EOL Docker images and DALI CUDA version

cli/jobs/pipelines-with-components/image_classification_with_densenet/apply_image_transformation/entry.spec.yaml

@@ -32,4 +32,4 @@ command: >-
   --output-path ${{outputs.output_path}}
 environment:
   conda_file: ./conda.yaml
-  image: mcr.microsoft.com/azureml/openmpi4.1.0-ubuntu20.04:latest
+  image: mcr.microsoft.com/azureml/openmpi4.1.0-ubuntu22.04:latest

cli/jobs/pipelines-with-components/image_classification_with_densenet/convert_to_image_directory/entry.spec.yaml

@@ -22,4 +22,4 @@ command: >-
   --output-path ${{outputs.output_path}}
 environment:
   conda_file: ./conda.yaml
-  image: mcr.microsoft.com/azureml/openmpi4.1.0-ubuntu20.04:latest
+  image: mcr.microsoft.com/azureml/openmpi4.1.0-ubuntu22.04:latest

cli/jobs/pipelines-with-components/image_classification_with_densenet/image_cnn_train/classify.py

@@ -22,11 +22,10 @@
 from image_classification.dataloaders import load_jpeg_from_file
 
 try:
-    from apex.fp16_utils import *
-    from apex import amp
+    from torch.amp import autocast, GradScaler
 except ImportError:
     raise ImportError(
-        "Please install apex from https://www.github.com/nvidia/apex to run this example."
+        "PyTorch with CUDA support is required to run this example."
     )
 
 
@@ -68,7 +67,7 @@ def main(args):
     model = model.cuda()
 
     if args.precision == "FP16":
-        model = network_to_half(model)
+        model = model.half()
 
     model.eval()
 

cli/jobs/pipelines-with-components/image_classification_with_densenet/image_cnn_train/conda.yaml

@@ -3,19 +3,18 @@ channels:
   - defaults
   - conda-forge
 dependencies:
-  - python=3.8.12
-  - pip=21.2.2
+  - python=3.10
+  - pip=23.3
   - pip:
       - mldesigner==0.1.0b4
-      - watchdog==0.10.3
-      - torch==1.8.1
-      - torchvision==0.9.1
-      - tensorboard==2.5.0
-      - pillow==8.2.0
-      - numpy==1.19.5
+      - watchdog==4.0.1
+      - torch==2.4.0
+      - torchvision==0.19.0
+      - tensorboard==2.17.0
+      - pillow==10.4.0
+      - numpy==1.26.4
       - --extra-index-url=https://developer.download.nvidia.com/compute/redist/
-      - nvidia-dali-cuda100
+      - nvidia-dali-cuda120
       - azureml-mlflow>=1.41.0
-      - protobuf==3.20.1
-      - pandas==1.2.1
-      - packaging==22.0 # fix for https://github.com/pypa/setuptools/issues/4483
+      - protobuf==4.25.3
+      - pandas==2.2.2

cli/jobs/pipelines-with-components/image_classification_with_densenet/image_cnn_train/entry.spec.yaml

@@ -111,7 +111,7 @@ outputs:
 code: ./
 
 environment:
-  image: mcr.microsoft.com/azureml/openmpi4.1.0-cuda11.8-cudnn8-ubuntu22.04:latest
+  image: mcr.microsoft.com/azureml/openmpi4.1.0-ubuntu22.04:latest
   conda_file: ./conda.yaml
 
 resources:
@@ -121,9 +121,8 @@ distribution:
   process_count_per_instance: 1
 
 command: >-
-  git clone https://github.com/NVIDIA/apex && cd apex && git checkout 3303b3e7174383312a3468ef390060c26e640cb1 
-  && python setup.py install && cd ..  
-  && mldesigner execute --source entry.py --name imagecnn_train --inputs train_data=${{inputs.train_data}} 
+  pip install setuptools &&
+  mldesigner execute --source entry.py --name imagecnn_train --inputs train_data=${{inputs.train_data}} 
   val_data=${{inputs.valid_data}} data_backend=${{inputs.data_backend}} arch=${{inputs.arch}} 
   model_config=${{inputs.model_config}} workers=${{inputs.workers}} epochs=${{inputs.epochs}} 
   batch_size=${{inputs.batch_size}} optimizer_batch_size=${{inputs.optimizer_batch_size}} lr=${{inputs.lr}} 

cli/jobs/pipelines-with-components/image_classification_with_densenet/image_cnn_train/image_classification/training.py

@@ -36,14 +36,8 @@
 from . import resnet as models
 from . import utils
 
-try:
-    from apex.parallel import DistributedDataParallel as DDP
-    from apex.fp16_utils import *
-    from apex import amp
-except ImportError:
-    raise ImportError(
-        "Please install apex from https://www.github.com/nvidia/apex to run this example."
-    )
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.amp import autocast, GradScaler
 
 from torch.utils.tensorboard import SummaryWriter
 
@@ -109,7 +103,7 @@ def __init__(self, arch, loss, pretrained_weights=None, cuda=True, fp16=False):
         if cuda:
             model = model.cuda()
         if fp16:
-            model = network_to_half(model)
+            model = model.half()
 
         # define loss function (criterion) and optimizer
         criterion = loss()
@@ -126,8 +120,8 @@ def forward(self, data, target):
 
         return loss, output
 
-    def distributed(self):
-        self.model = DDP(self.model)
+    def distributed(self, gpu=0):
+        self.model = DDP(self.model, device_ids=[gpu])
 
     def load_model_state(self, state):
         if not state is None:
@@ -172,13 +166,6 @@ def get_optimizer(
             weight_decay=weight_decay,
             nesterov=nesterov,
         )
-    if fp16:
-        optimizer = FP16_Optimizer(
-            optimizer,
-            static_loss_scale=static_loss_scale,
-            dynamic_loss_scale=dynamic_loss_scale,
-            verbose=False,
-        )
 
     if not state is None:
         optimizer.load_state_dict(state)
@@ -255,36 +242,39 @@ def _lr_fn(iteration, epoch):
 
 
 def get_train_step(
-    model_and_loss, optimizer, fp16, use_amp=False, batch_size_multiplier=1
+    model_and_loss, optimizer, fp16, use_amp=False, batch_size_multiplier=1, scaler=None
 ):
     def _step(input, target, optimizer_step=True):
         input_var = Variable(input)
         target_var = Variable(target)
-        loss, output = model_and_loss(input_var, target_var)
+
+        if use_amp:
+            with autocast('cuda'):
+                loss, output = model_and_loss(input_var, target_var)
+        else:
+            loss, output = model_and_loss(input_var, target_var)
+
         if torch.distributed.is_initialized():
             reduced_loss = utils.reduce_tensor(loss.data)
         else:
             reduced_loss = loss.data
 
-        if fp16:
-            optimizer.backward(loss)
-        elif use_amp:
-            with amp.scale_loss(loss, optimizer) as scaled_loss:
-                scaled_loss.backward()
+        if use_amp and scaler is not None:
+            scaler.scale(loss).backward()
         else:
             loss.backward()
 
         if optimizer_step:
-            opt = (
-                optimizer.optimizer
-                if isinstance(optimizer, FP16_Optimizer)
-                else optimizer
-            )
-            for param_group in opt.param_groups:
+            for param_group in optimizer.param_groups:
                 for param in param_group["params"]:
-                    param.grad /= batch_size_multiplier
-
-            optimizer.step()
+                    if param.grad is not None:
+                        param.grad /= batch_size_multiplier
+
+            if use_amp and scaler is not None:
+                scaler.step(optimizer)
+                scaler.update()
+            else:
+                optimizer.step()
             optimizer.zero_grad()
 
         torch.cuda.synchronize()
@@ -304,6 +294,7 @@ def train(
     epoch,
     detector,
     use_amp=False,
+    scaler=None,
     prof=-1,
     batch_size_multiplier=1,
     register_metrics=True,
@@ -319,6 +310,7 @@ def train(
         fp16,
         use_amp=use_amp,
         batch_size_multiplier=batch_size_multiplier,
+        scaler=scaler,
     )
 
     model_and_loss.train()
@@ -345,7 +337,7 @@ def train(
             if writer:
                 writer.add_scalar("train/summary/scalar/learning_rate", lr, epoch)
                 writer.add_scalar(
-                    "train/summary/scalar/loss", to_python_float(loss), total_train_step
+                    "train/summary/scalar/loss", float(loss), total_train_step
                 )
                 writer.add_scalar(
                     "perf/summary/scalar/compute_ips",
@@ -359,7 +351,7 @@ def train(
                 )
                 mlflow.log_metric("train/learning_rate", step=epoch, value=lr)
                 mlflow.log_metric(
-                    "train/loss", step=total_train_step, value=to_python_float(loss)
+                    "train/loss", step=total_train_step, value=float(loss)
                 )
                 mlflow.log_metric(
                     "perf/compute_ips",
@@ -448,7 +440,7 @@ def validate(
 
         it_time = time.time() - end
 
-        loss_sum += to_python_float(loss)
+        loss_sum += float(loss)
         total_val_step += 1
 
         end = time.time()
@@ -483,6 +475,7 @@ def train_loop(
     should_backup_checkpoint,
     save_checkpoint_epochs,
     use_amp=False,
+    scaler=None,
     batch_size_multiplier=1,
     best_prec1=0,
     start_epoch=0,
@@ -535,6 +528,7 @@ def train_loop(
                 epoch,
                 detector,
                 use_amp=use_amp,
+                scaler=scaler,
                 prof=prof,
                 register_metrics=epoch == start_epoch,
                 batch_size_multiplier=batch_size_multiplier,

cli/jobs/pipelines-with-components/image_classification_with_densenet/image_cnn_train/main.py

@@ -47,12 +47,11 @@
 import torchvision.datasets as datasets
 
 try:
-    from apex.parallel import DistributedDataParallel as DDP
-    from apex.fp16_utils import *
-    from apex import amp
+    from torch.nn.parallel import DistributedDataParallel as DDP
+    from torch.amp import autocast, GradScaler
 except ImportError:
     raise ImportError(
-        "Please install apex from https://www.github.com/nvidia/apex to run this example."
+        "PyTorch with CUDA support is required to run this example."
     )
 
 import image_classification.resnet as models
@@ -489,16 +488,12 @@ def _worker_init_fn(id):
     elif args.lr_schedule == "linear":
         lr_policy = lr_linear_policy(args.lr, args.warmup, args.epochs, logger=None)
 
+    scaler = None
     if args.amp:
-        model_and_loss, optimizer = amp.initialize(
-            model_and_loss,
-            optimizer,
-            opt_level="O2",
-            loss_scale="dynamic" if args.dynamic_loss_scale else args.static_loss_scale,
-        )
+        scaler = GradScaler('cuda')
 
     if args.distributed:
-        model_and_loss.distributed()
+        model_and_loss.distributed(args.gpu)
 
     model_and_loss.load_model_state(model_state)
 
@@ -514,6 +509,7 @@ def _worker_init_fn(id):
         should_backup_checkpoint(args),
         args.save_checkpoint_epochs,
         use_amp=args.amp,
+        scaler=scaler,
         batch_size_multiplier=batch_size_multiplier,
         start_epoch=start_epoch,
         best_prec1=best_prec1,

cli/jobs/pipelines-with-components/image_classification_with_densenet/init_image_transformation/entry.spec.yaml

@@ -124,4 +124,4 @@ command: >-
   
 environment:
   conda_file: ./conda.yaml
-  image: mcr.microsoft.com/azureml/openmpi4.1.0-cuda11.8-cudnn8-ubuntu22.04:latest
+  image: mcr.microsoft.com/azureml/openmpi4.1.0-ubuntu22.04:latest