Benchmarks: Add Benchmark - Add CNN model benchmarks. (#59)

* Benchmarks: Add Benchmark - Add CNN model benchmarks.

examples/benchmarks/pytorch_cnn_resnet101.py

+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+"""Model benchmark example for resnet101.
+
+Commands to run:
+  python3 examples/benchmarks/pytorch_cnn_resnet101.py (Single GPU)
+  python3 -m torch.distributed.launch --use_env --nproc_per_node=8 examples/benchmarks/pytorch_cnn_resnet101.py \
+      --distributed (Distributed)
+"""
+
+import argparse
+
+from superbench.benchmarks import Platform, Framework, BenchmarkRegistry
+from superbench.common.utils import logger
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        '--distributed', action='store_true', default=False, help='Whether to enable distributed training.'
+    )
+    args = parser.parse_args()
+
+    # Specify the model name and benchmark parameters.
+    model_name = 'resnet101'
+    parameters = '--batch_size 1 --image_size 224 --precision float32 --num_warmup 8 --num_steps 64 --run_count 2'
+    if args.distributed:
+        parameters += ' --distributed_impl ddp --distributed_backend nccl'
+
+    # Create context for resnet101 benchmark and run it for 64 steps.
+    context = BenchmarkRegistry.create_benchmark_context(
+        model_name, platform=Platform.CUDA, parameters=parameters, framework=Framework.PYTORCH
+    )
+
+    benchmark = BenchmarkRegistry.launch_benchmark(context)
+    if benchmark:
+        logger.info(
+            'benchmark: {}, return code: {}, result: {}'.format(
+                benchmark.name, benchmark.return_code, benchmark.result
+            )
+        )

superbench/benchmarks/model_benchmarks/__init__.py

@@ -6,5 +6,6 @@
 from superbench.benchmarks.model_benchmarks.model_base import ModelBenchmark
 from superbench.benchmarks.model_benchmarks.pytorch_bert import PytorchBERT
 from superbench.benchmarks.model_benchmarks.pytorch_gpt2 import PytorchGPT2
+from superbench.benchmarks.model_benchmarks.pytorch_cnn import PytorchCNN
 
-__all__ = ['ModelBenchmark', 'PytorchBERT', 'PytorchGPT2']
+__all__ = ['ModelBenchmark', 'PytorchBERT', 'PytorchGPT2', 'PytorchCNN']

superbench/benchmarks/model_benchmarks/pytorch_cnn.py

+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+"""Module of the Pytorch CNN models."""
+
+import time
+
+import torch
+from torchvision import models
+
+from superbench.common.utils import logger
+from superbench.benchmarks import BenchmarkRegistry, Precision
+from superbench.benchmarks.model_benchmarks.model_base import Optimizer
+from superbench.benchmarks.model_benchmarks.pytorch_base import PytorchBase
+from superbench.benchmarks.model_benchmarks.random_dataset import TorchRandomDataset
+
+
+class PytorchCNN(PytorchBase):
+    """The CNN benchmark class."""
+    def __init__(self, name, parameters=''):
+        """Constructor.
+
+        Args:
+            name (str): benchmark name.
+            parameters (str): benchmark parameters.
+        """
+        super().__init__(name, parameters)
+        self._supported_precision = [Precision.FLOAT32, Precision.FLOAT16]
+        self._optimizer_type = Optimizer.SGD
+        self._loss_fn = torch.nn.CrossEntropyLoss()
+
+    def add_parser_arguments(self):
+        """Add the CNN-specified arguments."""
+        super().add_parser_arguments()
+
+        self._parser.add_argument('--model_type', type=str, required=True, help='The cnn benchmark to run.')
+        self._parser.add_argument('--image_size', type=int, default=224, required=False, help='Image size.')
+        self._parser.add_argument('--num_classes', type=int, default=1000, required=False, help='Num of class.')
+
+    def _generate_dataset(self):
+        """Generate dataset for benchmarking according to shape info.
+
+        Return:
+            True if dataset is created successfully.
+        """
+        self._dataset = TorchRandomDataset(
+            [self._args.sample_count, 3, self._args.image_size, self._args.image_size],
+            self._world_size,
+            dtype=torch.long
+        )
+        if len(self._dataset) == 0:
+            logger.error('Generate random dataset failed - model: {}'.format(self._name))
+            return False
+
+        return True
+
+    def _create_model(self, precision):
+        """Construct the model for benchmarking.
+
+        Args:
+            precision (Precision): precision of model and input data, such as float32, float16.
+        """
+        try:
+            self._model = getattr(models, self._args.model_type)()
+            self._model = self._model.to(dtype=getattr(torch, precision.value))
+            if self._gpu_available:
+                self._model = self._model.cuda()
+        except BaseException as e:
+            logger.error(
+                'Create model with specified precision failed - model: {}, precision: {}, message: {}.'.format(
+                    self._name, precision, str(e)
+                )
+            )
+            return False
+
+        self._target = torch.LongTensor(self._args.batch_size).random_(self._args.num_classes)
+        if self._gpu_available:
+            self._target = self._target.cuda()
+
+        return True
+
+    def _train_step(self, precision):
+        """Define the training process.
+
+        Args:
+            precision (Precision): precision of model and input data, such as float32, float16.
+
+        Return:
+            The step-time list of every training step.
+        """
+        duration = []
+        curr_step = 0
+        while True:
+            for idx, sample in enumerate(self._dataloader):
+                start = time.time()
+                sample = sample.to(dtype=getattr(torch, precision.value))
+                if self._gpu_available:
+                    sample = sample.cuda()
+                self._optimizer.zero_grad()
+                output = self._model(sample)
+                loss = self._loss_fn(output, self._target)
+                loss.backward()
+                self._optimizer.step()
+                end = time.time()
+                curr_step += 1
+                if curr_step > self._args.num_warmup:
+                    # Save the step time of every training/inference step, unit is millisecond.
+                    duration.append((end - start) * 1000)
+                if self._is_finished(curr_step, end):
+                    return duration
+
+    def _inference_step(self, precision):
+        """Define the inference process.
+
+        Args:
+            precision (Precision): precision of model and input data,
+              such as float32, float16.
+
+        Return:
+            The latency list of every inference operation.
+        """
+        duration = []
+        curr_step = 0
+        with torch.no_grad():
+            self._model.eval()
+            while True:
+                for idx, sample in enumerate(self._dataloader):
+                    start = time.time()
+                    sample = sample.to(dtype=getattr(torch, precision.value))
+                    if self._gpu_available:
+                        sample = sample.cuda()
+                    self._model(sample)
+                    if self._gpu_available:
+                        torch.cuda.synchronize()
+                    end = time.time()
+                    curr_step += 1
+                    if curr_step > self._args.num_warmup:
+                        # Save the step time of every training/inference step, unit is millisecond.
+                        duration.append((end - start) * 1000)
+                    if self._is_finished(curr_step, end):
+                        return duration
+
+
+# Register CNN benchmarks.
+# Reference: https://pytorch.org/vision/stable/models.html
+#            https://github.com/pytorch/vision/tree/master/torchvision/models
+MODELS = [
+    'alexnet', 'densenet121', 'densenet169', 'densenet201', 'densenet161', 'googlenet', 'inception_v3', 'mnasnet0_5',
+    'mnasnet0_75', 'mnasnet1_0', 'mnasnet1_3', 'mobilenet_v2', 'mobilenet_v3_large', 'mobilenet_v3_small', 'resnet18',
+    'resnet34', 'resnet50', 'resnet101', 'resnet152', 'resnext50_32x4d', 'resnext101_32x8d', 'wide_resnet50_2',
+    'wide_resnet101_2', 'shufflenet_v2_x0_5', 'shufflenet_v2_x1_0', 'shufflenet_v2_x1_5', 'shufflenet_v2_x2_0',
+    'squeezenet1_0', 'squeezenet1_1', 'vgg11', 'vgg11_bn', 'vgg13', 'vgg13_bn', 'vgg16', 'vgg16_bn', 'vgg19_bn', 'vgg19'
+]
+
+for model in MODELS:
+    if hasattr(models, model):
+        BenchmarkRegistry.register_benchmark('pytorch-' + model, PytorchCNN, parameters='--model_type ' + model)
+    else:
+        logger.warning('model missing in torchvision.models - model: {}'.format(model))

tests/benchmarks/model_benchmarks/test_pytorch_cnn.py

+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+
+"""Tests for CNN model benchmarks."""
+
+from tests.helper import decorator
+from superbench.benchmarks import BenchmarkRegistry, Platform, Framework, BenchmarkType, ReturnCode
+from superbench.benchmarks.model_benchmarks.pytorch_cnn import PytorchCNN
+
+
+@decorator.cuda_test
+@decorator.pytorch_test
+def test_pytorch_cnn_with_gpu():
+    """Test pytorch cnn benchmarks with GPU."""
+    run_pytorch_cnn(
+        models=['resnet50', 'resnet101', 'resnet152', 'densenet169', 'densenet201', 'vgg11', 'vgg13', 'vgg16', 'vgg19'],
+        parameters='--batch_size 1 --image_size 224 --num_classes 5 --num_warmup 2 --num_steps 4 \
+            --model_action train inference',
+        check_metrics=[
+            'steptime_train_float32', 'throughput_train_float32', 'steptime_train_float16', 'throughput_train_float16',
+            'steptime_inference_float32', 'throughput_inference_float32', 'steptime_inference_float16',
+            'throughput_inference_float16'
+        ]
+    )
+
+
+@decorator.pytorch_test
+def test_pytorch_cnn_no_gpu():
+    """Test pytorch cnn benchmarks with CPU."""
+    run_pytorch_cnn(
+        models=['resnet50'],
+        parameters='--batch_size 1 --image_size 224 --num_classes 5 --num_warmup 2 --num_steps 4 \
+                --model_action train inference --precision float32 --no_gpu',
+        check_metrics=[
+            'steptime_train_float32', 'throughput_train_float32', 'steptime_inference_float32',
+            'throughput_inference_float32'
+        ]
+    )
+
+
+def run_pytorch_cnn(models=[], parameters='', check_metrics=[]):
+    """Run pytorch cnn benchmarks."""
+    for model in models:
+        context = BenchmarkRegistry.create_benchmark_context(
+            model, platform=Platform.CUDA, parameters=parameters, framework=Framework.PYTORCH
+        )
+
+        assert (BenchmarkRegistry.is_benchmark_context_valid(context))
+
+        benchmark = BenchmarkRegistry.launch_benchmark(context)
+
+        # Check basic information.
+        assert (benchmark)
+        assert (isinstance(benchmark, PytorchCNN))
+        assert (benchmark.name == 'pytorch-' + model)
+        assert (benchmark.type == BenchmarkType.MODEL)
+
+        # Check predefined parameters of resnet101 model.
+        assert (benchmark._args.model_type == model)
+
+        # Check parameters specified in BenchmarkContext.
+        assert (benchmark._args.batch_size == 1)
+        assert (benchmark._args.image_size == 224)
+        assert (benchmark._args.num_classes == 5)
+        assert (benchmark._args.num_warmup == 2)
+        assert (benchmark._args.num_steps == 4)
+
+        # Check Dataset.
+        assert (len(benchmark._dataset) == benchmark._args.sample_count * benchmark._world_size)
+
+        # Check results and metrics.
+        assert (benchmark.run_count == 1)
+        assert (benchmark.return_code == ReturnCode.SUCCESS)
+        for metric in check_metrics:
+            assert (len(benchmark.raw_data[metric]) == benchmark.run_count)
+            assert (len(benchmark.raw_data[metric][0]) == benchmark._args.num_steps)
+            assert (len(benchmark.result[metric]) == benchmark.run_count)