diff --git a/hpvm/test/dnn_benchmarks/pytorch/dnn/__init__.py b/hpvm/test/dnn_benchmarks/pytorch/dnn/__init__.py
index f4a16c6a2c0767efcac372006abf88f9919a75a9..22e7bfadbc9157814195d4b71d913ffa869a0fc2 100644
--- a/hpvm/test/dnn_benchmarks/pytorch/dnn/__init__.py
+++ b/hpvm/test/dnn_benchmarks/pytorch/dnn/__init__.py
@@ -1,6 +1,62 @@
+from pathlib import Path
+from typing import Union
+
+import torch
+from torch2hpvm import BinDataset, ModelExporter
+
from .alexnet import AlexNet, AlexNet2, AlexNetImageNet
from .datasets import CIFAR, MNIST, ImageNet
from .lenet import LeNet
-from .vgg16 import VGG16Cifar10, VGG16Cifar100, VGG16ImageNet
from .mobilenet import MobileNet
from .resnet import ResNet18, ResNet50
+from .vgg16 import VGG16Cifar10, VGG16Cifar100, VGG16ImageNet
+
+# DNN name -> (DNN class, input_channel, input_size, suggested_batchsize)
+benchmarks = {
+ "lenet_mnist": (LeNet, 1, 28, 1000),
+ "alexnet_cifar10": (AlexNet, 3, 32, 500),
+ "alexnet2_cifar10": (AlexNet2, 3, 32, 500),
+ "alexnet_imagenet": (AlexNetImageNet, 3, 224, 500),
+ "mobilenet_cifar10": (MobileNet, 3, 32, 500),
+ "resnet18_cifar10": (ResNet18, 3, 32, 500),
+ "resnet50_imagenet": (ResNet50, 3, 224, 25),
+ "vgg16_cifar10": (VGG16Cifar10, 3, 32, 500),
+ "vgg16_cifar100": (VGG16Cifar100, 3, 32, 500),
+ "vgg16_imagenet": (VGG16ImageNet, 3, 224, 10),
+}
+
+
+def export_example_dnn(
+ dnn_name: str, output_dir: Union[Path, str], generate_for_tuning: bool
+):
+ self_folder = Path(__file__).parent.absolute()
+ dnn_bench_dir = self_folder / "../.."
+
+ model_cls, nch, img_size, batch_size = benchmarks[dnn_name]
+ dataset_shape = 5000, nch, img_size, img_size
+ params = dnn_bench_dir / "model_params" / dnn_name
+ bin_tuneset = BinDataset(
+ params / "tune_input.bin", params / "tune_labels.bin", dataset_shape
+ )
+ bin_testset = BinDataset(
+ params / "test_input.bin", params / "test_labels.bin", dataset_shape
+ )
+ model: Module = model_cls()
+ checkpoint = dnn_bench_dir / f"model_params/pytorch/{dnn_name}.pth.tar"
+ model.load_state_dict(torch.load(checkpoint.as_posix()))
+
+ build_dir = output_dir / "build"
+ target_binary = build_dir / dnn_name
+ if generate_for_tuning:
+ exporter = ModelExporter(
+ model, bin_tuneset, bin_testset, output_dir, target="hpvm_tensor_inspect"
+ )
+ else:
+ conf_file = (
+ dnn_bench_dir / "hpvm-c/benchmarks" / dnn_name / "data/tuner_confs.txt"
+ ).absolute()
+ exporter = ModelExporter(
+ model, bin_tuneset, bin_testset, output_dir, config_file=conf_file
+ )
+ exporter.generate(batch_size=batch_size).compile(target_binary, build_dir)
+ return target_binary, exporter
diff --git a/hpvm/test/dnn_benchmarks/pytorch/test_frontend/test_frontend.py b/hpvm/test/dnn_benchmarks/pytorch/test_frontend/test_frontend.py
index 56161495d5baa9d570a5d2f36188fe437255de57..f248b7bc259e9b6ebd0ef304991e8df988836010 100755
--- a/hpvm/test/dnn_benchmarks/pytorch/test_frontend/test_frontend.py
+++ b/hpvm/test/dnn_benchmarks/pytorch/test_frontend/test_frontend.py
@@ -5,51 +5,14 @@ from pathlib import Path
from subprocess import run
from sys import argv
-import torch
-from torch2hpvm import BinDataset, ModelExporter
-from torch.nn import Module
-
-site.addsitedir(Path(__file__).parent.parent.absolute())
+self_folder = Path(__file__).parent.absolute()
+site.addsitedir(self_folder.parent)
import dnn
-benchmarks = {
- "lenet_mnist": (dnn.LeNet, 1, 28, 1000),
- "alexnet_cifar10": (dnn.AlexNet, 3, 32, 500),
- "alexnet2_cifar10": (dnn.AlexNet2, 3, 32, 500),
- "alexnet_imagenet": (dnn.AlexNetImageNet, 3, 224, 500),
- "mobilenet_cifar10": (dnn.MobileNet, 3, 32, 500),
- "resnet18_cifar10": (dnn.ResNet18, 3, 32, 500),
- "resnet50_imagenet": (dnn.ResNet50, 3, 224, 25),
- "vgg16_cifar10": (dnn.VGG16Cifar10, 3, 32, 500),
- "vgg16_cifar100": (dnn.VGG16Cifar100, 3, 32, 500),
- "vgg16_imagenet": (dnn.VGG16ImageNet, 3, 224, 10),
-}
-self_folder = Path(__file__).parent
netname = argv[1]
-model_cls, nch, img_size, batch_size = benchmarks[netname]
codegen_dir = Path(f"./{netname}")
print(f"Generating {netname} to {codegen_dir}")
if codegen_dir.exists():
shutil.rmtree(codegen_dir)
-
-params = self_folder / "../../model_params" / netname
-dataset_shape = 5000, nch, img_size, img_size
-bin_tuneset = BinDataset(
- params / "tune_input.bin", params / "tune_labels.bin", dataset_shape
-)
-bin_testset = BinDataset(
- params / "test_input.bin", params / "test_labels.bin", dataset_shape
-)
-model: Module = model_cls()
-checkpoint = self_folder / "../model_params/pytorch" / f"{netname}.pth.tar"
-model.load_state_dict(torch.load(checkpoint.as_posix()))
-print(model)
-
-build_dir = codegen_dir / "build"
-target_binary = build_dir / netname
-conf_file = self_folder / "../hpvm-c/benchmarks" / netname / "data/tuner_confs.txt"
-exporter = ModelExporter(
- model, bin_tuneset, bin_testset, codegen_dir, config_file=conf_file
-)
-exporter.generate(batch_size=batch_size).compile(target_binary, build_dir)
+target_binary, _ = dnn.export_example_dnn(netname, codegen_dir, False)
run([str(target_binary), "test"], check=True)
diff --git a/hpvm/test/dnn_benchmarks/pytorch/test_tuning/test_tuning.py b/hpvm/test/dnn_benchmarks/pytorch/test_tuning/test_tuning.py
index 644ce0a2176a4e9a257c6618761eb75da34c1697..746a6d8679f06e1e3aeb14bcdf81432f16a28359 100755
--- a/hpvm/test/dnn_benchmarks/pytorch/test_tuning/test_tuning.py
+++ b/hpvm/test/dnn_benchmarks/pytorch/test_tuning/test_tuning.py
@@ -4,61 +4,33 @@ import site
from pathlib import Path
from sys import argv
-import torch
-from predtuner import config_pylogger
-from predtuner.pipedbin import PipedBinaryApp
-from torch2hpvm import BinDataset, ModelExporter
-from torch.nn import Module
+from predtuner import PipedBinaryApp, config_pylogger
-site.addsitedir(Path(__file__).parent.parent.absolute())
+self_folder = Path(__file__).parent.absolute()
+site.addsitedir(self_folder.parent)
import dnn
# Set up logger
msg_logger = config_pylogger(output_dir=".", verbose=True)
-benchmarks = {
- "lenet_mnist": (dnn.LeNet, 1, 28, 1000),
- "alexnet_imagenet": (dnn.AlexNetImageNet, 3, 224, 100),
- "mobilenet_cifar10": (dnn.MobileNet, 3, 32, 500),
- "resnet18_cifar10": (dnn.ResNet18, 3, 32, 500),
- "vgg16_cifar10": (dnn.VGG16Cifar10, 3, 32, 500),
-}
-model_param = Path(__file__).parent / "../../model_params"
def generate(model_cls, nch, img_size, batch_size, netname):
codegen_dir = Path(f"./{netname}")
- build_dir = codegen_dir / "build"
- metadata_file = codegen_dir / "ops.json"
- binary_file = build_dir / netname
- build_dir = codegen_dir / "build"
-
- if codegen_dir.exists():
- shutil.rmtree(codegen_dir)
- params = model_param / netname
- dataset_shape = 5000, nch, img_size, img_size
- bin_tuneset = BinDataset(
- params / "tune_input.bin", params / "tune_labels.bin", dataset_shape
- )
- bin_testset = BinDataset(
- params / "test_input.bin", params / "test_labels.bin", dataset_shape
- )
- model: Module = model_cls()
- checkpoint = model_param / f"pytorch/{netname}.pth.tar"
- model.load_state_dict(torch.load(checkpoint.as_posix()))
- exporter = ModelExporter(
- model, bin_tuneset, bin_testset, codegen_dir, target="hpvm_tensor_inspect"
- )
- exporter.generate(batch_size=batch_size).compile(binary_file, build_dir)
+ binary_file, exporter = dnn.export_example_dnn(netname, codegen_dir, True)
+ metadata_file = codegen_dir / exporter.metadata_file
return binary_file, metadata_file
def main():
netname, is_pred = argv[1:]
is_pred = int(is_pred)
- model_cls, nch, img_size, batch_size = benchmarks[netname]
- binary_file, metadata_file = generate(
- model_cls, nch, img_size, batch_size, netname
- )
+ # Generating tunable binary
+ codegen_dir = Path(f"./{netname}")
+ if codegen_dir.exists():
+ shutil.rmtree(codegen_dir)
+ binary_file, exporter = dnn.export_example_dnn(netname, codegen_dir, True)
+ metadata_file = codegen_dir / exporter.metadata_file_name
+ # Tuning
app = PipedBinaryApp("test", binary_file, metadata_file)
tuner = app.get_tuner()
tuner.tune(