102-pytorch-onnx-to-openvino: PyTorch to ONNX and OpenVINO IR Tutorial

# Imports
import sys
import time
from pathlib import Path

import cv2
import numpy as np
import torch
from IPython.display import Markdown, display
from fastseg import MobileV3Large
from openvino.runtime import Core

sys.path.append("../utils")
from notebook_utils import CityScapesSegmentation, segmentation_map_to_image, viz_result_image

# Settings
IMAGE_WIDTH = 1024  # Suggested values: 2048, 1024 or 512. The minimum width is 512.
# Set IMAGE_HEIGHT manually for custom input sizes. Minimum height is 512
IMAGE_HEIGHT = 1024 if IMAGE_WIDTH == 2048 else 512
DIRECTORY_NAME = "model"
BASE_MODEL_NAME = DIRECTORY_NAME + f"/fastseg{IMAGE_WIDTH}"

# Paths where PyTorch, ONNX and OpenVINO IR models will be stored
model_path = Path(BASE_MODEL_NAME).with_suffix(".pth")
onnx_path = model_path.with_suffix(".onnx")
ir_path = model_path.with_suffix(".xml")

# Download the Fastseg Model
print("Downloading the Fastseg model (if it has not been downloaded before)....")
model = MobileV3Large.from_pretrained().cpu().eval()
print("Loaded PyTorch Fastseg model")

# Save the model
model_path.parent.mkdir(exist_ok=True)
torch.save(model.state_dict(), str(model_path))
print(f"Model saved at {model_path}")

# Convert PyTorch model to ONNX
if not onnx_path.exists():
    dummy_input = torch.randn(1, 3, IMAGE_HEIGHT, IMAGE_WIDTH)

    # For the Fastseg model, setting do_constant_folding to False is required
    # for PyTorch>1.5.1
    torch.onnx.export(
        model,
        dummy_input,
        onnx_path,
        opset_version=11,
        do_constant_folding=False,
    )
    print(f"ONNX model exported to {onnx_path}.")
else:
    print(f"ONNX model {onnx_path} already exists.")

# Convert ONNX Model to OpenVINO IR Format
# Construct the command for Model Optimizer
mo_command = f"""mo
                 --input_model "{onnx_path}"
                 --input_shape "[1,3, {IMAGE_HEIGHT}, {IMAGE_WIDTH}]"
                 --mean_values="[123.675, 116.28 , 103.53]"
                 --scale_values="[58.395, 57.12 , 57.375]"
                 --data_type FP16
                 --output_dir "{model_path.parent}"
                 """
mo_command = " ".join(mo_command.split())
print("Model Optimizer command to convert the ONNX model to OpenVINO:")
display(Markdown(f"`{mo_command}`"))

if not ir_path.exists():
    print("Exporting ONNX model to IR... This may take a few minutes.")
    mo_result = %sx $mo_command
    print("\n".join(mo_result))
else:
    print(f"IR model {ir_path} already exists.")

# Show results: Load and Preprocess an Input Image
def normalize(image: np.ndarray) -> np.ndarray:
    """
    Normalize the image to the given mean and standard deviation
    for CityScapes models.
    """
    image = image.astype(np.float32)
    mean = (0.485, 0.456, 0.406)
    std = (0.229, 0.224, 0.225)
    image /= 255.0
    image -= mean
    image /= std
    return image

image_filename = "data/street.jpg"
image = cv2.cvtColor(cv2.imread(image_filename), cv2.COLOR_BGR2RGB)

resized_image = cv2.resize(image, (IMAGE_WIDTH, IMAGE_HEIGHT))
normalized_image = normalize(resized_image)

# Convert the resized images to network input shape
input_image = np.expand_dims(np.transpose(resized_image, (2, 0, 1)), 0)
normalized_input_image = np.expand_dims(np.transpose(normalized_image, (2, 0, 1)), 0)

# ONNX Model in Inference Engine
# Load network to Inference Engine
ie = Core()
model_onnx = ie.read_model(model=onnx_path)
compiled_model_onnx = ie.compile_model(model=model_onnx, device_name="CPU")

output_layer_onnx = compiled_model_onnx.output(0)

# Run inference on the input image
res_onnx = compiled_model_onnx([normalized_input_image])[output_layer_onnx]

# Convert network result to segmentation map and display the result
result_mask_onnx = np.squeeze(np.argmax(res_onnx, axis=1)).astype(np.uint8)
viz_result_image(
    image,
    segmentation_map_to_image(result_mask_onnx, CityScapesSegmentation.get_colormap()),
    resize=True,
)

# IR Model in Inference Engine
# Load the network in Inference Engine
ie = Core()
model_ir = ie.read_model(model=ir_path)
compiled_model_ir = ie.compile_model(model=model_ir, device_name="CPU")

# Get input and output layers
output_layer_ir = compiled_model_ir.output(0)

# Run inference on the input image
res_ir = compiled_model_ir([input_image])[output_layer_ir]

result_mask_ir = np.squeeze(np.argmax(res_ir, axis=1)).astype(np.uint8)
viz_result_image(
    image,
    segmentation_map_to_image(result=result_mask_ir, colormap=CityScapesSegmentation.get_colormap()),
    resize=True,
)

# PyTorch Comparison
with torch.no_grad():
    result_torch = model(torch.as_tensor(normalized_input_image).float())

result_mask_torch = torch.argmax(result_torch, dim=1).squeeze(0).numpy().astype(np.uint8)
viz_result_image(
    image,
    segmentation_map_to_image(result=result_mask_torch, colormap=CityScapesSegmentation.get_colormap()),
    resize=True,
)

# Performance Comparison
num_images = 20

start = time.perf_counter()
for _ in range(num_images):
    compiled_model_onnx([normalized_input_image])
end = time.perf_counter()
time_onnx = end - start
print(
    f"ONNX model in Inference Engine/CPU: {time_onnx/num_images:.3f} "
    f"seconds per image, FPS: {num_images/time_onnx:.2f}"
)

start = time.perf_counter()
for _ in range(num_images):
    compiled_model_ir([input_image])
end = time.perf_counter()
time_ir = end - start
print(
    f"IR model in Inference Engine/CPU: {time_ir/num_images:.3f} "
    f"seconds per image, FPS: {num_images/time_ir:.2f}"
)

with torch.no_grad():
    start = time.perf_counter()
    for _ in range(num_images):
        model(torch.as_tensor(input_image).float())
    end = time.perf_counter()
    time_torch = end - start
print(
    f"PyTorch model on CPU: {time_torch/num_images:.3f} seconds per image, "
    f"FPS: {num_images/time_torch:.2f}"
)

if "GPU" in ie.available_devices:
    compiled_model_onnx_gpu = ie.compile_model(model=model_onnx, device_name="GPU")
    start = time.perf_counter()
    for _ in range(num_images):
        compiled_model_onnx_gpu([input_image])
    end = time.perf_counter()
    time_onnx_gpu = end - start
    print(
        f"ONNX model in Inference Engine/GPU: {time_onnx_gpu/num_images:.3f} "
        f"seconds per image, FPS: {num_images/time_onnx_gpu:.2f}"
    )

    compiled_model_ir_gpu = ie.compile_model(model=model_ir, device_name="GPU")
    start = time.perf_counter()
    for _ in range(num_images):
        compiled_model_ir_gpu([input_image])
    end = time.perf_counter()
    time_ir_gpu = end - start
    print(
        f"IR model in Inference Engine/GPU: {time_ir_gpu/num_images:.3f} "
        f"seconds per image, FPS: {num_images/time_ir_gpu:.2f}"
    )

# Show Device Information
devices = ie.available_devices
for device in devices:
    device_name = ie.get_property(device_name=device, name="FULL_DEVICE_NAME")
    print(f"{device}: {device_name}")