Update sample code for AI Hub

samples/python/aotgan/README.md

@@ -0,0 +1,59 @@
+# aotgan Sample Code
+
+## Introduction
+This is sample code for using AppBuilder to load aotgan QNN model to HTP and execute inference to erase and in-paint part of given input image.
+
+## Setup AppBuilder environment and prepare QNN SDK libraries by referring to the links below: 
+https://github.com/quic/ai-engine-direct-helper/blob/main/README.md
+https://github.com/quic/ai-engine-direct-helper/blob/main/Docs/User_Guide.md
+
+Copy the QNN libraries from QNN SDK to below path:
+```
+C:\ai-hub\aotgan\qnn\libQnnHtpV73Skel.so
+C:\ai-hub\aotgan\qnn\QnnHtp.dll
+C:\ai-hub\aotgan\qnn\QnnHtpV73Stub.dll
+C:\ai-hub\aotgan\qnn\QnnSystem.dll
+C:\ai-hub\aotgan\qnn\libqnnhtpv73.cat
+```
+
+## aotgan QNN models
+Download the quantized aotgan QNN models from Qualcomm® AI Hub:
+https://aihub.qualcomm.com/compute/models/aotgan
+
+After downloaded the model, copy it to the following path:
+```
+"C:\ai-hub\aotgan\models\aotgan.bin"
+```
+
+## Run the sample code
+Download the sample code from the following link:
+https://github.com/quic/ai-engine-direct-helper/blob/main/Samples/aotgan/aotgan.py
+
+After downloaded the sample code, please copy it to the following path:
+```
+C:\ai-hub\aotgan\
+```
+
+Copy one sample 512x512 image and mask to following path:
+```
+C:\ai-hub\aotgan\test_input_image.png
+C:\ai-hub\aotgan\test_input_mask.png
+```
+
+Run the sample code:	
+```
+python aotgan.py
+```
+
+## Output
+The output image will be saved to the following path:
+```
+C:\ai-hub\aotgan\out.png
+```
+
+## Reference
+You need to setup the AppBuilder environment before you run the sample code. Below is the guide on how to setup the AppBuilder environment:
+https://github.com/quic/ai-engine-direct-helper/blob/main/README.md
+https://github.com/quic/ai-engine-direct-helper/blob/main/Docs/User_Guide.md
+
+

samples/python/aotgan/aotgan.py

@@ -0,0 +1,123 @@
+# ---------------------------------------------------------------------
+# Copyright (c) 2024 Qualcomm Innovation Center, Inc. All rights reserved.
+# SPDX-License-Identifier: BSD-3-Clause
+# ---------------------------------------------------------------------
+
+import os
+import numpy as np
+import torch
+import torchvision.transforms as transforms
+
+from PIL import Image
+from PIL.Image import fromarray as ImageFromArray
+from torch.nn.functional import interpolate, pad
+from torchvision import transforms
+from typing import Callable, Dict, List, Tuple
+
+from qai_appbuilder import (QNNContext, Runtime, LogLevel, ProfilingLevel, PerfProfile, QNNConfig)
+
+image_size = 512
+aotgan = None
+image_buffer = None
+
+
+def preprocess_PIL_image(image: Image) -> torch.Tensor:
+    """Convert a PIL image into a pyTorch tensor with range [0, 1] and shape NCHW."""
+    transform = transforms.Compose([transforms.Resize(image_size),      # bgr image
+                                    transforms.CenterCrop(image_size),
+                                    transforms.PILToTensor()])
+    img: torch.Tensor = transform(image)  # type: ignore
+    img = img.float().unsqueeze(0) / 255.0  # int 0 - 255 to float 0.0 - 1.0
+    return img
+
+def torch_tensor_to_PIL_image(data: torch.Tensor) -> Image:
+    """
+    Convert a Torch tensor (dtype float32) with range [0, 1] and shape CHW into PIL image CHW
+    """
+    out = torch.clip(data, min=0.0, max=1.0)
+    np_out = (out.detach().numpy() * 255).astype(np.uint8)
+    return ImageFromArray(np_out)
+
+def preprocess_inputs(
+    pixel_values_or_image: Image,
+    mask_pixel_values_or_image: Image,
+) -> Dict[str, torch.Tensor]:
+
+    NCHW_fp32_torch_frames = preprocess_PIL_image(pixel_values_or_image)
+    NCHW_fp32_torch_masks = preprocess_PIL_image(mask_pixel_values_or_image)
+
+    # The number of input images should equal the number of input masks.
+    if NCHW_fp32_torch_masks.shape[0] != 1:
+        NCHW_fp32_torch_masks = NCHW_fp32_torch_masks.tile(
+            (NCHW_fp32_torch_frames.shape[0], 1, 1, 1)
+        )
+
+    # Mask input image
+    image_masked = (
+        NCHW_fp32_torch_frames * (1 - NCHW_fp32_torch_masks) + NCHW_fp32_torch_masks
+    )
+
+    return {"image": image_masked, "mask": NCHW_fp32_torch_masks}
+
+# AotGan class which inherited from the class QNNContext.
+class AotGan(QNNContext):
+    def Inference(self, input_data, input_mask):
+        input_datas=[input_data, input_mask]
+        output_data = super().Inference(input_datas)[0]
+        return output_data
+
+def Init():
+    global aotgan
+
+    # Config AppBuilder environment.
+    QNNConfig.Config(os.getcwd() + "\\qnn", Runtime.HTP, LogLevel.WARN, ProfilingLevel.BASIC)
+
+    # Instance for AotGan objects.
+    aotgan_model = "models\\aotgan.bin"
+    aotgan = AotGan("aotgan", aotgan_model)
+
+def Inference(input_image_path, input_mask_path, output_image_path):
+    global image_buffer
+
+    # Read and preprocess the image&mask.
+    image = Image.open(input_image_path)
+    mask = Image.open(input_mask_path)   
+    inputs = preprocess_inputs(image, mask)
+    image_masked, mask_torch = inputs["image"], inputs["mask"]         
+    image_masked = image_masked.numpy()
+    mask_torch = mask_torch.numpy()
+
+    image_masked = np.transpose(image_masked, (0, 2, 3, 1)) 
+    mask_torch = np.transpose(mask_torch, (0, 2, 3, 1)) 
+
+    # Burst the HTP.
+    PerfProfile.SetPerfProfileGlobal(PerfProfile.BURST)
+
+    # Run the inference.
+    output_image = aotgan.Inference([image_masked], [mask_torch])
+
+    # Reset the HTP.
+    PerfProfile.RelPerfProfileGlobal()
+
+    # show%save the result
+    output_image = torch.from_numpy(output_image)    
+    output_image = output_image.reshape(image_size, image_size, 3)  
+    output_image = torch.unsqueeze(output_image, 0)      
+    output_image = [torch_tensor_to_PIL_image(img) for img in output_image]
+    image_buffer = output_image[0]
+    image_buffer.save(output_image_path)
+    image_buffer.show()  
+
+
+def Release():
+    global aotgan
+
+    # Release the resources.
+    del(aotgan)
+
+
+Init()
+
+Inference("input.png", "mask.png", "output.png")
+
+Release()

samples/python/fastsam_x/README.md

@@ -0,0 +1,58 @@
+# unet_segmentation Sample Code
+
+## Introduction
+This is sample code for using AppBuilder to load unet_segmentation QNN model to HTP and execute inference to produce a segmentation mask for an image.
+
+## Setup AppBuilder environment and prepare QNN SDK libraries by referring to the links below: 
+https://github.com/quic/ai-engine-direct-helper/blob/main/README.md
+https://github.com/quic/ai-engine-direct-helper/blob/main/Docs/User_Guide.md
+
+Copy the QNN libraries from QNN SDK to below path:
+```
+C:\ai-hub\unet_segmentation\qnn\libQnnHtpV73Skel.so
+C:\ai-hub\unet_segmentation\qnn\QnnHtp.dll
+C:\ai-hub\unet_segmentation\qnn\QnnHtpV73Stub.dll
+C:\ai-hub\unet_segmentation\qnn\QnnSystem.dll
+C:\ai-hub\unet_segmentation\qnn\libqnnhtpv73.cat
+```
+
+## unet_segmentation QNN models
+Download the quantized unet_segmentation QNN models from Qualcomm® AI Hub:
+https://aihub.qualcomm.com/compute/models/unet_segmentation
+
+After downloaded the model, copy it to the following path:
+```
+"C:\ai-hub\unet_segmentation\models\unet_segmentation.bin"
+```
+
+## Run the sample code
+Download the sample code from the following link:
+https://github.com/quic/ai-engine-direct-helper/blob/main/Samples/unet_segmentation/unet_segmentation.py
+
+After downloaded the sample code, please copy it to the following path:
+```
+C:\ai-hub\unet_segmentation\
+```
+
+Copy one sample image to following path:
+```
+C:\ai-hub\unet_segmentation\in.jpg
+```
+
+Run the sample code:	
+```
+python unet_segmentation.py
+```
+
+## Output
+The output image will be saved to the following path:
+```
+C:\ai-hub\unet_segmentation\out.jpg
+```
+
+## Reference
+You need to setup the AppBuilder environment before you run the sample code. Below is the guide on how to setup the AppBuilder environment:
+https://github.com/quic/ai-engine-direct-helper/blob/main/README.md
+https://github.com/quic/ai-engine-direct-helper/blob/main/Docs/User_Guide.md
+
+