Notebook

Object segmentations with EfficientSAM and OpenVINO¶

Segment Anything Model (SAM) has emerged as a powerful tool for numerous vision applications. A key component that drives the impressive performance for zero-shot transfer and high versatility is a super large Transformer model trained on the extensive high-quality SA-1B dataset. While beneficial, the huge computation cost of SAM model has limited its applications to wider real-world applications. To address this limitation, EfficientSAMs, light-weight SAM models that exhibit decent performance with largely reduced complexity, were proposed. The idea behind EfficientSAM is based on leveraging masked image pretraining, SAMI, which learns to reconstruct features from SAM image encoder for effective visual representation learning.

More details about model can be found in paper, model web page and original repository

In this tutorial we consider how to convert and run EfficientSAM using OpenVINO. We also demonstrate how to quantize model using NNCF

Table of contents:¶

Prerequisites
Load PyTorch model
Run PyTorch model inference
- Prepare input data
- Define helpers for input and output processing
Convert model to OpenVINO IR format
Run OpenVINO model inference
Quantization
- Prepare calibration datasets
- Run Model Quantization
Verify quantized model inference
Interactive segmentation demo

Installation Instructions¶

This is a self-contained example that relies solely on its own code.

We recommend running the notebook in a virtual environment. You only need a Jupyter server to start. For details, please refer to Installation Guide.

Prerequisites¶

In [1]:

%pip install -q "openvino>=2023.3.0" "nncf>=2.7.0" opencv-python "gradio>=4.13" "matplotlib>=3.4" torch torchvision tqdm  --extra-index-url https://download.pytorch.org/whl/cpu

In [ ]:

import requests
from pathlib import Path


r = requests.get(
    url="https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/latest/utils/cmd_helper.py",
)
open("cmd_helper.py", "w").write(r.text)

In [ ]:

from cmd_helper import clone_repo

repo_dir = clone_repo("https://github.com/yformer/EfficientSAM.git")

%cd $repo_dir

r = requests.get(
    url="https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/latest/utils/notebook_utils.py",
)
open("notebook_utils.py", "w").write(r.text)

from notebook_utils import download_file, device_widget, quantization_widget  # noqa: F401

Load PyTorch model¶

There are several models available in the repository:

efficient-sam-vitt - EfficientSAM with Vision Transformer Tiny (VIT-T) as image encoder. The smallest and fastest model from EfficientSAM family.
efficient-sam-vits - EfficientSAM with Vision Transformer Small (VIT-S) as image encoder. Heavier than efficient-sam-vitt, but more accurate model.

EfficientSAM provides a unified interface for interaction with models. It means that all provided steps in the notebook for conversion and running the model will be the same for all models. Below, you can select one of them as example.

In [ ]:

from efficient_sam.build_efficient_sam import (
    build_efficient_sam_vitt,
    build_efficient_sam_vits,
)
import zipfile

MODELS_LIST = {
    "efficient-sam-vitt": build_efficient_sam_vitt,
    "efficient-sam-vits": build_efficient_sam_vits,
}

# Since EfficientSAM-S checkpoint file is >100MB, we store the zip file.
with zipfile.ZipFile("weights/efficient_sam_vits.pt.zip", "r") as zip_ref:
    zip_ref.extractall("weights")

Select one from supported models:

In [4]:

import ipywidgets as widgets

model_ids = list(MODELS_LIST)

model_id = widgets.Dropdown(
    options=model_ids,
    value=model_ids[0],
    description="Model:",
    disabled=False,
)

model_id

Out[4]:

Dropdown(description='Model:', options=('efficient-sam-vitt', 'efficient-sam-vits'), value='efficient-sam-vitt…

build PyTorch model

In [5]:

pt_model = MODELS_LIST[model_id.value]()

pt_model.eval();

Run PyTorch model inference¶

Prepare input data¶

First of all, we should prepare input data for model. Model has 3 inputs:

image tensor - tensor with normalized input image.
input points - tensor with user provided points. It maybe just some specific points on the image (e.g. provided by user clicks on the screen) or bounding box coordinates in format left-top angle point and right-bottom angle pint.
input labels - tensor with definition of point type for each provided point, 1 - for regular point, 2 - left-top point of bounding box, 3 - right-bottom point of bounding box.

In [6]:

from PIL import Image

image_path = "figs/examples/dogs.jpg"

image = Image.open(image_path)
image

Out[6]:

Define helpers for input and output processing¶

The code below defines helpers for preparing model input and postprocess inference results. The input format is accepted by the model described above. The model predicts mask logits for each pixel on the image and intersection over union score for each area, how close it is to provided points. We also provided some helper function for results visualization.

In [7]:

import torch
import matplotlib.pyplot as plt
import numpy as np


def prepare_input(input_image, points, labels, torch_tensor=True):
    img_tensor = np.ascontiguousarray(input_image)[None, ...].astype(np.float32) / 255
    img_tensor = np.transpose(img_tensor, (0, 3, 1, 2))
    pts_sampled = np.reshape(np.ascontiguousarray(points), [1, 1, -1, 2])
    pts_labels = np.reshape(np.ascontiguousarray(labels), [1, 1, -1])
    if torch_tensor:
        img_tensor = torch.from_numpy(img_tensor)
        pts_sampled = torch.from_numpy(pts_sampled)
        pts_labels = torch.from_numpy(pts_labels)
    return img_tensor, pts_sampled, pts_labels


def postprocess_results(predicted_iou, predicted_logits):
    sorted_ids = np.argsort(-predicted_iou, axis=-1)
    predicted_iou = np.take_along_axis(predicted_iou, sorted_ids, axis=2)
    predicted_logits = np.take_along_axis(predicted_logits, sorted_ids[..., None, None], axis=2)

    return predicted_logits[0, 0, 0, :, :] >= 0


def show_points(coords, labels, ax, marker_size=375):
    pos_points = coords[labels == 1]
    neg_points = coords[labels == 0]
    ax.scatter(
        pos_points[:, 0],
        pos_points[:, 1],
        color="green",
        marker="*",
        s=marker_size,
        edgecolor="white",
        linewidth=1.25,
    )
    ax.scatter(
        neg_points[:, 0],
        neg_points[:, 1],
        color="red",
        marker="*",
        s=marker_size,
        edgecolor="white",
        linewidth=1.25,
    )


def show_box(box, ax):
    x0, y0 = box[0], box[1]
    w, h = box[2] - box[0], box[3] - box[1]
    ax.add_patch(plt.Rectangle((x0, y0), w, h, edgecolor="yellow", facecolor=(0, 0, 0, 0), lw=5))


def show_anns(mask, ax):
    ax.set_autoscale_on(False)
    img = np.ones((mask.shape[0], mask.shape[1], 4))
    img[:, :, 3] = 0
    # for ann in mask:
    #     m = ann
    color_mask = np.concatenate([np.random.random(3), [0.5]])
    img[mask] = color_mask
    ax.imshow(img)

The complete model inference example demonstrated below

In [8]:

input_points = [[580, 350], [650, 350]]
input_labels = [1, 1]

example_input = prepare_input(image, input_points, input_labels)

predicted_logits, predicted_iou = pt_model(*example_input)

predicted_mask = postprocess_results(predicted_iou.detach().numpy(), predicted_logits.detach().numpy())

In [9]:

image = Image.open(image_path)

plt.figure(figsize=(20, 20))
plt.axis("off")
plt.imshow(image)
show_points(np.array(input_points), np.array(input_labels), plt.gca())
plt.figure(figsize=(20, 20))
plt.axis("off")
plt.imshow(image)
show_anns(predicted_mask, plt.gca())
plt.title(f"PyTorch {model_id.value}", fontsize=18)
plt.show()

Convert model to OpenVINO IR format¶

OpenVINO supports PyTorch models via conversion in Intermediate Representation (IR) format using OpenVINO Model Conversion API. openvino.convert_model function accepts instance of PyTorch model and example input (that helps in correct model operation tracing and shape inference) and returns openvino.Model object that represents model in OpenVINO framework. This openvino.Model is ready for loading on the device using ov.Core.compile_model or can be saved on disk using openvino.save_model.

In [10]:

import openvino as ov

core = ov.Core()

ov_model_path = Path(f"{model_id.value}.xml")

if not ov_model_path.exists():
    ov_model = ov.convert_model(pt_model, example_input=example_input)
    ov.save_model(ov_model, ov_model_path)
else:
    ov_model = core.read_model(ov_model_path)

Run OpenVINO model inference¶

In [11]:

device = device_widget()

device

Out[11]:

Dropdown(description='Device:', index=3, options=('CPU', 'GPU.0', 'GPU.1', 'AUTO'), value='AUTO')

Compile OpenVINO model¶

In [12]:

compiled_model = core.compile_model(ov_model, device.value)

Inference and visualize result¶

Now, we can take a look on OpenVINO model prediction

In [13]:

example_input = prepare_input(image, input_points, input_labels, torch_tensor=False)
result = compiled_model(example_input)

predicted_logits, predicted_iou = result[0], result[1]

predicted_mask = postprocess_results(predicted_iou, predicted_logits)

plt.figure(figsize=(20, 20))
plt.axis("off")
plt.imshow(image)
show_points(np.array(input_points), np.array(input_labels), plt.gca())
plt.figure(figsize=(20, 20))
plt.axis("off")
plt.imshow(image)
show_anns(predicted_mask, plt.gca())
plt.title(f"OpenVINO {model_id.value}", fontsize=18)
plt.show()

Quantization¶

NNCF enables post-training quantization by adding the quantization layers into the model graph and then using a subset of the training dataset to initialize the parameters of these additional quantization layers. The framework is designed so that modifications to your original training code are minor.

The optimization process contains the following steps:

Create a calibration dataset for quantization.
Run nncf.quantize to obtain quantized encoder and decoder models.
Serialize the INT8 model using openvino.save_model function.

Note: Quantization is time and memory consuming operation. Running quantization code below may take some time.

Please select below whether you would like to run EfficientSAM quantization.

In [14]:

to_quantize = quantization_widget()

to_quantize

Out[14]:

Checkbox(value=True, description='Quantization')

In [15]:

# Fetch `skip_kernel_extension` module
import requests

r = requests.get(
    url="https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/latest/utils/skip_kernel_extension.py",
)
open("skip_kernel_extension.py", "w").write(r.text)

%load_ext skip_kernel_extension

Prepare calibration datasets¶

The first step is to prepare calibration datasets for quantization. We will use coco128 dataset for quantization. Usually, this dataset used for solving object detection task and its annotation provides box coordinates for images. In our case, box coordinates will serve as input points for object segmentation, the code below downloads dataset and creates DataLoader for preparing inputs for EfficientSAM model.

In [16]:

%%skip not $to_quantize.value

from zipfile import ZipFile

DATA_URL = "https://ultralytics.com/assets/coco128.zip"
OUT_DIR = Path('.')

download_file(DATA_URL, directory=OUT_DIR, show_progress=True)

if not (OUT_DIR / "coco128/images/train2017").exists():
    with ZipFile('coco128.zip' , "r") as zip_ref:
        zip_ref.extractall(OUT_DIR)

'coco128.zip' already exists.

In [17]:

%%skip not $to_quantize.value

import torch.utils.data as data

class COCOLoader(data.Dataset):
    def __init__(self, images_path):
        self.images = list(Path(images_path).iterdir())
        self.labels_dir = images_path.parents[1] / 'labels' / images_path.name

    def get_points(self, image_path, image_width, image_height):
        file_name = image_path.name.replace('.jpg', '.txt')
        label_file =  self.labels_dir / file_name
        if not label_file.exists():
            x1, x2 = np.random.randint(low=0, high=image_width, size=(2, ))
            y1, y2 = np.random.randint(low=0, high=image_height, size=(2, ))
        else:    
            with label_file.open("r") as f:
                box_line = f.readline()
            _, x1, y1, x2, y2 = box_line.split()
            x1 = int(float(x1) * image_width)
            y1 = int(float(y1) * image_height)
            x2 = int(float(x2) * image_width)
            y2 = int(float(y2) * image_height)
        return [[x1, y1], [x2, y2]]

    def __getitem__(self, index):
        image_path = self.images[index]
        image = Image.open(image_path)
        image = image.convert('RGB')
        w, h = image.size
        points = self.get_points(image_path, w, h)
        labels = [1, 1] if index % 2 == 0 else [2, 3]
        batched_images, batched_points, batched_point_labels = prepare_input(image, points, labels, torch_tensor=False)
        return {'batched_images': np.ascontiguousarray(batched_images)[0], 'batched_points': np.ascontiguousarray(batched_points)[0], 'batched_point_labels': np.ascontiguousarray(batched_point_labels)[0]}
    
    def __len__(self):
        return len(self.images)

In [18]:

%%skip not $to_quantize.value

coco_dataset = COCOLoader(OUT_DIR / 'coco128/images/train2017')
calibration_loader = torch.utils.data.DataLoader(coco_dataset)

Run Model Quantization¶

The nncf.quantize function provides an interface for model quantization. It requires an instance of the OpenVINO Model and quantization dataset. Optionally, some additional parameters for the configuration quantization process (number of samples for quantization, preset, ignored scope, etc.) can be provided. EfficientSAM contains non-ReLU activation functions, which require asymmetric quantization of activations. To achieve a better result, we will use a mixed quantization preset. Model encoder part is based on Vision Transformer architecture for activating special optimizations for this architecture type, we should specify transformer in model_type.

In [19]:

%%skip not $to_quantize.value

import nncf

calibration_dataset = nncf.Dataset(calibration_loader)

model = core.read_model(ov_model_path)
quantized_model = nncf.quantize(model,
                                calibration_dataset,
                                model_type=nncf.parameters.ModelType.TRANSFORMER,
                                subset_size=128)
print("model quantization finished")

INFO:nncf:NNCF initialized successfully. Supported frameworks detected: torch, onnx, openvino

Output()

Output()

INFO:nncf:57 ignored nodes were found by name in the NNCFGraph
INFO:nncf:92 ignored nodes were found by name in the NNCFGraph

Output()

/home/ea/work/genai_env/lib/python3.8/site-packages/nncf/experimental/tensor/tensor.py:80: RuntimeWarning: invalid value encountered in multiply
  return Tensor(self.data * unwrap_tensor_data(other))

Output()

model quantization finished

Verify quantized model inference¶

In [20]:

%%skip not $to_quantize.value

compiled_model = core.compile_model(quantized_model, device.value)

result = compiled_model(example_input)

predicted_logits, predicted_iou = result[0], result[1]

predicted_mask = postprocess_results(predicted_iou, predicted_logits)

plt.figure(figsize=(20, 20))
plt.axis("off")
plt.imshow(image)
show_points(np.array(input_points), np.array(input_labels), plt.gca())
plt.figure(figsize=(20, 20))
plt.axis("off")
plt.imshow(image)
show_anns(predicted_mask, plt.gca())
plt.title(f"OpenVINO INT8 {model_id.value}", fontsize=18)
plt.show()

Save quantize model on disk¶

In [21]:

%%skip not $to_quantize.value

quantized_model_path = Path(f"{model_id.value}_int8.xml")
ov.save_model(quantized_model, quantized_model_path)

Compare quantized model size¶

In [22]:

%%skip not $to_quantize.value

fp16_weights = ov_model_path.with_suffix('.bin')
quantized_weights = quantized_model_path.with_suffix('.bin')

print(f"Size of FP16 model is {fp16_weights.stat().st_size / 1024 / 1024:.2f} MB")
print(f"Size of INT8 quantized model is {quantized_weights.stat().st_size / 1024 / 1024:.2f} MB")
print(f"Compression rate for INT8 model: {fp16_weights.stat().st_size / quantized_weights.stat().st_size:.3f}")

Size of FP16 model is 21.50 MB
Size of INT8 quantized model is 10.96 MB
Compression rate for INT8 model: 1.962

Compare inference time of the FP16 and INT8 models¶

To measure the inference performance of the FP16 and INT8 models, we use bencmark_app.

NOTE: For the most accurate performance estimation, it is recommended to run benchmark_app in a terminal/command prompt after closing other applications.

In [23]:

!benchmark_app -m $ov_model_path -d $device.value -data_shape "batched_images[1,3,512,512],batched_points[1,1,2,2],batched_point_labels[1,1,2]" -t 15

[Step 1/11] Parsing and validating input arguments
[ INFO ] Parsing input parameters
[Step 2/11] Loading OpenVINO Runtime
[ INFO ] OpenVINO:
[ INFO ] Build ................................. 2023.3.0-13739-294cc6668c4
[ INFO ] 
[ INFO ] Device info:
[ INFO ] AUTO
[ INFO ] Build ................................. 2023.3.0-13739-294cc6668c4
[ INFO ] 
[ INFO ] 
[Step 3/11] Setting device configuration
[ WARNING ] Performance hint was not explicitly specified in command line. Device(AUTO) performance hint will be set to PerformanceMode.THROUGHPUT.
[Step 4/11] Reading model files
[ INFO ] Loading model files
[ INFO ] Read model took 49.73 ms
[ INFO ] Original model I/O parameters:
[ INFO ] Model inputs:
[ INFO ]     batched_images (node: batched_images) : f32 / [...] / [?,?,?,?]
[ INFO ]     batched_points (node: batched_points) : i64 / [...] / [?,?,?,?]
[ INFO ]     batched_point_labels (node: batched_point_labels) : i64 / [...] / [?,?,?]
[ INFO ] Model outputs:
[ INFO ]     146 (node: aten::take_along_dim/GatherElements_2824) : f32 / [...] / [?,?,?,?,?]
[ INFO ]     140 (node: aten::take_along_dim/GatherElements) : f32 / [...] / [?,?,?]
[Step 5/11] Resizing model to match image sizes and given batch
[ INFO ] Model batch size: 1
[Step 6/11] Configuring input of the model
[ INFO ] Model inputs:
[ INFO ]     batched_images (node: batched_images) : f32 / [...] / [?,?,?,?]
[ INFO ]     batched_points (node: batched_points) : i64 / [...] / [?,?,?,?]
[ INFO ]     batched_point_labels (node: batched_point_labels) : i64 / [...] / [?,?,?]
[ INFO ] Model outputs:
[ INFO ]     146 (node: aten::take_along_dim/GatherElements_2824) : f32 / [...] / [?,?,?,?,?]
[ INFO ]     140 (node: aten::take_along_dim/GatherElements) : f32 / [...] / [?,?,?]
[Step 7/11] Loading the model to the device
[ INFO ] Compile model took 1512.47 ms
[Step 8/11] Querying optimal runtime parameters
[ INFO ] Model:
[ INFO ]   NETWORK_NAME: Model0
[ INFO ]   EXECUTION_DEVICES: ['CPU']
[ INFO ]   PERFORMANCE_HINT: PerformanceMode.THROUGHPUT
[ INFO ]   OPTIMAL_NUMBER_OF_INFER_REQUESTS: 12
[ INFO ]   MULTI_DEVICE_PRIORITIES: CPU
[ INFO ]   CPU:
[ INFO ]     AFFINITY: Affinity.CORE
[ INFO ]     CPU_DENORMALS_OPTIMIZATION: False
[ INFO ]     CPU_SPARSE_WEIGHTS_DECOMPRESSION_RATE: 1.0
[ INFO ]     ENABLE_CPU_PINNING: True
[ INFO ]     ENABLE_HYPER_THREADING: True
[ INFO ]     EXECUTION_DEVICES: ['CPU']
[ INFO ]     EXECUTION_MODE_HINT: ExecutionMode.PERFORMANCE
[ INFO ]     INFERENCE_NUM_THREADS: 36
[ INFO ]     INFERENCE_PRECISION_HINT: <Type: 'float32'>
[ INFO ]     NETWORK_NAME: Model0
[ INFO ]     NUM_STREAMS: 12
[ INFO ]     OPTIMAL_NUMBER_OF_INFER_REQUESTS: 12
[ INFO ]     PERFORMANCE_HINT: THROUGHPUT
[ INFO ]     PERFORMANCE_HINT_NUM_REQUESTS: 0
[ INFO ]     PERF_COUNT: NO
[ INFO ]     SCHEDULING_CORE_TYPE: SchedulingCoreType.ANY_CORE
[ INFO ]   MODEL_PRIORITY: Priority.MEDIUM
[ INFO ]   LOADED_FROM_CACHE: False
[Step 9/11] Creating infer requests and preparing input tensors
[ WARNING ] No input files were given for input 'batched_images'!. This input will be filled with random values!
[ WARNING ] No input files were given for input 'batched_points'!. This input will be filled with random values!
[ WARNING ] No input files were given for input 'batched_point_labels'!. This input will be filled with random values!
[ INFO ] Fill input 'batched_images' with random values 
[ INFO ] Fill input 'batched_points' with random values 
[ INFO ] Fill input 'batched_point_labels' with random values 
[Step 10/11] Measuring performance (Start inference asynchronously, 12 inference requests, limits: 15000 ms duration)
[ INFO ] Benchmarking in full mode (inputs filling are included in measurement loop).
[ INFO ] First inference took 897.36 ms
[Step 11/11] Dumping statistics report
[ INFO ] Execution Devices:['CPU']
[ INFO ] Count:            61 iterations
[ INFO ] Duration:         17901.14 ms
[ INFO ] Latency:
[ INFO ]    Median:        3359.11 ms
[ INFO ]    Average:       3383.00 ms
[ INFO ]    Min:           1396.16 ms
[ INFO ]    Max:           4409.79 ms
[ INFO ] Throughput:   3.41 FPS

In [24]:

if to_quantize.value:
    !benchmark_app -m $quantized_model_path -d $device.value -data_shape "batched_images[1,3,512,512],batched_points[1,1,2,2],batched_point_labels[1,1,2]" -t 15

[Step 1/11] Parsing and validating input arguments
[ INFO ] Parsing input parameters
[Step 2/11] Loading OpenVINO Runtime
[ INFO ] OpenVINO:
[ INFO ] Build ................................. 2023.3.0-13739-294cc6668c4
[ INFO ] 
[ INFO ] Device info:
[ INFO ] AUTO
[ INFO ] Build ................................. 2023.3.0-13739-294cc6668c4
[ INFO ] 
[ INFO ] 
[Step 3/11] Setting device configuration
[ WARNING ] Performance hint was not explicitly specified in command line. Device(AUTO) performance hint will be set to PerformanceMode.THROUGHPUT.
[Step 4/11] Reading model files
[ INFO ] Loading model files
[ INFO ] Read model took 79.61 ms
[ INFO ] Original model I/O parameters:
[ INFO ] Model inputs:
[ INFO ]     batched_images (node: batched_images) : f32 / [...] / [?,?,?,?]
[ INFO ]     batched_points (node: batched_points) : i64 / [...] / [?,?,?,?]
[ INFO ]     batched_point_labels (node: batched_point_labels) : i64 / [...] / [?,?,?]
[ INFO ] Model outputs:
[ INFO ]     146 (node: aten::take_along_dim/GatherElements_2824) : f32 / [...] / [?,?,?,?,?]
[ INFO ]     140 (node: aten::take_along_dim/GatherElements) : f32 / [...] / [?,?,?]
[Step 5/11] Resizing model to match image sizes and given batch
[ INFO ] Model batch size: 1
[Step 6/11] Configuring input of the model
[ INFO ] Model inputs:
[ INFO ]     batched_images (node: batched_images) : f32 / [...] / [?,?,?,?]
[ INFO ]     batched_points (node: batched_points) : i64 / [...] / [?,?,?,?]
[ INFO ]     batched_point_labels (node: batched_point_labels) : i64 / [...] / [?,?,?]
[ INFO ] Model outputs:
[ INFO ]     146 (node: aten::take_along_dim/GatherElements_2824) : f32 / [...] / [?,?,?,?,?]
[ INFO ]     140 (node: aten::take_along_dim/GatherElements) : f32 / [...] / [?,?,?]
[Step 7/11] Loading the model to the device
[ INFO ] Compile model took 1999.37 ms
[Step 8/11] Querying optimal runtime parameters
[ INFO ] Model:
[ INFO ]   NETWORK_NAME: Model0
[ INFO ]   EXECUTION_DEVICES: ['CPU']
[ INFO ]   PERFORMANCE_HINT: PerformanceMode.THROUGHPUT
[ INFO ]   OPTIMAL_NUMBER_OF_INFER_REQUESTS: 12
[ INFO ]   MULTI_DEVICE_PRIORITIES: CPU
[ INFO ]   CPU:
[ INFO ]     AFFINITY: Affinity.CORE
[ INFO ]     CPU_DENORMALS_OPTIMIZATION: False
[ INFO ]     CPU_SPARSE_WEIGHTS_DECOMPRESSION_RATE: 1.0
[ INFO ]     ENABLE_CPU_PINNING: True
[ INFO ]     ENABLE_HYPER_THREADING: True
[ INFO ]     EXECUTION_DEVICES: ['CPU']
[ INFO ]     EXECUTION_MODE_HINT: ExecutionMode.PERFORMANCE
[ INFO ]     INFERENCE_NUM_THREADS: 36
[ INFO ]     INFERENCE_PRECISION_HINT: <Type: 'float32'>
[ INFO ]     NETWORK_NAME: Model0
[ INFO ]     NUM_STREAMS: 12
[ INFO ]     OPTIMAL_NUMBER_OF_INFER_REQUESTS: 12
[ INFO ]     PERFORMANCE_HINT: THROUGHPUT
[ INFO ]     PERFORMANCE_HINT_NUM_REQUESTS: 0
[ INFO ]     PERF_COUNT: NO
[ INFO ]     SCHEDULING_CORE_TYPE: SchedulingCoreType.ANY_CORE
[ INFO ]   MODEL_PRIORITY: Priority.MEDIUM
[ INFO ]   LOADED_FROM_CACHE: False
[Step 9/11] Creating infer requests and preparing input tensors
[ WARNING ] No input files were given for input 'batched_images'!. This input will be filled with random values!
[ WARNING ] No input files were given for input 'batched_points'!. This input will be filled with random values!
[ WARNING ] No input files were given for input 'batched_point_labels'!. This input will be filled with random values!
[ INFO ] Fill input 'batched_images' with random values 
[ INFO ] Fill input 'batched_points' with random values 
[ INFO ] Fill input 'batched_point_labels' with random values 
[Step 10/11] Measuring performance (Start inference asynchronously, 12 inference requests, limits: 15000 ms duration)
[ INFO ] Benchmarking in full mode (inputs filling are included in measurement loop).
[ INFO ] First inference took 793.99 ms
[Step 11/11] Dumping statistics report
[ INFO ] Execution Devices:['CPU']
[ INFO ] Count:            61 iterations
[ INFO ] Duration:         16503.21 ms
[ INFO ] Latency:
[ INFO ]    Median:        3173.26 ms
[ INFO ]    Average:       3110.94 ms
[ INFO ]    Min:           1304.14 ms
[ INFO ]    Max:           3936.69 ms
[ INFO ] Throughput:   3.70 FPS

Interactive segmentation demo¶

In [ ]:

import numpy as np
from PIL import Image
import cv2
import matplotlib.pyplot as plt


def sigmoid(x):
    return 1 / (1 + np.exp(-x))


def format_results(masks, scores, logits, filter=0):
    annotations = []
    n = len(scores)
    for i in range(n):
        annotation = {}

        mask = masks[i]
        tmp = np.where(mask != 0)
        if np.sum(mask) < filter:
            continue
        annotation["id"] = i
        annotation["segmentation"] = mask
        annotation["bbox"] = [
            np.min(tmp[0]),
            np.min(tmp[1]),
            np.max(tmp[1]),
            np.max(tmp[0]),
        ]
        annotation["score"] = scores[i]
        annotation["area"] = annotation["segmentation"].sum()
        annotations.append(annotation)
    return annotations


def point_prompt(masks, points, point_label, target_height, target_width):  # numpy
    h = masks[0]["segmentation"].shape[0]
    w = masks[0]["segmentation"].shape[1]
    if h != target_height or w != target_width:
        points = [[int(point[0] * w / target_width), int(point[1] * h / target_height)] for point in points]
    onemask = np.zeros((h, w))
    for i, annotation in enumerate(masks):
        if isinstance(annotation, dict):
            mask = annotation["segmentation"]
        else:
            mask = annotation
        for i, point in enumerate(points):
            if point[1] < mask.shape[0] and point[0] < mask.shape[1]:
                if mask[point[1], point[0]] == 1 and point_label[i] == 1:
                    onemask += mask
                if mask[point[1], point[0]] == 1 and point_label[i] == 0:
                    onemask -= mask
    onemask = onemask >= 1
    return onemask, 0


def show_mask(
    annotation,
    ax,
    random_color=False,
    bbox=None,
    retinamask=True,
    target_height=960,
    target_width=960,
):
    mask_sum = annotation.shape[0]
    height = annotation.shape[1]
    weight = annotation.shape[2]
    # annotation is sorted by area
    areas = np.sum(annotation, axis=(1, 2))
    sorted_indices = np.argsort(areas)[::1]
    annotation = annotation[sorted_indices]

    index = (annotation != 0).argmax(axis=0)
    if random_color:
        color = np.random.random((mask_sum, 1, 1, 3))
    else:
        color = np.ones((mask_sum, 1, 1, 3)) * np.array([30 / 255, 144 / 255, 255 / 255])
    transparency = np.ones((mask_sum, 1, 1, 1)) * 0.6
    visual = np.concatenate([color, transparency], axis=-1)
    mask_image = np.expand_dims(annotation, -1) * visual

    mask = np.zeros((height, weight, 4))

    h_indices, w_indices = np.meshgrid(np.arange(height), np.arange(weight), indexing="ij")
    indices = (index[h_indices, w_indices], h_indices, w_indices, slice(None))

    mask[h_indices, w_indices, :] = mask_image[indices]
    if bbox is not None:
        x1, y1, x2, y2 = bbox
        ax.add_patch(plt.Rectangle((x1, y1), x2 - x1, y2 - y1, fill=False, edgecolor="b", linewidth=1))

    if not retinamask:
        mask = cv2.resize(mask, (target_width, target_height), interpolation=cv2.INTER_NEAREST)

    return mask


def process(
    annotations,
    image,
    scale,
    better_quality=False,
    mask_random_color=True,
    bbox=None,
    points=None,
    use_retina=True,
    withContours=True,
):
    if isinstance(annotations[0], dict):
        annotations = [annotation["segmentation"] for annotation in annotations]

    original_h = image.height
    original_w = image.width
    if better_quality:
        if isinstance(annotations[0], torch.Tensor):
            annotations = np.array(annotations)
        for i, mask in enumerate(annotations):
            mask = cv2.morphologyEx(mask.astype(np.uint8), cv2.MORPH_CLOSE, np.ones((3, 3), np.uint8))
            annotations[i] = cv2.morphologyEx(mask.astype(np.uint8), cv2.MORPH_OPEN, np.ones((8, 8), np.uint8))
    annotations = np.array(annotations)
    inner_mask = show_mask(
        annotations,
        plt.gca(),
        random_color=mask_random_color,
        bbox=bbox,
        retinamask=use_retina,
        target_height=original_h,
        target_width=original_w,
    )

    if isinstance(annotations, torch.Tensor):
        annotations = annotations.cpu().numpy()

    if withContours:
        contour_all = []
        temp = np.zeros((original_h, original_w, 1))
        for i, mask in enumerate(annotations):
            if isinstance(mask, dict):
                mask = mask["segmentation"]
            annotation = mask.astype(np.uint8)
            if not use_retina:
                annotation = cv2.resize(
                    annotation,
                    (original_w, original_h),
                    interpolation=cv2.INTER_NEAREST,
                )
            contours, _ = cv2.findContours(annotation, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
            for contour in contours:
                contour_all.append(contour)
        cv2.drawContours(temp, contour_all, -1, (255, 255, 255), 2 // scale)
        color = np.array([0 / 255, 0 / 255, 255 / 255, 0.9])
        contour_mask = temp / 255 * color.reshape(1, 1, -1)

    image = image.convert("RGBA")
    overlay_inner = Image.fromarray((inner_mask * 255).astype(np.uint8), "RGBA")
    image.paste(overlay_inner, (0, 0), overlay_inner)

    if withContours:
        overlay_contour = Image.fromarray((contour_mask * 255).astype(np.uint8), "RGBA")
        image.paste(overlay_contour, (0, 0), overlay_contour)

    return image


def segment_with_boxs(
    image,
    seg_image,
    global_points,
    global_point_label,
    input_size=1024,
    better_quality=False,
    withContours=True,
    use_retina=True,
    mask_random_color=True,
):
    if global_points is None or len(global_points) < 2 or global_points[0] is None:
        return image, global_points, global_point_label

    input_size = int(input_size)
    w, h = image.size
    scale = input_size / max(w, h)
    new_w = int(w * scale)
    new_h = int(h * scale)
    image = image.resize((new_w, new_h))

    scaled_points = np.array([[int(x * scale) for x in point] for point in global_points])
    scaled_points = scaled_points[:2]
    scaled_point_label = np.array(global_point_label)[:2]

    if scaled_points.size == 0 and scaled_point_label.size == 0:
        return image, global_points, global_point_label

    nd_image = np.array(image)
    img_tensor = nd_image.astype(np.float32) / 255
    img_tensor = np.transpose(img_tensor, (2, 0, 1))

    pts_sampled = np.reshape(scaled_points, [1, 1, -1, 2])
    pts_sampled = pts_sampled[:, :, :2, :]
    pts_labels = np.reshape(np.array([2, 3]), [1, 1, 2])

    results = compiled_model([img_tensor[None, ...], pts_sampled, pts_labels])
    predicted_logits = results[0]
    predicted_iou = results[1]
    all_masks = sigmoid(predicted_logits[0, 0, :, :, :]) >= 0.5
    predicted_iou = predicted_iou[0, 0, ...]

    max_predicted_iou = -1
    selected_mask_using_predicted_iou = None
    selected_predicted_iou = None

    for m in range(all_masks.shape[0]):
        curr_predicted_iou = predicted_iou[m]
        if curr_predicted_iou > max_predicted_iou or selected_mask_using_predicted_iou is None:
            max_predicted_iou = curr_predicted_iou
            selected_mask_using_predicted_iou = all_masks[m : m + 1]
            selected_predicted_iou = predicted_iou[m : m + 1]

    results = format_results(selected_mask_using_predicted_iou, selected_predicted_iou, predicted_logits, 0)

    annotations = results[0]["segmentation"]
    annotations = np.array([annotations])
    fig = process(
        annotations=annotations,
        image=image,
        scale=(1024 // input_size),
        better_quality=better_quality,
        mask_random_color=mask_random_color,
        use_retina=use_retina,
        bbox=scaled_points.reshape([4]),
        withContours=withContours,
    )

    global_points = []
    global_point_label = []
    return fig, global_points, global_point_label


def segment_with_points(
    image,
    global_points,
    global_point_label,
    input_size=1024,
    better_quality=False,
    withContours=True,
    use_retina=True,
    mask_random_color=True,
):
    input_size = int(input_size)
    w, h = image.size
    scale = input_size / max(w, h)
    new_w = int(w * scale)
    new_h = int(h * scale)
    image = image.resize((new_w, new_h))

    if global_points is None or len(global_points) < 1 or global_points[0] is None:
        return image, global_points, global_point_label
    scaled_points = np.array([[int(x * scale) for x in point] for point in global_points])
    scaled_point_label = np.array(global_point_label)

    if scaled_points.size == 0 and scaled_point_label.size == 0:
        return image, global_points, global_point_label

    nd_image = np.array(image)
    img_tensor = (nd_image).astype(np.float32) / 255
    img_tensor = np.transpose(img_tensor, (2, 0, 1))

    pts_sampled = np.reshape(scaled_points, [1, 1, -1, 2])
    pts_labels = np.reshape(np.array(global_point_label), [1, 1, -1])

    results = compiled_model([img_tensor[None, ...], pts_sampled, pts_labels])
    predicted_logits = results[0]
    predicted_iou = results[1]
    all_masks = sigmoid(predicted_logits[0, 0, :, :, :]) >= 0.5
    predicted_iou = predicted_iou[0, 0, ...]

    results = format_results(all_masks, predicted_iou, predicted_logits, 0)
    annotations, _ = point_prompt(results, scaled_points, scaled_point_label, new_h, new_w)
    annotations = np.array([annotations])

    fig = process(
        annotations=annotations,
        image=image,
        scale=(1024 // input_size),
        better_quality=better_quality,
        mask_random_color=mask_random_color,
        points=scaled_points,
        bbox=None,
        use_retina=use_retina,
        withContours=withContours,
    )

    global_points = []
    global_point_label = []
    # return fig, None
    return fig, global_points, global_point_label

In [ ]:

# Go back to the efficient-sam notebook directory
%cd ..

if not Path("gradio_helper.py").exists():
    r = requests.get(url="https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/latest/notebooks/efficient-sam/gradio_helper.py")
    open("gradio_helper.py", "w").write(r.text)

from gradio_helper import make_demo

demo = make_demo(segment_with_point_fn=segment_with_points, segment_with_box_fn=segment_with_boxs)

try:
    demo.queue().launch(debug=True)
except Exception:
    demo.queue().launch(share=True, debug=True)
# if you are launching remotely, specify server_name and server_port
# demo.launch(server_name='your server name', server_port='server port in int')
# Read more in the docs: https://gradio.app/docs/

Object segmentations with EfficientSAM and OpenVINO¶

Table of contents:¶

Installation Instructions¶

Prerequisites¶

Load PyTorch model¶

Run PyTorch model inference¶

Prepare input data¶

Define helpers for input and output processing¶

Convert model to OpenVINO IR format¶

Run OpenVINO model inference¶

Select inference device from dropdown list¶

Compile OpenVINO model¶

Inference and visualize result¶

Quantization¶

Prepare calibration datasets¶

Run Model Quantization¶

Verify quantized model inference¶

Save quantize model on disk¶

Compare quantized model size¶

Compare inference time of the FP16 and INT8 models¶

Interactive segmentation demo¶