app.py

from pickle import HIGHEST_PROTOCOL
# Import necessary libraries
import numpy as np
import math
import matplotlib.pyplot as plt
import cv2
import json
import gradio as gr
from huggingface_hub import hf_hub_download
from onnx import hub
import onnxruntime as ort
import tempfile
import onnx

# Load the ONNX model from ONNX Model Zoo
model = hub.load("efficientnet-lite4")

# Save the ModelProto object to a temporary file
with tempfile.NamedTemporaryFile(suffix=".onnx", delete=False) as temp_file:
    onnx.save(model, temp_file.name)
    model_path = temp_file.name

# Load the labels from a text file
labels = json.load(open("/content/drive/MyDrive/labels_map.txt", "r"))

# Define a function to preprocess the image for the EfficientNet-Lite4 model
def pre_process_edgetpu(img, dims):
    # Unpack the dimensions
    output_height, output_width, _ = dims
    # Resize the image while maintaining aspect ratio
    img = resize_with_aspectratio(
        img,
        output_height,
        output_width,
        inter_pol=cv2.INTER_LINEAR
    )
    # Crop the image from the center
    img = center_crop(img, output_height, output_width)
    # Convert image to float32 numpy array
    img = np.asarray(img, dtype='float32')
    # Normalize pixel values from [0-255] to [-1.0, 1.0]
    img -= [127.0, 127.0, 127.0]
    img /= [128.0, 128.0, 128.0]
    return img

# Define a function to resize the image while maintaining aspect ratio
def resize_with_aspectratio(
        img,
        out_height,
        out_width,
        scale=87.5,
        inter_pol=cv2.INTER_LINEAR):
    # Get original image dimensions
    height, width, _ = img.shape
    # Calculate new dimensions
    new_height = int(100. * out_height / scale)
    new_width = int(100. * out_width / scale)
    # Determine which dimension to scale based on aspect ratio
    if height > width:
        w = new_width
        h = int(new_height * height / width)
    else:
        h = new_height
        w = int(new_width * width / height)
    # Resize the image
    img = cv2.resize(img, (w, h), interpolation=inter_pol)
    return img

# Define a function to crop the image from the center
def center_crop(img, out_height, out_width):
    # Get image dimensions
    height, width, _ = img.shape
    # Calculate crop coordinates
    left = int((width - out_width) / 2)
    right = int((width + out_width) / 2)
    top = int((height - out_height) / 2)
    bottom = int((height + out_height) / 2)
    # Crop the image
    img = img[top:bottom, left:right]
    return img

# Create an ONNX Runtime inference session
sess = ort.InferenceSession(model_path)

# Define the main inference function
def inference(img):
    # Read the image file
    img = cv2.imread(img)
    # Convert BGR to RGB color space
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

    # Preprocess the image
    img = pre_process_edgetpu(img, (224, 224, 3))

    # Add batch dimension to the image
    img_batch = np.expand_dims(img, axis=0)

    # Run inference using the ONNX model
    results = sess.run(["Softmax:0"], {"images:0": img_batch})[0]
    # Get the top 5 predictions
    result = reversed(results[0].argsort()[-5:])
    # Create a dictionary to store results
    resultdic = {}
    for r in result:
        resultdic[labels[str(r)]] = float(results[0][r])
    return resultdic

# Set up the Gradio interface
title = "EfficientNet-Lite4"
description = """EfficientNet-Lite 4 is the largest variant and most accurate of the set of
EfficientNet-Lite model. It is an integer-only quantized model that produces the HIGHEST_PROTOCOL
accuracy of all of the EfficientNet models. It achieves 80.4% ImageNet top-1 accuracy, while
still running in real-time (e.g. 30ms/image) on a Pixel 4 CPU."""

examples = [['catonnx.jpg']]

# Launch the Gradio interface
gr.Interface(
    inference,
    gr.Image(type="filepath"),
    "label",
    title=title,
    description=description,
    examples=examples).launch()