Create app.py

app.py

@@ -0,0 +1,126 @@
+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()
+
+