Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -4,112 +4,99 @@ import json
|
|
4 |
import os
|
5 |
from PIL import Image
|
6 |
import onnxruntime as rt
|
|
|
7 |
class ONNXModel:
|
8 |
def __init__(self, dir_path) -> None:
|
9 |
-
"""
|
10 |
model_dir = os.path.dirname(dir_path)
|
11 |
with open(os.path.join(model_dir, "signature.json"), "r") as f:
|
12 |
self.signature = json.load(f)
|
|
|
13 |
self.model_file = os.path.join(model_dir, self.signature.get("filename"))
|
14 |
if not os.path.isfile(self.model_file):
|
15 |
-
raise FileNotFoundError(
|
16 |
-
|
17 |
self.signature_inputs = self.signature.get("inputs")
|
18 |
self.signature_outputs = self.signature.get("outputs")
|
19 |
-
|
20 |
if "Image" not in self.signature_inputs:
|
21 |
-
raise ValueError("ONNX model
|
22 |
-
|
23 |
-
#
|
24 |
version = self.signature.get("export_model_version")
|
25 |
if version is None or version != EXPORT_MODEL_VERSION:
|
26 |
-
print(
|
27 |
-
|
28 |
-
|
29 |
|
30 |
def load(self) -> None:
|
31 |
-
"""Load the model
|
32 |
-
|
33 |
-
self.session = rt.InferenceSession(path_or_bytes=self.model_file)
|
34 |
|
35 |
def predict(self, image: Image.Image) -> dict:
|
36 |
-
"""
|
37 |
-
|
38 |
-
"""
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
fetches = [(key, value.get("name")) for key, value in self.signature_outputs.items()]
|
43 |
-
# make the image a batch of 1
|
44 |
-
feed = {self.signature_inputs.get("Image").get("name"): [img]}
|
45 |
-
outputs = self.session.run(output_names=[name for (_, name) in fetches], input_feed=feed)
|
46 |
-
return self.process_output(fetches, outputs)
|
47 |
|
48 |
def process_image(self, image: Image.Image, input_shape: list) -> np.ndarray:
|
49 |
-
"""
|
50 |
-
Given a PIL Image, center square crop and resize to fit the expected model input, and convert from [0,255] to [0,1] values.
|
51 |
-
"""
|
52 |
width, height = image.size
|
53 |
-
# ensure image type is compatible with model and convert if not
|
54 |
if image.mode != "RGB":
|
55 |
image = image.convert("RGB")
|
56 |
-
# center crop image (you can substitute any other method to make a square image, such as just resizing or padding edges with 0)
|
57 |
-
if width != height:
|
58 |
-
square_size = min(width, height)
|
59 |
-
left = (width - square_size) / 2
|
60 |
-
top = (height - square_size) / 2
|
61 |
-
right = (width + square_size) / 2
|
62 |
-
bottom = (height + square_size) / 2
|
63 |
-
# Crop the center of the image
|
64 |
-
image = image.crop((left, top, right, bottom))
|
65 |
-
# now the image is square, resize it to be the right shape for the model input
|
66 |
-
input_width, input_height = input_shape[1:3]
|
67 |
-
if image.width != input_width or image.height != input_height:
|
68 |
-
image = image.resize((input_width, input_height))
|
69 |
|
70 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
71 |
image = np.asarray(image) / 255.0
|
72 |
-
# format input as model expects
|
73 |
return image.astype(np.float32)
|
74 |
|
75 |
-
def process_output(self,
|
76 |
-
|
77 |
-
# convert to normal python types with tolist(), and convert any byte strings to normal strings with .decode()
|
78 |
out_keys = ["label", "confidence"]
|
79 |
-
results = {}
|
80 |
-
for i, (key, _) in enumerate(fetches):
|
81 |
-
val = outputs[i].tolist()[0]
|
82 |
-
if isinstance(val, bytes):
|
83 |
-
val = val.decode()
|
84 |
-
results[key] = val
|
85 |
confs = results["Confidences"]
|
86 |
-
labels = self.signature
|
87 |
output = [dict(zip(out_keys, group)) for group in zip(labels, confs)]
|
88 |
-
|
89 |
-
|
90 |
-
EXPORT_MODEL_VERSION=1
|
91 |
model = ONNXModel(dir_path="model.onnx")
|
92 |
model.load()
|
93 |
|
94 |
-
|
95 |
def predict(image):
|
96 |
-
|
|
|
97 |
prediction = model.predict(image)
|
98 |
for output in prediction["predictions"]:
|
99 |
output["confidence"] = round(output["confidence"], 4)
|
100 |
return prediction
|
101 |
|
102 |
-
inputs = gr.Image(
|
103 |
outputs = gr.JSON()
|
104 |
|
105 |
-
description =
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
114 |
|
|
|
115 |
|
|
|
|
4 |
import os
|
5 |
from PIL import Image
|
6 |
import onnxruntime as rt
|
7 |
+
|
8 |
class ONNXModel:
|
9 |
def __init__(self, dir_path) -> None:
|
10 |
+
"""Load model metadata and initialize ONNX session."""
|
11 |
model_dir = os.path.dirname(dir_path)
|
12 |
with open(os.path.join(model_dir, "signature.json"), "r") as f:
|
13 |
self.signature = json.load(f)
|
14 |
+
|
15 |
self.model_file = os.path.join(model_dir, self.signature.get("filename"))
|
16 |
if not os.path.isfile(self.model_file):
|
17 |
+
raise FileNotFoundError("Model file does not exist.")
|
18 |
+
|
19 |
self.signature_inputs = self.signature.get("inputs")
|
20 |
self.signature_outputs = self.signature.get("outputs")
|
21 |
+
|
22 |
if "Image" not in self.signature_inputs:
|
23 |
+
raise ValueError("ONNX model must have an 'Image' input. Check signature.json.")
|
24 |
+
|
25 |
+
# Check export version
|
26 |
version = self.signature.get("export_model_version")
|
27 |
if version is None or version != EXPORT_MODEL_VERSION:
|
28 |
+
print(f"Warning: Expected model version {EXPORT_MODEL_VERSION}, but found {version}.")
|
29 |
+
|
30 |
+
self.session = None
|
31 |
|
32 |
def load(self) -> None:
|
33 |
+
"""Load the ONNX model with execution providers."""
|
34 |
+
self.session = rt.InferenceSession(self.model_file, providers=["CPUExecutionProvider"])
|
|
|
35 |
|
36 |
def predict(self, image: Image.Image) -> dict:
|
37 |
+
"""Process image and run ONNX model inference."""
|
38 |
+
img = self.process_image(image, self.signature_inputs["Image"]["shape"])
|
39 |
+
feed = {self.signature_inputs["Image"]["name"]: [img]}
|
40 |
+
output_names = [self.signature_outputs[key]["name"] for key in self.signature_outputs]
|
41 |
+
outputs = self.session.run(output_names=output_names, input_feed=feed)
|
42 |
+
return self.process_output(outputs)
|
|
|
|
|
|
|
|
|
|
|
43 |
|
44 |
def process_image(self, image: Image.Image, input_shape: list) -> np.ndarray:
|
45 |
+
"""Resize and normalize the image."""
|
|
|
|
|
46 |
width, height = image.size
|
|
|
47 |
if image.mode != "RGB":
|
48 |
image = image.convert("RGB")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
49 |
|
50 |
+
square_size = min(width, height)
|
51 |
+
left = (width - square_size) / 2
|
52 |
+
top = (height - square_size) / 2
|
53 |
+
right = (width + square_size) / 2
|
54 |
+
bottom = (height + square_size) / 2
|
55 |
+
image = image.crop((left, top, right, bottom))
|
56 |
+
|
57 |
+
input_width, input_height = input_shape[1:3]
|
58 |
+
image = image.resize((input_width, input_height))
|
59 |
image = np.asarray(image) / 255.0
|
|
|
60 |
return image.astype(np.float32)
|
61 |
|
62 |
+
def process_output(self, outputs: list) -> dict:
|
63 |
+
"""Format the model output."""
|
|
|
64 |
out_keys = ["label", "confidence"]
|
65 |
+
results = {key: outputs[i].tolist()[0] for i, key in enumerate(self.signature_outputs)}
|
|
|
|
|
|
|
|
|
|
|
66 |
confs = results["Confidences"]
|
67 |
+
labels = self.signature["classes"]["Label"]
|
68 |
output = [dict(zip(out_keys, group)) for group in zip(labels, confs)]
|
69 |
+
return {"predictions": sorted(output, key=lambda x: x["confidence"], reverse=True)}
|
70 |
+
|
71 |
+
EXPORT_MODEL_VERSION = 1
|
72 |
model = ONNXModel(dir_path="model.onnx")
|
73 |
model.load()
|
74 |
|
|
|
75 |
def predict(image):
|
76 |
+
"""Run inference on the given image."""
|
77 |
+
image = Image.fromarray(np.uint8(image), "RGB")
|
78 |
prediction = model.predict(image)
|
79 |
for output in prediction["predictions"]:
|
80 |
output["confidence"] = round(output["confidence"], 4)
|
81 |
return prediction
|
82 |
|
83 |
+
inputs = gr.Image(image_mode="RGB")
|
84 |
outputs = gr.JSON()
|
85 |
|
86 |
+
description = (
|
87 |
+
"This is a web interface for the Naked Detector model. "
|
88 |
+
"Upload an image and get predictions for the presence of nudity.\n"
|
89 |
+
"Model and website created by KUO SUKO, C110156115 NKUST."
|
90 |
+
)
|
|
|
|
|
|
|
91 |
|
92 |
+
interface = gr.Interface(
|
93 |
+
fn=predict,
|
94 |
+
inputs=inputs,
|
95 |
+
outputs=outputs,
|
96 |
+
title="Naked Detector",
|
97 |
+
description=description
|
98 |
+
)
|
99 |
|
100 |
+
interface.launch()
|
101 |
|
102 |
+
# this is changed by ChatGPT, if it run like a shit. don't blame me ><
|