Create infer-rotation.py
Browse files- infer-rotation.py +57 -0
infer-rotation.py
ADDED
@@ -0,0 +1,57 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from PIL import Image
|
2 |
+
import numpy as np
|
3 |
+
|
4 |
+
from transformers import ViTImageProcessor, TFViTModel
|
5 |
+
import keras
|
6 |
+
import argparse
|
7 |
+
|
8 |
+
VIT_WEIGHTS_PATH = "model-vit-ang-loss.h5"
|
9 |
+
BASE_MODEL = "google/vit-base-patch16-224"
|
10 |
+
IMAGE_SIZE = 224
|
11 |
+
|
12 |
+
class Inference:
|
13 |
+
def __init__(self):
|
14 |
+
self.vit_model = self._load_vit_model()
|
15 |
+
self.image_preprocessor = self._load_image_preprocessor()
|
16 |
+
|
17 |
+
def predict_rotation(self, image_path):
|
18 |
+
X = self._preprocess(image_path)
|
19 |
+
y = self.vit_model.predict(X)[0][0]
|
20 |
+
return y
|
21 |
+
|
22 |
+
def _preprocess(self, image_path):
|
23 |
+
img = Image.open(image_path)
|
24 |
+
img = img.resize((IMAGE_SIZE, IMAGE_SIZE))
|
25 |
+
img = np.array(img)
|
26 |
+
|
27 |
+
X_vit = self.image_preprocessor.preprocess(images=[img], return_tensors="pt")["pixel_values"]
|
28 |
+
return np.array(X_vit)
|
29 |
+
|
30 |
+
def _load_image_preprocessor(self):
|
31 |
+
print("Loading Image Preprocessor")
|
32 |
+
return ViTImageProcessor.from_pretrained(BASE_MODEL)
|
33 |
+
|
34 |
+
def _load_vit_model(self):
|
35 |
+
print("Loading Model")
|
36 |
+
vit_base = TFViTModel.from_pretrained(BASE_MODEL)
|
37 |
+
|
38 |
+
img_input = keras.layers.Input(shape=(3,IMAGE_SIZE, IMAGE_SIZE))
|
39 |
+
x = vit_base.vit(img_input)
|
40 |
+
y = keras.layers.Dense(1, activation="linear")(x[-1])
|
41 |
+
|
42 |
+
model = keras.Model(inputs=img_input, outputs=y)
|
43 |
+
print(model.summary())
|
44 |
+
|
45 |
+
print("Loading Weights")
|
46 |
+
model.load_weights(VIT_WEIGHTS_PATH)
|
47 |
+
|
48 |
+
return model
|
49 |
+
|
50 |
+
if __name__=="__main__":
|
51 |
+
parser = argparse.ArgumentParser()
|
52 |
+
parser.add_argument("--image-path", type=str, required=True)
|
53 |
+
args = parser.parse_args()
|
54 |
+
|
55 |
+
model = Inference()
|
56 |
+
expected_angle = model.predict_rotation(args.image_path)
|
57 |
+
print(f"Predicted angle is about '{expected_angle}' degrees")
|