Create test_infer_onnx.py
#6
by
wangfangyuan
- opened
- test_infer_onnx.py +154 -0
test_infer_onnx.py
ADDED
@@ -0,0 +1,154 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import onnxruntime
|
4 |
+
import numpy as np
|
5 |
+
import argparse
|
6 |
+
from utils import (
|
7 |
+
LoadImages,
|
8 |
+
non_max_suppression,
|
9 |
+
plot_images,
|
10 |
+
output_to_target,
|
11 |
+
)
|
12 |
+
import sys
|
13 |
+
import pathlib
|
14 |
+
CURRENT_DIR = pathlib.Path(__file__).parent
|
15 |
+
sys.path.append(str(CURRENT_DIR))
|
16 |
+
from optimum.amd.ryzenai import RyzenAIModelForObjectDetection
|
17 |
+
|
18 |
+
def preprocess(img):
|
19 |
+
img = torch.from_numpy(img)
|
20 |
+
img = img.float() # uint8 to fp16/32
|
21 |
+
img /= 255 # 0 - 255 to 0.0 - 1.0
|
22 |
+
return img
|
23 |
+
|
24 |
+
|
25 |
+
class DFL(nn.Module):
|
26 |
+
# Integral module of Distribution Focal Loss (DFL) proposed in Generalized Focal Loss https://ieeexplore.ieee.org/document/9792391
|
27 |
+
def __init__(self, c1=16):
|
28 |
+
super().__init__()
|
29 |
+
self.conv = nn.Conv2d(c1, 1, 1, bias=False).requires_grad_(False)
|
30 |
+
x = torch.arange(c1, dtype=torch.float)
|
31 |
+
self.conv.weight.data[:] = nn.Parameter(x.view(1, c1, 1, 1))
|
32 |
+
self.c1 = c1
|
33 |
+
|
34 |
+
def forward(self, x):
|
35 |
+
b, c, a = x.shape # batch, channels, anchors
|
36 |
+
return self.conv(x.view(b, 4, self.c1, a).transpose(2, 1).softmax(1)).view(
|
37 |
+
b, 4, a
|
38 |
+
)
|
39 |
+
|
40 |
+
|
41 |
+
def dist2bbox(distance, anchor_points, xywh=True, dim=-1):
|
42 |
+
"""Transform distance(ltrb) to box(xywh or xyxy)."""
|
43 |
+
lt, rb = torch.split(distance, 2, dim)
|
44 |
+
x1y1 = anchor_points - lt
|
45 |
+
x2y2 = anchor_points + rb
|
46 |
+
if xywh:
|
47 |
+
c_xy = (x1y1 + x2y2) / 2
|
48 |
+
wh = x2y2 - x1y1
|
49 |
+
return torch.cat((c_xy, wh), dim) # xywh bbox
|
50 |
+
return torch.cat((x1y1, x2y2), dim) # xyxy bbox
|
51 |
+
|
52 |
+
|
53 |
+
def post_process(x):
|
54 |
+
dfl = DFL(16)
|
55 |
+
anchors = torch.tensor(
|
56 |
+
np.load(
|
57 |
+
"./anchors.npy",
|
58 |
+
allow_pickle=True,
|
59 |
+
)
|
60 |
+
)
|
61 |
+
strides = torch.tensor(
|
62 |
+
np.load(
|
63 |
+
"./strides.npy",
|
64 |
+
allow_pickle=True,
|
65 |
+
)
|
66 |
+
)
|
67 |
+
box, cls = torch.cat([xi.view(x[0].shape[0], 144, -1) for xi in x], 2).split(
|
68 |
+
(16 * 4, 80), 1
|
69 |
+
)
|
70 |
+
dbox = dist2bbox(dfl(box), anchors.unsqueeze(0), xywh=True, dim=1) * strides
|
71 |
+
y = torch.cat((dbox, cls.sigmoid()), 1)
|
72 |
+
return y, x
|
73 |
+
|
74 |
+
|
75 |
+
def make_parser():
|
76 |
+
parser = argparse.ArgumentParser("onnxruntime inference sample")
|
77 |
+
parser.add_argument(
|
78 |
+
"-m",
|
79 |
+
"--onnx_model",
|
80 |
+
type=str,
|
81 |
+
default="./yolov8m.onnx",
|
82 |
+
help="input your onnx model.",
|
83 |
+
)
|
84 |
+
parser.add_argument(
|
85 |
+
"-i",
|
86 |
+
"--image_path",
|
87 |
+
type=str,
|
88 |
+
default='./demo.jpg',
|
89 |
+
help="path to your input image.",
|
90 |
+
)
|
91 |
+
parser.add_argument(
|
92 |
+
"-o",
|
93 |
+
"--output_path",
|
94 |
+
type=str,
|
95 |
+
default='./demo_infer.jpg',
|
96 |
+
help="path to your output directory.",
|
97 |
+
)
|
98 |
+
parser.add_argument(
|
99 |
+
"--ipu", action='store_true', help='flag for ryzen ai'
|
100 |
+
)
|
101 |
+
parser.add_argument(
|
102 |
+
"--provider_config", default='', type=str, help='provider config for ryzen ai'
|
103 |
+
)
|
104 |
+
return parser
|
105 |
+
|
106 |
+
classnames = ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light',
|
107 |
+
'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow',
|
108 |
+
'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee',
|
109 |
+
'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard',
|
110 |
+
'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple',
|
111 |
+
'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch',
|
112 |
+
'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone',
|
113 |
+
'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear',
|
114 |
+
'hair drier', 'toothbrush']
|
115 |
+
names = {k: classnames[k] for k in range(80)}
|
116 |
+
imgsz = [640, 640]
|
117 |
+
|
118 |
+
|
119 |
+
if __name__ == '__main__':
|
120 |
+
args = make_parser().parse_args()
|
121 |
+
source = args.image_path
|
122 |
+
dataset = LoadImages(
|
123 |
+
source, imgsz=imgsz, stride=32, auto=False, transforms=None, vid_stride=1
|
124 |
+
)
|
125 |
+
onnx_weight = args.onnx_model
|
126 |
+
if args.ipu:
|
127 |
+
onnx_model = RyzenAIModelForObjectDetection.from_pretrained(".\\", vaip_config=args.provider_config)
|
128 |
+
# providers = ["VitisAIExecutionProvider"]
|
129 |
+
# provider_options = [{"config_file": args.provider_config}]
|
130 |
+
# onnx_model = onnxruntime.InferenceSession(onnx_weight, providers=providers, provider_options=provider_options)
|
131 |
+
else:
|
132 |
+
onnx_model = onnxruntime.InferenceSession(onnx_weight)
|
133 |
+
for batch in dataset:
|
134 |
+
path, im, im0s, vid_cap, s = batch
|
135 |
+
im = preprocess(im)
|
136 |
+
if len(im.shape) == 3:
|
137 |
+
im = im[None]
|
138 |
+
# outputs = onnx_model.run(None, {onnx_model.get_inputs()[0].name: im.cpu().numpy()})
|
139 |
+
# outputs = [torch.tensor(item) for item in outputs]
|
140 |
+
# outputs = onnx_model.run(None, {onnx_model.get_inputs()[0].name: im.permute(0, 2, 3, 1).cpu().numpy()})
|
141 |
+
# outputs = [torch.tensor(item).permute(0, 3, 1, 2) for item in outputs]
|
142 |
+
outputs = onnx_model(im.permute(0, 2, 3, 1))
|
143 |
+
outputs = [outputs[0].permute(0, 3, 1, 2), outputs[1].permute(0, 3, 1, 2), outputs[2].permute(0, 3, 1, 2)]
|
144 |
+
preds = post_process(outputs)
|
145 |
+
preds = non_max_suppression(
|
146 |
+
preds, 0.25, 0.7, agnostic=False, max_det=300, classes=None
|
147 |
+
)
|
148 |
+
plot_images(
|
149 |
+
im,
|
150 |
+
*output_to_target(preds, max_det=15),
|
151 |
+
source,
|
152 |
+
fname=args.output_path,
|
153 |
+
names=names,
|
154 |
+
)
|