转自：https://zhen8838.github.io/2019/07/07/tflite-kmodel/
我的yolo3模型在k210里面输出结果完全不对，所以我十分怀疑是量化出了问题，但是我又找不到问题。还好昨天case小姐姐帮忙更新了nncase，可以在pc上推理kmdoel.然后我推理了几个图像，这次就是记录一下这个脚本，免得下次要用找不到了。。。
代码：
import numpy as np
from scipy.special import expit, softmax
from tensorflow import lite
from pathlib import Path
from skimage.io import imread
from termcolor import colored
np.set_printoptions(suppress=True)
""" 加载tflite """
interpreter = lite.Interpreter(model_path=str('mobile_yolo.tflite'))
interpreter.allocate_tensors()
input_index = [details["index"] for details in interpreter.get_input_details()]
output_index = [details["index"] for details in interpreter.get_output_details()]
""" 加载图片 """
img_paths = list(Path('test_logs').glob('*.jpg'))
img_paths.sort()
imgs = np.array([imread(str(path)) for path in img_paths])
imgs = imgs / 255
""" 推理 """
def infer(img: np.ndarray) -> [np.ndarray, np.ndarray]:
inp = img[np.newaxis, ...].astype('float32')
interpreter.set_tensor(input_index[0], inp)
interpreter.invoke()
predictions = [interpreter.get_tensor(idx)[0] for idx in output_index]
return predictions
tf_res = [infer(img) for img in imgs]
""" 加载bin """
bin_paths = list(Path('output').glob('*.bin'))
bin_paths.sort()
def parser_bin(path: Path) -> [np.ndarray, np.ndarray]:
content = path.open('rb').read() # type:bytes
assert len(content) / 4 == 7 * 10 * 75 + 14 * 20 * 75
# out = np.fromstring(content, dtype='<f4')
out = np.array(np.frombuffer(content, '<f4'))
# out = [out[:7 * 10 * 75].reshape(7, 10, 75),
# out[7 * 10 * 75:].reshape(14, 20, 75)]
out = [np.transpose(out[:7 * 10 * 75].reshape(75, 7, 10), (1, 2, 0)),
np.transpose(out[7 * 10 * 75:].reshape(75, 14, 20), (1, 2, 0))]
return out
kmd_res = [parser_bin(path) for path in bin_paths]
""" 解析输出 """
inshape = (224, 320)
anchors = np.array([[[81, 82], [135, 169], [344, 319]], [[10, 14], [23, 27], [37, 58]]])
for i in range(len(kmd_res)):
for j in range(len(output_index)):
grid_shape = tf_res[i][j].shape[0:2]
a = tf_res[i][j].reshape(grid_shape + (3, 25))
b = kmd_res[i][j].reshape(grid_shape + (3, 25))
""" 解析xy """
grid_y = np.tile(np.reshape(np.arange(0, grid_shape[0]), [-1, 1, 1, 1]), [1, grid_shape[1], 1, 1])
grid_x = np.tile(np.reshape(np.arange(0, grid_shape[1]), [1, -1, 1, 1]), [grid_shape[0], 1, 1, 1])
grid = np.concatenate([grid_x, grid_y], axis=-1)
a[..., 0:2] = (expit(a[..., 0:2]) + grid) / grid_shape[::-1] * inshape[::-1]
b[..., 0:2] = (expit(b[..., 0:2]) + grid) / grid_shape[::-1] * inshape[::-1]
""" 解析wh """
a[..., 2:4] = np.exp(a[..., 2:4]) * anchors[j]
b[..., 2:4] = np.exp(b[..., 2:4]) * anchors[j]
""" 解析confendice """
a[..., 4:5] = expit(a[..., 4:5])
b[..., 4:5] = expit(b[..., 4:5])
""" 解析类别 """
a[..., 5] = np.argmax(softmax(a[..., 5:], axis=-1), axis=-1)
b[..., 5] = np.argmax(softmax(b[..., 5:], axis=-1), axis=-1)
print(colored(f'img {i} layer {j} tflite :\n', 'blue'), a[np.where(a[..., 4] > .6)][:, :5])
print(colored(f'img {i} layer {j} kmodel :\n', 'green'), b[np.where(b[..., 4] > .6)][:, :5])
结果
INFO: Initialized TensorFlow Lite runtime.
img 0 layer 0 tflite :
[]
img 0 layer 0 kmodel :
[]
img 0 layer 1 tflite :
[[ 50.671265 184.66817 16.581987 14.750296 0.7672671]
[135.05923 188.98267 14.629576 9.189983 0.7180456]
[180.27417 193.83612 18.99109 10.469457 0.7107621]
[258.86105 197.36963 41.65087 16.02682 0.8539901]
[259.66782 197.74883 33.983326 23.298958 0.720852 ]]
img 0 layer 1 kmodel :
[[ 51.26226 184. 13.131495 14. 0.69366026]]
img 1 layer 0 tflite :
[[172.43805 147.88477 210.21072 149.22597 0.7295683]]
img 1 layer 0 kmodel :
[[173.1715 147.51495 211.01398 141.3503 0.72762936]]
img 1 layer 1 tflite :
[]
img 1 layer 1 kmodel :
[[107.401764 186.7845 63.801384 52.96505 0.73084366]
[117.872925 187.09857 63.801384 52.96505 0.6740316 ]]
img 2 layer 0 tflite :
[[181.06418 138.47464 224.90767 197.36978 0.9352028 ]
[172.32611 137.85072 261.53143 188.90297 0.7145112 ]
[195.62141 136.05687 213.14252 194.8197 0.85339344]]
img 2 layer 0 kmodel :
[[180.8456 137.29376 230.73112 184.79173 0.87710017]
[170.5144 137.29376 263.13138 186.64618 0.6714251 ]
[195.9328 135.123 211.01398 184.79173 0.87710017]]
img 2 layer 1 tflite :
[]
img 2 layer 1 kmodel :
[]
img 3 layer 0 tflite :
[[200.14052 100.00058 52.676952 117.090515 0.75216126]]
img 3 layer 0 kmodel :
[[199.123 100.954346 56.663612 107.2014 0.7615662]]
img 3 layer 1 tflite :
[[197.36374 98.02013 22.453339 80.50932 0.63156927]
[244.84764 98.41268 23.07575 50.217842 0.63718444]]
img 3 layer 1 kmodel :
[[131.26227 100.90143 19.180124 66.949135 0.6537735]
[197.21548 98.61127 19.594486 83.40284 0.8488212]
[245.87292 98.81582 23. 50.983635 0.7126105]
[281.7856 98.23825 21.457172 91.33108 0.7650425]]
分析
现在看起来量化应该是没什么问题，差距不是很大。我得去找找c代码是不是出错了。。好累。。 