我已经训练了一个具有四个全连接层和 44 个神经元输出的 Resnet 神经网络。我想在 Coral.ai TPU 上运行它，并使用8bits 的训练后量化进行量化。
在这种情况下，每个关节的位置由对应于神经元 i 和 i+1 的两个坐标（x 和 y）组成。在这种情况下有 22 个关节。使用 TFLite 模型，预测坐标（红色）与真实坐标（蓝色）非常相似。
但是当我量化时，这些结果太糟糕了
]
这些是每个坐标的输出（预测的，基本事实）
X0 = (75.85,64.24) Diff = 11.61
X1 = (55.79,43.36) Diff = 12.43
X2 = (80.86,68.24) Diff = 12.62
X3 = (52.66,39.51) Diff = 13.15
X4 = (82.12,69.49) Diff = 12.63
X5 = (50.15,37.42) Diff = 12.72
X6 = (85.25,70.42) Diff = 14.84
X7 = (48.27,35.87) Diff = 12.40
X8 = (83.37,68.90) Diff = 14.47
X9 = (48.27,37.08) Diff = 11.18
X10 = (82.74,67.45) Diff = 15.30
X11 = (50.78,38.98) Diff = 11.79
X12 = (72.09,59.65) Diff = 12.44
X13 = (60.18,48.35) Diff = 11.83
X14 = (73.34,60.95) Diff = 12.40
X15 = (60.81,49.41) Diff = 11.40
X16 = (73.34,60.60) Diff = 12.74
X17 = (62.06,51.45) Diff = 10.61
X18 = (73.34,59.57) Diff = 13.78
X19 = (63.31,52.89) Diff = 10.42
X20 = (74.60,62.95) Diff = 11.65
X21 = (59.55,48.71) Diff = 10.84
Y0 = (47.01,53.75) Diff = -6.74
Y1 = (49.52,52.70) Diff = -3.18
Y2 = (59.55,67.89) Diff = -8.34
Y3 = (59.55,65.91) Diff = -6.36
Y4 = (76.48,81.06) Diff = -4.58
Y5 = (71.46,79.27) Diff = -7.81
Y6 = (80.86,85.55) Diff = -4.68
Y7 = (74.60,83.44) Diff = -8.84
Y8 = (80.86,85.61) Diff = -4.74
Y9 = (75.22,84.10) Diff = -8.88
Y10 = (78.98,83.94) Diff = -4.96
Y11 = (73.97,82.61) Diff = -8.64
Y12 = (73.97,81.12) Diff = -7.15
Y13 = (75.22,81.12) Diff = -5.90
Y14 = (93.40,100.35) Diff = -6.95
Y15 = (93.40,101.04) Diff = -7.64
Y16 = (110.33,117.44) Diff = -7.12
Y17 = (109.70,117.75) Diff = -8.05
Y18 = (112.21,119.59) Diff = -7.39
Y19 = (110.95,119.97) Diff = -9.02
Y20 = (117.85,126.34) Diff = -8.49
Y21 = (116.60,126.85) Diff = -10.26
你知道发生了什么吗？图像的宽度为 160，高度为 120。
谢谢！！