使用设备: cpu
POT库已安装 ✓

=== 加载数据 ===

=== 准备教师模型 ===
加载预训练教师模型: resnet18_cifar100_teacher.pth

教师模型准确率: 56.20%

=== 准备学生模型 ===
学生模型初始准确率: 0.99%

============================================================
精度恢复方法对比
============================================================

--- 直接微调（基线） ---
Epoch 1/10: Train Loss=3.7123, Train Acc=13.07%, Test Acc=19.15%, Best=19.15%
Epoch 2/10: Train Loss=2.8968, Train Acc=26.60%, Test Acc=32.44%, Best=32.44%
Epoch 3/10: Train Loss=2.3692, Train Acc=37.20%, Test Acc=41.88%, Best=41.88%
Epoch 4/10: Train Loss=2.0078, Train Acc=45.36%, Test Acc=47.93%, Best=47.93%
Epoch 5/10: Train Loss=1.7330, Train Acc=51.67%, Test Acc=49.58%, Best=49.58%
Epoch 6/10: Train Loss=1.5201, Train Acc=56.89%, Test Acc=54.56%, Best=54.56%
Epoch 7/10: Train Loss=1.3298, Train Acc=61.68%, Test Acc=58.29%, Best=58.29%
Epoch 8/10: Train Loss=1.1740, Train Acc=66.01%, Test Acc=62.42%, Best=62.42%
Epoch 9/10: Train Loss=1.0520, Train Acc=69.51%, Test Acc=63.79%, Best=63.79%
Epoch 10/10: Train Loss=0.9729, Train Acc=71.98%, Test Acc=65.44%, Best=65.44%
微调后准确率: 65.44%

--- 传统知识蒸馏 ---

=== 传统知识蒸馏 ===
  Batch 0: Loss=5.0183 (CE=4.7827, KD=5.2540)
  Batch 50: Loss=4.1108 (CE=3.9583, KD=4.2633)
  Batch 100: Loss=4.5539 (CE=3.8664, KD=5.2415)
  Batch 150: Loss=4.0187 (CE=3.7289, KD=4.3086)
  Batch 200: Loss=3.9860 (CE=3.8240, KD=4.1480)
  Batch 250: Loss=3.7358 (CE=3.5183, KD=3.9534)
  Batch 300: Loss=3.4208 (CE=3.2846, KD=3.5569)
  Batch 350: Loss=3.4670 (CE=3.6129, KD=3.3211)
  Batch 400: Loss=3.1738 (CE=3.3758, KD=2.9719)
  Batch 450: Loss=3.4111 (CE=3.0153, KD=3.8070)
  Batch 500: Loss=3.1516 (CE=3.1998, KD=3.1033)
  Batch 550: Loss=3.0850 (CE=3.2364, KD=2.9336)
  Batch 600: Loss=2.8657 (CE=3.0175, KD=2.7139)
  Batch 650: Loss=2.6300 (CE=3.3387, KD=1.9214)
  Batch 700: Loss=2.5572 (CE=2.9761, KD=2.1383)
  Batch 750: Loss=2.9700 (CE=2.8484, KD=3.0915)
Epoch 1/10: Train Loss=3.4305, Train Acc=16.81%, Test Acc=26.86%
  Batch 0: Loss=2.7601 (CE=2.8982, KD=2.6220)
  Batch 50: Loss=4.6639 (CE=4.5115, KD=4.8163)
  Batch 100: Loss=4.1238 (CE=4.2748, KD=3.9728)
  Batch 150: Loss=4.5374 (CE=4.3484, KD=4.7263)
  Batch 200: Loss=4.3018 (CE=4.5894, KD=4.0141)
  Batch 250: Loss=4.9478 (CE=4.3503, KD=5.5453)
  Batch 300: Loss=3.7958 (CE=4.0943, KD=3.4973)
  Batch 350: Loss=4.3999 (CE=4.0410, KD=4.7587)
  Batch 400: Loss=4.0095 (CE=4.1461, KD=3.8730)
  Batch 450: Loss=3.4567 (CE=3.7638, KD=3.1495)
  Batch 500: Loss=3.9148 (CE=3.8625, KD=3.9672)
  Batch 550: Loss=3.7732 (CE=3.9506, KD=3.5958)
  Batch 600: Loss=3.8006 (CE=3.5445, KD=4.0566)
  Batch 650: Loss=3.9685 (CE=3.8712, KD=4.0658)
  Batch 700: Loss=3.3830 (CE=3.2701, KD=3.4960)
  Batch 750: Loss=3.8383 (CE=3.7010, KD=3.9755)
Epoch 2/10: Train Loss=4.1919, Train Acc=8.00%, Test Acc=13.28%
  Batch 0: Loss=3.7167 (CE=3.6558, KD=3.7775)
  Batch 50: Loss=4.0605 (CE=3.5526, KD=4.5684)
  Batch 100: Loss=3.2704 (CE=3.5968, KD=2.9440)
  Batch 150: Loss=3.7179 (CE=3.5580, KD=3.8778)
  Batch 200: Loss=3.1127 (CE=3.1047, KD=3.1207)
  Batch 250: Loss=3.3052 (CE=3.4613, KD=3.1492)
  Batch 300: Loss=3.0807 (CE=3.3123, KD=2.8492)
  Batch 350: Loss=3.1264 (CE=3.2329, KD=3.0199)
  Batch 400: Loss=3.1604 (CE=3.2048, KD=3.1161)
  Batch 450: Loss=3.0289 (CE=2.9329, KD=3.1249)
  Batch 500: Loss=3.0930 (CE=3.3270, KD=2.8590)
  Batch 550: Loss=2.7821 (CE=3.0099, KD=2.5543)
  Batch 600: Loss=2.8526 (CE=3.2882, KD=2.4171)
  Batch 650: Loss=2.6083 (CE=3.0583, KD=2.1583)
  Batch 700: Loss=2.7597 (CE=2.6209, KD=2.8985)
  Batch 750: Loss=3.1402 (CE=3.0181, KD=3.2623)
Epoch 3/10: Train Loss=3.2299, Train Acc=19.65%, Test Acc=23.10%
  Batch 0: Loss=2.9955 (CE=3.1412, KD=2.8498)
  Batch 50: Loss=2.8270 (CE=3.2667, KD=2.3873)
  Batch 100: Loss=2.8996 (CE=2.9802, KD=2.8190)
  Batch 150: Loss=2.4598 (CE=2.9407, KD=1.9790)
  Batch 200: Loss=2.7226 (CE=3.2012, KD=2.2439)
  Batch 250: Loss=3.1146 (CE=3.1324, KD=3.0969)
  Batch 300: Loss=2.3553 (CE=2.5486, KD=2.1620)
  Batch 350: Loss=2.3246 (CE=2.7039, KD=1.9453)
  Batch 400: Loss=2.2565 (CE=2.4489, KD=2.0641)
  Batch 450: Loss=2.4075 (CE=3.0087, KD=1.8063)
  Batch 500: Loss=2.7450 (CE=2.8466, KD=2.6434)
  Batch 550: Loss=2.3138 (CE=2.7028, KD=1.9247)
  Batch 600: Loss=2.2444 (CE=2.7289, KD=1.7599)
  Batch 650: Loss=2.7216 (CE=3.1779, KD=2.2653)
  Batch 700: Loss=2.3435 (CE=2.7048, KD=1.9821)
  Batch 750: Loss=2.3709 (CE=2.8599, KD=1.8820)
Epoch 4/10: Train Loss=2.5785, Train Acc=29.05%, Test Acc=32.21%
  Batch 0: Loss=2.3170 (CE=2.5289, KD=2.1050)
  Batch 50: Loss=2.1349 (CE=2.3400, KD=1.9298)
  Batch 100: Loss=2.2076 (CE=2.1362, KD=2.2790)
  Batch 150: Loss=2.2187 (CE=2.6572, KD=1.7802)
  Batch 200: Loss=2.4240 (CE=2.5813, KD=2.2666)
  Batch 250: Loss=2.3798 (CE=2.4706, KD=2.2891)
  Batch 300: Loss=2.2341 (CE=2.7766, KD=1.6915)
  Batch 350: Loss=2.0899 (CE=2.3567, KD=1.8231)
  Batch 400: Loss=2.4057 (CE=2.4768, KD=2.3345)
  Batch 450: Loss=2.1244 (CE=2.5681, KD=1.6806)
  Batch 500: Loss=2.1201 (CE=2.4778, KD=1.7624)
  Batch 550: Loss=1.9415 (CE=2.3576, KD=1.5253)
  Batch 600: Loss=2.1330 (CE=2.4328, KD=1.8331)
  Batch 650: Loss=2.0516 (CE=2.2522, KD=1.8511)
  Batch 700: Loss=2.0108 (CE=2.2317, KD=1.7899)
  Batch 750: Loss=1.9839 (CE=2.4454, KD=1.5223)
Epoch 5/10: Train Loss=2.1207, Train Acc=36.43%, Test Acc=39.85%
  Batch 0: Loss=1.7537 (CE=2.1174, KD=1.3901)
  Batch 50: Loss=1.9359 (CE=2.5136, KD=1.3582)
  Batch 100: Loss=2.0187 (CE=2.6363, KD=1.4010)
  Batch 150: Loss=1.6899 (CE=1.9779, KD=1.4018)
  Batch 200: Loss=1.6386 (CE=2.0544, KD=1.2227)
  Batch 250: Loss=1.6194 (CE=2.0276, KD=1.2113)
  Batch 300: Loss=1.9380 (CE=2.5003, KD=1.3757)
  Batch 350: Loss=1.9153 (CE=2.2764, KD=1.5543)
  Batch 400: Loss=1.7259 (CE=2.0041, KD=1.4476)
  Batch 450: Loss=1.7193 (CE=2.1752, KD=1.2634)
  Batch 500: Loss=1.5616 (CE=1.8173, KD=1.3059)
  Batch 550: Loss=1.8885 (CE=2.3089, KD=1.4680)
  Batch 600: Loss=1.5839 (CE=2.0612, KD=1.1065)
  Batch 650: Loss=1.7325 (CE=2.2751, KD=1.1899)
  Batch 700: Loss=1.6740 (CE=2.1340, KD=1.2140)
  Batch 750: Loss=1.5653 (CE=1.8819, KD=1.2487)
Epoch 6/10: Train Loss=1.7898, Train Acc=42.84%, Test Acc=45.96%
  Batch 0: Loss=1.5308 (CE=1.9380, KD=1.1235)
  Batch 50: Loss=1.5747 (CE=1.6250, KD=1.5244)
  Batch 100: Loss=1.6782 (CE=2.1622, KD=1.1943)
  Batch 150: Loss=1.5403 (CE=1.8710, KD=1.2095)
  Batch 200: Loss=1.7323 (CE=2.0077, KD=1.4570)
  Batch 250: Loss=1.7112 (CE=2.1310, KD=1.2914)
  Batch 300: Loss=1.5546 (CE=1.9862, KD=1.1230)
  Batch 350: Loss=1.5751 (CE=1.9803, KD=1.1699)
  Batch 400: Loss=1.5400 (CE=1.8949, KD=1.1851)
  Batch 450: Loss=1.3885 (CE=1.4770, KD=1.3000)
  Batch 500: Loss=1.5917 (CE=1.9958, KD=1.1875)
  Batch 550: Loss=1.4714 (CE=2.0049, KD=0.9378)
  Batch 600: Loss=1.5905 (CE=2.2076, KD=0.9735)
  Batch 650: Loss=1.4415 (CE=1.7306, KD=1.1524)
  Batch 700: Loss=1.3361 (CE=1.6627, KD=1.0094)
  Batch 750: Loss=1.7157 (CE=1.9249, KD=1.5065)
Epoch 7/10: Train Loss=1.5536, Train Acc=47.90%, Test Acc=48.90%
  Batch 0: Loss=1.4256 (CE=1.8058, KD=1.0455)
  Batch 50: Loss=1.4889 (CE=1.8660, KD=1.1118)
  Batch 100: Loss=1.1927 (CE=1.3361, KD=1.0493)
  Batch 150: Loss=1.2697 (CE=1.5698, KD=0.9696)
  Batch 200: Loss=1.3173 (CE=1.6835, KD=0.9511)
  Batch 250: Loss=1.5889 (CE=1.9675, KD=1.2103)
  Batch 300: Loss=1.4321 (CE=1.9220, KD=0.9423)
  Batch 350: Loss=1.4086 (CE=1.9232, KD=0.8941)
  Batch 400: Loss=1.3692 (CE=1.7519, KD=0.9865)
  Batch 450: Loss=1.4084 (CE=1.7169, KD=1.0999)
  Batch 500: Loss=1.3768 (CE=1.6100, KD=1.1436)
  Batch 550: Loss=1.4191 (CE=1.7003, KD=1.1379)
  Batch 600: Loss=1.4603 (CE=1.9130, KD=1.0076)
  Batch 650: Loss=1.2884 (CE=1.5445, KD=1.0322)
  Batch 700: Loss=1.6138 (CE=2.1758, KD=1.0519)
  Batch 750: Loss=1.2673 (CE=1.5472, KD=0.9874)
Epoch 8/10: Train Loss=1.3940, Train Acc=51.53%, Test Acc=51.59%
  Batch 0: Loss=1.3569 (CE=1.5216, KD=1.1922)
  Batch 50: Loss=1.3343 (CE=1.7438, KD=0.9249)
  Batch 100: Loss=1.2788 (CE=1.6996, KD=0.8579)
  Batch 150: Loss=1.1513 (CE=1.4476, KD=0.8549)
  Batch 200: Loss=1.2513 (CE=1.6475, KD=0.8550)
  Batch 250: Loss=1.3734 (CE=1.8533, KD=0.8935)
  Batch 300: Loss=1.3630 (CE=1.6417, KD=1.0843)
  Batch 350: Loss=1.2402 (CE=1.5417, KD=0.9386)
  Batch 400: Loss=1.4652 (CE=1.9963, KD=0.9342)
  Batch 450: Loss=1.1742 (CE=1.4693, KD=0.8790)
  Batch 500: Loss=1.2672 (CE=1.5318, KD=1.0026)
  Batch 550: Loss=1.2336 (CE=1.5234, KD=0.9438)
  Batch 600: Loss=0.9807 (CE=1.0998, KD=0.8616)
  Batch 650: Loss=1.4775 (CE=1.9481, KD=1.0070)
  Batch 700: Loss=1.3328 (CE=1.8787, KD=0.7869)
  Batch 750: Loss=0.9853 (CE=1.2482, KD=0.7224)
Epoch 9/10: Train Loss=1.2615, Train Acc=55.10%, Test Acc=55.00%
  Batch 0: Loss=1.1358 (CE=1.4841, KD=0.7875)
  Batch 50: Loss=1.1863 (CE=1.4977, KD=0.8749)
  Batch 100: Loss=1.0318 (CE=1.4264, KD=0.6373)
  Batch 150: Loss=1.1579 (CE=1.4973, KD=0.8184)
  Batch 200: Loss=1.1088 (CE=1.2616, KD=0.9560)
  Batch 250: Loss=0.9713 (CE=1.2445, KD=0.6981)
  Batch 300: Loss=1.1360 (CE=1.4680, KD=0.8041)
  Batch 350: Loss=1.1538 (CE=1.3785, KD=0.9290)
  Batch 400: Loss=1.0809 (CE=1.3545, KD=0.8074)
  Batch 450: Loss=1.1954 (CE=1.5444, KD=0.8465)
  Batch 500: Loss=1.1173 (CE=1.4480, KD=0.7867)
  Batch 550: Loss=1.0788 (CE=1.3832, KD=0.7743)
  Batch 600: Loss=1.2807 (CE=1.7888, KD=0.7727)
  Batch 650: Loss=1.2754 (CE=1.6636, KD=0.8872)
  Batch 700: Loss=1.1726 (CE=1.5244, KD=0.8208)
  Batch 750: Loss=1.3476 (CE=1.6257, KD=1.0695)
Epoch 10/10: Train Loss=1.1513, Train Acc=58.18%, Test Acc=56.92%
KD恢复后准确率: 56.92%
比微调提升: +-8.52%

--- 最优传输恢复 ---
对齐层: ['layer4.0.shortcut.0', 'layer4.1.conv1', 'layer4.1.conv2']
  Batch 0: Loss=2.7353 (CE=4.7337, OT=0.736829)
  Batch 50: Loss=2.0350 (CE=3.9769, OT=0.093022)
  Batch 100: Loss=2.0211 (CE=3.9889, OT=0.053283)
  Batch 150: Loss=1.9209 (CE=3.7947, OT=0.047028)
  Batch 200: Loss=1.8906 (CE=3.7380, OT=0.043234)
  Batch 250: Loss=1.9734 (CE=3.9081, OT=0.038697)
  Batch 300: Loss=1.7340 (CE=3.4291, OT=0.039035)
  Batch 350: Loss=1.8690 (CE=3.7011, OT=0.036796)
  Batch 400: Loss=1.8418 (CE=3.6490, OT=0.034610)
  Batch 450: Loss=1.7498 (CE=3.4644, OT=0.035124)
  Batch 500: Loss=1.7963 (CE=3.5586, OT=0.034072)
  Batch 550: Loss=1.8321 (CE=3.6322, OT=0.031953)
  Batch 600: Loss=1.8022 (CE=3.5702, OT=0.034215)
  Batch 650: Loss=1.4908 (CE=2.9420, OT=0.039606)
  Batch 700: Loss=1.6684 (CE=3.3019, OT=0.034941)
  Batch 750: Loss=1.6391 (CE=3.2465, OT=0.031770)
Epoch 1/10: Loss=1.8500, CE=3.6458, OT=0.054204, Acc=14.15%
  Batch 0: Loss=1.5480 (CE=3.0622, OT=0.033789)
  Batch 50: Loss=1.5804 (CE=3.1203, OT=0.040456)
  Batch 100: Loss=1.6282 (CE=3.2197, OT=0.036800)
  Batch 150: Loss=1.3888 (CE=2.7429, OT=0.034648)
  Batch 200: Loss=1.3974 (CE=2.7587, OT=0.036165)
  Batch 250: Loss=1.6395 (CE=3.2438, OT=0.035213)
  Batch 300: Loss=1.6202 (CE=3.2048, OT=0.035593)
  Batch 350: Loss=1.5310 (CE=3.0242, OT=0.037911)
  Batch 400: Loss=1.7330 (CE=3.4279, OT=0.038218)
  Batch 450: Loss=1.4496 (CE=2.8651, OT=0.034163)
  Batch 500: Loss=1.4020 (CE=2.7632, OT=0.040781)
  Batch 550: Loss=1.2774 (CE=2.5168, OT=0.037891)
  Batch 600: Loss=1.3227 (CE=2.6127, OT=0.032706)
  Batch 650: Loss=1.4501 (CE=2.8664, OT=0.033858)
  Batch 700: Loss=1.4336 (CE=2.8305, OT=0.036772)
  Batch 750: Loss=1.2499 (CE=2.4655, OT=0.034268)
Epoch 2/10: Loss=1.4370, CE=2.8384, OT=0.035673, Acc=28.10%
  Batch 0: Loss=1.3024 (CE=2.5687, OT=0.036195)
  Batch 50: Loss=1.3513 (CE=2.6660, OT=0.036540)
  Batch 100: Loss=1.2598 (CE=2.4898, OT=0.029750)
  Batch 150: Loss=1.2257 (CE=2.4135, OT=0.037896)
  Batch 200: Loss=1.0201 (CE=2.0070, OT=0.033186)
  Batch 250: Loss=1.2474 (CE=2.4577, OT=0.037054)
  Batch 300: Loss=1.3668 (CE=2.7015, OT=0.032128)
  Batch 350: Loss=0.9837 (CE=1.9342, OT=0.033154)
  Batch 400: Loss=1.3480 (CE=2.6589, OT=0.037070)
  Batch 450: Loss=1.2882 (CE=2.5365, OT=0.039931)
  Batch 500: Loss=1.4716 (CE=2.9083, OT=0.034808)
  Batch 550: Loss=1.3644 (CE=2.6885, OT=0.040405)
  Batch 600: Loss=1.1138 (CE=2.1910, OT=0.036552)
  Batch 650: Loss=1.1229 (CE=2.2102, OT=0.035685)
  Batch 700: Loss=1.1683 (CE=2.3014, OT=0.035192)
  Batch 750: Loss=1.0619 (CE=2.0841, OT=0.039600)
Epoch 3/10: Loss=1.1929, CE=2.3497, OT=0.036135, Acc=37.89%
  Batch 0: Loss=1.0050 (CE=1.9753, OT=0.034653)
  Batch 50: Loss=1.1721 (CE=2.3069, OT=0.037320)
  Batch 100: Loss=1.0503 (CE=2.0611, OT=0.039560)
  Batch 150: Loss=0.9930 (CE=1.9439, OT=0.042020)
  Batch 200: Loss=1.1078 (CE=2.1811, OT=0.034486)
  Batch 250: Loss=1.1526 (CE=2.2698, OT=0.035464)
  Batch 300: Loss=1.2903 (CE=2.5372, OT=0.043429)
  Batch 350: Loss=1.0469 (CE=2.0569, OT=0.036913)
  Batch 400: Loss=1.0696 (CE=2.0943, OT=0.044832)
  Batch 450: Loss=0.9699 (CE=1.9070, OT=0.032821)
  Batch 500: Loss=0.9537 (CE=1.8748, OT=0.032533)
  Batch 550: Loss=1.0302 (CE=2.0238, OT=0.036575)
  Batch 600: Loss=1.1579 (CE=2.2795, OT=0.036155)
  Batch 650: Loss=1.0601 (CE=2.0832, OT=0.036988)
  Batch 700: Loss=0.9888 (CE=1.9409, OT=0.036608)
  Batch 750: Loss=1.1161 (CE=2.1958, OT=0.036390)
Epoch 4/10: Loss=1.0246, CE=2.0133, OT=0.035963, Acc=45.28%
  Batch 0: Loss=1.0477 (CE=2.0624, OT=0.032991)
  Batch 50: Loss=0.9698 (CE=1.8994, OT=0.040249)
  Batch 100: Loss=0.9587 (CE=1.8826, OT=0.034794)
  Batch 150: Loss=0.7798 (CE=1.5233, OT=0.036255)
  Batch 200: Loss=0.7009 (CE=1.3674, OT=0.034442)
  Batch 250: Loss=0.9235 (CE=1.8070, OT=0.040036)
  Batch 300: Loss=0.7281 (CE=1.4176, OT=0.038674)
  Batch 350: Loss=0.7849 (CE=1.5355, OT=0.034355)
  Batch 400: Loss=0.9909 (CE=1.9471, OT=0.034786)
  Batch 450: Loss=0.7666 (CE=1.4998, OT=0.033376)
  Batch 500: Loss=1.0200 (CE=2.0042, OT=0.035751)
  Batch 550: Loss=0.9607 (CE=1.8862, OT=0.035215)
  Batch 600: Loss=0.8311 (CE=1.6297, OT=0.032482)
  Batch 650: Loss=0.7918 (CE=1.5486, OT=0.034972)
  Batch 700: Loss=0.8764 (CE=1.7208, OT=0.031988)
  Batch 750: Loss=0.8929 (CE=1.7492, OT=0.036579)
Epoch 5/10: Loss=0.9063, CE=1.7767, OT=0.035883, Acc=51.14%
  Batch 0: Loss=0.7309 (CE=1.4277, OT=0.034072)
  Batch 50: Loss=0.8797 (CE=1.7234, OT=0.035906)
  Batch 100: Loss=0.7745 (CE=1.5117, OT=0.037293)
  Batch 150: Loss=0.7407 (CE=1.4450, OT=0.036373)
  Batch 200: Loss=0.8269 (CE=1.6140, OT=0.039692)
  Batch 250: Loss=0.6584 (CE=1.2809, OT=0.035827)
  Batch 300: Loss=0.9341 (CE=1.8270, OT=0.041239)
  Batch 350: Loss=1.0843 (CE=2.1330, OT=0.035499)
  Batch 400: Loss=0.7723 (CE=1.5110, OT=0.033646)
  Batch 450: Loss=0.9275 (CE=1.8205, OT=0.034491)
  Batch 500: Loss=0.7827 (CE=1.5266, OT=0.038874)
  Batch 550: Loss=0.7928 (CE=1.5511, OT=0.034441)
  Batch 600: Loss=0.8288 (CE=1.6236, OT=0.033962)
  Batch 650: Loss=0.7225 (CE=1.4028, OT=0.042245)
  Batch 700: Loss=0.7003 (CE=1.3652, OT=0.035268)
  Batch 750: Loss=0.7099 (CE=1.3828, OT=0.037057)
Epoch 6/10: Loss=0.8200, CE=1.6042, OT=0.035761, Acc=54.99%
  Batch 0: Loss=0.7130 (CE=1.3905, OT=0.035456)
  Batch 50: Loss=0.6734 (CE=1.3144, OT=0.032384)
  Batch 100: Loss=0.6588 (CE=1.2815, OT=0.036165)
  Batch 150: Loss=0.6273 (CE=1.2182, OT=0.036357)
  Batch 200: Loss=0.6612 (CE=1.2894, OT=0.032999)
  Batch 250: Loss=0.8242 (CE=1.6152, OT=0.033197)
  Batch 300: Loss=0.8001 (CE=1.5672, OT=0.033020)
  Batch 350: Loss=0.6576 (CE=1.2794, OT=0.035741)
  Batch 400: Loss=0.9104 (CE=1.7828, OT=0.037918)
  Batch 450: Loss=0.8125 (CE=1.5901, OT=0.035039)
  Batch 500: Loss=0.9628 (CE=1.8939, OT=0.031756)
  Batch 550: Loss=0.6430 (CE=1.2487, OT=0.037330)
  Batch 600: Loss=0.6607 (CE=1.2870, OT=0.034509)
  Batch 650: Loss=0.7474 (CE=1.4579, OT=0.036902)
  Batch 700: Loss=0.7652 (CE=1.4975, OT=0.032926)
  Batch 750: Loss=0.8076 (CE=1.5805, OT=0.034781)
Epoch 7/10: Loss=0.7453, CE=1.4554, OT=0.035074, Acc=58.71%
  Batch 0: Loss=0.6226 (CE=1.2133, OT=0.031889)
  Batch 50: Loss=0.6895 (CE=1.3431, OT=0.035893)
  Batch 100: Loss=0.7626 (CE=1.4900, OT=0.035285)
  Batch 150: Loss=0.7193 (CE=1.4062, OT=0.032421)
  Batch 200: Loss=0.7088 (CE=1.3832, OT=0.034417)
  Batch 250: Loss=0.7488 (CE=1.4631, OT=0.034442)
  Batch 300: Loss=0.7054 (CE=1.3781, OT=0.032751)
  Batch 350: Loss=0.6430 (CE=1.2554, OT=0.030628)
  Batch 400: Loss=0.5948 (CE=1.1552, OT=0.034435)
  Batch 450: Loss=0.7515 (CE=1.4686, OT=0.034396)
  Batch 500: Loss=0.7125 (CE=1.3920, OT=0.032919)
  Batch 550: Loss=0.7533 (CE=1.4725, OT=0.034109)
  Batch 600: Loss=0.6719 (CE=1.3109, OT=0.032902)
  Batch 650: Loss=0.6606 (CE=1.2885, OT=0.032733)
  Batch 700: Loss=0.7737 (CE=1.5119, OT=0.035439)
  Batch 750: Loss=0.8512 (CE=1.6622, OT=0.040093)
Epoch 8/10: Loss=0.6865, CE=1.3380, OT=0.034913, Acc=61.39%
  Batch 0: Loss=0.6973 (CE=1.3589, OT=0.035680)
  Batch 50: Loss=0.4637 (CE=0.8904, OT=0.037031)
  Batch 100: Loss=0.6247 (CE=1.2133, OT=0.036122)
  Batch 150: Loss=0.5719 (CE=1.1057, OT=0.038030)
  Batch 200: Loss=0.6783 (CE=1.3216, OT=0.035003)
  Batch 250: Loss=0.5742 (CE=1.1169, OT=0.031632)
  Batch 300: Loss=0.6312 (CE=1.2285, OT=0.033855)
  Batch 350: Loss=0.6128 (CE=1.1929, OT=0.032700)
  Batch 400: Loss=0.5536 (CE=1.0734, OT=0.033753)
  Batch 450: Loss=0.6686 (CE=1.3040, OT=0.033301)
  Batch 500: Loss=0.6121 (CE=1.1904, OT=0.033916)
  Batch 550: Loss=0.5776 (CE=1.1199, OT=0.035280)
  Batch 600: Loss=0.7529 (CE=1.4712, OT=0.034529)
  Batch 650: Loss=0.5801 (CE=1.1232, OT=0.037013)
  Batch 700: Loss=0.5533 (CE=1.0709, OT=0.035663)
  Batch 750: Loss=0.6059 (CE=1.1789, OT=0.032922)
Epoch 9/10: Loss=0.6404, CE=1.2459, OT=0.034830, Acc=63.82%
  Batch 0: Loss=0.5967 (CE=1.1561, OT=0.037287)
  Batch 50: Loss=0.5469 (CE=1.0541, OT=0.039689)
  Batch 100: Loss=0.5520 (CE=1.0674, OT=0.036591)
  Batch 150: Loss=0.4398 (CE=0.8441, OT=0.035466)
  Batch 200: Loss=0.4584 (CE=0.8817, OT=0.035166)
  Batch 250: Loss=0.6723 (CE=1.3107, OT=0.033904)
  Batch 300: Loss=0.4842 (CE=0.9365, OT=0.031872)
  Batch 350: Loss=0.5154 (CE=1.0000, OT=0.030893)
  Batch 400: Loss=0.7751 (CE=1.5155, OT=0.034691)
  Batch 450: Loss=0.6380 (CE=1.2422, OT=0.033806)
  Batch 500: Loss=0.6183 (CE=1.2019, OT=0.034725)
  Batch 550: Loss=0.6927 (CE=1.3506, OT=0.034690)
  Batch 600: Loss=0.6425 (CE=1.2543, OT=0.030624)
  Batch 650: Loss=0.6149 (CE=1.1965, OT=0.033339)
  Batch 700: Loss=0.5976 (CE=1.1620, OT=0.033256)
  Batch 750: Loss=0.5158 (CE=0.9988, OT=0.032844)
Epoch 10/10: Loss=0.5961, CE=1.1580, OT=0.034302, Acc=66.25%
OT恢复后准确率: 60.01%
比微调提升: +-5.43%
比KD提升: +3.09%


============================================================
结果汇总
============================================================
教师模型准确率: 56.20%
学生初始准确率: 0.99%

方法                   准确率          提升(vs微调)       
-----------------------------------------------
直接微调                 65.44        -              
传统KD                 56.92        +-8.52         
最优传输恢复               60.01        +-5.43         

OT vs KD 提升: +3.09%

结论：如果OT恢复准确率高于传统KD，
     则验证了几何保真度在精度恢复中的重要性。

注意：压缩率越高，OT的优势通常越明显，
     因为深度压缩会严重破坏特征几何结构。

结果已保存到 ot_recovery_results.npy