search

LeeJunHyun / Image_Segmentation

Pytorch implementation of U-Net, R2U-Net, Attention U-Net, and Attention R2U-Net.
2.66k stars 594 forks source link

the difference of train_result and valid_result! #45

Closed CN-zdy closed 4 years ago

CN-zdy commented 4 years ago

@LeeJunHyun Hi: thank you for the code, i have trained the unet with the data from ISIC2018 ,but the train_result and valid_result is different, could you give me some advise? Why is the evaluation high from the beginning? Epoch [1/100], Loss: 616.1876, [Training] Acc: 0.9034, SE: 0.7722, SP: 0.9790, PC: 0.8181, F1: 0.7387, JS: 0.6296, DC: 0.7387 [Validation] Acc: 0.8565, SE: 0.4684, SP: 0.9901, PC: 0.6672, F1: 0.4806, JS: 0.4005, DC: 0.4806 Best U_Net model score : 0.8811 Epoch [2/100], Loss: 398.6359, [Training] Acc: 0.9281, SE: 0.8297, SP: 0.9785, PC: 0.8587, F1: 0.8081, JS: 0.7108, DC: 0.8081 [Validation] Acc: 0.8758, SE: 0.5535, SP: 0.9764, PC: 0.7565, F1: 0.5566, JS: 0.4618, DC: 0.5566 Best U_Net model score : 1.0184 Epoch [3/100], Loss: 340.5711, [Training] Acc: 0.9354, SE: 0.8442, SP: 0.9802, PC: 0.8693, F1: 0.8246, JS: 0.7319, DC: 0.8246 [Validation] Acc: 0.8137, SE: 0.4210, SP: 0.9514, PC: 0.5229, F1: 0.3643, JS: 0.2888, DC: 0.3643 Epoch [4/100], Loss: 320.5319, [Training] Acc: 0.9374, SE: 0.8481, SP: 0.9785, PC: 0.8721, F1: 0.8273, JS: 0.7371, DC: 0.8273 [Validation] Acc: 0.8686, SE: 0.4233, SP: 0.9935, PC: 0.6778, F1: 0.4639, JS: 0.3859, DC: 0.4639 Epoch [5/100], Loss: 291.2066, [Training] Acc: 0.9426, SE: 0.8587, SP: 0.9787, PC: 0.8788, F1: 0.8391, JS: 0.7521, DC: 0.8391 [Validation] Acc: 0.8385, SE: 0.3736, SP: 0.9840, PC: 0.5603, F1: 0.3846, JS: 0.3200, DC: 0.3846 Epoch [6/100], Loss: 268.9106, [Training] Acc: 0.9467, SE: 0.8685, SP: 0.9784, PC: 0.8816, F1: 0.8483, JS: 0.7636, DC: 0.8483 [Validation] Acc: 0.8908, SE: 0.6225, SP: 0.9703, PC: 0.7103, F1: 0.6074, JS: 0.5185, DC: 0.6074 Best U_Net model score : 1.1259 Epoch [7/100], Loss: 268.7711, [Training] Acc: 0.9445, SE: 0.8677, SP: 0.9770, PC: 0.8796, F1: 0.8469, JS: 0.7619, DC: 0.8469 [Validation] Acc: 0.8663, SE: 0.4388, SP: 0.9958, PC: 0.7275, F1: 0.4932, JS: 0.4153, DC: 0.4932 Epoch [8/100], Loss: 254.9294, [Training] Acc: 0.9472, SE: 0.8714, SP: 0.9777, PC: 0.8798, F1: 0.8495, JS: 0.7647, DC: 0.8495 [Validation] Acc: 0.8714, SE: 0.6046, SP: 0.9706, PC: 0.7240, F1: 0.5918, JS: 0.4992, DC: 0.5918 Epoch [9/100], Loss: 239.5681, [Training] Acc: 0.9492, SE: 0.8778, SP: 0.9768, PC: 0.8858, F1: 0.8571, JS: 0.7746, DC: 0.8571 [Validation] Acc: 0.8870, SE: 0.6585, SP: 0.9685, PC: 0.7207, F1: 0.6181, JS: 0.5193, DC: 0.6181 Best U_Net model score : 1.1375 Epoch [10/100], Loss: 235.1525, [Training] Acc: 0.9498, SE: 0.8782, SP: 0.9764, PC: 0.8861, F1: 0.8595, JS: 0.7766, DC: 0.8595 [Validation] Acc: 0.8666, SE: 0.4735, SP: 0.9908, PC: 0.7191, F1: 0.5103, JS: 0.4270, DC: 0.5103 Epoch [11/100], Loss: 234.3776, [Training] Acc: 0.9506, SE: 0.8804, SP: 0.9757, PC: 0.8870, F1: 0.8615, JS: 0.7794, DC: 0.8615 [Validation] Acc: 0.8819, SE: 0.5309, SP: 0.9843, PC: 0.6968, F1: 0.5439, JS: 0.4530, DC: 0.5439 Epoch [12/100], Loss: 227.3490, [Training] Acc: 0.9525, SE: 0.8807, SP: 0.9762, PC: 0.8875, F1: 0.8605, JS: 0.7787, DC: 0.8605 [Validation] Acc: 0.8778, SE: 0.5388, SP: 0.9807, PC: 0.7516, F1: 0.5626, JS: 0.4659, DC: 0.5626 Epoch [13/100], Loss: 215.5042, [Training] Acc: 0.9541, SE: 0.8852, SP: 0.9766, PC: 0.8886, F1: 0.8659, JS: 0.7854, DC: 0.8659 [Validation] Acc: 0.8602, SE: 0.3876, SP: 0.9956, PC: 0.6710, F1: 0.4320, JS: 0.3634, DC: 0.4320 Epoch [14/100], Loss: 210.8664, [Training] Acc: 0.9544, SE: 0.8862, SP: 0.9762, PC: 0.8874, F1: 0.8650, JS: 0.7848, DC: 0.8650 [Validation] Acc: 0.8937, SE: 0.6567, SP: 0.9757, PC: 0.7512, F1: 0.6259, JS: 0.5324, DC: 0.6259 Best U_Net model score : 1.1583 Epoch [15/100], Loss: 210.2750, [Training] Acc: 0.9549, SE: 0.8889, SP: 0.9770, PC: 0.8924, F1: 0.8697, JS: 0.7906, DC: 0.8697 [Validation] Acc: 0.8631, SE: 0.5639, SP: 0.9737, PC: 0.7506, F1: 0.5617, JS: 0.4642, DC: 0.5617 Epoch [16/100], Loss: 202.6749, [Training] Acc: 0.9555, SE: 0.8905, SP: 0.9765, PC: 0.8926, F1: 0.8706, JS: 0.7909, DC: 0.8706 [Validation] Acc: 0.8478, SE: 0.3574, SP: 0.9961, PC: 0.5989, F1: 0.3882, JS: 0.3243, DC: 0.3882 Epoch [17/100], Loss: 191.1452, [Training] Acc: 0.9585, SE: 0.8938, SP: 0.9770, PC: 0.8916, F1: 0.8729, JS: 0.7938, DC: 0.8729 [Validation] Acc: 0.8814, SE: 0.7139, SP: 0.9476, PC: 0.7399, F1: 0.6435, JS: 0.5461, DC: 0.6435 Best U_Net model score : 1.1896 Epoch [18/100], Loss: 203.5803, [Training] Acc: 0.9555, SE: 0.8900, SP: 0.9771, PC: 0.8930, F1: 0.8703, JS: 0.7914, DC: 0.8703 [Validation] Acc: 0.8602, SE: 0.3996, SP: 0.9807, PC: 0.5780, F1: 0.4198, JS: 0.3436, DC: 0.4198 Epoch [19/100], Loss: 192.3391, [Training] Acc: 0.9580, SE: 0.8947, SP: 0.9761, PC: 0.8931, F1: 0.8754, JS: 0.7971, DC: 0.8754 [Validation] Acc: 0.8673, SE: 0.4138, SP: 0.9951, PC: 0.6683, F1: 0.4570, JS: 0.3880, DC: 0.4570 Epoch [20/100], Loss: 191.9962, [Training] Acc: 0.9582, SE: 0.8967, SP: 0.9770, PC: 0.8919, F1: 0.8743, JS: 0.7967, DC: 0.8743 [Validation] Acc: 0.8756, SE: 0.4957, SP: 0.9798, PC: 0.6777, F1: 0.5077, JS: 0.4248, DC: 0.5077 Epoch [21/100], Loss: 190.4904, [Training] Acc: 0.9585, SE: 0.8967, SP: 0.9761, PC: 0.8925, F1: 0.8759, JS: 0.7978, DC: 0.8759 [Validation] Acc: 0.8679, SE: 0.5535, SP: 0.9696, PC: 0.6921, F1: 0.5424, JS: 0.4508, DC: 0.5424 Epoch [22/100], Loss: 185.0105, [Training] Acc: 0.9600, SE: 0.9003, SP: 0.9770, PC: 0.8951, F1: 0.8798, JS: 0.8030, DC: 0.8798 [Validation] Acc: 0.8789, SE: 0.6378, SP: 0.9576, PC: 0.6817, F1: 0.5850, JS: 0.4878, DC: 0.5850 Epoch [23/100], Loss: 182.5417, [Training] Acc: 0.9596, SE: 0.8995, SP: 0.9767, PC: 0.8945, F1: 0.8790, JS: 0.8032, DC: 0.8790 [Validation] Acc: 0.8602, SE: 0.4950, SP: 0.9883, PC: 0.6783, F1: 0.4958, JS: 0.4141, DC: 0.4958 Epoch [24/100], Loss: 183.4632, [Training] Acc: 0.9601, SE: 0.9005, SP: 0.9768, PC: 0.8950, F1: 0.8798, JS: 0.8035, DC: 0.8798 [Validation] Acc: 0.8804, SE: 0.6130, SP: 0.9796, PC: 0.7332, F1: 0.5899, JS: 0.4913, DC: 0.5899 Epoch [25/100], Loss: 172.4488, [Training] Acc: 0.9622, SE: 0.9043, SP: 0.9758, PC: 0.8969, F1: 0.8846, JS: 0.8093, DC: 0.8846 [Validation] Acc: 0.8547, SE: 0.4404, SP: 0.9950, PC: 0.6456, F1: 0.4643, JS: 0.3986, DC: 0.4643 Epoch [26/100], Loss: 175.5711, [Training] Acc: 0.9611, SE: 0.9026, SP: 0.9763, PC: 0.8965, F1: 0.8827, JS: 0.8076, DC: 0.8827 [Validation] Acc: 0.8671, SE: 0.4905, SP: 0.9912, PC: 0.7205, F1: 0.5126, JS: 0.4334, DC: 0.5126 Epoch [27/100], Loss: 165.1405, [Training] Acc: 0.9633, SE: 0.9072, SP: 0.9766, PC: 0.8961, F1: 0.8853, JS: 0.8098, DC: 0.8853 [Validation] Acc: 0.8669, SE: 0.4838, SP: 0.9912, PC: 0.7026, F1: 0.5064, JS: 0.4288, DC: 0.5064 Epoch [28/100], Loss: 172.5805, [Training] Acc: 0.9619, SE: 0.9027, SP: 0.9771, PC: 0.8966, F1: 0.8828, JS: 0.8072, DC: 0.8828 Decay learning rate to lr: 0.00011421844870607387. [Validation] Acc: 0.8818, SE: 0.7281, SP: 0.9499, PC: 0.6922, F1: 0.6223, JS: 0.5203, DC: 0.6223 Epoch [29/100], Loss: 165.0072, [Training] Acc: 0.9633, SE: 0.9076, SP: 0.9763, PC: 0.8961, F1: 0.8857, JS: 0.8111, DC: 0.8857 Decay learning rate to lr: 0.00011263208136293396. [Validation] Acc: 0.8596, SE: 0.4434, SP: 0.9904, PC: 0.6226, F1: 0.4550, JS: 0.3834, DC: 0.4550 Epoch [30/100], Loss: 175.4593, [Training] Acc: 0.9617, SE: 0.9058, SP: 0.9756, PC: 0.8969, F1: 0.8852, JS: 0.8096, DC: 0.8852 Decay learning rate to lr: 0.00011104571401979405. [Validation] Acc: 0.8681, SE: 0.4951, SP: 0.9931, PC: 0.7088, F1: 0.5199, JS: 0.4469, DC: 0.5199 Epoch [31/100], Loss: 158.7435, [Training] Acc: 0.9644, SE: 0.9075, SP: 0.9757, PC: 0.8981, F1: 0.8878, JS: 0.8131, DC: 0.8878 Decay learning rate to lr: 0.00010945934667665414. [Validation] Acc: 0.8584, SE: 0.4345, SP: 0.9936, PC: 0.6534, F1: 0.4599, JS: 0.3897, DC: 0.4599 Epoch [32/100], Loss: 172.6409, [Training] Acc: 0.9624, SE: 0.9062, SP: 0.9756, PC: 0.8972, F1: 0.8852, JS: 0.8106, DC: 0.8852 Decay learning rate to lr: 0.00010787297933351423. [Validation] Acc: 0.8719, SE: 0.5285, SP: 0.9778, PC: 0.6969, F1: 0.5276, JS: 0.4429, DC: 0.5276 Epoch [33/100], Loss: 159.3884, [Training] Acc: 0.9639, SE: 0.9094, SP: 0.9750, PC: 0.8975, F1: 0.8878, JS: 0.8135, DC: 0.8878 Decay learning rate to lr: 0.00010628661199037432. [Validation] Acc: 0.8675, SE: 0.5368, SP: 0.9867, PC: 0.6657, F1: 0.5260, JS: 0.4449, DC: 0.5260 Epoch [34/100], Loss: 163.9595, [Training] Acc: 0.9643, SE: 0.9100, SP: 0.9760, PC: 0.8991, F1: 0.8896, JS: 0.8166, DC: 0.8896 Decay learning rate to lr: 0.00010470024464723441. [Validation] Acc: 0.8849, SE: 0.7706, SP: 0.9392, PC: 0.7240, F1: 0.6617, JS: 0.5625, DC: 0.6617 Best U_Net model score : 1.2242 Epoch [35/100], Loss: 155.4412, [Training] Acc: 0.9651, SE: 0.9101, SP: 0.9765, PC: 0.8995, F1: 0.8902, JS: 0.8169, DC: 0.8902 Decay learning rate to lr: 0.0001031138773040945. [Validation] Acc: 0.8626, SE: 0.4245, SP: 0.9861, PC: 0.6021, F1: 0.4391, JS: 0.3667, DC: 0.4391 Epoch [36/100], Loss: 166.6415, [Training] Acc: 0.9631, SE: 0.9060, SP: 0.9757, PC: 0.8972, F1: 0.8860, JS: 0.8118, DC: 0.8860 Decay learning rate to lr: 0.00010152750996095459. [Validation] Acc: 0.8849, SE: 0.6580, SP: 0.9658, PC: 0.7212, F1: 0.6112, JS: 0.5165, DC: 0.6112 Epoch [37/100], Loss: 156.8936, [Training] Acc: 0.9645, SE: 0.9084, SP: 0.9761, PC: 0.8982, F1: 0.8885, JS: 0.8148, DC: 0.8885 Decay learning rate to lr: 9.994114261781468e-05. [Validation] Acc: 0.8381, SE: 0.3261, SP: 0.9949, PC: 0.5530, F1: 0.3419, JS: 0.2840, DC: 0.3419 Epoch [38/100], Loss: 153.6247, [Training] Acc: 0.9656, SE: 0.9129, SP: 0.9736, PC: 0.9010, F1: 0.8931, JS: 0.8206, DC: 0.8931 Decay learning rate to lr: 9.835477527467477e-05. [Validation] Acc: 0.8735, SE: 0.5074, SP: 0.9846, PC: 0.6922, F1: 0.5077, JS: 0.4240, DC: 0.5077 Epoch [39/100], Loss: 155.5283, [Training] Acc: 0.9662, SE: 0.9132, SP: 0.9773, PC: 0.8989, F1: 0.8916, JS: 0.8187, DC: 0.8916 Decay learning rate to lr: 9.676840793153486e-05. [Validation] Acc: 0.8601, SE: 0.4192, SP: 0.9891, PC: 0.6524, F1: 0.4478, JS: 0.3752, DC: 0.4478 Epoch [40/100], Loss: 151.9283, [Training] Acc: 0.9651, SE: 0.9116, SP: 0.9748, PC: 0.8999, F1: 0.8910, JS: 0.8183, DC: 0.8910 Decay learning rate to lr: 9.518204058839495e-05. [Validation] Acc: 0.8474, SE: 0.3625, SP: 0.9926, PC: 0.6137, F1: 0.3866, JS: 0.3206, DC: 0.3866 Epoch [41/100], Loss: 149.1667, [Training] Acc: 0.9671, SE: 0.9157, SP: 0.9766, PC: 0.9018, F1: 0.8954, JS: 0.8234, DC: 0.8954 Decay learning rate to lr: 9.359567324525504e-05. [Validation] Acc: 0.8795, SE: 0.6205, SP: 0.9780, PC: 0.7092, F1: 0.5919, JS: 0.4968, DC: 0.5919 Epoch [42/100], Loss: 144.5024, [Training] Acc: 0.9674, SE: 0.9181, SP: 0.9759, PC: 0.9023, F1: 0.8974, JS: 0.8260, DC: 0.8974 Decay learning rate to lr: 9.200930590211513e-05. [Validation] Acc: 0.8732, SE: 0.5549, SP: 0.9854, PC: 0.7127, F1: 0.5510, JS: 0.4626, DC: 0.5510 Epoch [43/100], Loss: 151.6849, [Training] Acc: 0.9661, SE: 0.9147, SP: 0.9752, PC: 0.9010, F1: 0.8946, JS: 0.8226, DC: 0.8946 Decay learning rate to lr: 9.042293855897522e-05. [Validation] Acc: 0.8626, SE: 0.4704, SP: 0.9934, PC: 0.7154, F1: 0.4976, JS: 0.4180, DC: 0.4976 Epoch [44/100], Loss: 150.5371, [Training] Acc: 0.9662, SE: 0.9132, SP: 0.9762, PC: 0.9005, F1: 0.8925, JS: 0.8209, DC: 0.8925 Decay learning rate to lr: 8.883657121583531e-05. [Validation] Acc: 0.8754, SE: 0.5314, SP: 0.9848, PC: 0.7298, F1: 0.5441, JS: 0.4539, DC: 0.5441 Epoch [45/100], Loss: 148.0798, [Training] Acc: 0.9671, SE: 0.9177, SP: 0.9763, PC: 0.9024, F1: 0.8970, JS: 0.8265, DC: 0.8970 Decay learning rate to lr: 8.72502038726954e-05. [Validation] Acc: 0.8690, SE: 0.4646, SP: 0.9943, PC: 0.7114, F1: 0.4996, JS: 0.4236, DC: 0.4996 Epoch [46/100], Loss: 134.4733, [Training] Acc: 0.9695, SE: 0.9201, SP: 0.9762, PC: 0.9029, F1: 0.8991, JS: 0.8285, DC: 0.8991 Decay learning rate to lr: 8.566383652955549e-05. [Validation] Acc: 0.8679, SE: 0.5878, SP: 0.9576, PC: 0.6365, F1: 0.5286, JS: 0.4398, DC: 0.5286 Epoch [47/100], Loss: 142.5119, [Training] Acc: 0.9674, SE: 0.9185, SP: 0.9749, PC: 0.9015, F1: 0.8962, JS: 0.8254, DC: 0.8962 Decay learning rate to lr: 8.407746918641558e-05. [Validation] Acc: 0.8480, SE: 0.3821, SP: 0.9937, PC: 0.5798, F1: 0.3997, JS: 0.3380, DC: 0.3997 Epoch [48/100], Loss: 154.5033, [Training] Acc: 0.9666, SE: 0.9182, SP: 0.9748, PC: 0.9024, F1: 0.8973, JS: 0.8267, DC: 0.8973 Decay learning rate to lr: 8.249110184327567e-05. [Validation] Acc: 0.8576, SE: 0.3789, SP: 0.9932, PC: 0.6079, F1: 0.4055, JS: 0.3419, DC: 0.4055 Epoch [49/100], Loss: 132.7678, [Training] Acc: 0.9695, SE: 0.9216, SP: 0.9762, PC: 0.9037, F1: 0.9014, JS: 0.8320, DC: 0.9014 Decay learning rate to lr: 8.090473450013576e-05. [Validation] Acc: 0.8607, SE: 0.4490, SP: 0.9907, PC: 0.6172, F1: 0.4580, JS: 0.3901, DC: 0.4580 Epoch [50/100], Loss: 135.9399, [Training] Acc: 0.9695, SE: 0.9216, SP: 0.9764, PC: 0.9028, F1: 0.9005, JS: 0.8305, DC: 0.9006 Decay learning rate to lr: 7.931836715699585e-05. [Validation] Acc: 0.8806, SE: 0.5015, SP: 0.9903, PC: 0.7104, F1: 0.5225, JS: 0.4403, DC: 0.5225 Epoch [51/100], Loss: 132.9578, [Training] Acc: 0.9692, SE: 0.9218, SP: 0.9756, PC: 0.9048, F1: 0.9018, JS: 0.8317, DC: 0.9018 Decay learning rate to lr: 7.773199981385594e-05. [Validation] Acc: 0.8362, SE: 0.2915, SP: 0.9979, PC: 0.5544, F1: 0.3252, JS: 0.2741, DC: 0.3252 Epoch [52/100], Loss: 134.7240, [Training] Acc: 0.9691, SE: 0.9224, SP: 0.9756, PC: 0.9042, F1: 0.9017, JS: 0.8321, DC: 0.9017 Decay learning rate to lr: 7.614563247071603e-05. [Validation] Acc: 0.8628, SE: 0.4795, SP: 0.9916, PC: 0.6601, F1: 0.4853, JS: 0.4079, DC: 0.4853 Epoch [53/100], Loss: 137.5579, [Training] Acc: 0.9687, SE: 0.9198, SP: 0.9757, PC: 0.9035, F1: 0.8989, JS: 0.8291, DC: 0.8989 Decay learning rate to lr: 7.455926512757612e-05. [Validation] Acc: 0.8373, SE: 0.2815, SP: 0.9977, PC: 0.4900, F1: 0.3105, JS: 0.2634, DC: 0.3105 Epoch [54/100], Loss: 129.5814, [Training] Acc: 0.9698, SE: 0.9255, SP: 0.9757, PC: 0.9042, F1: 0.9041, JS: 0.8353, DC: 0.9041 Decay learning rate to lr: 7.297289778443621e-05. [Validation] Acc: 0.8768, SE: 0.5064, SP: 0.9883, PC: 0.7182, F1: 0.5200, JS: 0.4341, DC: 0.5200 Epoch [55/100], Loss: 128.9607, [Training] Acc: 0.9703, SE: 0.9237, SP: 0.9754, PC: 0.9052, F1: 0.9031, JS: 0.8337, DC: 0.9031 Decay learning rate to lr: 7.13865304412963e-05. [Validation] Acc: 0.8811, SE: 0.6047, SP: 0.9803, PC: 0.7549, F1: 0.5893, JS: 0.4919, DC: 0.5893 Epoch [56/100], Loss: 128.9532, [Training] Acc: 0.9700, SE: 0.9231, SP: 0.9772, PC: 0.9047, F1: 0.9020, JS: 0.8328, DC: 0.9020 Decay learning rate to lr: 6.980016309815639e-05. [Validation] Acc: 0.8550, SE: 0.4100, SP: 0.9934, PC: 0.6078, F1: 0.4294, JS: 0.3644, DC: 0.4294 Epoch [57/100], Loss: 129.5934, [Training] Acc: 0.9710, SE: 0.9257, SP: 0.9762, PC: 0.9044, F1: 0.9043, JS: 0.8368, DC: 0.9043 Decay learning rate to lr: 6.821379575501648e-05. [Validation] Acc: 0.8531, SE: 0.4024, SP: 0.9960, PC: 0.6601, F1: 0.4330, JS: 0.3699, DC: 0.4330 Epoch [58/100], Loss: 128.0230, [Training] Acc: 0.9706, SE: 0.9255, SP: 0.9778, PC: 0.9069, F1: 0.9050, JS: 0.8367, DC: 0.9050 Decay learning rate to lr: 6.662742841187657e-05. [Validation] Acc: 0.8653, SE: 0.4470, SP: 0.9906, PC: 0.6303, F1: 0.4607, JS: 0.3883, DC: 0.4607 Epoch [59/100], Loss: 123.0228, [Training] Acc: 0.9720, SE: 0.9290, SP: 0.9762, PC: 0.9065, F1: 0.9073, JS: 0.8401, DC: 0.9073 Decay learning rate to lr: 6.504106106873666e-05. [Validation] Acc: 0.8283, SE: 0.2809, SP: 0.9983, PC: 0.4528, F1: 0.3051, JS: 0.2593, DC: 0.3051 Epoch [60/100], Loss: 130.3902, [Training] Acc: 0.9706, SE: 0.9260, SP: 0.9770, PC: 0.9051, F1: 0.9045, JS: 0.8364, DC: 0.9045 Decay learning rate to lr: 6.345469372559675e-05. [Validation] Acc: 0.8359, SE: 0.2593, SP: 0.9984, PC: 0.5649, F1: 0.2984, JS: 0.2458, DC: 0.2984 Epoch [61/100], Loss: 122.7981, [Training] Acc: 0.9713, SE: 0.9274, SP: 0.9758, PC: 0.9058, F1: 0.9055, JS: 0.8380, DC: 0.9055 Decay learning rate to lr: 6.186832638245684e-05. [Validation] Acc: 0.8321, SE: 0.2448, SP: 0.9982, PC: 0.4488, F1: 0.2759, JS: 0.2312, DC: 0.2759 Epoch [62/100], Loss: 133.0434, [Training] Acc: 0.9709, SE: 0.9258, SP: 0.9767, PC: 0.9066, F1: 0.9051, JS: 0.8376, DC: 0.9051 Decay learning rate to lr: 6.028195903931692e-05. [Validation] Acc: 0.8631, SE: 0.4801, SP: 0.9922, PC: 0.7014, F1: 0.4991, JS: 0.4219, DC: 0.4991 Epoch [63/100], Loss: 117.8739, [Training] Acc: 0.9725, SE: 0.9315, SP: 0.9749, PC: 0.9069, F1: 0.9094, JS: 0.8429, DC: 0.9094 Decay learning rate to lr: 5.8695591696177006e-05. [Validation] Acc: 0.8605, SE: 0.3999, SP: 0.9923, PC: 0.6408, F1: 0.4280, JS: 0.3599, DC: 0.4280 Epoch [64/100], Loss: 127.6605, [Training] Acc: 0.9711, SE: 0.9297, SP: 0.9756, PC: 0.9067, F1: 0.9078, JS: 0.8405, DC: 0.9078 Decay learning rate to lr: 5.710922435303709e-05. [Validation] Acc: 0.8230, SE: 0.2021, SP: 0.9992, PC: 0.4158, F1: 0.2313, JS: 0.1916, DC: 0.2313 Epoch [65/100], Loss: 128.5371, [Training] Acc: 0.9711, SE: 0.9305, SP: 0.9759, PC: 0.9075, F1: 0.9088, JS: 0.8425, DC: 0.9088 Decay learning rate to lr: 5.552285700989717e-05. [Validation] Acc: 0.8552, SE: 0.3914, SP: 0.9823, PC: 0.5981, F1: 0.3986, JS: 0.3271, DC: 0.3986 Epoch [66/100], Loss: 118.5605, [Training] Acc: 0.9728, SE: 0.9315, SP: 0.9774, PC: 0.9074, F1: 0.9093, JS: 0.8431, DC: 0.9093 Decay learning rate to lr: 5.3936489666757256e-05. [Validation] Acc: 0.8552, SE: 0.3702, SP: 0.9854, PC: 0.6322, F1: 0.3886, JS: 0.3183, DC: 0.3886 Epoch [67/100], Loss: 120.1187, [Training] Acc: 0.9726, SE: 0.9323, SP: 0.9766, PC: 0.9069, F1: 0.9098, JS: 0.8438, DC: 0.9098 Decay learning rate to lr: 5.235012232361734e-05. [Validation] Acc: 0.8365, SE: 0.2735, SP: 0.9978, PC: 0.4829, F1: 0.3005, JS: 0.2550, DC: 0.3005 Epoch [68/100], Loss: 115.3787, [Training] Acc: 0.9732, SE: 0.9331, SP: 0.9780, PC: 0.9084, F1: 0.9115, JS: 0.8460, DC: 0.9115 Decay learning rate to lr: 5.076375498047742e-05. [Validation] Acc: 0.8570, SE: 0.4180, SP: 0.9947, PC: 0.6673, F1: 0.4447, JS: 0.3768, DC: 0.4447 Epoch [69/100], Loss: 112.7998, [Training] Acc: 0.9734, SE: 0.9329, SP: 0.9766, PC: 0.9083, F1: 0.9115, JS: 0.8459, DC: 0.9115 Decay learning rate to lr: 4.9177387637337506e-05. [Validation] Acc: 0.8590, SE: 0.4145, SP: 0.9868, PC: 0.5589, F1: 0.4077, JS: 0.3382, DC: 0.4077 Epoch [70/100], Loss: 117.0291, [Training] Acc: 0.9731, SE: 0.9315, SP: 0.9768, PC: 0.9082, F1: 0.9102, JS: 0.8444, DC: 0.9102 Decay learning rate to lr: 4.759102029419759e-05. [Validation] Acc: 0.8575, SE: 0.3105, SP: 0.9938, PC: 0.5990, F1: 0.3534, JS: 0.2878, DC: 0.3534 Epoch [71/100], Loss: 111.5511, [Training] Acc: 0.9739, SE: 0.9315, SP: 0.9786, PC: 0.9082, F1: 0.9103, JS: 0.8447, DC: 0.9103 Decay learning rate to lr: 4.600465295105767e-05. [Validation] Acc: 0.8644, SE: 0.4540, SP: 0.9872, PC: 0.6522, F1: 0.4710, JS: 0.3954, DC: 0.4710 Epoch [72/100], Loss: 111.7463, [Training] Acc: 0.9736, SE: 0.9351, SP: 0.9767, PC: 0.9070, F1: 0.9120, JS: 0.8468, DC: 0.9120 Decay learning rate to lr: 4.4418285607917756e-05. [Validation] Acc: 0.8690, SE: 0.4672, SP: 0.9919, PC: 0.7014, F1: 0.4944, JS: 0.4190, DC: 0.4944 Epoch [73/100], Loss: 113.9828, [Training] Acc: 0.9736, SE: 0.9342, SP: 0.9778, PC: 0.9080, F1: 0.9120, JS: 0.8468, DC: 0.9120 Decay learning rate to lr: 4.283191826477784e-05. [Validation] Acc: 0.8249, SE: 0.1878, SP: 0.9992, PC: 0.3749, F1: 0.2150, JS: 0.1798, DC: 0.2150 Epoch [74/100], Loss: 114.5390, [Training] Acc: 0.9743, SE: 0.9362, SP: 0.9768, PC: 0.9091, F1: 0.9138, JS: 0.8498, DC: 0.9138 Decay learning rate to lr: 4.124555092163792e-05. [Validation] Acc: 0.8589, SE: 0.4451, SP: 0.9888, PC: 0.6534, F1: 0.4581, JS: 0.3879, DC: 0.4581 Epoch [75/100], Loss: 114.1872, [Training] Acc: 0.9738, SE: 0.9342, SP: 0.9771, PC: 0.9079, F1: 0.9118, JS: 0.8476, DC: 0.9118 Decay learning rate to lr: 3.9659183578498005e-05. [Validation] Acc: 0.8792, SE: 0.4889, SP: 0.9894, PC: 0.6991, F1: 0.5079, JS: 0.4266, DC: 0.5079 Epoch [76/100], Loss: 104.4477, [Training] Acc: 0.9753, SE: 0.9391, SP: 0.9782, PC: 0.9095, F1: 0.9158, JS: 0.8525, DC: 0.9158 Decay learning rate to lr: 3.807281623535809e-05. [Validation] Acc: 0.8809, SE: 0.5931, SP: 0.9759, PC: 0.7317, F1: 0.5736, JS: 0.4811, DC: 0.5736 Epoch [77/100], Loss: 110.7846, [Training] Acc: 0.9747, SE: 0.9372, SP: 0.9758, PC: 0.9095, F1: 0.9145, JS: 0.8513, DC: 0.9145 Decay learning rate to lr: 3.648644889221817e-05. [Validation] Acc: 0.8463, SE: 0.3824, SP: 0.9961, PC: 0.6520, F1: 0.4095, JS: 0.3465, DC: 0.4095 Epoch [78/100], Loss: 111.5872, [Training] Acc: 0.9746, SE: 0.9378, SP: 0.9767, PC: 0.9086, F1: 0.9143, JS: 0.8507, DC: 0.9143 Decay learning rate to lr: 3.4900081549078255e-05. [Validation] Acc: 0.8217, SE: 0.2495, SP: 0.9973, PC: 0.4141, F1: 0.2590, JS: 0.2165, DC: 0.2590 Epoch [79/100], Loss: 107.1365, [Training] Acc: 0.9752, SE: 0.9386, SP: 0.9772, PC: 0.9084, F1: 0.9147, JS: 0.8512, DC: 0.9147 Decay learning rate to lr: 3.331371420593834e-05. [Validation] Acc: 0.8203, SE: 0.1925, SP: 0.9987, PC: 0.4340, F1: 0.2168, JS: 0.1793, DC: 0.2168 Epoch [80/100], Loss: 110.9553, [Training] Acc: 0.9747, SE: 0.9401, SP: 0.9775, PC: 0.9104, F1: 0.9167, JS: 0.8541, DC: 0.9167 Decay learning rate to lr: 3.172734686279842e-05. [Validation] Acc: 0.8400, SE: 0.3086, SP: 0.9943, PC: 0.5604, F1: 0.3288, JS: 0.2702, DC: 0.3288 Epoch [81/100], Loss: 107.6854, [Training] Acc: 0.9747, SE: 0.9379, SP: 0.9757, PC: 0.9097, F1: 0.9153, JS: 0.8520, DC: 0.9153 Decay learning rate to lr: 3.014097951965851e-05. [Validation] Acc: 0.8590, SE: 0.4347, SP: 0.9888, PC: 0.6858, F1: 0.4561, JS: 0.3730, DC: 0.4561 Epoch [82/100], Loss: 100.4412, [Training] Acc: 0.9763, SE: 0.9421, SP: 0.9773, PC: 0.9112, F1: 0.9192, JS: 0.8574, DC: 0.9192 Decay learning rate to lr: 2.8554612176518595e-05. [Validation] Acc: 0.8586, SE: 0.4027, SP: 0.9890, PC: 0.6283, F1: 0.4167, JS: 0.3446, DC: 0.4167 Epoch [83/100], Loss: 100.5839, [Training] Acc: 0.9765, SE: 0.9410, SP: 0.9769, PC: 0.9100, F1: 0.9177, JS: 0.8555, DC: 0.9177 Decay learning rate to lr: 2.6968244833378682e-05. [Validation] Acc: 0.8735, SE: 0.5046, SP: 0.9875, PC: 0.6966, F1: 0.5143, JS: 0.4310, DC: 0.5143 Epoch [84/100], Loss: 103.8184, [Training] Acc: 0.9758, SE: 0.9386, SP: 0.9776, PC: 0.9094, F1: 0.9157, JS: 0.8532, DC: 0.9157 Decay learning rate to lr: 2.538187749023877e-05. [Validation] Acc: 0.8493, SE: 0.3664, SP: 0.9941, PC: 0.6099, F1: 0.3903, JS: 0.3255, DC: 0.3903 Epoch [85/100], Loss: 99.9430, [Training] Acc: 0.9762, SE: 0.9425, SP: 0.9774, PC: 0.9101, F1: 0.9187, JS: 0.8567, DC: 0.9187 Decay learning rate to lr: 2.3795510147098855e-05. [Validation] Acc: 0.8771, SE: 0.5697, SP: 0.9862, PC: 0.7346, F1: 0.5689, JS: 0.4821, DC: 0.5689 Epoch [86/100], Loss: 98.2186, [Training] Acc: 0.9769, SE: 0.9432, SP: 0.9774, PC: 0.9120, F1: 0.9199, JS: 0.8589, DC: 0.9199 Decay learning rate to lr: 2.2209142803958942e-05. [Validation] Acc: 0.8345, SE: 0.2655, SP: 0.9970, PC: 0.5850, F1: 0.3023, JS: 0.2480, DC: 0.3023 Epoch [87/100], Loss: 96.0286, [Training] Acc: 0.9779, SE: 0.9459, SP: 0.9791, PC: 0.9114, F1: 0.9216, JS: 0.8610, DC: 0.9216 Decay learning rate to lr: 2.062277546081903e-05. [Validation] Acc: 0.8534, SE: 0.3963, SP: 0.9900, PC: 0.6184, F1: 0.4122, JS: 0.3423, DC: 0.4122 Epoch [88/100], Loss: 95.6034, [Training] Acc: 0.9776, SE: 0.9456, SP: 0.9782, PC: 0.9116, F1: 0.9217, JS: 0.8614, DC: 0.9217 Decay learning rate to lr: 1.9036408117679116e-05. [Validation] Acc: 0.8494, SE: 0.3946, SP: 0.9945, PC: 0.6363, F1: 0.4100, JS: 0.3464, DC: 0.4100 Epoch [89/100], Loss: 103.0424, [Training] Acc: 0.9763, SE: 0.9436, SP: 0.9764, PC: 0.9108, F1: 0.9194, JS: 0.8582, DC: 0.9194 Decay learning rate to lr: 1.7450040774539202e-05. [Validation] Acc: 0.8722, SE: 0.5715, SP: 0.9749, PC: 0.7317, F1: 0.5528, JS: 0.4574, DC: 0.5528 Epoch [90/100], Loss: 94.8561, [Training] Acc: 0.9779, SE: 0.9454, SP: 0.9788, PC: 0.9129, F1: 0.9223, JS: 0.8623, DC: 0.9223 Decay learning rate to lr: 1.586367343139929e-05. [Validation] Acc: 0.8499, SE: 0.3627, SP: 0.9857, PC: 0.5811, F1: 0.3696, JS: 0.2977, DC: 0.3696 Epoch [91/100], Loss: 97.1100, [Training] Acc: 0.9772, SE: 0.9441, SP: 0.9789, PC: 0.9123, F1: 0.9209, JS: 0.8604, DC: 0.9209 Decay learning rate to lr: 1.4277306088259374e-05. [Validation] Acc: 0.8549, SE: 0.3847, SP: 0.9918, PC: 0.6167, F1: 0.4026, JS: 0.3344, DC: 0.4026 Epoch [92/100], Loss: 97.6109, [Training] Acc: 0.9772, SE: 0.9463, SP: 0.9780, PC: 0.9128, F1: 0.9219, JS: 0.8623, DC: 0.9219 Decay learning rate to lr: 1.2690938745119459e-05. [Validation] Acc: 0.8810, SE: 0.5486, SP: 0.9763, PC: 0.6881, F1: 0.5342, JS: 0.4462, DC: 0.5342 Epoch [93/100], Loss: 98.1793, [Training] Acc: 0.9771, SE: 0.9475, SP: 0.9777, PC: 0.9124, F1: 0.9231, JS: 0.8634, DC: 0.9231 Decay learning rate to lr: 1.1104571401979544e-05. [Validation] Acc: 0.8654, SE: 0.5406, SP: 0.9752, PC: 0.6586, F1: 0.5128, JS: 0.4222, DC: 0.5128 Epoch [94/100], Loss: 95.9679, [Training] Acc: 0.9780, SE: 0.9477, SP: 0.9769, PC: 0.9127, F1: 0.9232, JS: 0.8641, DC: 0.9232 Decay learning rate to lr: 9.518204058839629e-06. [Validation] Acc: 0.8524, SE: 0.3428, SP: 0.9933, PC: 0.6244, F1: 0.3769, JS: 0.3094, DC: 0.3769 Epoch [95/100], Loss: 97.0966, [Training] Acc: 0.9774, SE: 0.9460, SP: 0.9774, PC: 0.9115, F1: 0.9209, JS: 0.8613, DC: 0.9209 Decay learning rate to lr: 7.931836715699714e-06. [Validation] Acc: 0.8344, SE: 0.2619, SP: 0.9978, PC: 0.4991, F1: 0.2908, JS: 0.2430, DC: 0.2908 Epoch [96/100], Loss: 93.9771, [Training] Acc: 0.9782, SE: 0.9485, SP: 0.9789, PC: 0.9114, F1: 0.9232, JS: 0.8641, DC: 0.9232 Decay learning rate to lr: 6.345469372559799e-06. [Validation] Acc: 0.8573, SE: 0.4120, SP: 0.9915, PC: 0.6316, F1: 0.4266, JS: 0.3564, DC: 0.4266 Epoch [97/100], Loss: 90.2515, [Training] Acc: 0.9787, SE: 0.9472, SP: 0.9778, PC: 0.9126, F1: 0.9225, JS: 0.8633, DC: 0.9225 Decay learning rate to lr: 4.759102029419884e-06. [Validation] Acc: 0.8519, SE: 0.3352, SP: 0.9952, PC: 0.6188, F1: 0.3719, JS: 0.3101, DC: 0.3719 Epoch [98/100], Loss: 94.7197, [Training] Acc: 0.9784, SE: 0.9477, SP: 0.9788, PC: 0.9125, F1: 0.9234, JS: 0.8642, DC: 0.9234 Decay learning rate to lr: 3.172734686279969e-06. [Validation] Acc: 0.8474, SE: 0.3480, SP: 0.9933, PC: 0.6090, F1: 0.3693, JS: 0.3056, DC: 0.3693 Epoch [99/100], Loss: 95.0584, [Training] Acc: 0.9779, SE: 0.9487, SP: 0.9775, PC: 0.9103, F1: 0.9225, JS: 0.8632, DC: 0.9225 Decay learning rate to lr: 1.5863673431400541e-06. [Validation] Acc: 0.8265, SE: 0.1964, SP: 0.9988, PC: 0.4323, F1: 0.2260, JS: 0.1865, DC: 0.2260 Epoch [100/100], Loss: 92.9643, [Training] Acc: 0.9780, SE: 0.9449, SP: 0.9783, PC: 0.9149, F1: 0.9227, JS: 0.8639, DC: 0.9227 Decay learning rate to lr: 1.393369198233324e-19. [Validation] Acc: 0.8399, SE: 0.2861, SP: 0.9969, PC: 0.6008, F1: 0.3243, JS: 0.2663, DC: 0.3243

Yanxingang commented 4 years ago

I think this is because the dataset is in different size.
ISIC2018 dataset I downloaded has 16072 images.

CN-zdy commented 4 years ago

@Yanxingang Thank you. I don't know why. ISIC2018 dataset I downloaded only has 2594 images in train data. Other data don't have GT

Yanxingang commented 4 years ago

Sorry, I make a mistake. The dataset is correct. I'm training. I will show you the result later.

Here's my result. I think, firstly, training acc is growing up while training. But it doesn't mean the val acc must grow up while training. Although the learning rate is reduced. The validation dataset is small. So the weight whose validation acc is largest is choosen to test the test dataset. The test acc is 0.8937

In this example, the batchsize is set to 1. This may cause Unstable convergence.

Epoch [1/150], Loss: 570.0526, [Training] Acc: 0.9019, SE: 0.7542, SP: 0.9784, PC: 0.8075, F1: 0.7205, JS: 0.6109, DC: 0.7205 [Validation] Acc: 0.8359, SE: 0.7846, SP: 0.8647, PC: 0.5920, F1: 0.5496, JS: 0.4421, DC: 0.5496 Best U_Net model score : 0.9917 Epoch [2/150], Loss: 376.0612, [Training] Acc: 0.9287, SE: 0.8176, SP: 0.9806, PC: 0.8476, F1: 0.7923, JS: 0.6946, DC: 0.7923 [Validation] Acc: 0.8772, SE: 0.7124, SP: 0.9163, PC: 0.5920, F1: 0.5305, JS: 0.4250, DC: 0.5305 Epoch [3/150], Loss: 329.8286, [Training] Acc: 0.9347, SE: 0.8346, SP: 0.9796, PC: 0.8544, F1: 0.8069, JS: 0.7119, DC: 0.8069 [Validation] Acc: 0.9079, SE: 0.5440, SP: 0.9886, PC: 0.8248, F1: 0.5621, JS: 0.4549, DC: 0.5621 Best U_Net model score : 1.0171 Epoch [4/150], Loss: 298.9687, [Training] Acc: 0.9397, SE: 0.8526, SP: 0.9789, PC: 0.8607, F1: 0.8259, JS: 0.7351, DC: 0.8259 [Validation] Acc: 0.8281, SE: 0.8838, SP: 0.8450, PC: 0.4564, F1: 0.5017, JS: 0.3838, DC: 0.5017 Epoch [5/150], Loss: 283.7692, [Training] Acc: 0.9426, SE: 0.8542, SP: 0.9784, PC: 0.8635, F1: 0.8272, JS: 0.7395, DC: 0.8272 [Validation] Acc: 0.8761, SE: 0.1143, SP: 0.9939, PC: 0.5202, F1: 0.1470, JS: 0.1064, DC: 0.1470

CN-zdy commented 4 years ago

@Yanxingang Thank you very much! Problem was solved when i set batchsize to 4.