Hi all,
I ran my code for a segmentation task and the resulting output of the validation loss and score compared to the training is not reasonable. Here’s the output:
Epoch: 1/10 Training | pathologies - loss: 1.4878 score: 0.7301 lunglobes - loss: 3.6833 score: 0.5535 Validation | pathologies - loss: 2.1480 score: 0.6627 lunglobes - loss: 4.7785 score: 0.3517 |Time: 0:03:06
Training | pathologies: [0.8714 0.1824 0.2388], Lunglobes: [0.8764 0.122 0.0881 0.1137 0.1 0.118 ]
Validation | pathologies: [0.9067 0.0004 0.0193], Lunglobes: [0.8082 0.0519 0.0288 0.0347 0.0811 0.0519]
Validation loss decreased (inf --> 3.321953). Saving model ...
Epoch: 2/10 Training | pathologies - loss: 0.9487 score: 0.9429 lunglobes - loss: 2.7038 score: 0.8808 Validation | pathologies - loss: 2.0427 score: 0.9951 lunglobes - loss: 4.6012 score: 0.9285 |Time: 0:02:43
Training | pathologies: [0.9637 0.3579 0.4035], Lunglobes: [0.9765 0.2197 0.1402 0.222 0.1157 0.257 ]
Validation | pathologies: [0.9872 0.0016 0.0278], Lunglobes: [0.9646 0.0958 0.0952 0.0498 0.0964 0.0754]
Validation loss decreased (3.321953 --> 3.277506). Saving model ...
Epoch: 3/10 Training | pathologies - loss: 0.7948 score: 0.9630 lunglobes - loss: 2.2768 score: 0.9454 Validation | pathologies - loss: 2.0350 score: 0.9624 lunglobes - loss: 4.5200 score: 0.9001 |Time: 0:02:43
Training | pathologies: [0.9771 0.4909 0.4688], Lunglobes: [0.9859 0.2769 0.2207 0.2997 0.1757 0.3169]
Validation | pathologies: [0.9839 0.0016 0.0473], Lunglobes: [0.961 0.1586 0.1008 0.0494 0.0877 0.1052]
Epoch: 4/10 Training | pathologies - loss: 0.7234 score: 0.9670 lunglobes - loss: 1.8668 score: 0.9572 Validation | pathologies - loss: 2.1716 score: 0.8341 lunglobes - loss: 4.8151 score: 0.7460 |Time: 0:02:42
Training | pathologies: [0.9806 0.5481 0.5065], Lunglobes: [0.9886 0.309 0.2769 0.3348 0.239 0.3806]
Validation | pathologies: [0.9201 0.0007 0.017 ], Lunglobes: [0.863 0.0596 0.0854 0.038 0.0362 0.0679]
Epoch: 5/10 Training | pathologies - loss: 0.6755 score: 0.9701 lunglobes - loss: 1.6758 score: 0.9631 Validation | pathologies - loss: 2.0456 score: 0.9831 lunglobes - loss: 4.5550 score: 0.9199 |Time: 0:02:42
Training | pathologies: [0.983 0.5677 0.536 ], Lunglobes: [0.9894 0.3563 0.3296 0.3845 0.277 0.3667]
Validation | pathologies: [0.9889 0.0019 0.0296], Lunglobes: [0.9713 0.1499 0.1192 0.0417 0.065 0.1274]
I am using a custom weighted dice + cross-entropy loss with dice metric as the evaluation metric as well as Adam optimizer. Your contributions are well appreciated.