herosan163/ageestimation

PyTorch Implementation of Mean-Variance Loss for age estimation.

ageestimation's Issues

Unable to reproduce results

Log from last two epochs

epoch: 97, mean_loss: 11.779, variance_loss: 81.855, softmax_loss: 4.167, loss: 97.801, mae: 4.265306
epoch: 97, test_mae: 3.533333
[98, 0] mean_loss: 0.015, variance_loss: 0.086, softmax_loss: 0.075, loss: 0.176
[98, 1] mean_loss: 0.074, variance_loss: 0.181, softmax_loss: 0.201, loss: 0.456
[98, 2] mean_loss: 0.042, variance_loss: 0.137, softmax_loss: 0.155, loss: 0.334
[98, 3] mean_loss: 0.018, variance_loss: 0.261, softmax_loss: 0.175, loss: 0.454
[98, 4] mean_loss: 0.003, variance_loss: 0.084, softmax_loss: 0.084, loss: 0.171
[98, 5] mean_loss: 0.032, variance_loss: 0.132, softmax_loss: 0.165, loss: 0.329
[98, 6] mean_loss: 0.029, variance_loss: 0.270, softmax_loss: 0.068, loss: 0.367
[98, 7] mean_loss: 0.142, variance_loss: 0.065, softmax_loss: 0.233, loss: 0.440
[98, 8] mean_loss: 0.085, variance_loss: 0.187, softmax_loss: 0.177, loss: 0.449
[98, 9] mean_loss: 0.002, variance_loss: 0.070, softmax_loss: 0.053, loss: 0.125
[98, 10] mean_loss: 0.034, variance_loss: 0.254, softmax_loss: 0.226, loss: 0.514
[98, 11] mean_loss: 0.003, variance_loss: 0.083, softmax_loss: 0.128, loss: 0.214
[98, 12] mean_loss: 0.012, variance_loss: 0.081, softmax_loss: 0.173, loss: 0.265
[98, 13] mean_loss: 0.007, variance_loss: 0.146, softmax_loss: 0.083, loss: 0.236
epoch: 98, mean_loss: 11.781, variance_loss: 81.849, softmax_loss: 4.172, loss: 97.801, mae: 4.724490
epoch: 98, test_mae: 2.800000
[99, 0] mean_loss: 0.020, variance_loss: 0.038, softmax_loss: 0.143, loss: 0.200
[99, 1] mean_loss: 0.017, variance_loss: 0.191, softmax_loss: 0.104, loss: 0.312
[99, 2] mean_loss: 0.005, variance_loss: 0.138, softmax_loss: 0.065, loss: 0.208
[99, 3] mean_loss: 0.001, variance_loss: 0.041, softmax_loss: 0.084, loss: 0.126
[99, 4] mean_loss: 0.036, variance_loss: 0.171, softmax_loss: 0.111, loss: 0.319
[99, 5] mean_loss: 0.003, variance_loss: 0.060, softmax_loss: 0.082, loss: 0.145
[99, 6] mean_loss: 0.007, variance_loss: 0.121, softmax_loss: 0.140, loss: 0.268
[99, 7] mean_loss: 0.060, variance_loss: 0.245, softmax_loss: 0.378, loss: 0.682
[99, 8] mean_loss: 0.026, variance_loss: 0.323, softmax_loss: 0.127, loss: 0.476
[99, 9] mean_loss: 0.001, variance_loss: 0.080, softmax_loss: 0.051, loss: 0.132
[99, 10] mean_loss: 0.028, variance_loss: 0.239, softmax_loss: 0.110, loss: 0.377
[99, 11] mean_loss: 0.102, variance_loss: 0.127, softmax_loss: 0.171, loss: 0.399
[99, 12] mean_loss: 0.032, variance_loss: 0.118, softmax_loss: 0.196, loss: 0.345
[99, 13] mean_loss: 0.021, variance_loss: 0.213, softmax_loss: 0.118, loss: 0.351
epoch: 99, mean_loss: 11.747, variance_loss: 82.039, softmax_loss: 4.170, loss: 97.956, mae: 4.540816
epoch: 99, test_mae: 4.000000
best_loss_epoch: 2, best_val_loss: 96.452576, best_mae_epoch: 43, best_val_mae: 3.683673

As you can see it doesn't seem to be converging and the best loss was at epoch 2

LAMBDA_1 and LAMBDA_2 switched?

in main.py

criterion1 = MeanVarianceLoss(LAMBDA_2, LAMBDA_1, START_AGE, END_AGE).cuda()

it looks like LAMBDA_2 and LAMBDA_1 are switched.

anyhow I couldn't replicate your result so it might be because of the above. I am not sure.

I had to go through 1000+ epochs to reach MAE of 3 or 4.

Can you share the model pretrained on the IMDB-WIKI dataset?

Thanks very much for your work!
Can you share the model pretrained on the IMDB-WIKI dataset?

Pre-trained model

Thank you for your implementation Herosan163.

Would it be possible for you to share the trained model?

Thank you

The softmax loss becomes bigger during iterates.

Hello, when I run your code, the softmax loss starts to increase during the 16th epoch. Does this happen to you?

Code meaning

Can you explain to me what this sentence means? Thank you.
self.labels.append(int(basename[4:6]))
Is it to get the age label of FGNET?and When I run in morph2, how should I get the label?

What is your final loss and mae，please？