single-image-prediction__w-pretrained-models

Single Image Predictions

Due to the large file sizes of the model, the pre-trained models are not available in this GitHub repository but are shared via Google Drive instead. Below are instructions for using these pre-trained models to make predictions on .jpg files.

1) Downloading the models

Visit the Google Drive link via your web browser at https://drive.google.com/drive/folders/168ijUQyvGLhHoQUQMlFS2fVt2p5ZV2bD?usp=sharing. There should be 3 subfolders,

• morph2 (725 Mb)
• cacd (744 Mb)
• afad (730 Mb)

which contain the models for each respective dataset.

2) Making predictions

There are three Python script files in this directory

• ce.py (regular cross entropy)
• ordinal.py (Niu et al. ordinal regression)
• coral.py (CORAL ordinal regression)

The models can be executed via the following shell (terminal) commands:

Cross entropy-based ResNet-34 classifier

python ce.py --dataset afad \
--image_path example-images/afad/18_years__948-0.jpg \
--state_dict_path ../afad/afad-ce__seed1/best_model.pt              

Output:

Class probabilities: tensor([[6.9999e-03, 9.3442e-03, 2.6857e-02, 7.9845e-02, 1.9216e-01, 4.3945e-01,
         1.1403e-01, 5.1849e-02, 2.2000e-02, 1.7506e-02, 1.9361e-02, 1.0340e-02,
         4.2141e-03, 1.8779e-03, 1.6173e-03, 5.1615e-08, 8.1854e-04, 4.1623e-06,
         4.9016e-04, 4.3236e-04, 2.8226e-04, 2.7673e-04, 5.1856e-08, 2.2791e-04,
         2.1825e-08, 1.8387e-05]])
Predicted class label: 5
Predicted age in years: 20

Note the class labels in the training sets start at 0, which is why the true age (Predicted age) is larger than the predicted label (Predicted class label).

Niu et al. Ordinal Regression w. ResNet-34

python ordinal.py --dataset afad \
--image_path example-images/afad/18_years__948-0.jpg \
--state_dict_path ../afad/afad-ordinal__seed1/best_model.pt    

Output:

Class probabilities: tensor([[9.9470e-01, 9.8194e-01, 9.6384e-01, 9.1268e-01, 7.6474e-01, 5.9748e-01,
         4.3897e-01, 2.9948e-01, 2.0706e-01, 1.3781e-01, 8.4818e-02, 4.4908e-02,
         3.0308e-02, 1.9418e-02, 1.2340e-02, 1.3033e-02, 9.0761e-03, 8.7384e-03,
         5.9033e-03, 3.4582e-03, 2.3147e-03, 6.9869e-04, 6.6803e-04, 1.1985e-04,
         1.3445e-04]])
Predicted class label: 6
Predicted age in years: 21

CORAL Ordinal Regression w. ResNet-34

python coral.py --dataset afad \
--image_path example-images/afad/18_years__948-0.jpg \
--state_dict_path ../afad/afad-coral__seed1/best_model.pt

Output:

Class probabilities: tensor([[7.9409e-01, 6.6242e-01, 5.0478e-01, 2.5930e-01, 6.9571e-02, 2.1320e-02,
         7.6356e-03, 2.5884e-03, 1.1306e-03, 3.9728e-04, 1.4751e-04, 6.7989e-05,
         3.1884e-05, 1.5259e-05, 6.9953e-06, 6.9953e-06, 3.0060e-06, 2.9677e-06,
         1.4319e-06, 6.5273e-07, 2.5818e-07, 8.4094e-08, 8.4094e-08, 2.2461e-08,
         2.2461e-08]])
Predicted class label: 3
Predicted age in years: 18

Note that if you would like to try out CACD or MORPH2 images, you need to change all three arguments accordingly, for example

python coral.py --dataset cacd \
--image_path example-images/cacd/41_Jason_Statham_0003.jpg \
--state_dict_path ../cacd/cacd-coral__seed2/best_model.pt

Output:

Class probabilities: tensor([[1.0000e+00, 1.0000e+00, 1.0000e+00, 1.0000e+00, 1.0000e+00, 1.0000e+00,
         1.0000e+00, 1.0000e+00, 1.0000e+00, 9.9999e-01, 9.9998e-01, 9.9996e-01,
         9.9991e-01, 9.9981e-01, 9.9961e-01, 9.9918e-01, 9.9837e-01, 9.9683e-01,
         9.9390e-01, 9.8826e-01, 9.7705e-01, 9.5824e-01, 9.2423e-01, 8.6348e-01,
         7.7465e-01, 6.3956e-01, 4.8542e-01, 3.2803e-01, 2.0232e-01, 1.2037e-01,
         6.7961e-02, 3.7153e-02, 1.9279e-02, 1.0134e-02, 5.1062e-03, 2.5095e-03,
         1.2001e-03, 5.4535e-04, 2.3731e-04, 9.9439e-05, 4.1659e-05, 1.6209e-05,
         6.1432e-06, 2.3529e-06, 8.2844e-07, 2.3559e-07, 6.8404e-08, 1.2028e-08]])
Predicted class label: 26
Predicted age in years: 40