DL-DIY potential project ideas

• posing this problem as regression or classification and compare them
• implementing the DEX method and Residual DEX method
• implement and test label smoothing for classification
• use Gaussian/Laplace likelihood loss (aleatoric loss) for regression having the variance as a network parameter (homoscedastic) or prediction from the input sample (heteroscedastic)
• find strategies for dealing with imbalanced data
• use other attributes from extended APPA-REAL dataset, e.g. ethnic, makeup, gender, expression, to train in a multi-task setting
• test generalization of a model trained on APPA-REAL for other datasets (check this paper for references on other datasets). Can you think of some ways of cheap domain adaptation, e.g. leveraging BatchNorm layers?

---

Age Estimation PyTorch

PyTorch-based CNN implementation for estimating age from face images. Currently only the APPA-REAL dataset is supported. Similar Keras-based project can be found here.

Requirements

pip install -r requirements.txt

Demo

Webcam is required. See python demo.py -h for detailed options.

python demo.py

Using --img_dir argument, images in that directory will be used as input:

python demo.py --img_dir [PATH/TO/IMAGE_DIRECTORY]

Further using --output_dir argument, resulting images will be saved in that directory (no resulting image window is displayed in this case):

python demo.py --img_dir [PATH/TO/IMAGE_DIRECTORY] --output_dir [PATH/TO/OUTPUT_DIRECTORY]

Train

Download Dataset

Download and extract the APPA-REAL dataset.

The APPA-REAL database contains 7,591 images with associated real and apparent age labels. The total number of apparent votes is around 250,000. On average we have around 38 votes per each image and this makes the average apparent age very stable (0.3 standard error of the mean).

wget http://158.109.8.102/AppaRealAge/appa-real-release.zip
unzip appa-real-release.zip

Train Model

Train a model using the APPA-REAL dataset. See python train.py -h for detailed options.

python train.py --data_dir [PATH/TO/appa-real-release] --tensorboard tf_log

Check training progress:

tensorboard --logdir=tf_log

Training Options

You can change training parameters including model architecture using additional arguments like this:

python train.py --data_dir [PATH/TO/appa-real-release] --tensorboard tf_log MODEL.ARCH se_resnet50 TRAIN.OPT sgd TRAIN.LR 0.1

All default parameters defined in defaults.py can be changed using this style.

Test Trained Model

Evaluate the trained model using the APPA-REAL test dataset.

python test.py --data_dir [PATH/TO/appa-real-release] --resume [PATH/TO/BEST_MODEL.pth]

After evaluation, you can see something like this:

100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 16/16 [00:08<00:00,  1.28it/s]
test mae: 4.800