Age and Gender Prediction using Deep cnns and Transfer Learning

Download 299,02 Kb. Pdf ko'rish
bet	6/13
Sana	22.12.2023
Hajmi	299,02 Kb.
	#127116

1 2 3 4 5 6 7 8 9 ... 13

Training & Testing. For age classification, our model classifies ages into 5 groups (0-
24, 25-49, 50-74, 75-99, and 100-124). For this, we had to perform integer division (by
25) on the age values and later one hot encodes them before feeding them into the
model. Similarly, gender also had to be one hot encoded for gender classification into
male and female. The loss function chosen for age estimation was mean-squared error
(MSE) as it is a regression task, whereas for age and gender classification it was cate-
gorical-cross entropy. For training, each model was trained using a custom data gener-
ator that allows training in mini-batches. Learning rate decay was used during training
as it allowed the learning rate to decrease after a fixed number of epochs. This is essen-
tial as the learning rate becomes very precarious during the latter stages of training,
when approaching convergence. Various experiments with different optimizers were
conducted, the results of which have been summarized in section 5.
Each model was trained between 30 to 50 epochs on average. Initial learning rate
was set of the order 1e-3. Batch size of 32 was used. The learning rate was changed to
0.6 times the current learning rate after about 9 epochs (on average) to ensure that by
the end of training, the learning rate is small enough for the model to converge to the
local minimum. Fig. 4 showcases the training plots of our models. In all graphs, the
blue line denotes the training and the red line denotes the validation result. It is very
evident that the training for gender classification was the noisiest whereas the training
for age estimation was the smoothest.

6
(a) (b) (c)
Fig. 2. Training plots depicting loss by the epoch of a) Age Estimation b) Age Classification
c) Gender Classification
Table 4 shows the lowest loss value to which our model could converge while training.
Table 4. Minimum loss value
Age
Estimation
Age
Classification
Gender
Classification
Train
Validation
Train
Validation
Train
Validation
56.8346
56.7397
0.3886
0.4009
0.1037
0.1864
The next subsection explores our work using transfer learning.

Download 299,02 Kb.

1 2 3 4 5 6 7 8 9 ... 13

Download 299,02 Kb.

Pdf ko'rish