Use CNN models to analyze the CelebFaces Attributes Dataset.
-
Ratio of each face attribute.
-
Use correlation matrix and PCA to analyze the highly connected face attributes.
- Correlation matrix implies the highly connected attributes.
- PCA implies different aspects of connections in the attributes.
- Demonstrate the PCA loading plots and a table to summarise the study.
-
Strategic
a. Input shape
b. Convolutional layer
i. number of block ii. number of layers in each block iii. dropout iv. hyperparameters
c. Fully connected layer
i. number of layers ii. units iii. dropout
-
Purpose
a. Check the effect of the hyperparameters
b. Find the best combination of convolutional layer and fully connected layer
c. Hint to futher adjust CNN models
-
Predict the rank of the candidate models
a. Motivation
i. Time consuming ii. Memory consuming iii. Highly restricted to PC's specifications iv. Sampling unbiased data is suitable for representing the population
b. Method
i. Sampling a properly small dataset for model training process ii. Use metrics in each epoch to compare the performance with other models iii. Hypotheses testing result will suggest the rank of the performance of the candidate models
c. Validation
i. Whether the proposed rank of the candidate models matches the original dataset training result.
There are 83.49% faces without beard and there are 47.24% faces wearing lipstick.
In case of using 200k images for training models, we observe the performance of the accuracies according to all candidate models.
The training accuracies and the validation accuracies of all candidate models for no beard are almost the same as 0.8329 and 0.8531, respectively.
The minimum and maximum of the training accuracies of all candidate models are 0.8484 and 0.9288, respectively. As for the validation accuracy varies from 0.8699 to 0.9227.
In case of using sampling 2k images for the process of training model, we use unpaired two sample t-test for testing if the performance of each model has significantly difference from the others.
- Including the application of our CNN models
You can review our code and examine our data on GitHub.