Celebrity classification with transfer learning using inceptionV3
Celebrity classification using Convolutional Neural Networks
Used dataset from VGGFace2, you can find it here
I choosed 5 classes from the dataset with id 51-55
The dataset of 5 classes contains 1808 images totaly
The data was split in the following way:
1. 80% of data for training
2. 10% of data for testing
3. 10% of data for validation
We don’t train an entire Convolutional Network from scratch. We freeze some of the ConvNet layers and only fine-tune some higher-level portion of the network.
So we freeze all layers from InceptionV3 except the last 63 layer
This is the architecture of inceptionv3. I added two more layers. One GlobalAveragePooling and one Dense layer with 1024 neurons.
Firstly we have freezed all the layers of the network except the classifier and we trained the model for 50 epochs.Then we have unfreezed the last 63 layers of the network and trained the network for 50 epochs.
The accuracy of the model during the training
The loss of the model during the training
The accuracy of the model at test set is: 0.982
| precision | recall | f1-score | support | |
|---|---|---|---|---|
| Adam_Gilchrist | 1.00 | 0.94 | 0.97 | 32 |
| Adam_Housley | 0.95 | 1.00 | 0.98 | 21 |
| Adam_Irigoyen | 1.00 | 1.00 | 1.00 | 38 |
| Adam_Johnson | 1.00 | 0.98 | 0.99 | 43 |
| Adam_Lallana | 0.96 | 1.00 | 0.98 | 45 |