The goal of this assignment is to create an interactive website where one can perform face recognition. We have used following classes to train our model:
- Donald Trump
- Elon Musk
- Emma Stone
- Emma Watson
- Geoffrey Hinton
- Hrithik Roshan
- Narendra Modi
- Priyanka Chopra
- Tom Cruise
- Will Smith
All the files and the models have to be deployed to AWS Lambda. The code to deploy them can be found here.
- Training Accuracy: 99.90%
- Loss: 0.00062
- Loss Function: Cross Entropy Loss
- Epochs: 15
- Optimizer: SGD
- Learning Rate: 0.01
- Batch Size: 128
- Scheduler: StepLR
| Input Image | Prediction |
|---|---|
 |
Emma Stone |
The dataset was aligned using dlib library. The code to do face alignment using dlib can be found here.
After alingment each image was set to a resolution of 600x600.
The images in the dataset was resized to a resolution of 224x224 and the transformation of RandomHorizontalFlip was applied, before training.
Since the number of images in each class was not equal, it led to class Imbalance problem. To solve this problem, a widely adopted technique is called resampling was implemented. It consists of less samples from the majority class and/or adding more examples from the minority class.
train_loader = torch.utils.data.DataLoader(dataset_train, batch_size, sampler=sampler)
The InceptionResnetV1 model was used for this classification which was pretrained on vggface2. The code for the model can be found here.
While training the model, we freezed the initial layers and fine-tuned last five layers of the model. To freeze the initial layers:
param.requires_grad = False