![]() So when app is run locally on a laptop Video Streaming through webcam is possible. Code in current state makes use of webcam if available on server side not client side. ![]() Opencv tries to open the camera on whatever device the app is running on.The app in current state can't be deployed on web as: Training took around 13 hours locally for 75 epochs with an accuracy of ~66 %.The images were normalised, resized to (48,48) and converted to grayscale in batches of 64 with help of 'ImageDataGenerator' in Keras API.A bit more tinkering with hyper parameters might lead to a better accuracy Image Processing and Training: This model architecture gives good enough accuracy. Note:- Tried Implementing various other models like VGG16 but accuracy was far too low. Used 'categorical_crossentropy' for loss with 'Adam' optimizer with 'accuracy' metric.Final Dense layer has 'softmax' activation for classifying 7 emotions.Dropout is set to 0.25 as anything above results in poor performance.Conv2D layers throughout the model have different filter size from 32 to 128, all with activation 'relu'.The model architecture is a sequential model consisting of Conv2d, Maxpool2d, Dropout and Dense layers:.This might be a factor resulting in okaysish accuracy after training. ![]() Note that the dataset is highly imbalanced with happy class having maxiumum representation. Models trained on this dataset can classify 7 emotions. The dataset used for this project is the famous FER2013 dataset.
0 Comments
Leave a Reply. |
Details
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |