amol2709/emotion-recogition-using-keras

EMOTION-RECOGITION-USING-KERAS

I strongly recomend you to follow exact directory structure the way I have folllowed.

Task 1 : Build Dataset

After downloading dataset create following folders:

--fer2013

| |--fer2013 {In this folder keep .csv file ,in my repositories I Have deliberately not kept ...I urge you to download dataset krom kaggle and paste csv file in this directory}

| |--hdf5 {We will convert images into hdf5 format and store here }

| |--output {we will save training plot into output directory}

Now open your command prompt go to directory where you keep build_dataset.py and write these:

python build_dataset.py

After executing it, you will find that in directory fer2013/hdf5 three hdf5 file(test.hdf5,train.hdf5,validation.hdf5) will be generated. I have not uploaded these three files(it's size is 700mb).... .

Task 2: Train Model from scratch

To train the model open command prompt and execute following command

let's say you want to train model for 20 epoch at some learning rate let's say 'LR1' and want to check training loss,validation loss ,training accuracy ,validation accuracy and their plot then follow these steps:

step 1: First open train_recognizer.py and change learning rate (at line 67 and 76) and also change epoch(number of epoch) at line 95.

step 2: Open command prompt and execute this command:

python train_recognizer.py --checkpoints checkpoints

After executing this you will see in directory fer2013/output one image will be formed titled as 'vggnet_emotion.png' it will have loss and accuracy plots.....u can analyze from this plot and make sure we are not overfitting or underfitting .

If you do not satisfied from your model perfomance you can again start training from where you have left it............HOW?

ANSWER:

In --checkpoints directory we are storing model weights as the network is trained. If you open directory checkpoints you will see that after every 5 epoch we are storing model weights .

Only thing we need to do is we need to load a specific epoch from disk and we can restart training, we can supply the --model path to the specific checkpoint and provide the --start-epoch of the associated checkpoint.

EXample: let's say you want to restart training after 20 epoch. open command prompt and write:

      python train_recognizer.py --checkpoints checkpoints --model checkpoints/epoch_20.hdf5 --start 20

At Last ....To see on your own images and videos follow these step:

1. If you want to detect emotion on images:

Run emotion_detector_image.py , write following line in command prompt:

python emotion_detector_image.py --cascade haarcascade_frontalface_default.xml --model checkpoints/epoch_75.hdf5 -- image NAME OF IMAGE

You can parse your best epoch....

Note: If you have single face in image then .....all probability correspond to each emotion will make sense,if you have multiple faces in image then showing probabilty correspond to each emotion will look very clumpsy....that's why I have not did that...Any way when you get window which shows probability of different emotion that window is only relevant when you haveonly one face in an image.

2. If you want to detect emotion on video:

=> for web camera:

python emotion_detector.py --cascade haarcascade_frontalface_default.xml --model checkpoints/epoch_75.hdf5

=> for mp4 video: python emotion_detector_.py --cascade haarcascade_frontalface_default.xml --model checkpoints/epoch_75.hdf5 -- video video.mp4