#IN THIS IPYTHON FILE 7 EMOTIONS OF HUMEN ARE DETECTED #THE CLASSES ARE [ANGRY,DISGUST,FEAR,HAPPY,NEUTRAL,SAD,SURPRISE] ---> fro this detection I am using convolution neural networks with keras ---> here i put 4 continue layers again and again for better results --->i am using imagedaragenrator class of keras to actually manuplate the image of input in this manuplation i am converting to gray then changing the size and adding a third dimention for neural network for adding third dimention i am using numpy function extend dims.
after this i am making the model in this i am defining convolution neural network in this i am putting these parameters: 1)conv2D :will create a convolution layer for model 2)activation : defining activation function(relu) 3)batch normalization : normalize activation of previous layer for each batch 4)maxpooling2D : taking the maximum value from a defined window size of full matrix 5)dropout: randomly select neurons and ignore during training (0.5/half)
here after defining function the filters were not enough so i am adding them again and testing for this 2nd time : 64 filters 3rd time : 128 filters 4th time : 256 filters then flatting them and then dence connected layers for connecting all the neurons. doing it again but without flatten layer because we have done it in previous layer and then putting activation layerwith softmax
After prapring the model i am also putting 3 main things with callback class of keras these things are following :
- checkpoint : it will detect the validation loss and minimize it. it is more likely a booster for neural network 2)early stopping : it helps to stop training earlier than usual the point behind using this is take the tracking of validation loss from checkpoint and if it has minimized and all the weights are allready restored then simply stop the trainining it helps a lot when we need our model quickly as possible
- reduce learing rate : Let's assume our model is in traing part and we find there is no improvment in learning rate, In this case we can reduce the learning rate. This function reduce it by a factor 2-10.
after these three important steps we can compile our model for this i am calling the compile function from my model class which is sequential class of keras then fitting the train and validation sets by model.fit_generator and we are done with our model
for testing the model i also provide a live mode where you can detect the emotion by running that and also a code which will detect emotion of saved image from your pc