In my dataset, I seperate it to two parts. One for test, the other for evaluation. Specifically, I take 85 percents images in each class for testing(train_data), 25 percents images for evaluating(val_data).
In each part, I divded images into 10 categories, they are laughing, pullinghair, quarrel, slapping, punching, stabbing, gossiping, strangle, isolation and nonbullying.
Note that :
- the dataset of unbulloying action comes from Standford 40 Action
- I don't have the Copy Right on the bully image datasets, so I can't upload these datasets yet. The Copy Right belongs to Dr. Feng Luo
In my model, I made two covolutional layers and one fully connected layer. In convolutional layer, I make inchannel in first layer to 3 dimension(Red, Green, Blue) and outchannel to 16. For the second convulotional layer, inchannel to 16 and outchannel to 32 Both of these convolutional layers are using 3X3 filter size, stride =2 , padding = 1 and 2x2 for max pooling and using ReLU for activation function. To prevent overfitting, I add BatchNormalization for each convlutional layer and Dropout 50 percents information after ending of each convolutional layer.
In fully connnected layers, I flated 565632 neurons to do 10 categories classification
To initilize weight, I use xavier to initial the weight of two convlutional layers, 1 for Batch Normalization layer, 0.01 for fully connected layer.
I have set learning rate to 0.005, batch_size to 8
In training part, I made epoch number to 35 and print currnt epoch , step and loss after each step
For testing, I have write a test script called 'test.py' and 'mycnn.ckpt' is my model.
Running the command of python test.py test.jpg(or any other test image) to test that script
I think my model still have overfiting problem, and my module's loss function is tending to stable after 20 epoch, the parameter is not carefully designed in this CNN Network. Also, in order to building the CNN structure by my own rather than using the existing CNN structure, I think my CNN architecture still unmature. This is another reason that my model encoutered overfitting problem.