*Unzip the folder gtsrb small, given here is small dataset if you want the large dataset you can download it here:https://cdn.cs50.net/ai/2020/x/projects/5/traffic.zip

• To run the program : python .\traffic.py .\gtsrb\

First I tried using 128 filters and 2x2 kernel for the convolution layer [tf.keras.layers.Conv2D( 128, (2,2), activation="relu", input_shape=(IMG_WIDTH, IMG_HEIGHT, 3) )] and with Max pooling (2, 2) [tf.keras.layers.MaxPooling2D(pool_size=(2, 2)), tf.keras.layers.Conv2D( 128, (2, 2), activation="relu", input_shape=(IMG_WIDTH, IMG_HEIGHT, 3) ] and a hidden layer with droupout 0.5 and 128 units, and also with another (2, 2) maxpooling layer and I got an Accuracy of 92.74% and loss of 2.52% Then I Changed the second convolution layer to 32, (2, 2), activation="relu", after that i got an accuracy of 80.83% and a loss of 6.3%. Then I changed both the convolution layer to 32, (3, 3) and I got an Accuracy of 94.33% and loss of 2.24%The best one Then I changed both max pooling layers to (4,4) and I got an Accuracy of 43.52% and loss of 17.02% (compilation time decreased) Then I changed back both the max pooling layers to (2, 2) and increased the both the convolution layers to 128(3, 3) and I got an Accuracy of 5.44% and loss of 34.98% (compilation time increased) Then I changed the first convolution layer to 32, (3, 3) and the 2nd to 64, (3,3) and got an Accuracy of 98.02% and loss of 0.791%

    Also I changed the dropout rate and found that 0.5 is good enough compared to other values. I found that having two 32,(3, 3) is optimal.