- Update your code in order to create the LeNet-5 convolutional neural network, perform forward and back-prop using PyTorch’s nn package.
- Compare speed and training charts with the fully connected network (HW4).
The following are the training time and training error for LeNet-5 model using MNIST dataset
The following are the training time and training error for homework 4 model using MNIST dataset
- Train a LeNet-5 that correctly classify the CIFAR-100 data set.
- Visualise one object and its caption.
- Visualise continuous captioning of images from a webcam.
For CIFAR-100, the accuracy is still only 18.56% after 150 training epochs in my implementation; For CIFAR-10, the accuracy can reach 59.73% after simply 50 training epochs.
The following is a running example of Img2obj.view() function. The function displayed a 32x32 image and predicted the image as a "cat" object.
The following is a running example of Img2obj.cam() function. The last line "dog" in the terminal means the function detected a "dog"-like object in front of the webcam.
