Week #30

03 Jun 2019

Goal:

1. re-establish the experiments on ImageNet with Resnet, using CIFAR 10 dataset, try to get the same result as the paper.

Using the Resnet 18, without the fully connected layer, the final output is bs * 512 * 16 * 16, so with

2. Using ResNet in our case.

After re-observing the Resnet, I find that the ResNet shrink the image after each conv layer, for the res-18, without the final average pooling layer, using our 128 * 128 image input, we will get a bs * 512 * 16 * 16 output.

Since the target image size is 128 * 128. With 3 classes, we need to transform the final output size to bs * 3 * 128 * 128

So I added a upsample layer, which convert the bs * 512 * 16 * 16 to size bs * 512 * 128 * 128, and use an additional 1 * 1 conv layer to convert the chanel to 3, so the final output of the network is bs * 3 * 128 * 128.