Computer Vision News - February 2020

3 Summary EfficientNet with PyTorch 15 We will use one image of a panda to test the model. We will also use the same code to classify the same image with ResNet-18 and Mobile-Net which are also implemented in PyTorch. For this, we only make the following change (respectively): We now run the code to check the results. We start with ResNet and Mobile- Net: As we can see, the baseline models achieved high accuracy with running time of 0.14 and 0.32 seconds. Now Let's run Efficient-Net. We start with Efficient- Net-B0: As expected, Efficient-Net-B0 achieved slightly worst accuracy but relatively comparable running time. Here we can see that although theoretically it computes less FLOPs, the running time is not dramatically faster. Let's check what happen when we take Efficient-Net-B7: model = models . resnet18(pretrained = True ) #ResNet-18 model = models . mobilenet_v2(pretrained = True ) #Mobile-Net-v2