We provide a Jupyter notebook to reproduce the results for Method 1 (ResNet-50) on the bird classification problem (CUB Dataset). We used a ResNet-50 network pretrained on the iNaturalist dataset as a feature extractor, and we fine-tuned it for the 200-way bird classification problem. For doing so, we use the layer-4 output of the network with the dimension of 2048x7x7 and added an average pool layer, followed by a simple linear layer for classification. We trained the linear layer using the Adam optimizer with all intermediate weights frozen. The resulting network achieves an 85.83% accuracy on the CUB test dataset.
You can download the weights from Google Drive