Fine-Tuning on ResNet50 pre-trained on PlantClef2016 Dataset.
Using Keras high API with TensorFlow backend. Freezing first 15 layers of ResNet50 model pretrained on ImageNet dataset. Dropout with rate 0.5 added before softmax Outoput layer.
PlantClef2015 1000 species of plants represented by images of the whole plant or different parts of the plants. Train set split 80:30 for Cross-Validaiton. Each image is identified based on taxonomic ClassId. Unique species ClassId mapped to indexes in range [0,999] representing the class id for model output. Mapping of ImageId (Image name) to generated classes mapping saved as a numpy array file. I rescalled the images to (224,224,3) seperatley. Possible to specify resizing using a Keras library. Test on selected set of the offered test dataset. Only 4633 images corresponding to one of the 1000 species. Source: http://www.imageclef.org/lifeclef/2016/plant
Loss:
Sparse Categorical Cross-Enrtropy
Accuracy:
Mean accuracy rate on all predictions
Adam with Learning rate 0.0001
Train set Accuracy 0.9875 Test set Accuracy 0.586 Top-3 test Accuracy 0.7335 Top-5 test Accuracy 0.785