Using Pretrained InceptionV3 network to predict employment labels of villages.
Model:
We used keras InceptionV3 network which is pretrained on imagenet dataset. We removed the top layer of network, added a fc layer with 1024 nodes and an output layer with 3 nodes. The input image size is 2242243. We used only R,G,B bands in LANDSAT data. The dataset is then split into training and test data with 80% as training data and 20% as test data.
Link to trained model:
https://drive.google.com/open?id=1tw7o5LMoPhBvfe6ztPinupxsLIy4kVUP
Training: In training, only the weight of FC layer is learned keeping weight of the remaining architecture fixed. The network is trained for 2 epochs.
to train network:
nohup python inceptionnet_pretrained.py > Results/pretrained_out1.txt 2>&1 &
To plot training loss:
python plot_loss.py
To plot confusion matrix:
python plot_confusion_matrix.py
Results:
Unnormalized confusion matrix:
Normalized confusion matrix:
Test accuracy is 47.11%