giukim/siamese-network

siamese-network

---

config.py

• data_dir : Dataset directory path. Must compose like {data_dir}/train, {data_dir}/val
• IMAGE_SIZE : Siamese network inference resolution (IMAGE_SIZE x IMAGE_SIZE)
• isColor : Siamese network inference channel
• min_lr, max_lr : cosine lr-scheduler min-max lr
• margin : Positive/Negative sample margin for calculate loss

---

model.py

• Network output channel = 30

---

train.py

• Training siamese network
• During training, Visualize evaluation samples inference compressed by t-SNE from 30 channel to 2 channel

---

predict.py

• Calculate mean of features from Train dataset path for pre-trained model
• Calculate features of predict source images
• Match features and save nearest object to 'pred_result/'