YOLOv2 Implementation in TF/Keras. Allowing to experiment on different feature detectors (MobileNet, Darknet-19). Paper: https://arxiv.org/abs/1612.08242
my training datasets contains 12300*8 data ,and my calsss is 3, my learning rate is 0.0001. I train it with yolo-coco.weights as initial weights,and just after two epoch ,the val_loss comes to impove while loss still decresce.,I think it comes out overfitting .
Did you meet this problem,and maybe add dropout into net would work?
Below is my training details
12299/12300 [============================>.] - ETA: 0s - loss: 1.6080Epoch 00001: val_loss improved from inf to 1.55635, saving model to ./backup/best_yolov2-01-1.56.weights
12300/12300 [==============================] - 2773s 225ms/step - loss: 1.6080 - val_loss: 1.5564
12299/12300 [============================>.] - ETA: 0s - loss: 1.5306Epoch 00002: val_loss improved from 1.55635 to 1.52850, saving model to ./backup/best_yolov2-02-1.53.weights
12300/12300 [==============================] - 2738s 223ms/step - loss: 1.5307 - val_loss: 1.5285
12299/12300 [============================>.] - ETA: 0s - loss: 1.4806Epoch 00003: val_loss did not improve
12300/12300 [==============================] - 2741s 223ms/step - loss: 1.4806 - val_loss: 1.5510
12299/12300 [============================>.] - ETA: 0s - loss: 1.4509Epoch 00004: val_loss did not improve
12300/12300 [==============================] - 2741s 223ms/step - loss: 1.4509 - val_loss: 1.5861
my training datasets contains 12300*8 data ,and my calsss is 3, my learning rate is 0.0001. I train it with yolo-coco.weights as initial weights,and just after two epoch ,the val_loss comes to impove while loss still decresce.,I think it comes out overfitting . Did you meet this problem,and maybe add dropout into net would work?
Below is my training details
12299/12300 [============================>.] - ETA: 0s - loss: 1.6080Epoch 00001: val_loss improved from inf to 1.55635, saving model to ./backup/best_yolov2-01-1.56.weights 12300/12300 [==============================] - 2773s 225ms/step - loss: 1.6080 - val_loss: 1.5564 12299/12300 [============================>.] - ETA: 0s - loss: 1.5306Epoch 00002: val_loss improved from 1.55635 to 1.52850, saving model to ./backup/best_yolov2-02-1.53.weights 12300/12300 [==============================] - 2738s 223ms/step - loss: 1.5307 - val_loss: 1.5285 12299/12300 [============================>.] - ETA: 0s - loss: 1.4806Epoch 00003: val_loss did not improve 12300/12300 [==============================] - 2741s 223ms/step - loss: 1.4806 - val_loss: 1.5510 12299/12300 [============================>.] - ETA: 0s - loss: 1.4509Epoch 00004: val_loss did not improve 12300/12300 [==============================] - 2741s 223ms/step - loss: 1.4509 - val_loss: 1.5861