search

jimgoo / caffe-oxford102

Caffe CNNs for the Oxford 102 flower dataset
127 stars 54 forks source link

so, you just used the training set(1020 images) to fine tuned the AlexNet and got a good result? #4

Closed huyangc closed 8 years ago

huyangc commented 8 years ago

I fine tuned the AlexNet with the training set, and the loss divergent, but when using the test set(6k images) instead, it converged. So I'm really curious about how you fine tuned the AlexNet and got a good result. : )

huyangc commented 8 years ago

However I don't know why but when I trained the net for the third time and it converged. However it is strange that the input form influence the result so much, raw image(jpeg) led to 78% precision, lmdb led to 86% precision. I can't get the 93% precision and I don't know why for now. maybe I can preprocess the input image such as flipping and so on to increase the numbers of input images to get a better precision.