Fine-tuning the inception v3 model

[StephaneZFR]

Hi,

First, thanks for all the available code for im2txt. I could make it work and generate some first captions. But now, I’m completely lost on a problem about fine-tuning im2txt, hoping it’s the good place to ask.

To summarize, I first fine-tuned an inception model using the method described in (https://github.com/tensorflow/models/tree/master/slim#Tuning). My training generates a bunch of files like model.meta, model.index and model.ckpt-xxx. I think the training is correct as I can evaluate it using eval_image_classifier.py available in slim git repo.

Then , what I would like to do is use this fine-tuned inception model as an input to train im2txt as in https://github.com/tensorflow/models/tree/master/im2txt#training-a-model.

1. But as training im2txt requires something like inception_v3.ckpt, what is the correct file to put as input?

I also tried fine-tuning the inception model in im2txt as in (https://github.com/tensorflow/models/tree/master/im2txt#fine-tune-the-inception-v3-model). But I don’t understand what it concretely does.

2. Does it fine-tune the CNN part of im2txt to adapt it to new classes? If so, how can I give it new data (e.g. the flowers dataset)?

Thanks in advance for your help. Regards, Stephane

[shengsr]

Hi Stephane,

I have the same questions, do you know the answers now?

Regards, Shurong