The code in /homework/AI_Training_1108 is not original one

[connie980149]

Hi all,

Today I found that the code in ./homework/AI_Training_1108/cp_lab2_image_caption was modified by someone and causes bugs.
If you need original codes, please unzip '' lab2 image caption.zip '' in the folder to your workspace.

Connie

[iiRobotTT]

Dear TA,

Can you please upload the "lab2 image caption.zip" to the GitHub as well? It would be convenient for us to access the codes when the server is down. Thanks.

Sunny

[connie980149]

I have upload the code to ''original branch''.
But the files in .\data is too big for GitHub, you can get dataset_coco.json from link

[bcpenggh]

Dear TA,
I noticed only "misc/init.py" and "opts.py" are different. It seems the major difference is line 51 in "opts.py." The original version disables "--finetune_cnn_after", and the modified copy fine tune cnn from the beginning. Will this significantly slow down the training? It takes ~2 hours to train each epoch in my case.

[bcpenggh]

Dear TA,
Another difference in "opts.py" is about the language level. In the original copy this feature is disabled. May I know what is this for?

[bcpenggh]

To answer my own question, disabling CNN fine tuning makes the training almost 2x faster in my case.

[connie980149]

For your questions:

1. fine tuning means during training process, the weights of CNN will train together. If we disable fine tune, it won't update the weights of CNN. This will speed up the training process by reducing calculation.
2. language level is to calculate the score of the captions your model generated, but the score is not used during training. Disable this function is for speeding up your training process. You are welcome to enable language level for evaluating your captions.