search

fengyang0317 / unsupervised_captioning

Code for Unsupervised Image Captioning
MIT License
215 stars 51 forks source link

question about the training process #23

Open sakura2233565548 opened 4 years ago

sakura2233565548 commented 4 years ago

Hi Yang: I use the code that you release in the github, and I run the train code with the parameter: two GPUs, batch_size 256, lr_rate 0.0005, now training 26000 batch_size, but the found_object is so low, about 0.14. I observe your training process and your found_object is increasing so high after 14k batch_size, so can you have some answers for this symptom? I'd appreciate it if you can help solve this problem!

fengyang0317 commented 4 years ago

Are you using the MSCOCO dataset? Have you tried the data files in https://drive.google.com/drive/folders/1ol8gLj6hYgluldvdj9XFKm16TCqOr7EE?

sakura2233565548 commented 4 years ago

Are you using the MSCOCO dataset? Have you tried the data files in https://drive.google.com/drive/folders/1ol8gLj6hYgluldvdj9XFKm16TCqOr7EE?

yes,I have checked the dataset ,mscoco 2014, No problem~So I am confused with this problem.

fengyang0317 commented 4 years ago

Did you generate the training input by yourself? You may try to use the input files from google drive.

sakura2233565548 commented 4 years ago

Did you generate the training input by yourself? You may try to use the input files from google drive.

I didn't use the input files that you release on google drive~I use the code to preprocess the image dataset by myself.