eval_sentence_predictions is done, but when i try to run visualize_result_struct.html,

[evercherish]

eval_sentence_predictions is done. but when i try to run visualize_result_struct.html, the html couldnt display anything. no images and no candidate captions.
i'll appreciate any help .

[rakshithShetty]

Hi, The visualize_result_struct.html loads the json file which is output from eval_sentence_prediction.py... Make sure the name of the json here is correctly pointing to your result file... This should at least display the banner and details of result file.

Next thing to check is if the images are actually in the location pointed to in the result json file. It uses a relative path 'data// '

[evercherish]

thanks for your help! I've solved the problem. However, i get some new questions. TAT

i trained a model on coco dataset, but when testing the model, i found the predict sentences of each test image were almost the same. All the predict sentences are just like 'A man in blue shirt is sitting on a beach', whatever the test image is.
Were there some parameters I didn't set correctly? Such as max_epochs ,learning_rate. Or did I miss something during the training process?

Sorry for disturbing you again.

[rakshithShetty]

Hi,
It is unlikely that some training parameters can cause the learnt model to be so bad? Can you check what was the model perplexity after the traing process, even moderately good models on coco dataset tend to have perplexity < 18. If this holds training process is not the culprit

Then check if right model checkpoint is used when testing.

Can you also check if you are feeding the right image features during the test?

1. Features are extracted from the same source as used in training
2. If right feature indices are specified for each image if you are using the predict_on_images.py

[evercherish]

Thank for your reply.
The model's perplexity is 11.31.
If the model is used by predict_on_images.py to predict the coco val images, the predict captions are correct. However if it is used by eval_sentence_predictions.py to predict the same coco val images, the predict captions are incorrect, just like what I said before, they are all the same.
I tried to use the image features extracted by myself, but the results didn't change.
Please help me.

[xinghedyc]

I got the same problem when using eval_sentence_predictions.py.