The loss on COCO dataset of training stage1 dosen't decrease

[BruceW91]

Hi, I run the train_coco_word2_1_pool.m, but after more than 10 epochs, I found the train result is still bad(as below). I didn't change any hyper parameters but I don't know why it doesn't work. Can you tell me what make this result happen? And I found the learning rate in your code is 0.1, but it's reported as 0.001 in the paper. Which lr is correct and better in this task?

[layumi]

The learning rate = global learning rate * layer learning rate. For COCO, it may converge slowly. It may take about 20 epochs to start converging.

[layumi]

I also notice there is error in your code? @habbakuk1103 Do you prepare the dataset in the right way?

[BruceW91]

Thank you for response. I think I prepare the COCO dataset as you told, but the dimension of my 'coco_dictionary.mat' is 29141, but it's 29972 in your code. I don't know whether it's caused by I prepare the dataset in the matlab r2016b on Windows system of my computer because I can't install this version on our linux server. Could you give me a link to download the data after preprocessed as below if possible? By the way, I am not familiar with matconvnet. Could you tell me what the 'error' in the result mean? I check the code before and I think the data and label should be aligned.

[layumi]

I wonder whether you provided the right file path.

[BruceW91]

The path of the 'prepare_imdb.m' is as below in matlab on my server, but its version is matlab R2014b which doesn't have 'jsondecode' function So I implement the program on my computer, thus the path is changed as below I wonder whether this difference of paths cause the problem.

[BruceW91]

I have solved the problem. Thank you for your suggestion.