Not Able to Reproduce the results in the paper for FB15K dataset

[navdeepkjohal]

Hi
I am running your code with the settings provided in your DKRL paper. I am using pretrained TransE embeddings for FB15K dataset and providing pre-trained word vectors to the model in an effort to reproduce the results for DKRL(CNN)+TransE model. With the best efforts I was able to produce the following results:

       rank(raw)    hits@n(raw)   rank(filter)    hits@n(filter)

left: 367.469 0.413198 221.211 0.603189
right: 274.171 0.490494 181.965 0.65682
mid: 29.4202 0.482555 29.0556 0.61377

Could you please tell me if there are some special settings that one need to do to reproduce the results. Or kindly share the model parameters (final embeddings after the training) here on GitHub?

P.S. I used the following settings: 1000 iterations during learning, window_size = 2 (first layer), window_size = 2 (second layer), n_w = 100, n = 100, learning_rate = 0.001, margin = 1. Basically all the default settings in your github code except window_size of the second layer.

[chehefei]

Hi, I want to know how long does your training code run under this parameter? thank you