search

dalab / deep-ed

Source code for the EMNLP'17 paper "Deep Joint Entity Disambiguation with Local Neural Attention", https://arxiv.org/abs/1704.04920
Apache License 2.0
223 stars 50 forks source link

Entity vectors and model file #2

Open kanihal opened 7 years ago

kanihal commented 7 years ago

Hi, Could you please share the compressed version of the 'generated' folder along with trained entity vecs and model file? (No need for training csv files)

The reason I'm asking is that I don't have a good computing setup to train the entire thing, I would like to see how the model performs on my dataset.

octavian-ganea commented 7 years ago

Hi,

This cannot work as you would need to train entity vectors for all candidates in your dataset, which might be different from the candidates that appeared in the datasets that I used. Regarding to the computing setup, testing with a pre-trained model (via test_one_loaded_model.lua) will still use most of the resources used in training. But it should not use more than 10 GB of RAM if you do not load the entire training data to disk.

se4u commented 6 years ago

@octavian-ganea Just wanted to say that if will be great if you can upload the model file. Actually I personally am only interested in the weights for the f_network for the local ED model defined in the linear_layers.lua

f_network = nn.Sequential()
  :add(nn.Linear(2,opt.nn_pem_interm_size))
  :add(nn.ReLU())
  :add(nn.Linear(opt.nn_pem_interm_size,1))

Figure 3 in the paper already has a plot for the global ED model and I was interested in a similar plot for the local ED model.

octavian-ganea commented 6 years ago

That's right, figure 3 is plotting that function. Unfortunately I do not have the local model, so I would have to re-run the whole training again. Not sure I will have time before the ICML/ACL deadlines. However, I hope the instructions on how to reproduce our results are easy to follow, so feel free to try it out yourself :).