search

boathit / research-papers

9 stars 2 forks source link

best_model.pt #2

Open KeyBoarder8 opened 6 years ago

KeyBoarder8 commented 6 years ago

Because of my poor device or version of pytorch,it shows "CUDA error: out of memory" .Could you please share me the flie “best_model.pt” you get?Google cloud link or other may be OK.Thanks so much!

boathit commented 6 years ago

The minimal GPU memory is 6GB. As the work has been done one year ago, I have cleaned up all the intermediate files including best_model.pt in the server. I will manage to train a model for you in my machine but it takes time.

KeyBoarder8 commented 6 years ago

@boathit My GPU memory is just 4GB......Sorry to take your time and I will look forward to your reply.

boathit commented 6 years ago

@KeyBoarder8, our trained model is available here. It was tested with PyTorch 0.1.2 and cuda8.0.

KeyBoarder8 commented 6 years ago

@boathit Thank you for your kindness!

KeyBoarder8 commented 6 years ago

@boathit Excuse me,where can I find the skip gram model in your code

boathit commented 6 years ago

The main benefit of pre-training is accelerating the subsequent training process. But the training speed actually is not a big problem if you have a strong GPU so I remove it in this repository. If you want to add it, just call the load_pretrained_embedding function at here by referring the official tutorial.

KeyBoarder8 commented 6 years ago

@boathit OK,Thanks!

KeyBoarder8 commented 6 years ago

@boathit Sorry to bother you again.If I currently have a small number of trajectory data and the region is different from the trajectory data set you use, do I need to retrain the neural network model?

boathit commented 6 years ago

Yes, the testing data is supposed to own the same vocabulary as the training data.