acnn
acnn copied to clipboard
FrankWork
question for version
I noticed that in the log file, the version "baseline+attentive pooling" can get the result: 05-10 21:12 Epoch: 21 Train: 94.81% Test: 75.19%. What are the model configurations in details for this result? If possible, could you send me the model file for this result? My email is [email protected]. I have tried my best but cannot reach this performance. Thank you so much!
@jind11 the training is very fast even on cpu, so you can train you own model:
$ git pull [email protected]:FrankWork/acnn.git
$ git pull origin baseap
$ git checkout baseap
$ python main.py
I rerun the program, and got the following results:
12-08 14:22 Epoch: 16 Train: 74.76% Test: 73.93%
12-08 14:22 Epoch: 17 Train: 74.95% Test: 74.59%
12-08 14:22 Epoch: 18 Train: 75.59% Test: 74.56%
12-08 14:23 Epoch: 19 Train: 76.80% Test: 74.22%
12-08 14:23 Epoch: 20 Train: 77.56% Test: 73.93%
12-08 14:23 Epoch: 21 Train: 78.33% Test: 74.78%
12-08 14:23 Epoch: 22 Train: 78.45% Test: 75.04%
12-08 14:24 Epoch: 23 Train: 79.99% Test: 75.30%
12-08 14:24 Epoch: 24 Train: 80.29% Test: 75.44%
12-08 14:24 Epoch: 25 Train: 79.96% Test: 75.15%
I have one question: where do you get the pre-trained embedding file? Is it trained on English Wikipedia? Thanks!
@jind11 trained on Wikipedia
https://github.com/Hironsan/awesome-embedding-models
https://ronan.collobert.com/senna/senna-v3.0.tgz
I see. According to my experiment on different embedding sources, the pre-trained embeddings has influence on the performance. I am gonna train word2vec 300 and 400 dimension on English Wikipedia myself.