Open Bfzanchetta opened 5 years ago
which example were you running ?
I'm running Sentiment Analysis example on BigDL. I'm using 2 head nodes and 8 workers. Each node contains 16 vCPUs and 120GB of RAM. They are all designed into a Spark2 cluster with Hadoop and YARN.
Interesting points: I was having the same issue on the example Text Classifier for LSTM, where after 15 or 20 epochs it kept ending with <5% accuracy rates. However, I missed a passage on BigDL's wiki that said that training distributed LSTM networks demands higher number of epochs to achieve the same accuracy than a regular CNN model at epoch 15.
I will test this in Sentiment Analysis.
For sentiment analysis, you may take a look at https://github.com/intel-analytics/analytics-zoo/tree/master/apps/sentiment-analysis
Hello @Bfzanchetta, I'm now trying to build a textclassifier model as well. But I met RPC lost while training, but I only used a 9G training set, with a 300G*4 cluster.
Could you show me your "build model" and "optimizer" code so I can figure out weather it's spark config problem or my app problem?
Thanks!
I'm attempting to train the given Text Classifier with LSTM instead of CNN on a 8-workers BigDL's cluster. However, the training unveils a very low accuracy rate. Here's the print from the last attempt:
[EDIT] I posted the thread's model as Text Classifier, however I'm mentioning Sentiment Analysis. Has anyone ever had this low accuracy on LSTM for Text Classifier application? Thanks!