Liang Zhang comments

Results 8 comments of


                                            Liang Zhang

Does not work on Volta (V100) class hardware

Fine tuning XLNet on STS-B with V100 16GB GPU(s) does not converge

I also tested the STS-B result in a AWS p3.8xlarge(4 V100 GPUs) using exactly the same parameter for finetuning and evaluation as https://github.com/zihangdai/xlnet#1-sts-b-sentence-pair-relevance-regression-with-gpus. The Here are the evaluation result: ```...

Fine tuning XLNet on STS-B with V100 16GB GPU(s) does not converge

Just realized what I went wrong. I used python3 instead of python2. After switching to python2, I can get the claimed result. ``` I0704 01:11:24.028214 139952961881856 run_classifier.py:786] ================================================================================ I0704 01:11:24.028338...

Fine tuning XLNet on STS-B with V100 16GB GPU(s) does not converge

@huseinzol05 Yeah, I feel it is weird as well! The only thing I changed is rather than using `source activate tensorflow_p36` to activate the python3 env, `source activate tensorflow_p27` is...

Fine tuning XLNet on STS-B with V100 16GB GPU(s) does not converge

I think I closed this issue too early. I finetuned STS-B again in python3 with tensorflow 1.13.1 and this time it converged: ``` I0704 06:09:41.269094 140499994494720 estimator.py:2039] Saving 'checkpoint_path' summary...

Fine tuning XLNet on STS-B with V100 16GB GPU(s) does not converge

Another observation is, I fine tuned SST-2 twice, each with 10k steps in python 2 with tensorflow 1.13.1 in p3.8xlarge. The first attempt failed to converge and the second attempt...

I can't download datasets in China

I could, but the download process is extremely slow!

Error while using spark-redshift jar

I use Zeppelin to do ETL to redshift and encountered the same AbstractMethodError. By configuring the spark interpreter to exclude `com.databricks:spark-avro_2.11:3.0.0` while depending on `com.databricks:spark-redshift_2.11:2.0.1`, and then to specify another...