Mahasweta Chakraborti

Predict subset of variables

1

Under the option 'M', is it possible to minimize loss and predict only a subset of the variables over time? Is it possible to use a masked MSE loss for...

Issues with sentencepiece

2

Run-classifier.py was running fine. Till I pretrained model and tried to use a checkpoint for fine tuning. python3 run_classifier.py --use_tpu=True --tpu=$TPU_NAME --do_train=False --do_eval=True --eval_all_ckpt=False --task_name=imdb --data_dir=/home/user/xlnet/Data2 --output_dir=gs://user/data/output --model_dir=gs://luser/data/MODEL_DIR --uncased=False --spiece_model_path=/home/user/xlnet/$MODEL_DIR/spiece.model/...

My parameters: Used the same parameters suggested for xlnet pretrain. Gradually reduced learning rate when loss kept swinging between ~1.2 and ~0.2 after 40,000 epochs

What is the significance of num_predict in terms of number of tokens to be predicted? My data mostly comprises small snippets of social media text, and a few larger comprehensions....

Domain specific corpus

1

Hi Could explain a way to incorporate domain specific corpus to train the model? My work involves identifying n-grams prevalent in medical texts, such as "sudden infant death syndrome" which...