kyunghyuncho

kyunghyuncho

kyunghyuncho

Jason Eisner

Jason Eisner is a Professor of Computer Science at Johns Hopkins University (JHU). He is recognized for his significant research contributions to natural language processing (NLP), programming languages, and artificial intelligence. His work is characterized by its theoretical depth and practical application, aiming to bridge the gap between current language models and robust reasoning capabilities.

Academic Affiliations

Eisner is a prominent faculty member in the Department of Computer Science at Johns Hopkins University. His affiliations extend across several key research centers and departments within JHU, reflecting the interdisciplinary nature of his work:

• Department of Computer Science, Johns Hopkins University
• Center for Language and Speech Processing (CLSP), Johns Hopkins University

kyunghyuncho

# download `https://github.com/dwyl/english-words/raw/refs/heads/master/words_alpha.txt`
# and save it as `words_alpha.txt` in the same directory as this notebook
import os
import requests

if not os.path.exists('words_alpha.txt'):
    url = 'https://github.com/dwyl/english-words/raw/refs/heads/master/words_alpha.txt'

    r = requests.get(url, allow_redirects=True)
    open('words_alpha.txt', 'wb').write(r.content)

kyunghyuncho

n_months = 4 # let's say 4 months of continuous training
sec_per_grad = 1.7 # 1.7 seconds per gradient computation (a very rough number)

instance_price_per_hour = 24.01 # assuming AWS p4de instances with 1yr reservation
discount = 0.2 # assuming you can get some discount from AWS or whatever your cloud vendor is

n_gpu = 8 # 8x A100's with 80G each
n_instances = 4 # the number of AWS p4de instances
ex_per_gpu = 2 # the number of training examples processed per gpu
n_tokens_per_ex = 2000 # the length of each training example

kyunghyuncho

import numpy

'''
(1./n_classes) * (1.-label_smoothing) + 1. * label_smoothing = target_prob
1./n_classes - label_smoothing * (1./n_classes - 1.) = target_prob
label_smoothing * (1. - 1./n_classes) =  target_prob - 1./n_classes
label_smoothing * (n_classes - 1.) / n_classes = target_prob - 1./n_classes
label_smoothing = (target_prob * n_classes - 1.) /n_classes * n_classes / (n_classes - 1.)
label_smoothing = (target_prob * n_classes - 1.) / (n_classes - 1.)
'''

kyunghyuncho

import numpy

s = numpy.random.randn(10)

def clip_probability_logits(logit, epsilon=0.):
    s_max = numpy.max(logit)
    s_transf = logit - s_max

    normalization_const = numpy.exp(s_transf).sum()
    p_transf = numpy.exp(s_transf) / normalization_const

kyunghyuncho

import numpy

import os

os.environ["CUDA_VISIBLE_DEVICES"]="0"
import torch

os.environ["CUDA_VISIBLE_DEVICES"]="1"
import tensorflow as tf

kyunghyuncho

import copy
import re
import numpy

def align_sents(more_sentences, fewer_sentences):
    j_more = 0

    fewer2more = [None] * len(fewer_sentences)
    more_sentences = more_sentences

kyunghyuncho

import sys
import argparse
import logging

from collections import OrderedDict

import numpy

import torch
from torch import nn