armancohan

from transformers import (
    AutoTokenizer,
    LEDConfig,
    LEDForConditionalGeneration,
)

from datasets import load_dataset
import random

random.seed(2)

armancohan

```
g batch of size torch.Size([2407, 2]) because not full seq_len of 16384
----------------------------------------------------------------------------------------------------
| Eval 354 at step    88500 | time: 1345.55s | valid loss  0.74 | bpc   1.07357
----------------------------------------------------------------------------------------------------
| epoch 130 step    88510 |     16 batches | lr 0.000442 | ms/batch 11917.46 | loss  0.75 | bpc   1.07888
| epoch 130 step    88520 |     26 batches | lr 0.000442 | ms/batch 5110.18 | loss  0.78 | bpc   1.12858
| epoch 130 step    88530 |     36 batches | lr 0.000442 | ms/batch 5107.78 | loss  0.71 | bpc   1.02528
| epoch 130 step    88540 |     46 batches | lr 0.000442 | ms/batch 5109.07 | loss  0.74 | bpc   1.07031
| epoch 130 step    88550 |     56 batches | lr 0.000442 | ms/batch 5111.60 | loss  0.78 | bpc   1.12227

armancohan

import typing


def return_dict_structure(obj):
    """Return the structure of a possibly nested dictionary object."""
    new_obj = {}
    if isinstance(obj, typing.List) or isinstance(obj, typing.Tuple):
        if obj:
            return [return_dict_structure(obj[0]), '...']
        else:

armancohan

Things that are often needed and faster to find here than Google.

Create a directory and exist ok:

import pathlib
# create parent of a file
pathlib.Path('file_path').parent.mkdir(parents=True, exist_ok=True)
# create directory
pathlib.Path('directory_path').mkdir(parents=True, exist_ok=True)

armancohan

#!/usr/bin/python
import sys

def format_block(txt, maxlen=79, indent=2):
    lines = txt.replace('\n',' ')
    res = []
    line = ''
    tokens = [e for e in lines.split(' ') if e]
    i=0
    lineno=0

armancohan

# add this to ~/.ipython/profile_default/startup/start.py

import os
import sys
import re
from collections import Counter, defaultdict, namedtuple
import itertools
import json
import numpy as np
import gzip

armancohan

You need to have pv installed.

Command:

tar -cf - [Source directory] -P | pv -s $(du -sb [Source dir] | awk '{print $1}') | gzip > [Dest tar.gz file]

Example:

tar -cf - dir/ -P | pv -s $(du -sb dir/ | awk '{print $1}') | gzip > file.tar.gz

armancohan

Bash Completion on OS X With Brew

Install bash-completion with brew:

brew install bash-completion

Add to .bash_profile

if [ -f $(brew --prefix)/etc/bash_completion ]; then

armancohan

from argparse import ArgumentParser
ap = ArgumentParser()
ap.add_argument('--verbose', '-v', action='store_true', default=False)
args = ap.parse_args()

imoprt logging
log = logging.getLogger(__name__)

if args.verbose:
    log.basicConfig(format="%(levelname)s: %(message)s", level=log.DEBUG)

armancohan

import sys
import tensorflow as tf


def main(n):
  """ Calculate the n-th Fibonacci number """
  if n < 1:
    print('n should be greater than 0')
    sys.exit(1)
  elif n == 1 or n == 2: