mohanpedala

Table of Contents

• set -e, -u, -x, -o pipefail
• set -e
• set -x
• set -u
• set -o pipefail
• Setting IFS
• Original Reference

set -e, -u, -x, -o pipefail

wassname

from torch.nn.modules.utils import _pair

class CausalConv2d(nn.Conv2d):

    def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=None, dilation=1, groups=1, bias=True):
        kernel_size = _pair(kernel_size)
        stride = _pair(stride)
        dilation = _pair(dilation)
        if padding is None:
            padding = [int((kernel_size[i] -1) * dilation[i]) for i in range(len(kernel_size))]

wassname

def isfinite(x):
    """
    Quick pytorch test that there are no nan's or infs.
    
    note: torch now has torch.isnan
    url: https://gist.github.com/wassname/df8bc03e60f81ff081e1895aabe1f519
    """
    not_inf = ((x + 1) != x)
    not_nan = (x == x)
    return not_inf & not_nan

wassname

"""
Pytorch sampler that samples ordered indices from unordered sequences. 

Good for use with dask and RNN's, because 
1. Dask will slow down if sampling between chunks, so we must do one chunk at a time
2. RNN's need sequences so we must have seqences e.g. 1,2,3
3. But RNN's train better with batches that are uncorrelated so we want each batch to be sequence from a different part of a chunk.

For example, given each chunk is `range(12)`. Our seq_len is 3. We might end up with these indices:
- [[1,2,3],[9,10,11],[4,5,6]]

wassname

import torch.utils.data


class NumpyDataset(torch.utils.data.Dataset):
    """Dataset wrapping arrays.

    Each sample will be retrieved by indexing array along the first dimension.

    Arguments:
        *arrays (numpy.array): arrays that have the same size of the first dimension.

wassname

from dask.callbacks import Callback
from tqdm.auto import tqdm

class TQDMDaskProgressBar(Callback, object):
    """
    A tqdm progress bar for dask.
    Usage:
        ```
        with TQDMDaskProgressBar():

wassname

"""
In jupyter notebook simple logging to console
"""
import logging
import sys

logging.basicConfig(stream=sys.stdout, level=logging.INFO)

# Test
logger = logging.getLogger('LOGGER_NAME')

wassname

class AdamStepLR(torch.optim.Adam):
    """Combine Adam and lr_scheduler.StepLR so we can use it as a normal optimiser"""
    def __init__(self, params, lr=0.001, betas=(0.9, 0.999), eps=1e-08, weight_decay=0, step_size=50000, gamma=0.5):
        super().__init__(params, lr, betas, eps, weight_decay)
        self.scheduler = torch.optim.lr_scheduler.StepLR(self, step_size, gamma)
        
    def step(self):
        self.scheduler.step()
        return super().step()

wassname

"""
A noisy convolution 2d for pytorch
Adapted from:
- https://raw.githubusercontent.com/Scitator/Run-Skeleton-Run/master/common/modules/NoisyLinear.py
- https://github.com/pytorch/pytorch/pull/2103/files#diff-531f4c06f42260d699f43dabdf741b6d
More details can be found in the paper `Noisy Networks for Exploration`
Original: https://gist.github.com/wassname/001aff274c7c8196055fabfc06cf80c5
"""
import math

tmcw

Book Id,Title,Author,Author l-f,Additional Authors,ISBN,ISBN13,My Rating,Average Rating,Publisher,Binding,Number of Pages,Year Published,Original Publication Year,Date Read,Date Added,Bookshelves,Bookshelves with positions,Exclusive Shelf,My Review,Spoiler,Private Notes,Read Count,Recommended For,Recommended By,Owned Copies,Original Purchase Date,Original Purchase Location,Condition,Condition Description,BCID
13278990,The Housing Monster,prole.info,"prole.info, prole.info",,"=""160486530X""","=""9781604865301""",0,3.77,PM Press,Paperback,160,2012,2011,,2017/12/07,currently-reading,currently-reading (#3),currently-reading,,,,1,,,0,,,,,
7805,Pale Fire,Vladimir Nabokov,"Nabokov, Vladimir",,"=""0141185260""","=""9780141185262""",0,4.19,Penguin Books Ltd,Paperback,246,2000,1962,,2013/10/09,to-read,to-read (#26),to-read,,,,0,,,0,,,,,
34220725,Never Use Futura,Doug Thomas,"Thomas, Doug",Ellen Lupton,"=""1616895721""","=""9781616895723""",4,4.29,Princeton Architectural Press,Paperback,208,2017,,,2017/11/12,,,read,"Pr