wassname

%pylab inline
import pandas as pd
import dask.dataframe as dd

def get_unbal_df(size = 100, balance=None):
    """Get a randomly unbalanced df"""
    if balance is None:
        balance = np.random.randint(-100, 100)
    
    if balance<0:

wassname

from typing import Iterator, Collection
from fastai.data_block import CategoryListBase
from fastai.text import *


class BinaryProcessor(CategoryProcessor):
    def create_classes(self, classes):
        self.classes = classes
        if classes is not None: self.c2i = {0:0, 1:1}
    def generate_classes(self, items):

wassname

"""
Cache a torch dataset to npy files using dask

url:https://gist.github.com/wassname/f38f8774b6f97977b660d20dfa0f0036
lic:MIT
author:wassname

usage:    
    batch_size=16
    chunk_size=batch_size*4

wassname

#!/usr/bin/env python
# coding: utf-8
get_ipython().run_line_magic('pylab', 'inline')
import torch

def jaccard_distance_loss(y_true, y_pred, smooth=100):
    """
    Jaccard = (|X & Y|)/ (|X|+ |Y| - |X & Y|)
            = sum(|A*B|)/(sum(|A|)+sum(|B|)-sum(|A*B|))
    

wassname

def cache_load_utturances(ttl=360000):
    """
    Decorator for wrapping simple cache around load_utterances.

    Since some arguments are unhashable (tokenizer) or immutable (list) we need to make the key manually
    """

    def decorate(func):
        @simple_cache.wraps(func)
        def wrapper(**kwargs):

wassname

"""
@url: https://gist.github.com/wassname/f3cbdc14f379ba9ec2acfafe5c1db592
"""
import pandas as pd
import sklearn.metrics
import numpy as np

def classification_report(*args, **kwargs):
    """
    

wassname

import numpy as np 

# handle pytorch tensors etc, by using tensorboardX's method
try:
    from tensorboardX.x2num import make_np
except ImportError:
    def make_np(x):
        return np.array(x).copy().astype('float16')

class RunningStats(object):

wassname

class RunningMean(object):
    def __init__(self, sum=0, i=0):
        self.sum = sum
        self.i = i

    def __add__(self, other):
        return RunningMean(self.sum+other, self.i+1)

    def add(self, loss):

wassname

"""
Transformer in ~80 lines of code.

From Thomas Wolf's tweet https://twitter.com/Thom_Wolf/status/1129658539142766592.
"""

import torch
from torch import nn

class Transformer(nn.Module):

wassname

class JaccardDistanceLoss(torch.nn.Module):    

    def __init__(self, smooth=100, dim=1, size_average=True, reduce=True):
        """
        Jaccard = (|X & Y|)/ (|X|+ |Y| - |X & Y|)
                = sum(|A*B|)/(sum(|A|)+sum(|B|)-sum(|A*B|))

        The jaccard distance loss is usefull for unbalanced datasets. This has been
        shifted so it converges on 0 and is smoothed to avoid exploding or disapearing
        gradient.