(Tim Lee) #1

Hi All,

If you want the Jupyter notebook versions of the lecture notes, I have saved a versions on my github here.

Here’s lecture 4 notes. @vikbehal, I couldn’t edit your post, so I had to post another one. @jeremy, please turn this to a wiki thread so others can edit? Thank you!


Unofficial Deep Learning Lecture 4 Notes

Today’s topics - how to use the tools for these applications.

  1. Structured Learning
  2. Deep Learning for Language
  3. Deep Learning for Recommendation




activations - A number that’s been calculated.

In lesson 2, Jeremy said, “An activation is a number saying the confidence of some level of, say, eyeballness.”. Activations are also sometimes referred to as parameters or features.

Looking at the Dog Breeds CNN model, let’s look at the different layers

Looking at Dogbreeds lets’ look at the layers:

(1)-Dropout (p = 0.25) <----
(5)-Dropout (p = 0.5)

That means we go through and pick some activations and delete some of the calculated numbers.

p = whats the probability that you drop some cells.

Note: Each mini-batch we throw again a different set of activations.

  • p = 0.00 - No drop
  • p = 0.01 - will drop 1% of your activations (train well, but not very general)
  • p = 0.99 - will drop out 99% of your activations, will kill your accuracy (super general)

If you find it overfitting, increase p.

As you start using bigger models you will probably need to increase the amount of dropout.

You may notice we have two linear layers. We’ve been adding 2 linear layers, xtra_fc=[] can handle this additional layer.

If we pass in the blank xfra_fc, we get a minimum model, which is the simplest sequential model:


ps = [0, 0.2] can set the different drop out layers for the ConvLearner parameters

1. Structured and Time Series Data

%matplotlib inline
%reload_ext autoreload
%autoreload 2
import sys
import torch
from fastai.structured import *
from fastai.column_data import *
np.set_printoptions(threshold=50, edgeitems=20)


Create datasets

We will look at the rossman kaggle competition as an example of structured data

In addition to the provided data, we will be using external datasets put together by participants in the Kaggle competition. You can download all of them here.

For completeness, the implementation used to put them together is included below.

def concat_csvs(dirname):
    path = f'{PATH}{dirname}'

    wrote_header = False
    with open(f"{path}.csv","w") as outputfile:
        for filename in filenames:
            name = filename.split(".")[0]
            with open(filename) as f:
                line = f.readline()
                if not wrote_header:
                    wrote_header = True
                for line in f:
                     outputfile.write(name + "," + line)

Feature Space:

  • train: Training set provided by competition
  • store: List of stores
  • store_states: mapping of store to the German state they are in
  • List of German state names
  • googletrend: trend of certain google keywords over time, found by users to correlate well w/ given data
  • weather: weather
  • test: testing set
table_names = ['train', 'store', 'store_states', 'state_names', 
               'googletrend', 'weather', 'test']

We’ll be using the popular data manipulation framework pandas. Among other things, pandas allows you to manipulate tables/data frames in python as one would in a database.

We’re going to go ahead and load all of our csv’s as dataframes into the list tables.

tables = [pd.read_csv(f'{PATH}{fname}.csv', low_memory=False) for fname in table_names]
from IPython.display import HTML

We can use head() to get a quick look at the contents of each table:

  • train: Contains store information on a daily basis, tracks things like sales, customers, whether that day was a holdiay, etc.
  • store: general info about the store including competition, etc.
  • store_states: maps store to state it is in
  • state_names: Maps state abbreviations to names
  • googletrend: trend data for particular week/state
  • weather: weather conditions for each state
  • test: Same as training table, w/o sales and customers
for t in tables: display(t.head())
Store DayOfWeek Date Sales Customers Open Promo StateHoliday SchoolHoliday
0 1 5 2015-07-31 5263 555 1 1 0 1
1 2 5 2015-07-31 6064 625 1 1 0 1
2 3 5 2015-07-31 8314 821 1 1 0 1
3 4 5 2015-07-31 13995 1498 1 1 0 1
4 5 5 2015-07-31 4822 559 1 1 0 1

Store StoreType Assortment CompetitionDistance CompetitionOpenSinceMonth CompetitionOpenSinceYear Promo2 Promo2SinceWeek Promo2SinceYear PromoInterval
0 1 c a 1270.0 9.0 2008.0 0 NaN NaN NaN
1 2 a a 570.0 11.0 2007.0 1 13.0 2010.0 Jan,Apr,Jul,Oct
2 3 a a 14130.0 12.0 2006.0 1 14.0 2011.0 Jan,Apr,Jul,Oct
3 4 c c 620.0 9.0 2009.0 0 NaN NaN NaN
4 5 a a 29910.0 4.0 2015.0 0 NaN NaN NaN

Store State
0 1 HE
1 2 TH
2 3 NW
3 4 BE
4 5 SN

StateName State
0 BadenWuerttemberg BW
1 Bayern BY
2 Berlin BE
3 Brandenburg BB
4 Bremen HB

file week trend
0 Rossmann_DE_SN 2012-12-02 - 2012-12-08 96
1 Rossmann_DE_SN 2012-12-09 - 2012-12-15 95
2 Rossmann_DE_SN 2012-12-16 - 2012-12-22 91
3 Rossmann_DE_SN 2012-12-23 - 2012-12-29 48
4 Rossmann_DE_SN 2012-12-30 - 2013-01-05 67

file Date Max_TemperatureC Mean_TemperatureC Min_TemperatureC Dew_PointC MeanDew_PointC Min_DewpointC Max_Humidity Mean_Humidity ... Max_VisibilityKm Mean_VisibilityKm Min_VisibilitykM Max_Wind_SpeedKm_h Mean_Wind_SpeedKm_h Max_Gust_SpeedKm_h Precipitationmm CloudCover Events WindDirDegrees
0 NordrheinWestfalen 2013-01-01 8 4 2 7 5 1 94 87 ... 31.0 12.0 4.0 39 26 58.0 5.08 6.0 Rain 215
1 NordrheinWestfalen 2013-01-02 7 4 1 5 3 2 93 85 ... 31.0 14.0 10.0 24 16 NaN 0.00 6.0 Rain 225
2 NordrheinWestfalen 2013-01-03 11 8 6 10 8 4 100 93 ... 31.0 8.0 2.0 26 21 NaN 1.02 7.0 Rain 240
3 NordrheinWestfalen 2013-01-04 9 9 8 9 9 8 100 94 ... 11.0 5.0 2.0 23 14 NaN 0.25 7.0 Rain 263
4 NordrheinWestfalen 2013-01-05 8 8 7 8 7 6 100 94 ... 10.0 6.0 3.0 16 10 NaN 0.00 7.0 Rain 268

5 rows × 24 columns

Id Store DayOfWeek Date Open Promo StateHoliday SchoolHoliday
0 1 1 4 2015-09-17 1.0 1 0 0
1 2 3 4 2015-09-17 1.0 1 0 0
2 3 7 4 2015-09-17 1.0 1 0 0
3 4 8 4 2015-09-17 1.0 1 0 0
4 5 9 4 2015-09-17 1.0 1 0 0

Data Cleaning / Feature Engineering

As a structured data problem, we necessarily have to go through all the cleaning and feature engineering, even though we’re using a neural network.

These are the consolidated Data processing blocks. For the originals, please refer to lesson 4 on deep learning.

train, store, store_states, state_names, googletrend, weather, test = tables

We turn state Holidays to booleans, to make them more convenient for modeling. We can do calculations on pandas fields using notation very similar (often identical) to numpy.

Data Processing Blocks

train.StateHoliday = train.StateHoliday!='0'
test.StateHoliday = test.StateHoliday!='0'

def join_df(left, right, left_on, right_on=None, suffix='_y'):
    if right_on is None: right_on = left_on
    return left.merge(right, how='left', left_on=left_on, right_on=right_on, 
                      suffixes=("", suffix))

weather = join_df(weather, state_names, "file", "StateName")

googletrend['Date'] = googletrend.week.str.split(' - ', expand=True)[0]
googletrend['State'] = googletrend.file.str.split('_', expand=True)[2]
googletrend.loc[googletrend.State=='NI', "State"] = 'HB,NI'

add_datepart(weather, "Date", drop=False)
add_datepart(googletrend, "Date", drop=False)
add_datepart(train, "Date", drop=False)
add_datepart(test, "Date", drop=False)
add_datepart(googletrend, "Date", drop=False)

trend_de = googletrend[googletrend.file == 'Rossmann_DE']

store = join_df(store, store_states, "Store")

joined = join_df(train, store, "Store")

joined = join_df(joined, googletrend, ["State","Year", "Week"])

joined = joined.merge(trend_de, 'left', ["Year", "Week"], suffixes=('', '_DE'))

joined = join_df(joined, weather, ["State","Date"])

joined_test = test.merge(store, how='left', left_on='Store', right_index=True)

for c in joined.columns:
    if c.endswith('_y'):
        if c in joined.columns: joined.drop(c, inplace=True, axis=1)                                                     
joined.CompetitionOpenSinceYear = joined.CompetitionOpenSinceYear.fillna(1900).astype(np.int32)
joined.CompetitionOpenSinceMonth = joined.CompetitionOpenSinceMonth.fillna(1).astype(np.int32)
joined.Promo2SinceYear = joined.Promo2SinceYear.fillna(1900).astype(np.int32)
joined.Promo2SinceWeek = joined.Promo2SinceWeek.fillna(1).astype(np.int32)

joined["CompetitionOpenSince"] = pd.to_datetime(dict(year=joined.CompetitionOpenSinceYear, 
                                                     month=joined.CompetitionOpenSinceMonth, day=15))
joined["CompetitionDaysOpen"] = joined.Date.subtract(joined.CompetitionOpenSince).dt.days

joined.loc[joined.CompetitionDaysOpen<0, "CompetitionDaysOpen"] = 0
joined.loc[joined.CompetitionOpenSinceYear<1990, "CompetitionDaysOpen"] = 0

joined["CompetitionMonthsOpen"] = joined["CompetitionDaysOpen"]//30
joined.loc[joined.CompetitionMonthsOpen>24, "CompetitionMonthsOpen"] = 24

joined["Promo2Since"] = pd.to_datetime(joined.apply(lambda x: Week(
    x.Promo2SinceYear, x.Promo2SinceWeek).monday(), axis=1).astype(pd.datetime))
joined["Promo2Days"] = joined.Date.subtract(joined["Promo2Since"]).dt.days

joined.loc[joined.Promo2Days<0, "Promo2Days"] = 0
joined.loc[joined.Promo2SinceYear<1990, "Promo2Days"] = 0

joined["Promo2Weeks"] = joined["Promo2Days"]//7
joined.loc[joined.Promo2Weeks<0, "Promo2Weeks"] = 0
joined.loc[joined.Promo2Weeks>25, "Promo2Weeks"] = 25

def get_elapsed(fld, pre):
    day1 = np.timedelta64(1, 'D')
    last_date = np.datetime64()
    last_store = 0
    res = []

    for s,v,d in zip(df.Store.values,df[fld].values, df.Date.values):
        if s != last_store:
            last_date = np.datetime64()
            last_store = s
        if v: last_date = d
        res.append(((d-last_date).astype('timedelta64[D]') / day1).astype(int))
    df[pre+fld] = res
columns = ["Date", "Store", "Promo", "StateHoliday", "SchoolHoliday"]
df = train[columns]

fld = 'SchoolHoliday'
df = df.sort_values(['Store', 'Date'])
get_elapsed(fld, 'After')
df = df.sort_values(['Store', 'Date'], ascending=[True, False])
get_elapsed(fld, 'Before')

fld = 'StateHoliday'
df = df.sort_values(['Store', 'Date'])
get_elapsed(fld, 'After')
df = df.sort_values(['Store', 'Date'], ascending=[True, False])
get_elapsed(fld, 'Before')

fld = 'Promo'
df = df.sort_values(['Store', 'Date'])
get_elapsed(fld, 'After')
df = df.sort_values(['Store', 'Date'], ascending=[True, False])
get_elapsed(fld, 'Before')

df = df.set_index("Date")

columns = ['SchoolHoliday', 'StateHoliday', 'Promo']

for o in ['Before', 'After']:
    for p in columns:
        a = o+p
        df[a] = df[a].fillna(0)
bwd = df[['Store']+columns].sort_index().groupby("Store").rolling(7, min_periods=1).sum()
fwd = df[['Store']+columns].sort_index(ascending=False
                                      ).groupby("Store").rolling(7, min_periods=1).sum()




df = df.merge(bwd, 'left', ['Date', 'Store'], suffixes=['', '_bw'])
df = df.merge(fwd, 'left', ['Date', 'Store'], suffixes=['', '_fw'])

df["Date"] = pd.to_datetime(df.Date)
Index(['Date', 'Store', 'AfterSchoolHoliday', 'BeforeSchoolHoliday',
       'AfterStateHoliday', 'BeforeStateHoliday', 'AfterPromo', 'BeforePromo',
       'SchoolHoliday_bw', 'StateHoliday_bw', 'Promo_bw', 'SchoolHoliday_fw',
       'StateHoliday_fw', 'Promo_fw'],
joined = join_df(joined, df, ['Store', 'Date'])
joined = joined[joined.Sales!=0]

Check point! Save your feather file!

joined = pd.read_feather(f'{PATH}joined')

Discussion on Categorical

Which say which are categorical / continuous. This is a decision that we are going to make. Sometimes there are variables that are directly coded as categorical, there’s only one choice. But continuous in the data, you can make a decision.

For the 3rd place team, most of the continuous ones are mainly the floating point numbers. The number of levels in a category is the cardinality. We could bin variables 10 -> 100 into 10-20, 20-30, etc.

cat_vars = ['Store', 'DayOfWeek', 'Year', 'Month', 'Day', 'StateHoliday', 'CompetitionMonthsOpen',
    'Promo2Weeks', 'StoreType', 'Assortment', 'PromoInterval', 'CompetitionOpenSinceYear', 'Promo2SinceYear',
    'State', 'Week', 'Events', 'Promo_fw', 'Promo_bw', 'StateHoliday_fw', 'StateHoliday_bw',
    'SchoolHoliday_fw', 'SchoolHoliday_bw']

contin_vars = ['CompetitionDistance', 'Max_TemperatureC', 'Mean_TemperatureC', 'Min_TemperatureC',
   'Max_Humidity', 'Mean_Humidity', 'Min_Humidity', 'Max_Wind_SpeedKm_h', 
   'Mean_Wind_SpeedKm_h', 'CloudCover', 'trend', 'trend_DE',
   'AfterStateHoliday', 'BeforeStateHoliday', 'Promo', 'SchoolHoliday']

n = len(joined); n

Now we can loop through and change them all into categorical

Turn all the continuous ones into 32bit floating point for pytorch’s sake

for v in cat_vars: joined[v] = joined[v].astype('category').cat.as_ordered()
for v in contin_vars: joined[v] = joined[v].astype('float32')
dep = 'Sales'
joined = joined[cat_vars+contin_vars+[dep, 'Date']]
idxs = get_cv_idxs(n, val_pct=150000/n)
joined_samp = joined.iloc[idxs].set_index("Date")
samp_size = len(joined_samp); samp_size

Get a smaller sample size

samp_size = n
joined_samp = joined.set_index("Date")
Store DayOfWeek Year Month Day StateHoliday CompetitionMonthsOpen Promo2Weeks StoreType Assortment ... Max_Wind_SpeedKm_h Mean_Wind_SpeedKm_h CloudCover trend trend_DE AfterStateHoliday BeforeStateHoliday Promo SchoolHoliday Sales
2015-07-31 1 5 2015 7 31 False 24 0 c a ... 24.0 11.0 1.0 85.0 83.0 57.0 -9.223372e+18 1.0 1.0 5263
2015-07-31 2 5 2015 7 31 False 24 25 a a ... 14.0 11.0 4.0 80.0 83.0 67.0 -9.223372e+18 1.0 1.0 6064

2 rows × 39 columns

Fastai Process Dataframe

  1. proc_df - pulls out the target (y) and deletes from the original.
  2. Also scales the dataframe.
  3. Also creates another object to keep track of std and mean for changing the test set.
  4. Also handles missing values, fills with median.
df, y, nas, mapper = proc_df(joined_samp, 'Sales', do_scale=True)
yl = np.log(y)


train_ratio = 0.75
# train_ratio = 0.9
train_size = int(samp_size * train_ratio); train_size
val_idx = list(range(train_size, len(df)))
val_idx = np.flatnonzero(
    (df.index<=datetime.datetime(2014,9,17)) & (df.index>=datetime.datetime(2014,8,1)))

Validation Set - Last two weeks

Make it as similar dataset as possible. This won’t be a random set like other times because this is time series.

The Deep Learning Version

def inv_y(a): return np.exp(a)

def exp_rmspe(y_pred, targ):
    targ = inv_y(targ)
    pct_var = (targ - inv_y(y_pred))/targ
    return math.sqrt((pct_var**2).mean())

max_log_y = np.max(yl)
y_range = (0, max_log_y*1.2)

Very similar to the image recognition process

  • find training dataset
  • find test set
  • set a learner
  • fit and go

But since we are not doing images, but structured data: ColumnarModelData

  • PATH - where to store everything that you save later
  • val_idx - which we will put in the validation set
  • df - dataframe
  • yl - target (dependent variable)
  • cat_flds - which things we want to treat as categorical. At this time everything is a number
md = ColumnarModelData.from_data_frame(PATH, val_idx, df, yl, cat_flds=cat_vars, bs=128)

Jumping ahead we need to set the learner …

  • How much drop out
  • How many embeddings
m = md.get_learner(emb_szs, len(df.columns)-len(cat_vars),
                   0.04, 1, [1000,500], [0.001,0.01], y_range=y_range)
lr = 1e-3

New Topic: Embeddings

The continuous variables, we are going to grab them all.

What if you had days of the week, lets assign 4 random numbers.

  • 0-Sunday
  • 1-Mon
  • 2-Tue
  • 3-Wed
  • 4-Thu
  • 5-Fri
  • 6-Sat

Where did 4 come from for Day of Week?

What’s our Cardinality for our Categorical Variables?

cat_sz = [(c, len(joined_samp[c].cat.categories)+1) for c in cat_vars]
[('Store', 1116),
 ('DayOfWeek', 8),
 ('Year', 4),
 ('Month', 13),
 ('Day', 32),
 ('StateHoliday', 3),
 ('CompetitionMonthsOpen', 26),
 ('Promo2Weeks', 27),
 ('StoreType', 5),
 ('Assortment', 4),
 ('PromoInterval', 4),
 ('CompetitionOpenSinceYear', 24),
 ('Promo2SinceYear', 9),
 ('State', 13),
 ('Week', 53),
 ('Events', 22),
 ('Promo_fw', 7),
 ('Promo_bw', 7),
 ('StateHoliday_fw', 4),
 ('StateHoliday_bw', 4),
 ('SchoolHoliday_fw', 9),
 ('SchoolHoliday_bw', 9)]

Rough Formula : Take your cardinality, divide by 2, but not more than 50

emb_szs = [(c, min(50, (c+1)//2)) for _,c in cat_sz]
[(1116, 50),
 (8, 4),
 (4, 2),
 (13, 7),
 (32, 16),
 (3, 2),
 (26, 13),
 (27, 14),
 (5, 3),
 (4, 2),
 (4, 2),
 (24, 12),
 (9, 5),
 (13, 7),
 (53, 27),
 (22, 11),
 (7, 4),
 (7, 4),
 (4, 2),
 (4, 2),
 (9, 5),
 (9, 5)]

Embedding = Distributed Representation

Jumping ahead we need to set the learner …

  • emb_szs Our embeddings
  • len(df.columns)-len(cat_vars)- how many continous variables
  • [1000,500] activations in linear layer
  • [0.001,0.01] - drop outs per linear layer
  • 1 - our output
m = md.get_learner(emb_szs, len(df.columns)-len(cat_vars),
                   0.04, 1, [1000,500], [0.001,0.01], y_range=y_range)
lr = 1e-3
m = md.get_learner(emb_szs, len(df.columns)-len(cat_vars),
                   0.04, 1, [1000,500], [0.001,0.01], y_range=y_range)
lr = 1e-3, 3, metrics=[exp_rmspe])
[ 0.       0.02479  0.02205  0.19309]                          
[ 1.       0.02044  0.01751  0.18301]                          
[ 2.       0.01598  0.01571  0.17248], 1, metrics=[exp_rmspe], cycle_len=1)
[ 0.       0.00676  0.01041  0.09711]                             

Step Overall

  1. List the Categorical
  2. List of row indexes in your dataset
  3. Call the columnar dataset
  4. How big you want your embedding matrix
  5. Get learner
  6. Fit

2. Deep Learning for Language

Language notebook

Language Modeling

Can you predict what the next word is going to be? Swiftkey actually does this. Modeling predicts the next word.

IMDb movie review dataset

#!pip install spacy

To download the ‘en’ library needed, run:

#!python -m spacy download en

The large movie view dataset contains a collection of 50,000 reviews from IMDB. The dataset contains an even number of positive and negative reviews. The authors considered only highly polarized reviews. A negative review has a score ≤ 4 out of 10, and a positive review has a score ≥ 7 out of 10. Neutral reviews are not included in the dataset. The dataset is divided into training and test sets. The training set is the same 25,000 labeled reviews.

The sentiment classification task consists of predicting the polarity (positive or negative) of a given text.

However, before we try to classify sentiment, we will simply try to create a language model; that is, a model that can predict the next word in a sentence. Why? Because our model first needs to understand the structure of English, before we can expect it to recognize positive vs negative sentiment.

So our plan of attack is the same as we used for Dogs v Cats: pretrain a model to do one thing (predict the next word), and fine tune it to do something else (classify sentiment).

Unfortunately, there are no good pretrained language models available to download, so we need to create our own.

import torch
%reload_ext autoreload
%autoreload 2
%matplotlib inline

import sys

from fastai.imports import *
from fastai.torch_imports import *
from fastai.core import *
from fastai.model import fit
from fastai.dataset import *

import torchtext
from torchtext import vocab, data
from torchtext.datasets import language_modeling

from fastai.rnn_reg import *
from fastai.rnn_train import *
from fastai.nlp import *
from fastai.lm_rnn import *

import dill as pickle

Torchtext - is Pytorch’s NLP library


TRN_PATH = 'train/all/'
VAL_PATH = 'test/all/'

%ls {PATH}
imdbEr.txt  imdb.vocab  emodelse/  README  etest/  etmp/  etrain/
trn_files = !ls {TRN}
review = !cat {TRN}{trn_files[6]}
"I have to say when a name like Zombiegeddon and an atom bomb on the front cover I was expecting a flat out chop-socky fung-ku, but what I got instead was a comedy. So, it wasn't quite was I was expecting, but I really liked it anyway! The best scene ever was the main cop dude pulling those kids over and pulling a Bad Lieutenant on them!! I was laughing my ass off. I mean, the cops were just so bad! And when I say bad, I mean The Shield Vic Macky bad. But unlike that show I was laughing when they shot people and smoked dope.<br /><br />Felissa, oh man. What can you say about that hottie. She was great and put those other actresses to shame. She should work more often!!!!! I also really liked the fight scene outside of the building. That was done really well. Lots of fighting and people getting their heads banged up. FUN! Last, but not least Joe Estevez and William Smith were great as the...well, I wasn't sure what they were, but they seemed to be having fun and throwing out lines. I mean, some of it didn't make sense with the rest of the flick, but who cares when you're laughing so hard! All in all the film wasn't the greatest thing since sliced bread, but I wasn't expecting that. It was a Troma flick so I figured it would totally suck. It's nice when something surprises you but not totally sucking.<br /><br />Rent it if you want to get stoned on a Friday night and laugh with your buddies. Don't rent it if you are an uptight weenie or want a zombie movie with lots of flesh eating.<br /><br />P.S. Uwe Boil was a nice touch."
!find {TRN} -name '*.txt' | xargs cat | wc -w
!find {VAL} -name '*.txt' | xargs cat | wc -w
' '.join(spacy_tok(review[0]))
"I have to say when a name like Zombiegeddon and an atom bomb on the front cover I was expecting a flat out chop - socky fung - ku , but what I got instead was a comedy . So , it was n't quite was I was expecting , but I really liked it anyway ! The best scene ever was the main cop dude pulling those kids over and pulling a Bad Lieutenant on them ! ! I was laughing my ass off . I mean , the cops were just so bad ! And when I say bad , I mean The Shield Vic Macky bad . But unlike that show I was laughing when they shot people and smoked dope . \n\n Felissa Rose ... man , oh man . What can you say about that hottie . She was great and put those other actresses to shame . She should work more often ! ! ! ! ! I also really liked the fight scene outside of the building . That was done really well . Lots of fighting and people getting their heads banged up . FUN ! Last , but not least Joe Estevez and William Smith were great as the ... well , I was n't sure what they were , but they seemed to be having fun and throwing out lines . I mean , some of it did n't make sense with the rest of the flick , but who cares when you 're laughing so hard ! All in all the film was n't the greatest thing since sliced bread , but I was n't expecting that . It was a Troma flick so I figured it would totally suck . It 's nice when something surprises you but not totally sucking . \n\n Rent it if you want to get stoned on a Friday night and laugh with your buddies . Do n't rent it if you are an uptight weenie or want a zombie movie with lots of flesh eating . \n\n P.S. Uwe Boil was a nice touch ."

We use Pytorch’s torchtext library to pre-process our data, telling it to use the wonderful spacy library to handle tokenization.

First, we create a torchtext field, which describes how to pre-process a piece of text - in this case, we tell torchtext to make everything lowercase, and tokenize it with spacy.

we will lower, then we will spacy tokenize.

TEXT = data.Field(lower=True, tokenize=spacy_tok)

fastai library works closely with torchtext. We create a ModelData object for language modeling by taking advantage of LanguageModelData, passing it our torchtext field object, and the paths to our training, test, and validation sets. In this case, we don’t have a separate test set, so we’ll just use VAL_PATH for that too.

As well as the usual bs (batch size) parameter, we also not have bptt; this define how many words are processing at a time in each row of the mini-batch. More importantly, it defines how many ‘layers’ we will backprop through. Making this number higher will increase time and memory requirements, but will improve the model’s ability to handle long sentences.

  • min_freq - must occur with 10 times
  • bs - batch size
  • bptt - back propogation - how much of the sentence will sit at any time
bs=64; bptt=70

WARNING: takes a long time to run

FILES = dict(train=TRN_PATH, validation=VAL_PATH, test=VAL_PATH)
md = LanguageModelData(PATH, TEXT, **FILES, bs=bs, bptt=bptt, min_freq=10)

After building our ModelData object, it automatically fills the TEXT object with a very important attribute: TEXT.vocab. This is a vocabulary, which stores which words (or tokens) have been seen in the text, and how each word will be mapped to a unique integer id. We’ll need to use this information again later, so we save it.

(Technical note: python’s standard Pickle library can’t handle this correctly, so at the top of this notebook we used the dill library instead and imported it as pickle).

Save the model for later

pickle.dump(TEXT, open(f'{PATH}models/TEXT.pkl','wb'))

Here are the: # batches; # unique tokens in the vocab; # tokens in the training set; # sentences

len(md.trn_dl), md.nt, len(md.trn_ds), len(md.trn_ds[0].text)
(4603, 34933, 1, 20626672)

This is the start of the mapping from integer IDs to unique tokens.

# 'itos': 'int-to-string'
['<unk>', '<pad>', 'the', ',', '.', 'and', 'a', 'of', 'to', 'is', 'it', 'in']
# 'stoi': 'string to int'

Note that in a LanguageModelData object there is only one item in each dataset: all the words of the text joined together.

Variable containing:
[torch.cuda.LongTensor of size 12x1 (GPU 0)]

Our LanguageModelData object will create batches with 64 columns (that’s our batch size), and varying sequence lengths of around 80 tokens (that’s our bptt parameter - backprop through time).

Each batch also contains the exact same data as labels, but one word later in the text - since we’re trying to always predict the next word. The labels are flattened into a 1d array.

(Variable containing:
     37    409     22  ...      50     56      4
    866    232      6  ...     815     51     34
      3    584    655  ...       4     59    325
         ...            ⋱           ...         
     44     60   2332  ...      14    841   9681
   2768     10      9  ...      75      3    360
   1032    149    899  ...   12254      6     14
 [torch.cuda.LongTensor of size 81x64 (GPU 0)], Variable containing:
 [torch.cuda.LongTensor of size 5184 (GPU 0)])

Let’s make our own Embedding off of our data

  • md.nt = number of tokens = 34,495
  • md.trn_ds[0].text = 20,621,966


We have a number of parameters to set - we’ll learn more about these later, but you should find these values suitable for many problems.

em_sz = 200  # size of each embedding vector
nh = 500     # number of hidden activations per layer
nl = 3       # number of layers

Researchers have found that large amounts of momentum (which we’ll learn about later) don’t work well with these kinds of RNN models, so we create a version of the Adam optimizer with less momentum than it’s default of 0.9.

Adam Optimizer, the defaults don’t work very well

opt_fn = partial(optim.Adam, betas=(0.7, 0.99))

fastai library uses a variant of the state of the art AWD LSTM Language Model developed by Stephen Merity. A key feature of this model is that it provides excellent regularization through Dropout. There is no simple way known (yet!) to find the best values of the dropout parameters below - you just have to experiment…

However, the other parameters (alpha, beta, and clip) shouldn’t generally need tuning.

  • opt_fn - defaults for adam optimizer
  • em_sz - embedding size
  • nh - number of activations
  • nl - number of layers
  • dropout - how much drop out to add (won’t elaborate)
  • learner.clip=0.3 - don’t let the gradient go any lower than this number

This is a RNN using an LSTM

learner = md.get_model(opt_fn, em_sz, nh, nl,
               dropouti=0.05, dropout=0.05, wdrop=0.1, dropoute=0.02, dropouth=0.05)
learner.reg_fn = partial(seq2seq_reg, alpha=2, beta=1)

As you can see below, I gradually tuned the language model in a few stages. I possibly could have trained it further (it wasn’t yet overfitting), but I didn’t have time to experiment more. Maybe you can see if you can train it to a better accuracy! (I used lr_find to find a good learning rate, but didn’t save the output in this notebook. Feel free to try running it yourself now.)

Trying different settings and modeling, 1, wds=1e-6, cycle_len=20, cycle_save_name='adam3_20')
[ 0.       6.44582  6.40362]                                    
[ 1.       4.71655  4.61361]                                  
[ 2.       4.59772  4.49624]                                  
[ 3.       4.53644  4.44157]                                  
[ 4.       4.5055   4.40163]                                  
[ 5.       4.47234  4.37386]                                  
[ 6.       4.44514  4.35307]                                  
[ 7.       4.42154  4.33165]                                  
[ 8.       4.39408  4.31588]                                  
[ 9.       4.38333  4.30152]                                  
[ 10.        4.35194   4.28591]                               
[ 11.        4.35191   4.27355]                               
[ 12.        4.32064   4.25818]                               
[ 13.        4.29965   4.2474 ]                               
[ 14.        4.28177   4.23785]                               
[ 15.        4.28475   4.23243]                               
[ 16.        4.25953   4.22698]                               
[ 17.        4.23739   4.22239]                               
[ 18.        4.24809   4.22182]                               
[ 19.        4.25874   4.22156]                               
learner.load_cycle('adam3_20', 0)


We’ll need to the saved vocab from the language model, since we need to ensure the same words map to the same IDs.

TEXT = pickle.load(open(f'{PATH}models/TEXT.pkl','rb'))

sequential=False tells torchtext that a text field should be tokenized (in this case, we just want to store the ‘positive’ or ‘negative’ single label).

splits is a torchtext method that creates train, test, and validation sets. The IMDB dataset is built into torchtext, so we can take advantage of that. Take a look at lang_model-arxiv.ipynb to see how to define your own fastai/torchtext datasets.

IMDB_LABEL = data.Field(sequential=False)
splits = torchtext.datasets.IMDB.splits(TEXT, IMDB_LABEL, 'data/')
t = splits[0].examples[0]
t.label, ' '.join(t.text[:16])
 'i remember watching this movie with my friends when we were 4 years old , but')

fastai library can create a ModelData object directly from torchtext splits.

md2 = TextData.from_splits(PATH, splits, bs)
m3 = md2.get_model(opt_fn, 1500, bptt, emb_sz=em_sz, n_hid=nh, n_layers=nl, 
           dropout=0.1, dropouti=0.4, wdrop=0.5, dropoute=0.05, dropouth=0.3)
m3.reg_fn = partial(seq2seq_reg, alpha=2, beta=1)

Because we’re fine-tuning a pretrained model, we’ll use differential learning rates, and also increase the max gradient for clipping, to allow the SGDR to work better.


3. Deep Learning for Recommendation

Next week, we will be learning Collaborative Filtering and making this.

Wiki: Lesson 4
Deep Learning Brasília - Lição 4
Deep Learning Brasília - Revisão (lições 1, 2, 3 e 4)
(Jeremy Howard) #2

Now wikified!

(WG) #3

It feels like Christmas when these posts arrive @timlee !

(Vikram Kalabi) #4

This is wickedly amazing! You rock @timlee.

(Aditya) #5

Thank you…

(Ken) #6

@timlee, these notes are so amazing. Thank you for sharing!

(Allie Crevier) #7

Nice notes! I’m still working my way through them, but I wanted to mention that I think Jeremy mentioned that weights, not activations, are sometimes called parameters. I also wonder if it’s not better to think of weights as features, or as representing features we want to find, rather than activations as features.

(Surya K) #8

@timlee great notes! in the image for embeddings of categorical features, the embeddings are concatenated or are they summed? each categorical feature can have variable length embedding.

(Jeremy Howard) #9

Yes that’s correct, and it’s an important point. Since the notes post is a wiki, perhaps you could edit it to make that correction?

Generally we create features (with feature engineering) by doing computations to the variables. It’s the output of these computations that are the features, not the parameters to them. Therefore, it does seem appropriate to describe the activations as ‘features’, at least in some contexts.

(Allie Crevier) #10

+1 sounds good. made a pr on the notes.


Hi Jeremy - I have couple of questions - one is usage of 1d convolution instead of RNN - is it possible/advisable to make a language model using 1d convolution instead of RNN; and another one is about application of the language model.How useful or sane do you think using the language model in order kinds of data types - for example in genome sequencing - where the output would be malignant / benign tumor and the input might be short sequence of genomes considered to be as words.

(Jeremy Howard) #12

The current SoTA for language modeling is an RNN, but CNNs have shown good results in similar areas.

I think a language model for genomics would be fascinating…

(Rishav Kumar) #13

@jeremy Here

df["Promo2Since"] = pd.to_datetime(df.apply(lambda x: Week(
    x.Promo2SinceYear, x.Promo2SinceWeek).monday(), axis=1).astype(pd.datetime))

Is the Week method provided by pandas? All I understand here is we are converting given year and week number to a date. Also what does monday() do here?
I didn’t find any relevant information about it on google search.

(Jeremy Howard) #14

It’s coming from this package:

(Phong) #15

Hi, I just have a question. This language model is using the movie dataset from the IMDB where the review is negative or positive. But for other sentiments like whether a person like football or not, how could we get the training dataset for such that problem ?

(joost) #16

tells whether the datatype represents sequential data. If False, no tokenization is applied. Default: True.