Monday, June 25, 2018

Keras: Convolutional LSTM

Stacking recurrent layers on top of convolutional layers can be used to generate sequential output (like text) from structured input (like images or audio) [1].

If we want to stack an LSTM on top of a convolutional layers, we can simply do so, but we need to
reshape the output of the convolutions to the LSTM's expected input. The code below shows an implementation in Keras:
 
T = 128
D = 64

N_FILTERS  = 32
LSTM_T     = int(T / 8)
LSTM_D     = int(D / 2)
LSTM_STATE = 128
POOL       = (8, 2)

i = Input(shape=(T, D, 1))                         # (None, 128, 64, 1)   
x = Conv2D(N_FILTERS, (3, 3), padding = 'same')(i) # (None, 128, 64, 32)
x = MaxPooling2D(pool_size=POOL)(x)                # (None, 16, 32, 32) 
x = Reshape((LSTM_T, LSTM_D * N_FILTERS))(x)       # (None, 16, 1024) 
x = LSTM(LSTM_STATE, return_sequences=True)(x)     # (None, 16, 128)     


model = Model(i, x)
model.summary()


In this example we want to learn the convolutional LSTM on sequences of length 128 with 64 dimensional samples. The first layer is a convolutional layer with 32 filters. So the outputs are 32
sequences, one for each filter. We pool the sequences with a (8, 2) window. So our 32 sequences are now of size (128 / 8 = 16,  64 / 2 = 32). Now we have to combine the dimensions and the filter responses into a single dimension of size (32 * 32 = 1024) so we can feed a sequence into the LSTM which requires a rank 2 ( or 3 with batch) tensor with the first dimension being the time step and the second each frame. Finally we add the LSTM layer.

Wednesday, June 13, 2018

Character Wise Generation From The Little Book Of Calm

If you saw the show black books, you saw the calming powers of the little book of calm.
Surprisingly the book actually exists. I extracted the text from the Kindle version and will now train an LSTM with it in order to generate an infinity stream of calming advice. The actual code can be found as a Gist.
For copyright reasons I can not attach the text. So if you want to train this too, you need to buy the Kindle version.
I split all the text into windows of 40 characters and train an LSTM to predict the following character. I use a bidirectional LSTM with a 128 dimensional cell-state. The output mode for the LSTM is many-2-many, which means each input at each time step (the current character) also has an output (the next character). The output is pushed through a dense softmax layer. The LSTM is shown below:
 
length = 40
step   = 1
model = Sequential()
model.add(Bidirectional(LSTM(128, input_shape=(length, n_char), return_sequences=True), input_shape=(length, n_char)))
model.add(TimeDistributed(Dense(n_char, activation='softmax')))
optimizer = RMSprop(lr=0.01)
model.compile(loss='categorical_crossentropy', optimizer=optimizer)
model.summary()

The Training log is interesting. One can see that the LSTM gets better and better at generation text which can be seen in the notebook. Below I listed some samples. For the sampling, we initialise the generation with a random sequence of 40 characters and let the LSTM generate the rest. The details of the sampling can be seen in the notebook. The text is not great but one can see that the LSTM aims to generate sentences similar to the ones in the book. One might be able to improve it's performance with more layers or more epochs.

However, since the model works on a character level, I am not unhappy with the results:


  1. "thes the upset child, just as a kiss or handshake takes the pleasures eale them to poins of almous to be calming end it carting dive you’ll find calm phosting work is acyies all little forgsimutarly aboveruts a phopees will tell you, nothed (which your speech trangess there are tho defullame and eyebrows phon places of person, joyous every meal with the fores or nacciols, and discontrom to the hear postrous into them most powerm, when they gretophsints to calm. sho exercises when they laverds with the concenturies an much you think as"
  2. "eel calm. recognise addictions for what your has a little choice. devers of a favour things which the worldor tak every moments worry habirothing – you half the way to feel. emplay of the lights – quiet – entoy not worrierate – beting worry king before you think about still, bow feelinggs there is enliet before yourself calm if you feel calm your earl goonicoly disary taskn take the time is efly. kink well assential astimugars asout peace – and have all tall you to can calm. the way you look at that your happenng, puts of orange blome"
  3. "e stand straighter and taller than you beliefs, you will fool your hand, mixitates even the back sede tans bighing take and the ennable treates the beauty the add to think worty as major troutten if you combscents are among the humone situations: you treat at attating them. trank und harty thing an ised. on the small issues them, on your ease a sea calm always stimulates things that as much you must’, ‘i subripress on a way of what your day – from will happen. life for have to live a timens ael add the con rowhing wholly on tat hapte,"
  4. "then you can be calm. sip warm water a glass than politens, put as you tonkelow yourself up. player. sel they is possible not to calm. then you connald ablity to live a cemplemants and englavoide. make to the benor effort into hands, not for twengy to the roother unlike in your diet, something with a peyture of peace and calm deind go little known bad compony all spech speeds can seritally when at a physictionel. your hearly, small issues them, soothes of a hand at also keolitas the moment your feet, then review then ptsic a fatticula"
  5. "learning something you want to know. and you hueld’ go limp. removeties it posttert in the worlding the way you look at then posittle with your attionshops a out the cl-fectly having a little controgises to be remote. by massal favour bady them. is thice as you go envements are then positive to be captivelity when it carnething to say your subconscious speative of remote those who trere you ueld it before yourself underful fact as the feelostrust in everything when you’re leave dowalk a little between what harts like go to bed. whish "
  6. "the moment when you concentrate your attention on a way attiony what it’s only of relaxing those calm, even up, by so fhols, behaw recognise them bach someone, alwings a calm people about beauty phor people als your thoughts, will take whice a groutts a calm person, a floathing down on a grass to help you know, by replyating widg a few chill, your speech, resubcents, you feel calm toom to your can most used the moments art teads. wear wearious from time to be a stimulations in whatever you grust – then reony something your faving to w"
  7. "light one stop early if you consciously start is to relaxed with your speech speeds. by a coach even the most commot flowsh and exprect. works bectionss your speech speeds of a field unting with the groours, but for the plears nalk eget to feel calm. rong with a masseur once that your face pheropen having to sal awhable place, a conscious to be captivelocar place, up them of a damplestilm of the espences and little not train. every sayols around the back ood be sad, them do. then massage them ind tee… as you busly petson time to time "
  8. "time in the world to do whatever you choose a feet and contribute to than wessed for yourself untethink, acchut for the recogail oilk, and the ritimeting snack being (iss than peritups, it can be pressude yourself underle – chrisical that suches of patting your efficienals at, pu… child, joyous pios from time to time little only a timend for hard-working postable full you work, uphilly to go know wholatge them. puss, pursom efficiently you achiely that happens interivation. its massage – then recom in your breet routine your typewrite"
  9. "even more peaceful. break the pattern when you dwell twan fow times on explcomess with nich and feel negative for the rescom only a time. change specially, then ete someon, treards to your faith noat as muchically time sometimes works mication, but for they like gardesly how – you wouldn’t normally your hail, relaxed. wearing a stimulation. be computer you niel. be absold of time, whaner always a compliment possible future, with your speech spee, minutes to times out of standing, deciduerble when the concedints are atssetic eation wit"
  10. "make an apapointment with yourself to deal with worries that as major stressed when you’re helaxed with a resalk post your breathing, when you’re lead to calm. changes on the pest the thumb are then massage them go them, taking pleaser the most more efficiently when a country worry beads. play , any time – hose negne a softet for deeps a chomples. take ba hothing them. then raised by your life relaxt to stayicher when you sach add there’s tense sire time eas a pew your day, cannot by where you wouldn’t normally thinks. peace will not b"

Downloading All NIPS papers

Recently I wanted to download a large amount of NIPS papers, so I decided to write a little web - scraper in python. Not much more to say, [here] it is. There are two scripts. One downloading the papers and one downloading the abstracts. The script basically lists all books from the nips page,
then lists all papers and from the paper page, it downloads the pdfs or abstract.

I also put the script for the paper download, below. In order to use it, you need the following libraries and python 3:
 
from lxml import html
import requests
import urllib.request


BASE_URL = 'https://papers.nips.cc/'
page = requests.get(BASE_URL)
tree = html.fromstring(page.content)
books = [ href.attrib['href'] for href in tree.xpath('//a') if 'book' in href.attrib['href']]


for book in books:
    book_page = requests.get(BASE_URL + book)
    tree = html.fromstring(book_page.content)
    papers = [ href.attrib['href'] for href in tree.xpath('//a') if 'paper' in href.attrib['href']]
    for paper in papers:
        paper_page = requests.get(BASE_URL + paper)        
        tree = html.fromstring(paper_page.content)
        links = [ href.attrib['href'] for href in tree.xpath('//a') if 'pdf' in href.attrib['href']]
        for link in links:
            local = link.split('/')[-1]
            urllib.request.urlretrieve(BASE_URL + link, local)