ngrams.py

#!/usr/bin/env python3
# encoding: utf-8

"""Calculating ngram distributions (letters, bigrams, trigrams) from text or getting them from precomputed files."""

import collections
import hashlib

from config import MODIFIERS_PER_LAYER
from layout_base import POS_TO_FINGER, Layout, Layouts

from textcheck import occurrence_dict_difference as diffdict

### Constants

# shorthand: ngrams in these are more seldomly between word-parts.
#: shorthand pre and suffixes
_steno_fix = "all ant ander bis da dar ent er es ge gegen hab her hin hint ich in könn konnt keit kon lich mit rück schaft sei seid seine selbst so sollst sonders sondern tum tüm über übr un ung unser unter ver voll völl voll völl wenn werd wieder widr wiedr wider willst zer zu zurück zum zur"

#: shorthand words
_steno_short = "als also auf aus das dass deine dem den der des dessen deutsch die dies doch durch eine fort für hast hat haft hatt hätt heit ist kannst keine meine miss nicht nichts nieder noch nur ohne schon sich sie sind solch und usw vom woll wir wird wirst wo worden wurde würde"

#: shorthand, too much
_steno_short_too_long = "besonders zusammen zwischen vielleicht"

#: shorthand ngram-letters
_steno_letters = "pf cht ng h ch eit tr br gr kr rr dr fr pr pfr wr uhr ur au eu rs ss tz sch st nd ndr rd mp mpf sp spr str schr zw schw schm schn st cr qu ex ion"

#: shorthand, higher level letters
_steno_eil_letters = "en es ge  ber ger ter ker rer der fer per wer pfer igkeit ung du st est pro frag fall fahr zahl hand man männ dis bot isch istisch ie bar sam ial iell"

#: shorthand, higher level words → these are likely too much.
_steno_eil_short = "ismus ismen stadt statt nach richt bitt biet  trag jahr herr komm kömm schließ kauf käuf tisch immer nimmer gleich wachs  selb stell schreib schrieb schrift zeit tag groß etwa etwas wer jetzt größ geschäft mensch punkt tausend million mindestens wenigstens einzeln volk völk prozentual darauf gesamt insgesamt sonst forder förder konto selbstverständlich kunst zunächst gesetz allgemein angenehm ungefähr gesellschaft möglich wirtschaft versicher genossenschaft kapital sozial beschäftig person fabrik finanz organisation"

### Functions

def read_file(path):
    """Get the data from a file.

    >>> read_file("testfile")[:2]
    'ui'
    """

    f = open(path, encoding="utf-8")
    data = f.read()
    f.close()
    return data

def _split_uppercase_repeat(rep, num, layout=Layouts.NEO2,
                            mods=MODIFIERS_PER_LAYER):
    """Split a single bigram.

    >>> reps = _split_uppercase_repeat("AB", 5)
    >>> list(reversed(sorted([(j,i) for i, j in reps.items()])))
    [(10, 'a⇧'), (5, '⇧b'), (5, '⇗⇧'), (5, '⇗a'), (5, 'ab'), (2.5, '⇗b')]
    """
    # first check whether we really have a bigram and whether we need to split it.
    try:
        r2 = rep[1] # second first to error out early.
        r1 = rep[0]
    except IndexError:
        # this is no repeat but at most a single key.
        return {}
    pos1 = layout.char_to_pos(r1)
    pos2 = layout.char_to_pos(r2)
    # if any key isn’t found, the repeat doesn’t need splitting.
    if pos1 is None or pos2 is None:
        # caught all repeats for which one key isn’t in the layout. We
        # don’t need to change anything for these.
        return {rep: num}
    # same is true if all keys are layer 0.
    layer1 = pos1[2]
    layer2 = pos2[2]
    if layer1 == 0 and layer2 == 0:
        # caught all lowercase repeats. We don’t need to change
        # anything for these.
        return {rep: num}

    #: Adjustment of the weight of two modifiers on the same hand, because we can’t yet simulate moving the hand to use a different finger for M4/M3 when the pinky is needed on M3/shift. 2 * WEIGHT_FINGER_REPEATS * mods_on_same_hand_adjustment should be lower than (COST_PER_KEY_NOT_FOUND - max(COST_LAYER_ADDITION) - the most expensive key), because a key with M3-shift brings 2 finger repeats: one as first part in a bigram and the second as second part. 
    mods_on_same_hand_adjustment = 1/32
    repeats = collections.Counter()
    
    # now get the base keys.
    if layer1 == 0:
        base1 = r1
    else:
        base1 = layout.pos_to_char(pos1[:2] + (0, ))
    if layer2 == 0:
        base2 = r2
    else:
        base2 = layout.pos_to_char(pos2[:2] + (0, ))

    # add the base keys as repeat
    try: repeats[base1+base2] += num
    except KeyError: repeats[base1+base2] = num

    # now check for the mods which are needed to get the key
    # if the pos is left, we need the modifiers on the right.
    if layout.pos_is_left(pos1):
        mods1 = mods[layer1][1]
    else:
        mods1 = mods[layer1][0]
    if layout.pos_is_left(pos2): 
        mods2 = mods[layer2][1]
    else:
        mods2 = mods[layer2][0]

    # now we have the mods, so we do the splitting by mods.
    for m1 in mods1:
        # each of the mods with the key itself
        try: repeats[m1+base1] += num
        except KeyError: repeats[m1+base1] = num
        # each mod of the first with each mod of the second
        for m2 in mods2:
            try: repeats[m1+m2] += num
            except KeyError: repeats[m1+m2] = num
        # each of the first mods with the second base key
        # counted only 0.5 as strong, because the mod is normally hit and released short before the key is.
        try: repeats[m1+base2] += 0.5*num
        except KeyError: repeats[m1+base2] = 0.5*num 

    # the first base key with the second mods.
    # counted 2x as strong, because the mod is normally hit and released short before the key is.
    for m2 in mods2:
        try: repeats[base1+m2] += 2*num 
        except KeyError: repeats[base1+m2] = 2*num
        # also the second mod with the second base key
        try: repeats[m2+base2] += num
        except KeyError: repeats[m2+base2] = num

    # the mods of the first with each other
    # 0123 → 01 02 03 12 13 23
    if mods1[1:]: 
        for i in range(len(mods1)):
            for m2 in mods1[i+1:]:
                try: repeats[mods1[i]+m2] += num*mods_on_same_hand_adjustment
                except KeyError: repeats[mods1[i]+m2] = num*mods_on_same_hand_adjustment

    # the mods of the second with each other
    # 0123 → 01 02 03 12 13 23
    if mods2[1:]: 
        for i in range(len(mods2)):
            for m2 in mods2[i+1:]:
                try: repeats[mods2[i]+m2] += num*mods_on_same_hand_adjustment
                except KeyError: repeats[mods2[i]+m2] = num*mods_on_same_hand_adjustment

    return repeats

                    
def split_uppercase_repeats(reps, layout=Layouts.NEO2):
    """Split bigrams with uppercase letters (or others with any mod) into several lowercase bigrams by adding bigrams with mods and the base key. 

    Note: Using a collections.Counter() does not make this faster -- neither for pypy nor for cPython.

        Ab -> shift-a, shift-b, a-b.
        aB -> a-shift, shift-b, a-b.
        AB -> shift-a, shift-b, a-b, 0.5*(shift_L-shift_R, shift_R-shift_L)
        (a -> mod3-n, mod3-a, n-a.
        () -> mod3L-n, mod3L-r, n-r, 0.5*(mod3L-mod3L, mod3L-mod3L)
        (} -> mod3L-n, mod3R-e, n-e, 0.5*(mod3L-mod3R, mod3R-mod3L)
        ∃ℕ -> mod3R-e, mod4R-e, mod3L-n, mod4L-n,
              mod3R-n, mod4R-n, e-mod3L, e-mod4L,
              mod3R-mod3L, mod3R-mod4L, mod4R-mod3L, mod4R-mod4L,
              0.5*(mod3R-mod4R, mod4R-mod3R, mod3L-mod4L, mod4L-mod3L)

        Ψ∃: 

            # Modifiers from different keys
            '⇩⇙', M3L + M4R
            '⇩⇘', M3L + M3R
            '⇚⇙', M4L + M4R # TODO: M4L is hit with the ringfinger, here. Take that into account.
            '⇚⇘', M4L + M3R # TODO: M4L is hit with the ringfinger, here. Take that into account.

            # The different Modifiers of one of the keys with each other
            # sorted – should that be (m3-m4, m4-m3)/64?
            '⇩⇚', M3L + M4L / 32 (because this mistakenly gives a finger repeat, since we can’t yet simulate hitting M4 with the ringfinger and M3 with the pinky. # TODO: M4L is hit with the ringfinger, here. Take that into account.
            '⇘⇙', M3R + M4R / 32
            
            # Modifiers with the corresponding base keys
            '⇩h', M3L + h
            '⇚h', M4L + h
            '⇙e', M4R + e
            '⇘e', M3R + e
            
            # Modifiers with the other base key
            '⇩e', shiftL + e
            '⇚e', M4L + e # TODO: M4L is hit with the ringfinger, here. Take that into account.
            'h⇙', h + M4R
            'h⇘', h + M3R
            
            # The base keys on the base layer
            'he'
    
    >>> reps = [(36, "ab"), (24, "Ab"), (16, "aB"), (10, "AB"), (6, "¾2"), (4, "(}"), (2, "Ψ∃"), (1, "q")]
    >>> list(reversed(sorted([(j,i) for i, j in split_uppercase_repeats(reps).items()])))
    [(86, 'ab'), (52, 'a⇧'), (34, '⇗a'), (26, '⇧b'), (17.0, '⇗b'), (10, '⇗⇧'), (8, 'n⇘'), (6, '⇩⇘'), (6, '⇘e'), (6, '¾2'), (4, '⇩n'), (4, 'ne'), (4, 'h⇙'), (4, 'h⇘'), (3.0, '⇩e'), (2, '⇩⇙'), (2, '⇩h'), (2, '⇚⇙'), (2, '⇚⇘'), (2, '⇚h'), (2, '⇙e'), (2, 'he'), (1.0, '⇚e'), (0.0625, '⇩⇚'), (0.0625, '⇘⇙')]
    >>> reps = [(1, ", ")]
    >>> from layout_base import Layout, Layouts
    >>> layout = Layout.from_string("äuobp kglmfx+\\naietc hdnrsß\\n⇚.,üpö qyzwv", base_layout=Layouts.NEO2)
    >>> list(reversed(sorted([(j,i) for i, j in split_uppercase_repeats(reps, layout=layout).items()])))
    [(1, ', ')]
    """
    # replace uppercase by ⇧ + char1 and char1 + char2 and ⇧ + char2
    # char1 and shift are pressed at the same time
    #: The resulting bigrams after splitting.
    repeats = collections.Counter()
    _sur = _split_uppercase_repeat
    _mods = MODIFIERS_PER_LAYER
    for num, rep in reps:
        # this function gets called 100k times. Microoptimize the hell out of it.
        for key, val in _sur(rep, num, layout=layout, mods=_mods,).items():
            try:
                repeats[key] += val
            except KeyError:
                repeats[key] = val

    return repeats
    #reps = [(num, rep) for rep, num in repeats.items()]
    #reps.sort()
    #reps.reverse()
    #return reps

def repeats_in_file(data):
    """Sort the repeats in a file by the number of occurrances.

    >>> data = read_file("testfile")
    >>> repeats_in_file(data)[:3]
    [(2, 'a\\n'), (2, 'Aa'), (1, 'ui')]
    """
    repeats = collections.Counter()
    for i in range(len(data)-1):
        rep = data[i] + data[i+1]
        try:
            repeats[rep] += 1
        except KeyError:
            repeats[rep] = 1
    sorted_repeats = [(repeats[i], i) for i in repeats]
    sorted_repeats.sort(reverse=True)
    #reps = split_uppercase_repeats(sorted_repeats) # wrong place
    return sorted_repeats

def split_uppercase_letters(reps, layout):
    """Split uppercase letters (or others with any mod) into two lowercase letters (with the mod).

    >>> letters = [(4, "a"), (3, "A")]
    >>> split_uppercase_letters(letters, layout=Layouts.NEO2)
    [(4, 'a'), (3, '⇗'), (3, 'a')]
    """
    # replace uppercase by ⇧ and char1
    upper = []
    repeats = []
    for num, rep in reps:
        pos = layout.char_to_pos(rep)
        if pos and pos[2]:
            upper.append((num, rep, pos))
        else:
            repeats.append((num, rep))
    reps = repeats

    up = []
    
    for num, rep, pos in upper:
        layer_mods = MODIFIERS_PER_LAYER[pos[2]]
                                         
        if layout.pos_is_left(pos):
            for m in layer_mods[1]: # left keys use the right mods
                up.append((num, m)) 
        else:
            for m in layer_mods[0]:  # right keys use the left mods
                up.append((num, m))

        # also append the base layer key.
        up.append((num, layout.pos_to_char((pos[0], pos[1], 0),)))
                
    reps.extend(up)
    reps = [(int(num), r) for num, r in reps]
    reps.sort(reverse=True)
    return reps

def letters_in_file(data):
    """Sort the repeats in a file by the number of occurrances.

    >>> data = read_file("testfile")
    >>> letters_in_file(data)[:3]
    [(5, 'a'), (4, '\\n'), (2, 'r')]
    """
    letters = collections.Counter()
    for letter in data:
        if letter in letters:
            letters[letter] += 1
        else:
            letters[letter] = 1
    sort = [(letters[i], i) for i in letters]
    sort.sort(reverse=True)
    return sort

def unique_sort(liste):
    """Count the occurrence of each item in a list.

    >>> unique_sort([1, 2, 1])
    [(1, 2), (2, 1)]
    """
    counter = collections.Counter()
    for i in liste:
        if i in counter:
            counter[i] += 1
        else:
            counter[i] = 1

    sorted_repeats = [(counter[i], i) for i in counter]
    sorted_repeats.sort()
    return sorted_repeats   

def repeats_in_file_sorted(data):
    """Sort the repeats in a file by the number of occurrances.

    >>> data = read_file("testfile")
    >>> repeats_in_file_sorted(data)[:2]
    [(1, '\\na'), (1, '\\ne')]
    """
    repeats = repeats_in_file(data)
    repeats.reverse()
    return repeats

def _unescape_ngram_list(ngrams): 
    """unescape \n and \ in an ngram list."""
    for i in range(len(ngrams)):
        if "\\" in ngrams[i][1]: 
            ngrams[i] = (ngrams[i][0], ngrams[i][1].replace("\\\\", "\\"))
            ngrams[i] = (ngrams[i][0], ngrams[i][1].replace("\\n", "\n"))
    return ngrams
                  


def repeats_in_file_precalculated(data, only_existing=True):
    """Get the repeats from a precalculated file.

    >>> data = read_file("2gramme.txt")
    >>> repeats_in_file_precalculated(data)[:3]
    [(10159250, 'en'), (10024681, 'er'), (9051717, 'n ')]
    """
    md5 = hashlib.md5(data.encode("utf-8")).hexdigest()
    if ( 'reps_repeats_in_file_precalculated' not in repeats_in_file_precalculated.__dict__ or
         md5 not in repeats_in_file_precalculated.reps_repeats_in_file_precalculated):
        if 'reps_repeats_in_file_precalculated' not in repeats_in_file_precalculated.__dict__:
            repeats_in_file_precalculated.reps_repeats_in_file_precalculated = {}
        reps = [line.lstrip().split(" ", 1) for line in data.splitlines() if line.lstrip().split(" ", 1)[1:]]
        if only_existing: 
            all_keys = Layouts.NEO2.get_all_chars()
            try: 
                reps = [(int(num), r) for num, r in reps if r[1:] and r[0] in all_keys and r[1] in all_keys]
            except ValueError: # we got floats
                reps = [(float(num), r) for num, r in reps if r[1:] and r[0] in all_keys and r[1] in all_keys]
        else:
            try:
                reps = [(int(num), r) for num, r in reps if r[1:]]
            except ValueError: # we got floats
                reps = [(float(num), r) for num, r in reps if r[1:]]
    
        # cleanup
        _unescape_ngram_list(reps)
        repeats_in_file_precalculated.reps_repeats_in_file_precalculated[md5] = reps
        return repeats_in_file_precalculated.reps_repeats_in_file_precalculated[md5]
    else:
        return repeats_in_file_precalculated.reps_repeats_in_file_precalculated[md5]



def split_uppercase_trigrams(trigs):
    """Split uppercase repeats into two to three lowercase repeats.

    Here we don’t care about shift-collisions with the “second” letter, because we only use it for handswitching and the shift will always mean a handswitch afterwards (opposing shift). ⇒ Ab → Sh-ab, ignoring a-Sh-b. ⇒ for handswitching ignore trigrams with any of the shifts. 

    >>> trigs = [(8, "abc"), (7, "Abc"), (6, "aBc"), (5, "abC"), (4, "ABc"), (3, "aBC"), (2, "AbC"), (1, "ABC")]
    >>> split_uppercase_trigrams(trigs)
    [(15, 'abc'), (7, 'a⇧b'), (7, 'a⇗b'), (5, '⇧bc'), (5, '⇗bc'), (5, 'b⇧c'), (5, 'b⇗c'), (4, '⇧ab'), (4, '⇗ab'), (3, 'ab⇧'), (3, 'ab⇗'), (2, '⇧b⇧'), (2, '⇧b⇗'), (2, '⇧a⇧'), (2, '⇧a⇗'), (2, '⇗b⇧'), (2, '⇗b⇗'), (2, '⇗a⇧'), (2, '⇗a⇗')]
    """
    # replace uppercase by ⇧ + char1 and char1 + char2
    upper = [(num, trig) for num, trig in trigs if not trig == trig.lower() and trig[2:]]
    # and remove them temporarily from the list of trigrams - don’t compare list with list, else this takes ~20min!
    trigs = [(num, trig) for num, trig in trigs if trig == trig.lower() and trig[2:]]
    up = []
    # since this gets a bit more complex and the chance to err is high,
    # we do this dumbly, just checking for the exact cases.
    # TODO: Do it more elegantly: Replace every uppercase letter by "⇧"+lowercase
    #       and then turn the x-gram into multiple 3grams (four[:-1], four[1:]; five… ).
    for num, trig in upper: 
        # Abc
        if not trig[0] == trig[0].lower() and trig[1] == trig[1].lower() and trig[2] == trig[2].lower():
            up.append((max(1, num//2), "⇧"+trig[:2].lower()))
            up.append((max(1, num//2), "⇗"+trig[:2].lower()))
            up.append((num, trig.lower()))
        # aBc
        elif trig[0] == trig[0].lower() and not trig[1] == trig[1].lower() and trig[2] == trig[2].lower():
            up.append((max(1, num//2), "⇧"+trig[1:].lower()))
            up.append((max(1, num//2), "⇗"+trig[1:].lower()))
            up.append((max(1, num//2), trig[0].lower()+"⇧"+trig[1].lower()))
            up.append((max(1, num//2), trig[0].lower()+"⇗"+trig[1].lower()))
            
        # abC
        elif trig[0] == trig[0].lower() and trig[1] == trig[1].lower() and not trig[2] == trig[2].lower():
            up.append((max(1, num//2), trig[:2].lower() + "⇧"))
            up.append((max(1, num//2), trig[:2].lower() + "⇗"))
            up.append((max(1, num//2), trig[1].lower()+"⇧"+trig[2].lower()))
            up.append((max(1, num//2), trig[1].lower()+"⇗"+trig[2].lower()))
            
        # ABc (4, '⇧a⇧'), (4, 'a⇧b'), (4, '⇧bc')
        elif not trig[0] == trig[0].lower() and not trig[1] == trig[1].lower() and trig[2] == trig[2].lower():
            up.append((max(1, num//4), "⇧"+trig[0].lower()+"⇧"))
            up.append((max(1, num//2), trig[0].lower()+"⇧"+trig[1].lower()))
            up.append((max(1, num//2),  "⇧" + trig[1:].lower()))
            
            up.append((max(1, num//4), "⇗"+trig[0].lower()+"⇧"))
            up.append((max(1, num//4), "⇧"+trig[0].lower()+"⇗"))
            up.append((max(1, num//4), "⇗"+trig[0].lower()+"⇗"))

            up.append((max(1, num//2), trig[0].lower()+"⇗"+trig[1].lower()))
            up.append((max(1, num//2),  "⇗" + trig[1:].lower()))
            
        # aBC (3, 'a⇧b'), (3, '⇧b⇧'), (3, 'b⇧c')
        elif trig[0] == trig[0].lower() and not trig[1] == trig[1].lower() and not trig[2] == trig[2].lower():
            up.append((max(1, num//4), "⇧"+trig[1].lower()+"⇧"))
            up.append((max(1, num//2), trig[0].lower()+"⇧"+trig[1].lower()))
            up.append((max(1, num//2), trig[1].lower()+"⇧"+trig[2].lower()))
            
            up.append((max(1, num//4), "⇗"+trig[1].lower()+"⇧"))
            up.append((max(1, num//4), "⇧"+trig[1].lower()+"⇗"))
            up.append((max(1, num//4), "⇗"+trig[1].lower()+"⇗"))

            up.append((max(1, num//2), trig[0].lower()+"⇗"+trig[1].lower()))
            up.append((max(1, num//2), trig[1].lower()+"⇗"+trig[2].lower()))
            
        # AbC (2, '⇧ab'), (2, 'ab⇧'), (2, 'b⇧c')
        elif not trig[0] == trig[0].lower() and trig[1] == trig[1].lower() and not trig[2] == trig[2].lower():
            up.append((max(1, num//2),  "⇧" + trig[:2].lower()))
            up.append((max(1, num//2),  trig[:2].lower() + "⇧"))
            up.append((max(1, num//2), trig[1].lower()+"⇧"+trig[2].lower()))
            
            up.append((max(1, num//2),  "⇗" + trig[:2].lower()))
            up.append((max(1, num//2),  trig[:2].lower() + "⇗"))
            up.append((max(1, num//2), trig[1].lower()+"⇗"+trig[2].lower()))

        # ABC (1, '⇧a⇧'), (1, 'a⇧b'), (1, '⇧b⇧'), (1, 'b⇧c')
        elif not trig[0] == trig[0].lower() and not trig[1] == trig[1].lower() and not trig[2] == trig[2].lower():
            up.append((max(1, num//4), "⇧"+trig[0].lower()+"⇧"))
            up.append((max(1, num//2), trig[0].lower()+"⇧"+trig[1].lower()))
            up.append((max(1, num//4), "⇧"+trig[1].lower()+"⇧"))
            up.append((max(1, num//2), trig[1].lower()+"⇧"+trig[2].lower()))
            
            up.append((max(1, num//4), "⇗"+trig[0].lower()+"⇧"))
            up.append((max(1, num//4), "⇧"+trig[0].lower()+"⇗"))
            up.append((max(1, num//4), "⇗"+trig[0].lower()+"⇗"))

            up.append((max(1, num//4), "⇗"+trig[1].lower()+"⇧"))
            up.append((max(1, num//4), "⇧"+trig[1].lower()+"⇗"))
            up.append((max(1, num//4), "⇗"+trig[1].lower()+"⇗"))

            up.append((max(1, num//2), trig[0].lower()+"⇗"+trig[1].lower()))
            up.append((max(1, num//2), trig[1].lower()+"⇗"+trig[2].lower()))

    
    trigs.extend(up)
    trigs = [(num, r) for num, r in trigs if r[1:]]
    t = collections.Counter()
    for num, r in trigs:
        try: t[r] += num
        except KeyError: t[r] = num
    trigs = [(num, r) for r, num in t.items()]
    trigs.sort(reverse=True)
    return trigs


def split_uppercase_trigrams_correctly(trigs, layout, just_record_the_mod_key=False):
    """Split uppercase repeats into two to three lowercase repeats.

    Definition: 

        a → b → c
        | × | × |
        sa→ sb→ sc
        senkrechte nur nach oben. Kreuze und Pfeile nur nach vorne. Alle Trigramme, die du aus dem Bild basteln kannst.

    >>> trigs = [(8, "abc"), (7, "∀bC"), (6, "aBc"), (5, "abC"), (4, "ABc"), (3, "aBC"), (2, "AbC"), (1, "ABC")]
    >>> # split_uppercase_trigrams_correctly(trigs, Layouts.NEO2)
    """
    # kick out any who don’t have a position
    pos_trig = [(num, [layout.char_to_pos(char) for char in trig], trig) for num, trig in trigs]
    pos_trig = [(num, pos, trig) for num, pos, trig in pos_trig if not None in pos]
    
    # get all trigrams with non-baselayer-keys
    upper = [(num, pos, trig) for num, pos, trig in pos_trig if True in [p[2]>0 for p in pos]]
    # and remove them temporarily from the list of trigrams - don’t compare list with list, else this takes ~20min!
    trigs = [(num, trig) for num, pos, trig in pos_trig if not True in [p[2]>0 for p in pos]]

    #: The trigrams to add to the baselayer trigrams
    up = []

    mod = MODIFIERS_PER_LAYER
    for num, pos, trig in upper:
        print(trig)
        # lower letters
        l0 = layout.pos_to_char((pos[0][0], pos[0][1], 0))
        l1 = layout.pos_to_char((pos[1][0], pos[1][1], 0))
        l2 = layout.pos_to_char((pos[2][0], pos[2][1], 0))
        # mods
        m0 = mod[pos[0][2]]
        m1 = mod[pos[1][2]]
        m2 = mod[pos[2][2]]
        ### Algorithm
        algo = """
        a → b → c
        | × | × |
        sa→ sb→ sc
        | × | × |   ; seperate dimension. ma is connected to a and sa.
        ma→ mb→ mc
        """
        #: Matrix der Tasten und Modifikatoren
        m = []
        for p, c in zip(pos, (l0, l1, l2)):
            mx = mod[p[2]] # liste mit bis zu 2 mods
            if just_record_the_mod_key:
                mx = [i+c for i in mx[0]]
            elif layout.pos_is_left(p):
                mx = mx[1]
            else: mx = mx[0]
            col = [c]
            if mx: 
                col.append(mx[0])
            else: col.append(None)
            if mx[1:]: 
                col.append(mx[1])
            else: col.append(None)
            m.append(col)

        # Matrix created
        #: All possible starting positions for trigrams in that matrix
        sp = [(0,0), (0,1), (0,2), (1,0), (1,1), (1,2), (2,1), (2,2)] # not last letter
        # reduce the starting positions to the actually existing letters.
        sp = [p for p in sp if m[p[0]][p[1]] is not None]

        #: All possible paths in the matrix for letters
        paths = [(1,0), (1,1), (1,2)]
        #: Additional possible paths for modifiers
        mod_paths = [(0,1), (0,2)]

        #: The new trigrams which get created due to splitting.
        new_trigs = [] # option: take a set to avoid double entries.

        # move all paths
        for s in sp:
            #: trigrams of matrix positions [(p0, p1, p2), …]
            tri = []
            #: bigrams of matrix positions [(p0, p1), …]
            tr = []
            # try all possible path for two steps.
            #: the paths
            p = paths[:]
            # modifiers get extra options
            if s[1]: p.extend(mod_paths)

            # try all paths, append to tr if not None
            for n in p:
                new_pos = (s[0] + n[0], (s[1] + n[1])%3)
                try: 
                    if m[new_pos[0]][new_pos[1]] is not None:
                        tr.append((s, new_pos))
                except IndexError: # left the matrix
                    pass

            # now try all paths, starting from the positions in tr.
            for s,t in tr:
                #: the paths
                p = paths[:]
                # modifiers get extra options
                if t[1]: p.extend(mod_paths)
                for n in p:
                    new_pos = (t[0]+n[0], (t[1] + n[1])%3)
                    try: 
                        if m[new_pos[0]][new_pos[1]] is not None:
                            tri.append((s, t, new_pos))
                    except IndexError: # left the matrix
                        pass
            print([m[s[0]][s[1]]+m[t[0]][t[1]]+m[n[0]][n[1]] for s,t,n in tri])
            new_trigs.extend([m[s[0]][s[1]]+m[t[0]][t[1]]+m[n[0]][n[1]] for s,t,n in tri])
        for tri in new_trigs:
            up.append((num, tri))
            
    print (up)
    trigs.extend(up)
    trigs = [(int(num), r) for num, r in trigs if r[1:]]
    trigs.sort(reverse=True)
    return trigs


def trigrams_in_file(data, only_existing=True):
    """Sort the trigrams in a file by the number of occurrances.

    >>> data = read_file("testfile")
    >>> trigrams_in_file(data)[:4]
    [(4, '⇧aa'), (4, '⇗aa'), (2, 't⇧a'), (2, 't⇗a')]
    """
    trigs = collections.Counter()
    for i in range(len(data)-2):
        trig = data[i] + data[i+1] + data[i+2]
        if trig in trigs:
            trigs[trig] += 1
        else:
            trigs[trig] = 1

    if only_existing: 
        all_keys = Layouts.NEO2.get_all_chars()
        sorted_trigs = [(trigs[i], i) for i in trigs if i[2:] and i[0] in all_keys and i[1] in all_keys and i[2] in all_keys]
    else:
        sorted_trigs = [(trigs[i], i) for i in trigs if i[2:]]
    sorted_trigs.sort(reverse=True)
    trigs = split_uppercase_trigrams(sorted_trigs)
    return trigs

def ngrams_in_filepath(datapath, slicelength=1000000):
    """Sort the trigrams in a file by the number of occurrances.

    >>> lett, big, trig = ngrams_in_filepath("testfile")
    >>> lett[:3]
    [(5, 'a'), (4, '\\n'), (2, 'r')]
    >>> big[:3]
    [(2, 'a\\n'), (2, 'Aa'), (1, 'ui')]
    >>> trig[:10]
    [(4, '⇧aa'), (4, '⇗aa'), (2, 't⇧a'), (2, 't⇗a'), (2, 'a⇧a'), (2, 'a⇗a'), (2, 'aa\\n'), (1, 'uia'), (1, 'rt⇧'), (1, 'rt⇗')]
    """
    f = open(datapath, encoding="utf-8", errors="ignore")
    letters = collections.Counter()
    repeats = collections.Counter()
    trigs = collections.Counter()
    data = f.read(slicelength)
    step = 0
    while data[2:]:
        if step == 1: 
            print("reading ngrams from", datapath)
        if step:
            print("read ~", int(f.tell()/10000)/100, "MiB")
        step += 1
        
        for i in range(len(data)-2):
            letter = data[i]
            if letter in letters:
                letters[letter] += 1
            else:
                letters[letter] = 1
            
            rep = data[i] + data[i+1]
            if rep in repeats:
                repeats[rep] += 1
            else:
                repeats[rep] = 1
            
            trig = data[i] + data[i+1] + data[i+2]
            if trig in trigs:
                trigs[trig] += 1
            else:
                trigs[trig] = 1
                
        data = data[-2:] + f.read(slicelength)

    # final two letters
    for letter in data:
        if letter in letters:
            letters[letter] += 1
        else:
            letters[letter] = 1
    # final bigram
    rep = data[-2] + data[-1]
    if rep in repeats:
        repeats[rep] += 1
    else:
        repeats[rep] = 1
    
       
    
    letters = [(letters[i], i) for i in letters]
    letters.sort(reverse=True)

    repeats = [(repeats[i], i) for i in repeats]
    repeats.sort(reverse=True)

    trigs = [(trigs[i], i) for i in trigs]
    trigs.sort(reverse=True)
    # split uppercase trigrams here, because we really want to do that only *once*.
    trigs = split_uppercase_trigrams(trigs)
    return letters, repeats, trigs

def trigrams_in_file_precalculated(data, only_existing=True):
    """Get the repeats from a precalculated file.

    CAREFUL: SLOW!

    >>> data = read_file("3gramme.txt")
    >>> trigrams_in_file_precalculated(data)[:6]
    [(5678553, 'en '), (4467769, 'er '), (2891228, ' de'), (2493088, 'der'), (2304026, 'sch'), (2272028, 'ie ')]
    """
    md5 = hashlib.md5(data.encode("utf-8")).hexdigest()
    if ( 'trigs_trigrams_in_file_precalculated' not in trigrams_in_file_precalculated.__dict__ or
         md5 not in trigrams_in_file_precalculated.trigs_trigrams_in_file_precalculated):
        if 'trigs_trigrams_in_file_precalculated' not in trigrams_in_file_precalculated.__dict__:
            trigrams_in_file_precalculated.trigs_trigrams_in_file_precalculated = {}
        trigs = [line.lstrip().split(" ", 1) for line in data.splitlines() if line.split()[1:]]
    
        if only_existing: 
            all_keys = Layouts.NEO2.get_all_chars()
            try: 
                trigs = [(int(num), r) for num, r in trigs if r[2:] and r[0] in all_keys and r[1] in all_keys and r[2] in all_keys]
            except ValueError: # we got floats
                trigs = [(float(num), r) for num, r in trigs if r[2:] and r[0] in all_keys and r[1] in all_keys and r[2] in all_keys]
        else:
            try:
                trigs = [(int(num), r) for num, r in trigs if r[2:]]
            except ValueError: # we got floats
                trigs = [(float(num), r) for num, r in trigs if r[2:]]
        # cleanup
        _unescape_ngram_list(trigs)
        trigs = split_uppercase_trigrams(trigs)
        trigrams_in_file_precalculated.trigs_trigrams_in_file_precalculated[md5] = trigs
        return trigs
    else:
        return trigrams_in_file_precalculated.trigs_trigrams_in_file_precalculated[md5]

def letters_in_file_precalculated(data, only_existing=True):
    """Get the repeats from a precalculated file.

    >>> data = read_file("1gramme.txt")
    >>> letters_in_file_precalculated(data)[:3]
    [(46474641, ' '), (44021504, 'e'), (26999087, 'n')]
    """
    md5 = hashlib.md5(data.encode("utf-8")).hexdigest()
    if ( 'ret_letters_in_file_precalculated' not in letters_in_file_precalculated.__dict__ or
         md5 not in letters_in_file_precalculated.__dict__['ret_letters_in_file_precalculated']):
        if 'ret_letters_in_file_precalculated' not in letters_in_file_precalculated.__dict__:
            letters_in_file_precalculated.ret_letters_in_file_precalculated = {}
        letters = [line.lstrip().split(" ", 1) for line in data.splitlines() if line.split()[1:] or line[-2:] == "  "]
        # cleanup
        _unescape_ngram_list(letters)
        try:
            if only_existing:
                all_keys = Layouts.NEO2.get_all_chars()
                letters_in_file_precalculated.ret_letters_in_file_precalculated[md5] = [(int(num), let) for num, let in letters if let in all_keys]
                return letters_in_file_precalculated.ret_letters_in_file_precalculated[md5]
            else:
                letters_in_file_precalculated.ret_letters_in_file_precalculated[md5] = [(int(num), let) for num, let in letters if let]
                return letters_in_file_precalculated.ret_letters_in_file_precalculated[md5]
        except ValueError: # floats in there
            if only_existing:
                all_keys = Layouts.NEO2.get_all_chars()
                letters_in_file_precalculated.ret_letters_in_file_precalculated[md5] = [(float(num), let) for num, let in letters if let in all_keys]
                return letters_in_file_precalculated.ret_letters_in_file_precalculated[md5]
            else:
                letters_in_file_precalculated.ret_letters_in_file_precalculated[md5] = [(float(num), let) for num, let in letters if let]
                return letters_in_file_precalculated.ret_letters_in_file_precalculated[md5]
    else:
        return letters_in_file_precalculated.ret_letters_in_file_precalculated[md5]

def clean_data(data):
    """Remove cruft from loaded data"""
    if data[:1] == "\ufeff":
        data = data[1:]
    return data

def fix_impossible_ngrams(data, layout=None):
    """Make sure that we get the best possible data:

    1. Correct the representation of some symbols. (Tabs can be "⇥" or "\\t")
    2. Remove the ngrams that can't be typed using a specified layout

    If no layout is specified, the ngrams aren't changed.
    """
    if layout:
        # The list of possible inputs in a given layout
        keys = layout.get_all_chars()
        
        # Make sure the correct tab-syntax is shown in the ngrams
        if ("⇥" in keys) and not ("\t" in keys):
            # Replace all "\t" with "⇥"
            for tuple_idx, (num, ngram) in enumerate(data):
                if "\t" in ngram:
                    corrected_ngram = ngram.replace("\t", "⇥")
                    data[tuple_idx] = (num, corrected_ngram)
        elif ("\t" in keys) and not ("⇥" in keys):
            # Replace all "⇥" with "\t"
            for tuple_idx, (num, ngram) in enumerate(data):
                if "⇥" in ngram:
                    corrected_ngram = ngram.replace("⇥", "\t")
                    data[tuple_idx] = (num, corrected_ngram)

        # def can_be_typed(ngram_tuple):
        #     """Check if a ngram can be typed"""
        #     ngram = ngram_tuple[1]
        #     for letter in ngram:
        #         if letter not in keys:
        #             #print("filtered trig:", ngram)
        #             #print("crucial letter:", letter.encode("unicode_escape"))
        #             #print()
        #             return False
        #     return True
        # data = list(filter(can_be_typed, data))

        return data
    else:
        # Do nothing
        return data

def ngrams_in_data(data):
    """Get ngrams from a dataset."""
    letters = letters_in_file(data)
    repeats = bigrams = repeats_in_file(data)
    trigrams = trigrams_in_file(data)
    return letters, repeats, trigrams
    

### Classes

class FileNotFoundException(Exception):
    def __str__(self):
        return repr(self.parameter)

class ParseException(Exception):
    def __str__(self):
        return repr(self.parameter)


class NGrams(object):
    """
    NGrams contains ngrams from various sources in raw and weighted
    form and can export them to the simple (1gramme.txt, 2gramme.txt,
    3gramme.txt) form with a given number of total keystrokes.

    self.one, self.two and self.three are dictionaries of 1grams, 2grams and threegrams with their respective probability of occurring.
    """
    def __init__(self, config):
        """Create an ngrams object.

        config: plain text file (utf-8 encoded).
            # ngrams source v0.1
            weight type filepath
            # comment
            weight type filepath
            …

        >>> ngrams = NGrams('ngrams_test.config')
        Reading text testfile
        Reading bla errortest
        unrecognized filetype bla errortest
        >>> ngrams.raw
        [(1.0, ([(5, 'a'), (4, '\\n'), (2, 'r'), (2, 'e'), (2, 'A'), (1, 'u'), (1, 't'), (1, 'o'), (1, 'n'), (1, 'i'), (1, 'g'), (1, 'd')], [(2, 'a\\n'), (2, 'Aa'), (1, 'ui'), (1, 'tA'), (1, 'rt'), (1, 'rg'), (1, 'od'), (1, 'nr'), (1, 'ia'), (1, 'g\\n'), (1, 'eo'), (1, 'en'), (1, 'd\\n'), (1, 'ae'), (1, 'aa'), (1, 'aA'), (1, '\\nr'), (1, '\\ne'), (1, '\\na')], [(4, '⇧aa'), (4, '⇗aa'), (2, 't⇧a'), (2, 't⇗a'), (2, 'a⇧a'), (2, 'a⇗a'), (2, 'aa\\n'), (1, 'uia'), (1, 'rt⇧'), (1, 'rt⇗'), (1, 'rg\\n'), (1, 'od\\n'), (1, 'nrt'), (1, 'iae'), (1, 'g\\na'), (1, 'eod'), (1, 'enr'), (1, 'd\\nr'), (1, 'aen'), (1, 'aa⇧'), (1, 'aa⇗'), (1, 'a\\ne'), (1, '\\nrg'), (1, '\\neo'), (1, '\\naa')]))]
        """
        # read the config.
        try:
            with open(config, encoding="utf-8") as f:
                self.config = f.read()
        except IOError:
            raise FileNotFoundException("File", config, "can’t be read.")
        if self.config.startswith("# ngrams source v0.0"):
            self.parse_config_0_0()
        elif self.config.startswith("# ngrams source v0.1"):
            self.parse_config_0_1()
        else:
            raise ParseException("ngrams config has no version header. Please start it with # ngrams source v<version>")
        self.normalize_raw() # gets one, two and three

    def finalize(self):
        """Do ngrams adjustments for ngram-files which are going to be used directly, but which are not suitable for intermediate files (must be done only once)."""
        # increases the weight for ngrams which are shorts in steno: profit from 100 years of professional writing experience.
        self.weight_steno()
        self.weight_punctuation()

    def diff(self, other):
        """Compare two ngram distributions.

        >>> n = NGrams('ngrams_test_diff.config')
        Reading pregenerated 1-gramme.arne.txt;2-gramme.arne.txt;3-gramme.arne.txt
        >>> from copy import deepcopy
        >>> m = deepcopy(n)
        >>> m.one["a"] += 0
        >>> n.diff(m)
        [{}, {}, {}]
        >>> m.one["a"] += 3
        >>> n.diff(m)[1:]
        [{}, {}]
        >>> str(n.diff(m)[0]["a"])[:7]
        '-2.8652'

        @return the differences of the normalized 1-, 2- and 3grams as dicts.
        """
        def _normalize_dict(d):
            """Normalize a dict to have a valuesum of 1"""
            s = sum(d.values())
            # no zero division
            if s == 0:
                return d
            return {key: num/s for key, num in d.items()}

        def _diffdict(d1, d2):
            """All different keys with the difference."""
            diff = {key: d1[key] - d2[key] for key in d1 if key in d2 and abs(d1[key] - d2[key]) > 1.0e-14}
            # python floats normally should have binary precision 53,
            # but with 1.0e-15 I still get stray results when diffing
            # the same dictionary.
            for key in d1:
                if key in d2: continue
                diff[key] = d1[key]
            for key in d2:
                if key in d1: continue
                diff[key] = -d2[key]
            return diff

        # normalized
        one = _normalize_dict(self.one)
        two = _normalize_dict(self.two)
        three = _normalize_dict(self.three)

        o1 = _normalize_dict(other.one)
        o2 = _normalize_dict(other.two)
        o3 = _normalize_dict(other.three)

        return [_diffdict(s, o) for s, o in ((one, o1), (two, o2), (three, o3))]


    def parse_config_0_1(self):
        lines = self.config.splitlines()
        # the raw list of included ngrams.
        self.raw = []
        for l in lines[1:]:
            # # start a comment
            if l.startswith("#"):
                continue
            parts = l.split()
            weight, typ = parts[:2]
            weight = float(weight)
            datapath = l[l.index(typ)+len(typ)+1:]
            print ("Reading", typ, datapath)
            if typ=="text":
                one, two, three = ngrams_in_filepath(datapath=datapath)
                self.raw.append((weight, (one, two, three)))
            elif typ=="pykeylogger":
                one, two, three = self.read_pykeylogger_logfile(datapath)
                self.raw.append((weight, (one, two, three)))
            elif typ=="pregenerated":
                onegrams, twograms, threegrams = datapath.split(";")
                letterdata = clean_data(read_file(onegrams))
                one = letters_in_file_precalculated(letterdata)
                bigramdata = clean_data(read_file(twograms))
                two = repeats_in_file_precalculated(bigramdata)
                trigramdata = clean_data(read_file(threegrams))
                three = trigrams_in_file_precalculated(trigramdata)
                self.raw.append((weight, (one, two, three)))
            elif typ=="maildir":
                one, two, three = self.read_maildir_dir(datapath)
                self.raw.append((weight, (one, two, three)))
            elif typ=="chatlog":
                one, two, three = self.read_mirc_chatlog(datapath)
                self.raw.append((weight, (one, two, three)))
            else: print("unrecognized filetype", typ, datapath)

    def read_maildir_dir(self, folderpath):
        """Read all message contents from within a maildir folder.

        see http://docs.python.org/library/mailbox.html#mailbox.Maildir
        """
        text = ""
        from mailbox import Maildir
        from termctrl import home, erase
        m = Maildir(folderpath)
        num = 0
        for message in m.itervalues():
            home()
            print (num, "/", len(m), end="")
            text += self._maildir_message_own_content(message)
            num += 1
        home()
        erase()
        one, two, three = ngrams_in_data(data=text)
        return one, two, three

    def _maildir_message_own_content(self, message):
        """Remove all quotes and headers from the message.

        @return string of the content which was written by the author."""
        import quopri
        text = ""
        for t in message.walk():
            if t.get_content_type() == "text/plain":
                # remove quoted printable
                # the following should work, but does not. 
                # body = quopri.decodestring(t.as_string())
                # fix directly from quopri: 
                from io import BytesIO
                infp = BytesIO(t.as_string().encode("utf-8"))
                outfp = BytesIO()
                quopri.decode(infp, outfp, header=False)
                try: 
                    body = outfp.getvalue().decode("utf-8")
                except UnicodeDecodeError:
                    try: body = outfp.getvalue().decode("iso-8859-1")
                    except UnicodeDecodeError:                     
                        print (t.as_string())
                        body = t.as_string()

                try: 
                    body = body[body.index("\n\n"):]
                except ValueError: continue # no body
                # we only ever add the previous line, so we can remove fullquotes along with their data.
                previous_line = ""
                for l in body.splitlines():
                    # stop reading the message
                    if "Forwarded Message" in previous_line or previous_line.startswith("--"):
                        break
                    # skip quotes and stuff I want to ignore.
                    if (previous_line.startswith(">") or
                        previous_line.startswith(" >") or
                        not " " in previous_line or # most likely an escaped datablock
                        "http://" in previous_line): # an url
                        previous_line = l
                        continue
                    # add all lines which are not the beginning of a comment.
                    if not ">" in l and not previous_line.endswith(":"):
                        text += previous_line
                    previous_line = l
                    text += "\n"
#        print(text)
        return text

    def read_mirc_chatlog(self, datapath):
        """Read a chatlog and remove date and usernames."""
        text = ""
        # we use regexps
        import re
        with open(datapath, encoding="utf-8") as f:
            for line in f.readlines():
                # no lines without time (status stuff)
                if not line.startswith("["):
                    continue
                # kill time
                line = re.sub(r"^\[\d\d:\d\d\.\d\d\]\ ", "", line)
                # kill status
                if line.startswith("***"):
                    continue
                # remove username
                re.sub(r"^<\w>\ ", "", line)
                text += line
        one, two, three = ngrams_in_data(data=text)
        return one, two, three


    def read_pykeylogger_logfile(self, datapath):
        """Read a logfile from pykeylogger and extract all normal keys."""
        with open(datapath, encoding="utf-8") as f:
            data = f.read()
        text = ""
        for line in data.splitlines():
            # skip the beginning
            line = "|".join(line.split("|")[6:])
            # replace all special chars we know by the special chars used in our keyboard definition
            line = line.replace("[KeyName:Return]", "\n")
            line = line.replace("[KeyName:BackSpace]", "←")
            line = line.replace("[KeyName:Control_L]", "♕")
            line = line.replace("[KeyName:Control_R]", "♛")
            line = line.replace("[KeyName:Alt_L]", "♔")
            line = line.replace("[KeyName:Alt_R]", "♚")
            line = line.replace("[KeyName:Tab]", "⇥")
            m3 = "[KeyName:Mode_switch]"
            m4 = "[KeyName:[65027]]"
            # 3rd and 4th layer are not captured correctly: cut the rest of the line when they are pressed
            for cut in (m3, m4): 
                if cut in line:
                    line=line[:line.index(cut)]
            # remove all other special chars
            try: 
                text += line[:line.index("[KeyName")]
            except ValueError:
                # no key name in there.
                text += line
                continue
            for part in line.split("[KeyName:")[1:]:
                try: 
                    text += part[part.index("]")+1:]
                except ValueError:
                    print(part, line)
        
        one, two, three = ngrams_in_data(data=text)
        return one, two, three
        
    def parse_config_0_0(self): 
        lines = self.config.splitlines()
        # the raw list of included ngrams.
        self.raw = []
        for l in lines[1:]:
            spaceidx = l.index(" ")
            weight = float(l[:spaceidx])
            datapath = l[spaceidx+1:]
            one, two, three = ngrams_in_datapath(datapath=datapath)
            self.raw.append((weight, (one, two, three)))

    def normalize_raw(self):
        """Normalize the raw ngrams and store them in self.one, self.two and self.three."""
        
        # normalize ngrams
        def _normalize(ngramlist):
            """normalize a list of ngrams.

            @param ngramlist: [(num, "ngram"), …]
            @return list with sum([num for num, ngram in ngrams]) == 1
            """
            total = sum((num for num, ngram in ngramlist))
            return [(num / total, ngram) for num, ngram in ngramlist]

        normalized = []
        for weight, ngrams in self.raw:
            normalized.append((weight,
                               [_normalize(ngram) for ngram in ngrams]))

        # weight them.
        self.one = collections.Counter()
        self.two = collections.Counter()
        self.three = collections.Counter()
        for weight, ngrams in normalized:
            one = ngrams[0]
            two = ngrams[1]
            three = ngrams[2]
            for norm, weighted in ((one, self.one),
                                   (two, self.two),
                                   (three, self.three)): 
                for num, ngram in norm:
                    try: weighted[ngram] += num*weight
                    except KeyError: weighted[ngram] = num*weight

    def weight_steno(self, factor=0.25):
        """weight shorthand ngrams higher by the factor than others, because they are seldomly at the rim between word-segments.

        profit from 100 years of professional writing experience."""
        # get the steno
        _steno = _steno_fix.split() + _steno_short.split() + _steno_letters.split() + _steno_eil_letters.split()
        stenotwo = set()
        for ngram in _steno:
            while ngram[1:]:
                stenotwo.add(ngram[:2])
                ngram = ngram[1:]
        stenothree = set()
        for ngram in _steno:
            while ngram[2:]:
                stenothree.add(ngram[:3])
                ngram = ngram[1:]

        #print(set(self.two.keys()).difference(stenotwo))
        #print(len(stenotwo)/len(self.two)) # → 0.008315677966101696

        # weight the ngrams
        for two in stenotwo:
            try:
                self.two[two] *= 1+factor
            except KeyError: pass
        for three in stenothree:
            try:
                self.three[three] *= 1+factor
            except KeyError: pass

    def weight_punctuation(self, factor=0.5, punctuation=".,"):
        """reduce the weight of punctuation marks by the factor."""
        # one grams
        for p in punctuation:
            try: self.one[p] *= 1-factor
            except KeyError: pass
        # 2grams: every one containing them.
        for key in self.two:
            for p in punctuation:
                if key and p in key:
                    self.two[key] *= 1-factor
        # 3grams: all that have punctuation as second or third letter.
        for key in self.three:
            for p in punctuation:
                if key and p in key[1:]:
                    self.three[key] *= 1-factor

    def save(self, one, two, three):
        """save the data to the files one, two and three (i.e. 1gramme.txt, 2gramm…).

        Plan: Add a total number of keystrokes parameter.
        @param one: path to the 1grams file.
        """
        for p, ngrams in ((one, self.one),
                          (two, self.two),
                          (three, self.three)): 
            gramlist = [(num, ngram) for ngram, num in ngrams.items()] 
            gramlist.sort()
            data = ""
            for num, ngram in reversed(gramlist):
                ngram = ngram.replace("\\", "\\\\")
                ngram = ngram.replace("\n", "\\n")
                data += str(num) + " " + ngram + "\n"
            with open(p, "w", encoding="utf-8") as f:
                f.write(data)


### Get all data: the main interface to this module

def get_all_data(data=None, letters=None, repeats=None, number_of_letters=None, number_of_bigrams=None, trigrams=None, number_of_trigrams=None, datapath=None, ngram_config_path=None, layout=None): 
    """Get letters, bigrams and trigrams.

    @param data: a string of text.
    @param ngram_config_path: the path to an ngram config file. If this is specified, all other values are ignored.
    @return: letters, number_of_letters, bigrams, number_of_bigrams, trigrams, number_of_trigrams
    """
    if ngram_config_path:
        ngrams = NGrams(ngram_config_path)
        return [(num, ngram) for ngram, num in ngrams.one.items()], len(ngrams.one), [(num, ngram) for ngram, num in ngrams.two.items()], len(ngrams.two), [(num, ngram) for ngram, num in ngrams.three.items()], len(ngrams.three)

    # if we get a datastring, we use it for everything.
    if datapath is not None:
        all_letters, all_repeats, all_trigrams = ngrams_in_filepath(datapath)
        # remove the ngrams that can't be typed, if the is the layout=layout parameter was used.
        letters = fix_impossible_ngrams(all_letters, layout=layout)
        bigrams = fix_impossible_ngrams(all_repeats, layout=layout)
        trigrams = fix_impossible_ngrams(all_trigrams, layout=layout)
        
        number_of_letters = sum([i for i, s in letters])
        number_of_bigrams = sum([i for i, s in bigrams])
        number_of_trigrams = sum([i for i, s in trigrams])
    
    elif data is not None:
        all_letters, all_repeats, all_trigrams = ngrams_in_data(data)
        # remove the ngrams that can't be typed, if the is the layout=layout parameter was used.
        letters = fix_impossible_ngrams(all_letters, layout=layout)
        bigrams = fix_impossible_ngrams(all_repeats, layout=layout)
        trigrams = fix_impossible_ngrams(all_trigrams, layout=layout)
    
        number_of_letters = sum([i for i, s in letters])
        number_of_bigrams = sum([i for i, s in bigrams])
        number_of_trigrams = sum([i for i, s in trigrams])

    # otherwise we get the missing values from the predefined files. 
    if letters is None or number_of_letters is None: 
        letterdata = clean_data(read_file("1-gramme.arne.txt"))
        all_letters = letters_in_file_precalculated(letterdata)
        letters = fix_impossible_ngrams(all_letters, layout=layout)
        number_of_letters = sum([i for i, s in letters])

    if repeats is None or number_of_bigrams is None: 
        bigramdata = clean_data(read_file("2-gramme.arne.txt"))
        all_bigrams = repeats_in_file_precalculated(bigramdata)
        bigrams = fix_impossible_ngrams(all_bigrams, layout=layout)
        number_of_bigrams = sum([i for i, s in bigrams])
    else: bigrams = repeats

    if trigrams is None or number_of_trigrams is None:
        trigramdata = clean_data(read_file("3-gramme.arne.txt"))
        all_trigrams = trigrams_in_file_precalculated(trigramdata)
        trigrams = fix_impossible_ngrams(all_trigrams, layout=layout)
        number_of_trigrams = sum([i for i, s in trigrams])

    letters_set = set(letters)
    bigrams_set = set(bigrams)
    trigrams_set = set(trigrams)
    return letters_set, number_of_letters, bigrams_set, number_of_bigrams, trigrams_set, number_of_trigrams


### Self test

def _test():
    from doctest import testmod
    testmod()

if __name__ == "__main__":
    import argparse
    parser = argparse.ArgumentParser(description="Read nGrams from files.")
    parser.add_argument("args")
    parser.add_argument("--conf", dest="config", help="The ngram config file to use.")
    parser.add_argument("--final", dest="final", action="store_true", help="Do ngrams adjustments for ngram-files which are going to be used directly, but which are not suitable for intermediate files (must be done only once).")
    parser.add_argument("--test", dest="test", action="store_true", help="Run the self-test.")
    parser.add_argument("--save1", dest="one", help="File for the 1-grams. You need conf, save1, save2 AND save3 for saving.")
    parser.add_argument("--save2", dest="two", help="File for the 2-grams.")
    parser.add_argument("--save3", dest="three", help="File for the 3-grams.")
    from sys import argv
    args = parser.parse_args(argv)
    if args.config:
        ngrams = NGrams(args.config)
    if args.final:
        ngrams.finalize()
    if args.config and args.one and args.two and args.three:
        ngrams.save(args.one, args.two, args.three)
    if args.test:
        _test()