""" A generator for reading CSV files.

Format is restricted to proper (quoted and escaped) CSV files

with a column for label and another for the text. The file may

contain other fields, but only the class label and the text is

used.

Args:

header: None or e sequence with two elements, specifying

the headers that correspond to label and

the text respectively. If None, header is

assumed to contain no headers.

"""

if path.endswith('.gz'):

import gzip

fp = gzip.open(path, 'rt')

elif path.endswith('.xz'):

import lzma

fp = lzma.open(path, 'rt')

elif path.endswith('.bz2'):

import bz2

fp = bz2.open(path, 'rt')

else:

fp = open(path, 'rt')

if header is None:

csvfp = csv.DictReader(fp, fieldnames=('label', 'text'),

delimiter=sep)

label_h, text_h = 'label', 'text'

else:

label_h, text_h = header

csvfp = csv.DictReader(fp, delimiter=sep)

for row in csvfp:

yield row[label_h], row[text_h]

bos="<", eos=">", suffix="", flatten=True):

""" For the given sequence s. Return all ngrams in range ngmin-ngmax.

spearator is useful for readability

bos/eos symbols are added to indicate beginning and end of seqence

suffix is an arbitrary string useful for distinguishing

in case differernt types of ngrams

if flatten is false, a list of lists is returned where the

first element contains the ngrams of size ngmin

and the last contains the ngrams of size ngmax

"""

# return a single dummy feature if there are no applicable ngrams

# probably resulting in a mojority-class classifier

if ngmax == 0 or (ngmax - ngmin < 0) :

return ['__dummy__']

ngrams = [[] for x in range(1, ngmax + 1)]

s = [bos] + s + [eos]

for i, ch in enumerate(s):

for ngsize in range(ngmin, ngmax + 1):

if (i + ngsize) <= len(s):

ngrams[ngsize - 1].append(

separator.join(s[i:i+ngsize]) + suffix)

if flatten:

ngrams = [ng for nglist in ngrams for ng in nglist]

def __init__(self, path=None,

num_data=None,

cat_data=None,

tokenizer=_TOKENIZER,

negative_class=None,

labels=[],

maxlen=_MAX_LEN, minlen=_MIN_LEN,

text_label="text",

class_label="label",

sep='\t'):

# _id is a dirty hack to identify the objec quickly

self._id = random.getrandbits(64) ^ int(10000*time.time())

self.maxlen = maxlen

self.text_label = text_label

self.class_label = class_label

self.delimiter = sep

self.minlen = minlen

self.texts = []

self.labels = []

self.num_data = num_data

self.cat_data = cat_data

self.num_features = None

self.cat_features = None

self.label_names = OrderedDict()

self.negative_class = negative_class

if negative_class:

self.label_names[negative_class] = 0

for l in labels:

if l not in self.labels:

self.label_names[l] = len(self.label_names)

self.tokenizer = tokenizer

if path:

self.load(path, num_data, cat_data)

def __eq__(self, other):

if isinstance(other, TextCData):

return self._id == other._id

else:

return False

def symbolic_labels(self, index=None):

label_names = np.array(list(self.label_names.keys()))

if index is None:

index = self.labels

return label_names[index]

def __len__(self):

return len(self.labels)

def update(self, texts, labels=None, labelstr=None, num_feats=None):

""" Update the data with given texts, labels and optionall

numeric features.

TODO: [cleanup] code is replicated in load()

"""

assert (labels is not None) or (labelstr is not None)

if self.num_features is not None:

assert num_feats is not None

self.num_features = np.vstack((self.num_features, num_feats))

self.texts.extend(texts)

if labelstr is None:

self.labels.extend(labels)

else:

for l in labelstr:

if l not in self.label_names:

self.label_names[l] = len(self.label_names)

self.labels.extend(

[self.label_names.get(l) for l in labelstr])

def load(self, path, num_data=None, cat_data=None):

labels_str = []

linen = 0

for l, t in read_csv(path,

header=(self.class_label, self.text_label),

sep=self.delimiter):

linen += 1

if len(t) < self.minlen or len(t) > self.maxlen:

warning("Skipping: line {}...".format(linen))

continue

labels_str.append(l)

self.texts.append(t)

if l not in self.label_names:

self.label_names[l] = len(self.label_names)

self.labels.extend(

[self.label_names.get(l) for l in labels_str])

if num_data is not None:

self.num_features = []

open_file = open

if num_data.endswith('.gz'):

open_file = gzip.open

with open_file(num_data, 'rt') as fp:

for line in fp:

if line.startswith('### '): continue

self.num_features.append([float(x)

for x in line.strip().split()])

assert len(self.labels) == len(self.num_features)

scaler = StandardScaler()

self.num_features = scaler.fit_transform(

np.array(self.num_features))

debug("Loaded extra features {} from {}.".format(

self.num_features.shape, self.num_data))

def stats(self, histogram=None, most_common=0):

labels_int = list(self.label_names.values())

labels_str = list(self.label_names.keys())

label_dist = Counter(self.labels)

fmt = '{:30s} {:>6d}' + ''.join(['{:10.2f}{:10.2f}{:>7d}{:>7d}']*2)

wlen_dist_all = []

clen_dist_all = []

for li in labels_int:

clen_dist = np.array([len(x) for i, x in enumerate(self.texts)\

if self.labels[i] == li])

wlen_dist = np.array([len(self.tokenizer(x)) \

for i, x in enumerate(self.texts) if self.labels[i] == li])

clen_dist_all.extend(clen_dist)

wlen_dist_all.extend(wlen_dist)

print(fmt.format('{}({}):'.format(labels_str[li], li),

label_dist[li],

clen_dist.mean(), clen_dist.std(),

clen_dist.min(), clen_dist.max(),

wlen_dist.mean(), wlen_dist.std(),

wlen_dist.min(), wlen_dist.max()))

clen_dist_all = np.array(clen_dist_all)

wlen_dist_all = np.array(wlen_dist_all)

print(fmt.format('Total:', sum(label_dist.values()),

clen_dist_all.mean(), clen_dist_all.std(),

clen_dist_all.min(), clen_dist_all.max(),

wlen_dist_all.mean(), wlen_dist_all.std(),

wlen_dist_all.min(), wlen_dist_all.max()))

if most_common:

tok_counter = [Counter() for _ in labels_int]

ch_counter = [Counter() for _ in labels_int]

for i, txt in enumerate(self.texts):

tok_counter[self.labels[i]].update(self.tokenizer(txt.lower()))

# only char bigrams - generally useful for detecting odd things

# at te beginning or end of documents.

ch_counter[self.labels[i]].update(get_ngrams(list(txt.lower()), 2,2))

for li in labels_int:

lname = list(self.label_names.keys())[li]

print(lname, 'tokens',

[x[0] for x in tok_counter[li].most_common(most_common)])

print(lname, 'chars',

[x[0] for x in ch_counter[li].most_common(most_common)])

if histogram:

import matplotlib.pyplot as plt

_, plts = plt.subplots(nrows=1,ncols=2)

plts[0].hist(clen_dist_all, bins=100)

plts[0].set_title("Char")

plts[1].hist(wlen_dist_all, bins=100)

plts[1].set_title("Word")

if isinstance(histogram, str):

fig = plt.gcf()

fig.savefig(histogram)

else:

plt.show()

def copy(self):

return self.subset(range(len(self)))

def subset(self, index):

subset = TextCData()

for attr in vars(self):

if not attr.startswith('_'):

setattr(subset, attr, getattr(self, attr))

subset.texts = [self.texts[i] for i in index]

subset.labels = [self.labels[i] for i in index]

if self.num_features is not None:

subset.num_features = self.num_features[index]

""" A generator that returns random drwas from a parameter space.

param_space is a sequence (name, type, range)

where type is either 'numeric' or 'categorical',

and range is a triple (start, stop, step) for

numeric parameters, and another sequence of

parameter values to explore.

the function keeps the set of returned parameter values, and

if an already returned parameter set is drawn max_iter times,

it terminates.

"""

seen = set()

rejected = 0

while True:

params = []

for param, type_, seq in param_space:

if 'numeric'.startswith(type_):

try:

start, stop, step = seq

p_range = np.arange(start, stop + step, step).tolist()

except:

start, stop = seq

p_range = np.arange(start, stop + 1, 1).tolist()

rval = random.choice(p_range)

params.append((param, rval))

elif 'categorical'.startswith(type_):

params.append((param, random.choice(seq)))

param_hash = md5(str(params).encode()).digest()

if param_hash not in seen:

seen.add(param_hash)

rejected = 0

yield dict(params)

else:

rejected += 1

if rejected == max_iter:

info("More than {} iterations with already drawn parameters. "

"The search space is probably exhausted".format(max_iter))

p_vals = []

for param, type_, seq in param_space:

if 'numeric'.startswith(type_):

try:

start, stop, step = seq

p_range = np.arange(start, stop + step, step).tolist()

except:

start, stop = seq

p_range = np.arange(start, stop + 1, 1).tolist()

p_vals.append(p_range)

elif 'categorical'.startswith(type_):

p_vals.append(seq)

for params in itertools.product(*p_vals):

paramstr = str(paramstr)

try:

with open(paramster, 'r') as fp:

params = eval(fp.read().strip())

except:

try:

params = eval(paramstr)

except:

params = '(' + paramstr + ')'

with open(filename, 'r') as fp:

for line in fp:

if (len(line) > 1 and line[0] != '#'):

log_data = json.loads(line)

tuned_params = log_data['params']

model_params = log_data['model_params']

scores = log_data['scores']

PARAMS = {

'name': 'textc',

'baseline': 'random',

'adapt_thresh': 0.0,

}

""" Base text classifier class.

"""

def __init__(self, arg_parser=True, **params):

self._set_defaults()

if arg_parser:

self.arg_parser = self._setup_arg_parser()

self._trained = True # Always for baselines

self.set_params(**params)

def _set_defaults(self):

for k,v in self.PARAMS.items():

setattr(self, k, v)

def get_params(self):

return {k:getattr(self, k) for k in self.PARAMS \

if not k.startswith('_')}

def set_params_str(self, s):

val_dict = dict()

for pval in s.split(','):

p, val = pval.split('=')

try:

val_dict[p] = eval(val)

except:

val_dict[p] = val

self.set_params(**val_dict)

def set_params(self, **kwargs):

for k, v in kwargs.items():

if k not in self.PARAMS:

warning("Ignoring unknown parameter {}.".format(k))

else:

old_v = getattr(self, k)

if v != k:

setattr(self, k, v)

self._trained = False

def fit(self, train=None):

""" Fit the model.

This should be overridden in the real text classifer classes.

"""

self._trained = True

def _predict_k_fold(self, train, k=10, decision_func=False):

#TODO: pass the parameter k

splits = StratifiedKFold(n_splits=k, shuffle=True)

folds = list(splits.split(train.texts, train.labels))

predictions = [None for _ in range(len(train.labels))]

if decision_func:

decision_val = [None for _ in range(len(train.labels))]

for ti, vi in folds:

if decision_func:

pred, dec, _, _ = self.predict(test=train.subset(vi),

train=train.subset(ti), decision_func=True)

else:

pred = self.predict(test=train.subset(vi),

train=train.subset(ti))

for i, j in enumerate(vi):

predictions[j] = pred[i]

if decision_func:

decision_val[j] = dec[i]

if decision_func:

return predictions, decision_val

return predictions

def _predict(self, test, train=None, decision_func=False):

""" Return predictions for the testset. Implements baselines.

Here we only return the numeric indeices for the labels. To

obtain symbolic names, use predict().

This should be overridden in the real text classifer classes.

"""

if train and not self._trained:

self.fit(train)

label_set = list(train.label_names.values())

test_len = len(test.texts)

if self.baseline == 'random':

predictions = np.random.choice(label_set, size=test_len)

elif self.baseline == 'majority':

predictions = test_len * [Counter(train.labels).most_common(1)[0][0]]

elif self.baseline == 'random_sample':

prob = Counter(train.labels)

prob = np.array([prob[l] for l in label_set]) / sum(prob.values())

predictions = np.random.choice(label_set,

size=test_len,

p=prob)

if decision_func:

return predictions, None

return predictions

def predict(self, test=None, train=None, label_names=False,

decision_func=False,

score=False, conf_mat=False):

if test is None:

predictions = self._predict_k_fold(train, decision_func=decision_func)

else:

if self.adapt_thresh != 0.0:

info("Adaptive predictions enabled")

if train is None: train = self._training_data

assert train is not None

pred, dec_val = self._predict(test, train, decision_func=True)

texts_add, labels_add, num_add = [], [], []

for i, v in enumerate(dec_val):

if len(dec_val.shape) == 1: # binary

pick = abs(v) > self.adapt_thresh

else:

pick = len(np.argwhere(v > self.adapt_thresh).flatten()) == 1

if pick:

texts_add.append(test.texts[i])

labels_add.append(pred[i])

if test.num_features is not None:

num_add.append(test.num_features[i])

if not num_add: num_add = None

train_aug = train.copy()

train_aug.update(texts_add, labels_add, num_feats=num_add)

info("Retraining with {} new trainng instances".format(len(texts_add)))

predictions = self._predict(test, train_aug, decision_func=decision_func)

else:

predictions = self._predict(test, train, decision_func=decision_func)

decision_val = None

if decision_func:

predictions, decision_val = predictions

if label_names and self._training_data.label_names:

label_names = list(self._training_data.label_names)

predictions = [label_names[i] for i in predictions]

sc = None

if score:

sc = self._score(test.symbolic_labels(), predictions,

negative_class=train.negative_class)

cf = None

if conf_mat:

from sklearn.metrics import confusion_matrix

cf = confusion_matrix(test.symbolic_labels(), predictions)

return predictions, decision_val, sc, cf

def _score(self, gold, pred,

scores={'precision', 'recall', 'f1-score'},

negative_class=None,

average=None):

""" Return the score for the testset.

"""

from sklearn.metrics import precision_recall_fscore_support as prfs

if average is None:

average = 'macro'

if negative_class:

average = 'binary'

scores = [sc \

if ':' in sc or \

sc not in {'precision', 'recall', 'f1-score'}\

else ':'.join((sc, average))\

for sc in scores]

scores = {k:None for k in scores}

for sc_avg in list(scores):

if ':' in sc_avg:

sc, avg = sc_avg.split(':')

else:

sc = sc_avg

avg = None

if scores[sc_avg] is not None:

continue

if sc not in {'precision', 'recall', 'f1-score', 'accuracy'}:

warning("Skipping unknown score `{}'.".format(sc))

continue

if sc in {'precision', 'recall', 'f1-score'}:

if avg not in {'binary', 'micro', 'macro'}:

warning("Skipping `{}': unknown avgeraging method."

.format(sc_avg))

continue

p, r, f, _ = prfs(gold, pred, average=avg)

scores[':'.join(('precision', avg))] = p

scores[':'.join(('recall', avg))] = r

scores[':'.join(('f1-score', avg))] = f

if sc == 'accuracy':

from sklearn.metrics import accuracy_score

scores['accuracy'] = accuracy_score(gold, pred)

return {k:v for k,v in scores.items()}

def score(self, test=None, train=None,

scores={'precision', 'recall', 'f1-score'},

average=None):

""" Return the score for the testset.

"""

from sklearn.metrics import precision_recall_fscore_support as prfs

pred, _, _, _ = self.predict(test, train)

if test:

gold = test.labels

else:

gold = train.labels

return self._score(gold, pred, scores=scores,

negative_class=train.negative_class,

average=average)

def ttt(self, method):

if method == 'grid':

param_iter = grid_iter(params)

def tune(self, param_space,

train, dev=None,

method=None,

round_digits=None,

max_iter=-1,

optimize=None,

scores=None,

k=None, split_r=0.2, n_splits=1,

save=sys.stdout, skip_params=None):

""" Fit and evaluate a model repeatedly,

the best one based on params.

Args:

param_space: a dict-like object whose keys are the

parameter names and values are tuples of

(type, range) or (type, list). 'type' is one of

'numeric', 'categorical'. For numeric

parameters the range is defined as [begin, end]

or [begin, end, step] is required. The former

is interpreted as range of integers in the

range [begin, end]. For 'categorical'

parameters, a list of values is required.

train: TextCData object used for training. If no test

set is given, it is used both for training

and testing

dev: same as 'train', used as development set

method: search method: 'grid', 'random' or a iterable

that returns values from the option range.

max_iter: maximum number of fit/predict/eval iterations

k: Use k-fold CV.

split_r: ratio of the held-out set, ignored if test set or

argument `k' is given

n_splits: number of splits, sklearn StratifiedShuffleSplit is

used for n-splits

save: file-like object save the results after each iteration

skip_params: A sequence of dictionaries containing individual

parameter values to skip. Useful for resuming a search.

optimize: optimize using given metric.

scores: the scores to report/log.

"""

if method is None: method = 'grid'

if optimize is None: optimize = 'f1-score:macro'

if scores is None: scores = ['precision', 'recall', 'f1-score']

param_space = [p for p in _str_to_param_space(param_space) \

if p[0] in self.get_params()]

# if not param_space:

# info('Tunable parameter list is empty')

# return

if method == 'grid':

param_iter = grid_iter(param_space)

elif method == 'random':

param_iter = random_iter(param_space)

else:

param_iter = method(params)

if dev is not None:

tune_str = "development data".format()

else:

if k:

splits = StratifiedKFold(n_splits=k, shuffle=True)

tune_str = "{}-fold CV".format(k)

else:

splits = StratifiedShuffleSplit(

n_splits=n_splits, test_size=split_r)

tune_str = "{} splits of {} ratio".format(n_splits, split_r)

trn_splits, val_splits = [], []

for ti, vi in splits.split(train.texts, train.labels):

trn_splits.append(train.subset(ti))

val_splits.append(train.subset(vi))

def param_to_str(p_dict):

p_list = sorted(tuple(p_dict.items()))

p_fmt = "{}={} " * len(p_list)

return p_fmt.format(*[x for t in p_list for x in t])

if skip_params:

skip_params = set([md5(param_to_str(p).encode()).digest() \

for p in skip_params])

else:

skip_params = set()

best_mean = 0.0

best_sc = None

best_param = None

for param in param_iter:

scores = []

p_str = param_to_str(param)

p_hash = md5(p_str.encode()).digest()

if p_hash in skip_params:

info('Skipping: {}'.format(p_str))

continue

info('Tuning with {}: {}'.format(tune_str, p_str))

self.set_params(**param)

if dev is not None:

sc = self.score(dev, train=train)

scores.append(sc)

else:

for i in range(len(trn_splits)):

sc = self.score(val_splits[i], train=trn_splits[i])

scores.append(sc)

sc_names = scores[0].keys()

scores = {k:[sc[k] for sc in scores] for k in sc_names}

sc_mean = np.array(scores[optimize]).mean()

if sc_mean > best_mean:

best_mean = sc_mean

best_sc = scores

best_param = param

if save:

json.dump({'params': param,

'model_params': self.get_params(),

'scores': scores},

save, ensure_ascii=False)

print('', file=save, flush=True)

max_iter -= 1

if max_iter == 0:

break

stop_file = '.stop-tune' + str(os.getpid())

if os.path.isfile(stop_file):

os.remove(stop_file)

break

return best_param, best_sc

def _setup_arg_parser(self):

import argparse

ap = argparse.ArgumentParser()

ap.add_argument('--input', '-i', help="Path to the training data")

ap.add_argument('--test', '-t', help="Path to the testing data")

ap.add_argument('--unlabeled-input', '-u',

help="Path to unlabeled training data")

ap.add_argument('--unlabeled-num-input', '-U',

help="Path to numeric features for the unlabeled data")

ap.add_argument('--input-numeric', '-N',

help="Path to (optional) additional numeric features")

ap.add_argument('--test-numeric', '-M',

help="Path to (optional) additional numeric features")

ap.add_argument('--class-label', '-L', default='label',

help="Label of the column corrsponding to the class.")

ap.add_argument('--text-label', '-T', default='text',

help="Label of the column corrsponding to the text.")

ap.add_argument('--delimiter', '-D', default='\t',

help="Delimiter used in input files")

ap.add_argument('--output', '-o', default='-',

help="Output file. `-' means stdout.")

ap.add_argument('--negative-class',

help="The negative class label.")

ap.set_defaults(command='tune')

subp = ap.add_subparsers(help="Command")

tunep = subp.add_parser('tune')

tunep.set_defaults(command='tune')

tunep.add_argument('params',

help=('A string or a filename defining parameter space '

'to be searched.'

'String must be interpretable by python eval() '

'as a sequence whose members are triples of '

'(name, type, values). '

'If "type" is "numeric", values should specify a range '

'(start, stop, step), otherwise ("categorical"), '

'a sequence of values. '

'Example: (("C", "real", (0.5, 1.5, 0.1)), '

'("lowercase", "cat", ("word", "char", "both"))). '

'If "params" is a readable file, the string is read '

'from the file.'

))

tunep.add_argument('--search-method', '-s', choices=('grid', 'random'),

default='grid', help="Method used for hyper parmeter search")

tunep.add_argument('--optimize',

help=('A string of the form "score" or "score:averaging" '

'Currently supported scores are '

'accuracy, precision, recall, f1-score, '

'and supproted averaging methods are '

'micro, macro and binary. '

'If averaging is not specified it is set to '

'macro or binary depending on the classification task.'

))

tunep.add_argument('--max-iter', '-m', type=int, default=-1,

help=('Maximum number of hyperparameter combinations '

'to compare. Default (-1) means until '

'the search space is exhausted'))

tunep.add_argument('--k-folds', '-k', type=int, default=None,

help=('Use k-fold cross validation. '

'Ignored if -t is given.'))

tunep.add_argument('--test-ratio', '-r', type=float, default=0.2,

help=('Ratio of held-out data. '

'Ignored if --test option is given'))

tunep.add_argument('--n-splits', '-n', type=int, default=1,

help=('Number of splits. Ignored if -t or -k is given'))

tunep.add_argument('--save', '-S', metavar='LOGFILE',

help=('Save intermediate parameter values and scores '

'to given log file. '

'Use - for standard output'))

tunep.add_argument('--resume-from', '-R', metavar='LOGFILE',

help=('Resume tuning, skipping the parameters that '

'are logged in the log file.'))

predp = subp.add_parser('predict')

predp.set_defaults(command='predict')

predp.add_argument('params',

help=('A string that can be interpreted by python eval() '

'as a sequence whose members are pairs '

'of, parameter=value'))

predp.add_argument('--score', action='store_true',

help='Also print out the scores.')

predp.add_argument('--only-score', action='store_true',

help='Print out only the scores.')

predp.add_argument('--conf-matrix', action='store_true',

help='Also print out the confusion matrix.')

predp.add_argument('--output-decision-value', action='store_true',

help=('Also return decision function values'))

testp = subp.add_parser('score')

testp.set_defaults(command='score')

testp.add_argument('params',

help=('A string that can be interpreted by python eval() '

'as a sequence whose members are pairs '

'of, parameter=value'))

#TODO: this should not be here

statsp = subp.add_parser('stats')

statsp.set_defaults(command='stats')

statsp.add_argument('--histogram', '-H',

const=True, metavar='FILE', nargs='?',

help="Plot a histogram, optionally to FILE.")

statsp.add_argument('--most-common', type=int, default=0,

help="Also print most-common tokens")