def train(self,trainfile, devfile, multifolds=1,as_text=False):
p = StdOutFilter()
p.start()
train_feats, _, _, _, _ = read_conll(trainfile,genre_pat=self.genre_pat,as_text=as_text)
train_for_rnn, scalers = feats2rnn(train_feats)
dev_feats, _, _, _, _ = read_conll(devfile,genre_pat=self.genre_pat,as_text=as_text)
dev_for_rnn, scalers = feats2rnn(dev_feats)
with io.open(self.corpus_dir + os.sep + "dev."+self.ext,'w',encoding="utf8",newline="\n") as f:
f.write(dev_for_rnn)
# NCRFpp expects a test file, we reuse dev
with io.open(self.corpus_dir + os.sep + "test."+self.ext,'w',encoding="utf8",newline="\n") as f:
f.write(dev_for_rnn)
if multifolds > 1:
train_chunks, test_chunks = self.split_dataset(train_for_rnn,multifolds=multifolds)
all_preds = []
all_labs = []
for i in range(opts.multifolds):
with io.open(self.corpus_dir + os.sep + "train."+self.ext,'w',encoding="utf8",newline="\n") as f:
f.write(train_chunks[i])
with io.open(self.corpus_dir + os.sep + "raw."+self.ext,'w',encoding="utf8",newline="\n") as f:
f.write(test_chunks[i])
# TRAIN ON FOLD
sys.stderr.write("\no Training on fold " + str(i+1)+"/" + str(multifolds) + "\n")
config = ncrf_dir + os.sep + self.segtype + ".train"+self.auto+".config"
ncrf(config)
# PREDICT
config = ncrf_dir + os.sep + self.segtype + ".decode"+self.auto+".config"
ncrf(config,status="decode")
labs, scores = self.read_preds()
all_labs += labs
all_preds += scores
# SERIALIZE MULTITRAIN PREDS
with io.open(self.corpus_dir + os.sep + corpus + self.auto + "_multitrain.tab",'w',newline='\n') as f:
for j, pred in enumerate(all_preds):
if self.conn:
lab = all_labs[j]
else:
lab = 1 if pred >= 0.15 else 0 # Arbitrary threshold, we will only use prob as feature for metalearner
f.write(str(lab) + "\t" + str(pred) + "\n")
# TRAIN MODEL ON ALL TRAIN
sys.stderr.write("\no Training on full train set\n")
with io.open(self.corpus_dir + os.sep + "train."+self.ext,'w',encoding="utf8",newline="\n") as f:
f.write(train_for_rnn)
config = ncrf_dir + os.sep + self.segtype + ".train"+self.auto+".config"
ncrf(config)
p.end()
def predict(self,testfile,as_text=True): test_feats, _, _, _, _ = read_conll(testfile,genre_pat=self.genre_pat,as_text=as_text) test_for_rnn, scalers = feats2rnn(test_feats) with io.open(self.corpus_dir + os.sep + "raw."+self.ext,'w',encoding="utf8",newline="\n") as f: f.write(test_for_rnn) p = StdOutFilter() p.start() config = ncrf_dir + os.sep + self.segtype+".decode"+self.auto+".config" ncrf(config,status="decode") p.end() labs, probas = self.read_preds() return zip(labs,probas)
udpipe = NLTKSentencer(lang=corpus[:3])
conllu = io.open(topred_file, encoding="utf8").read()
filtered = []
for line in conllu.split("\n"):
if "\t" in line:
fields = line.split("\t")
if "-" in fields[0]:
continue
filtered.append(line)
conllu = "\n".join(filtered) + "\n"
pred = [x[0] for x in udpipe.predict(conllu)]
# pred = udpipe.predict(conllu)
gold_feats, _, toks, _, _ = read_conll(gold_file,
genre_pat="^(..)",
mode="sent",
as_text=False)
gold = [int(t['wid'] == 1) for t in gold_feats]
conf_mat = confusion_matrix(gold, pred)
sys.stderr.write("Evaluating on " + corpus + "\n")
sys.stderr.write(str(conf_mat) + "\n")
true_positive = conf_mat[1][1]
false_positive = conf_mat[0][1]
false_negative = conf_mat[1][0]
prec = true_positive / (true_positive + false_positive)
rec = true_positive / (true_positive + false_negative)
f1 = 2 * prec * rec / (prec + rec)
sys.stderr.write("P: " + str(prec) + "\n")
sys.stderr.write("R: " + str(rec) + "\n")
for corpus in corpora:
corpus_start_time = time.time()
train = glob(opts.data_dir + os.sep + corpus + os.sep + corpus +
"_train.conll")[0]
dev = glob(opts.data_dir + os.sep + corpus + os.sep + corpus +
"_dev.conll")[0]
test = glob(opts.data_dir + os.sep + corpus + os.sep + corpus +
"_test.conll")[0]
feats = DepFeatures()
from conll_reader import read_conll
outputs = read_conll(dev)
train = outputs[0]
# Extract features from file
sys.stderr.write("\no Extracting features from training corpus " +
corpus + "\n")
train_feats = feats.extract_depfeatures(train)
elapsed = time.time() - corpus_start_time
sys.stderr.write("\nTime training on corpus:\n")
sys.stderr.write(str(timedelta(seconds=elapsed)) + "\n\n")
sys.stderr.write("\nTotal time:\n")
elapsed = time.time() - start_time
sys.stderr.write(str(timedelta(seconds=elapsed)) + "\n\n")
if opts.mode == "train":
# Get Connectives: Raw frequency and ratio of the connectives found in the training data
conn = connective.train(train_path)
if opts.eval_test:
dev_path = test_path
with open(dev_path, 'r', encoding='utf-8') as f:
lines = f.read()
# Prediction on the devset
pred_labels_probs = connective.predict(lines)
resps = [
tok[1] if float(tok[2]) > 0.5 else "_" for tok in pred_labels_probs
]
train_feats, _, _, _, _ = read_conll(dev_path, as_text=False)
gold = io.open(dev_path, encoding="utf8").read()
lines = gold.split("\n")
processed = []
i = 0
for line in lines:
if "\t" in line:
fields = line.split('\t')
if "-" in fields[0]:
processed.append(line)
continue
else:
fields[-1] = resps[i]
processed.append("\t".join(fields))
i += 1
else:
def read_conll_sentbreak(self, infile, neighborwindowsize=5,as_text=True,cut=True,do_tag=True,multitrain=False):
global TRAIN_LIMIT
# read data from conll_reader
vowels = "AEIOUaeiouéèàáíìúùòóаэыуояеёюи"
numeric_entries = []
nonnumeric_entries = []
goldseg_entries = []
if as_text:
conllu_in = infile
else:
conllu_in = io.open(infile, 'r', encoding='utf8').read()
# # Reduce data if too large
# if not cut and conllu_in.count("\n") > 100000 and multitrain:
# sys.stderr.write("o Data too large; forcing cut and turning off multitraining\n")
# cut = True
# TRAIN_LIMIT = 100000
# if cut:
# conllu_in = shuffle_cut_conllu(conllu_in,limit=TRAIN_LIMIT)
train_feats,_,_,_,_ = read_conll(conllu_in,mode="seg",genre_pat=None,as_text=True,cap=None,char_bytes=False)
featurekeys = ["label", "word", "pos", "cpos", "head", "head_dist", "deprel", "case", "tok_len", "depchunk", "conj", "s_len", "s_type", "sent_doc_percentile", "parentclauses"]
for lnum, line in enumerate(train_feats):
lfeatures = [line[x] for x in featurekeys]
lfeatures[0] = int(lfeatures[0]!="_")
firstletter = str(lfeatures[1][0].encode("utf8")[0]) if self.lang == "zho" else lfeatures[1][0]
firstisupper = int (firstletter.upper() == firstletter)
firstisconsonant = len(re.findall('[^'+vowels+']', firstletter))
firstisvowel = len(re.findall('['+vowels+']', firstletter))
firstisdigit = len(re.findall('[0-9]', firstletter))
firstisspecial = len(re.findall('[^A-Za-z0-9]', firstletter))
lastletter = str(lfeatures[1][-1].encode("utf8")[-1]) if self.lang == "zho" else lfeatures[1][-1]
lastisupper = int(lastletter.upper() == lastletter)
lastisconsonant = len(re.findall('[^'+vowels+']', lastletter))
lastisvowel = len(re.findall('['+vowels+']', lastletter))
lastisdigit = len(re.findall('[0-9]', lastletter))
lastisspecial = len(re.findall('[^A-Za-z0-9]', lastletter))
numconsonants = len(re.findall('[^'+vowels+']', lfeatures[1]))
numvowels = len(re.findall('['+vowels+']', lfeatures[1]))
numdigits = len(re.findall('[0-9]', lfeatures[1]))
numspecials = len(re.findall('[^A-Za-z0-9]', lfeatures[1]))
numeric_entries.append(
[
numconsonants, numvowels, numdigits, numspecials,
firstisupper, firstisconsonant, firstisvowel, firstisdigit, firstisspecial,
lastisupper, lastisconsonant, lastisspecial,
lfeatures[4], lfeatures[5], lfeatures[8], lfeatures[11], lfeatures[13]
])
nonnumeric_entries.append([lfeatures[2],lfeatures[3], firstletter,lastletter,
lfeatures[6], lfeatures[7], lfeatures[9], lfeatures[10], lfeatures[12], re.sub(r'^([^\|]*\|[^\|]*)\|.*', r'\1', lfeatures[14])
])
goldseg_entries.append(lfeatures[0])
# featurekeys = ["label", "word", "pos", "cpos", "head", "head_dist", "deprel", "case", "tok_len", 9 "depchunk", "conj", 11"s_len", "s_type", "sent_doc_percentile", "parentclauses"]
numeric_colnames = ['numconsonants', 'numvowels', 'numdigits', 'numspecials',
'firstisupper', 'firstisconsonant', 'firstisvowel', 'firstisdigit','firstisspecial',
'lastisupper', 'lastisconsonant','lastisspecial',
featurekeys[4], featurekeys[5], featurekeys[8], featurekeys[11], featurekeys[13]
]
nonnumeric_colnames = ['gold_pos', 'gold_cpos','firstletter', 'lastletter',
featurekeys[6], featurekeys[7], featurekeys[9], featurekeys[10], featurekeys[12], featurekeys[14]
]
numeric_entries = np.array(numeric_entries, dtype=np.float32)
nonnumeric_entries = np.array(nonnumeric_entries)
# Dummy multi vectors cattodummy is much better
unigram_entries, unigram_colnames = self.cattodummy(nonnumeric_entries, nonnumeric_colnames, numeric_entries, numeric_colnames)
sys.stderr.write("o unigram dataframe ready\n")
if neighborwindowsize >= 3:
sys.stderr.write("o duplicating to %d-gram...\n" %neighborwindowsize)
prev_entries, prev_colnames = self.copyforprevnext(unigram_entries, unigram_colnames, 'prev')
next_entries, next_colnames = self.copyforprevnext(unigram_entries, unigram_colnames, 'next')
ngram_entries, ngram_colnames = self.numpyconcat2df(prev_entries, prev_colnames, unigram_entries, unigram_colnames)
ngram_entries, ngram_colnames = self.numpyconcat2df(ngram_entries,ngram_colnames, next_entries, next_colnames)
del prev_entries, prev_colnames, next_colnames, next_entries
if neighborwindowsize >=5:
prevprev_entries, prevprev_colnames = self.copyforprevnext(unigram_entries, unigram_colnames, 'prevprev')
nextnext_entries, nextnext_colnames = self.copyforprevnext(unigram_entries, unigram_colnames, 'nextnext')
ngram_entries, ngram_colnames = self.numpyconcat2df(prevprev_entries, prevprev_colnames, ngram_entries,
ngram_colnames)
ngram_entries, ngram_colnames = self.numpyconcat2df(ngram_entries, ngram_colnames, nextnext_entries,
nextnext_colnames)
del prevprev_entries, prevprev_colnames, nextnext_colnames, nextnext_entries
if neighborwindowsize == 7:
prevprevprev_entries, prevprevprev_colnames = self.copyforprevnext(unigram_entries, unigram_colnames, 'prevprevprev')
nextnextnext_entries, nextnextnext_colnames = self.copyforprevnext(unigram_entries, unigram_colnames, 'nextnextnext')
ngram_entries, ngram_colnames = self.numpyconcat2df(prevprevprev_entries, prevprevprev_colnames, ngram_entries,
ngram_colnames)
ngram_entries, ngram_colnames = self.numpyconcat2df(ngram_entries, ngram_colnames, nextnextnext_entries,
nextnextnext_colnames)
del prevprevprev_entries, prevprevprev_colnames, nextnextnext_colnames, nextnextnext_entries
else:
ngram_entries, ngram_colnames = unigram_entries, unigram_colnames
del unigram_entries, numeric_colnames, numeric_entries
return ngram_entries, ngram_colnames, goldseg_entries, unigram_colnames
# Run test
segmenter = LRSegmenter(lang=lang,model=corpusname,windowsize=args.windowsize)
if args.verbose:
segmenter.verbose = True
if segmenter.verbose:
sys.stderr.write("o Processing corpus "+corpusname+"\n")
if args.mode == "train":
# When running from CLI, we always train (predict mode is done on imported class)
segmenter.train(data_folder + os.sep+ corpusname + os.sep + corpusname + "_train.conll",as_text=False,standardization=args.standardization,multitrain=args.multitrain)
# Now evaluate model
predictions, probas = zip(*segmenter.predict(data_folder + os.sep+ corpusname + os.sep +corpusname + "_dev.conll",
as_text=False,standardization=args.standardization,do_tag=True))
# Get gold labels for comparison
conllu_in = io.open(data_folder + os.sep+ corpusname + os.sep +corpusname + "_dev.conll", 'r', encoding='utf8').read()
devfeats,_,_,_,_ = read_conll(conllu_in, mode="seg", genre_pat=None, as_text=True, cap=None, char_bytes=False)
labels = [int(x["label"]!='_') for x in devfeats]
# give dev F1 score
from sklearn.metrics import classification_report, confusion_matrix
print(classification_report(labels, predictions, digits=6))
print(confusion_matrix(labels, predictions))
elapsed = time.time() - start_time
sys.stderr.write(str(timedelta(seconds=elapsed)) + "\n\n")
data_dir = os.path.abspath(lib + os.sep + ".." + os.sep + ".." + os.sep +
"data_parsed")
corpora = os.listdir(data_dir)
corpora = [c for c in corpora if os.path.isdir(os.path.join(c, data_dir))]
log = io.open("baseline_parsed.log", 'w', encoding="utf8")
table_out = []
all_f = []
for corpus in corpora:
if "pdtb" in corpus:
continue
dev = glob(data_dir + os.sep + corpus + os.sep + "*_test.conll")[0]
train_feats, vocab, toks, firsts, lasts = read_conll(dev)
labels = [int(t["label"] != "_") for t in train_feats]
baseline_preds = [int(t["wid"] == 1) for t in train_feats]
conf_mat = confusion_matrix(labels, baseline_preds)
true_positive = conf_mat[1][1]
false_positive = conf_mat[0][1]
false_negative = conf_mat[1][0]
prec = true_positive / (true_positive + false_positive)
rec = true_positive / (true_positive + false_negative)
f1 = 2 * prec * rec / (prec + rec)
log.write("corpus: " + corpus + "\n")
log.write("=" * 10 + "\n")
log.write(str(confusion_matrix(labels, baseline_preds)) + "\n")
log.write("P: " + str(prec) + "\n")
log.write("R: " + str(rec) + "\n")
log.write("F1: " + str(f1) + "\n\n")
def read_data(self, infile, size, as_text, rare_thresh, chosen_feats=None):
cap = 3 * size if size is not None else None
train_feats, vocab, toks, firsts, lasts = read_conll(
infile,
genre_pat=self.genre_pat,
mode="seg",
cap=cap,
char_bytes=self.lang == "zho",
as_text=as_text)
vocab = Counter(vocab)
top_n_words = vocab.most_common(rare_thresh)
top_n_words, _ = zip(*top_n_words)
for tok in train_feats:
if tok["word"] not in top_n_words:
tok["word"] = tok["pos"]
tokens_by_abs_id = self.traverse_trees(train_feats)
data, headers = self.n_gram(train_feats, tokens_by_abs_id)
# Features to use for all n-gram tokens
num_labels = [
"head_dist", "left_span", "right_span", "samepar_left", "tok_len"
]
cat_labels = [
"case", "closest_left", "closest_right", "deprel", "farthest_left",
"farthest_right", "pos", "word", "morph", "cpos", "depchunk"
]
pref_cat = []
pref_num = []
for pref in ["mn2", "mn1", "par", "par_par", "pl1", "pl2"]:
pref_cat += [pref + "_" + h for h in cat_labels]
pref_num += [pref + "_" + h for h in num_labels]
# Features only needed for node token
cat_labels += [
"genre"
] + pref_cat #+ ["heading_first","heading_last"]#+ ["s_type"]
num_labels += [
"dist2end", "sent_doc_percentile", "tok_id", "wid", "quote", "rank"
] + pref_num # + ["bracket"]
num_labels += ["par_quote",
"par_par_quote"] #,"par_bracket","par_par_bracket"]
# Use specific feature subset
if chosen_feats is not None:
new_cat = []
new_num = []
for feat in chosen_feats:
if feat in cat_labels:
new_cat.append(feat)
elif feat in num_labels:
new_num.append(feat)
cat_labels = new_cat
num_labels = new_num
data = pd.DataFrame(data, columns=headers)
data_encoded, multicol_dict = self.multicol_fit_transform(
data, pd.Index(cat_labels))
data_x = data_encoded[cat_labels + num_labels].values
data_y = np.where(data_encoded['label'] == "_", 0, 1)
return data_encoded, data_x, data_y, cat_labels, num_labels, multicol_dict, firsts, lasts, top_n_words
def predict(self, infile, eval_gold=False, as_text=True):
"""
Predict sentence splits using an existing model
:param infile: File in DISRPT shared task *.tok or *.conll format (sentence breaks will be ignored in .conll)
:param eval_gold: Whether to score the prediction; only applicable if using a gold .conll file as input
:param genre_pat: A regex pattern identifying the document genre from document name comments
:param as_text: Boolean, whether the input is a string, rather than a file name to read
:return: tokenwise binary prediction vector if eval_gold is False, otherwise prints evaluation metrics and diff to gold
"""
if self.model is None: # Try default model location
model_path = ".." + os.sep + ".." + os.sep + "models" + os.sep + "subtreeseg.pkl"
else:
model_path = self.model
clf, num_labels, cat_labels, multicol_dict, top_n_words, firsts, lasts = joblib.load(
model_path)
feats, _, toks, _, _ = read_conll(infile,
genre_pat=self.genre_pat,
mode="seg",
as_text=as_text,
char_bytes=self.lang == "zho")
tokens_by_abs_id = self.traverse_trees(feats)
feats, headers = self.n_gram(feats, tokens_by_abs_id, dummies=False)
temp = []
headers_with_oov = [
"first", "last", "deprel", "closest_left", "closest_right",
"farthest_left", "farthest_right", "pos", "cpos", "morph",
"s_type", "depchunk"
]
for pref in ["mn2", "mn1", "par", "par_par", "pl1", "pl2"]:
temp += [pref + "_" + h for h in headers_with_oov]
headers_with_oov += temp
genre_warning = False
for i, header in enumerate(headers):
if header in headers_with_oov and header in cat_labels:
for item in feats:
if item[i] not in multicol_dict["encoder_dict"][
header].classes_:
item[i] = "_"
elif header == "genre" and "genre" in cat_labels:
for item in feats:
if item[i] not in multicol_dict["encoder_dict"][
"genre"].classes_: # New genre not in training data
if not genre_warning:
sys.stderr.write(
"! WARN: Genre not in training data: " +
item[i] + "; suppressing further warnings\n")
genre_warning = True
item[i] = "_"
elif header.endswith("word") and header in cat_labels:
for item in feats:
# Replace rare words and words never seen before in this position with POS
if item[i] not in top_n_words or item[
i] not in multicol_dict["encoder_dict"][
header].classes_:
pos_col = headers.index(header.replace("word", "pos"))
if item[pos_col] in multicol_dict["encoder_dict"][
header].classes_:
item[i] = item[pos_col]
else:
item[i] = "_"
data = feats
data = pd.DataFrame(data, columns=headers)
data_encoded = self.multicol_transform(
data,
columns=multicol_dict["columns"],
all_encoders_=multicol_dict["all_encoders_"])
data_x = data_encoded[cat_labels + num_labels].values
probas = clf.predict_proba(data_x)
probas = [p[1] for p in probas]
preds = [int(p > 0.5) for p in probas]
for i, p in enumerate(preds):
if data["tok_id"].values[
i] == 1: # Ensure tok_id 1 is always a segment start
preds[i] = 1
if eval_gold:
gold = np.where(data_encoded['label'] == "_", 0, 1)
conf_mat = confusion_matrix(gold, preds)
sys.stderr.write(str(conf_mat) + "\n")
true_positive = conf_mat[1][1]
false_positive = conf_mat[0][1]
false_negative = conf_mat[1][0]
prec = true_positive / (true_positive + false_positive)
rec = true_positive / (true_positive + false_negative)
f1 = 2 * prec * rec / (prec + rec)
sys.stderr.write("P: " + str(prec) + "\n")
sys.stderr.write("R: " + str(rec) + "\n")
sys.stderr.write("F1: " + str(f1) + "\n")
with io.open("diff.tab", 'w', encoding="utf8") as f:
for i in range(len(gold)):
f.write("\t".join([toks[i],
str(gold[i]),
str(preds[i])]) + "\n")
return conf_mat, prec, rec, f1
else:
return zip(preds, probas)
def tune(self, trainfile, devfile, max_evals=10, as_text=False):
train_feats, _, _, _, _ = read_conll(trainfile,genre_pat=self.genre_pat,as_text=as_text)
train_for_rnn, scalers = feats2rnn(train_feats)
dev_feats, _, _, _, _ = read_conll(devfile,genre_pat=self.genre_pat,as_text=as_text)
dev_for_rnn, scalers = feats2rnn(dev_feats)
with io.open(self.corpus_dir + os.sep + "dev."+self.ext,'w',encoding="utf8",newline="\n") as f:
f.write(dev_for_rnn)
# NCRFpp expects a test file, we reuse dev
with io.open(self.corpus_dir + os.sep + "test."+self.ext,'w',encoding="utf8",newline="\n") as f:
f.write(dev_for_rnn)
# TRAIN MODEL ON ALL TRAIN
with io.open(self.corpus_dir + os.sep + "train."+self.ext,'w',encoding="utf8",newline="\n") as f:
f.write(train_for_rnn)
config = ncrf_dir + os.sep + self.segtype + ".train"+self.auto+".config"
def objective(params):
# sys.stderr.write(str(params))
data = Data()
data.read_config(config)
data.HP_batch_size = int(params['batch_size'])
data.HP_lr = float(params['lr'])
# data.word_emb_dim=int(params['word_emb_dim'])
data.char_emb_dim=int(params['char_emb_dim'])
data.word_feature_extractor=params['word_seq_feature']
data.char_feature_extractor=params['char_seq_feature']
#data.optimizer=params['optimizer']
data.HP_cnn_layer=int(params['cnn_layer'])
#data.HP_char_hidden_dim=int(params['char_hidden_dim'])
data.HP_hidden_dim=int(params['hidden_dim'])
data.HP_dropout=float(params['dropout'])
data.HP_lstm_layer=int(params['lstm_layer'])
data.average_batch_loss=str2bool(params['ave_batch_loss'])
p = StdOutFilter4Tune()
p.start()
ret, best_dev = ncrf(config=None, data=data)
p.end()
sys.stdout.write("F1 {:.3f} params {}".format(-best_dev, params))
if ret == 1:
return {'loss': -best_dev, 'status': STATUS_OK }
else:
return {'status': STATUS_FAIL }
space = {
'batch_size': scope.int(hp.quniform('batch_size', 10, 100, 10)),
'lr': hp.quniform('lr', 0.003, 0.18, 0.001),
# 'word_emb_dim': scope.int(hp.quniform('word_emb_dim', 100, 300, 10)),
'char_emb_dim': scope.int(hp.quniform('char_emb_dim', 30, 70, 10)),
'word_seq_feature': hp.choice('word_seq_feature', ["LSTM","CNN"]),
'char_seq_feature': hp.choice('char_seq_feature', ["LSTM","CNN"]),
#'optimizer': hp.choice('optimizer', ["SGD","AdaGrad","AdaDelta","RMSProp","Adam"]),
'optimizer': hp.choice('optimizer', ["AdaGrad"]),
'cnn_layer': scope.int(hp.quniform('cnn_layer', 1, 8, 1)),
'char_hidden_dim': scope.int(hp.quniform('char_hidden_dim', 50, 200, 10)),
'hidden_dim': scope.int(hp.quniform('hidden_dim', 100, 300, 20)),
'dropout': hp.quniform('dropout', 0.2, 0.8, 0.1),
'lstm_layer': scope.int(hp.quniform('lstm_layer', 1, 5, 1)),
'ave_batch_loss': hp.choice('ave_batch_loss', ["True","False"])
}
best_params = fmin(fn=objective, space=space, algo=tpe.suggest, max_evals=max_evals)
best_params = space_eval(space,best_params)
with io.open(script_dir + os.sep + "params" + os.sep + "RNNSegmenter"+self.auto+"_best_params.tab",'a',encoding="utf8") as bp:
corpus = os.path.basename(trainfile).split("_")[0]
for k, v in best_params.items():
bp.write("\t".join([corpus, 'RNNClassifier', k, str(v)])+"\n")
return best_params