def post(self):
"""
Parse multiple string and return the associated entity for each token in each string.
"""
args = self.parser.parse_args()
ref_strings = args.get('strings')
tokens = [[[token] for token in ref_string.split(" ")]
for ref_string in ref_strings]
data = prepare_dataset(tokens, current_app.word_to_id,
current_app.char_to_id, {},
current_app.model.parameters['lower'], True)
tagged = []
for index, datum in enumerate(data):
model_inputs = create_input(datum, current_app.model.parameters,
False)
y_pred = np.array(current_app.inference[1](*model_inputs))[1:-1]
tags = [
current_app.model.id_to_tag[y_pred[i]]
for i in range(len(y_pred))
]
tagged.append([
Entity(term=term, entity=entity)
for term, entity in zip(ref_strings[index].split(" "), tags)
])
response = ParseBatchResponse(reference_strings=ref_strings,
data=tagged)
return response
def training_loop():
taskB_out_label = []
for i, file in enumerate(taskB_in):
output = []
for j, sent in enumerate(file):
with torch.no_grad():
precheck_sent = utils.create_input(taskB_in[i][j].split(),
tokenizer).to(device)
sent_out_label = model(precheck_sent)[1]
sent_str_label = [biluo_decode[t] for t in sent_out_label]
output.append(sent_str_label)
taskB_out_label.append(output)
true_pos = 0
total_ph_pred = 0
for i, file in enumerate(taskB_label):
file_true_pos, file_total_ph_pred = eval_func(
taskB_out_label[i],
[["O"] + s.split() + ["O"] for s in taskB_label[i]])
true_pos = true_pos + file_true_pos
total_ph_pred = total_ph_pred + file_total_ph_pred
precision = 0
recall = 0
F1score = 0
if (total_ph_pred != 0):
precision = true_pos / total_ph_pred
if (total_phrases_truth != 0):
recall = true_pos / total_phrases_truth
if ((precision + recall) != 0):
F1score = 2 * precision * recall / (precision + recall)
print("Precision : {} | Recall : {} | F1 Score : {}".format(
precision, recall, F1score))
def post(self):
"""
Parse a single string and return the associated entity for each token in the string.
"""
args = self.parser.parse_args()
ref_string = args.get('string')
if ref_string is None or ref_string == "":
# Hackish way as reqparse can't catch empty string
abort(400, description='string is empty or not provided.')
tokens = ref_string.split(" ")
data = prepare_dataset([[[token] for token in tokens]],
current_app.word_to_id, current_app.char_to_id,
{}, current_app.model.parameters['lower'], True)
model_inputs = create_input(data[0], current_app.model.parameters,
False)
y_pred = np.array(current_app.inference[1](*model_inputs))[1:-1]
tags = [
current_app.model.id_to_tag[y_pred[i]] for i in range(len(y_pred))
]
response = ParseResponse(reference_string=ref_string,
data=[
Entity(term=term, entity=entity)
for term, entity in zip(tokens, tags)
])
return response
def next_batch(self, bz):
index = np.random.choice(len(self.memory), bz)
memory = [self.memory[i] for i in index]
state = [translate_state(i.get("state")) for i in memory]
state_next = [translate_state(i.get("state_next")) for i in memory]
action = [i.get("action") for i in memory]
finish = [int(i.get("finish")) for i in memory]
reward = [i.get("reward") for i in memory]
result = {
"state": create_input(state),
"state_next": create_input(state_next),
"action": action,
"finish": finish,
"reward": reward
}
return result
def validation_loop():
valid_taskB_out_label = []
for i, file in enumerate(valid_taskB_in):
output = []
for j, sent in enumerate(file):
with torch.no_grad():
precheck_sent = utils.create_input(
valid_taskB_in[i][j].split(), tokenizer).to(device)
sent_out_label = model(precheck_sent)[1]
sent_str_label = [biluo_decode[t] for t in sent_out_label]
output.append(sent_str_label)
valid_taskB_out_label.append(output)
valid_true_pos = 0
valid_total_ph_pred = 0
for i, file in enumerate(valid_taskB_label):
valid_file_true_pos, valid_file_total_ph_pred = eval_func(
valid_taskB_out_label[i],
[["O"] + s.split() + ["O"] for s in valid_taskB_label[i]])
valid_true_pos = valid_true_pos + valid_file_true_pos
valid_total_ph_pred = valid_total_ph_pred + valid_file_total_ph_pred
valid_precision = 0
valid_recall = 0
valid_F1score = 0
if (valid_total_ph_pred != 0):
valid_precision = valid_true_pos / valid_total_ph_pred
if (valid_total_phrases_truth != 0):
valid_recall = valid_true_pos / valid_total_phrases_truth
if ((valid_precision + valid_recall) != 0):
valid_F1score = 2 * valid_precision * \
valid_recall/(valid_precision+valid_recall)
print("Precision : {} | Recall : {} | F1 Score : {}".format(
valid_precision, valid_recall, valid_F1score))
def prepare_dataset(sentences,
word_to_id,
char_to_id,
gazetteer_list,
brown_dict,
tag_to_id,
l1_model,
l1_f_eval,
lower=False):
"""
Prepare the dataset. Return a list of lists of dictionaries containing:
- word indexes
- word char indexes
- tag indexes
"""
def f(x):
return x.lower() if lower else x
data = []
for s in sentences:
str_words = [w[0] for w in s]
words = [
word_to_id[f(w) if f(w) in word_to_id else '<UNK>']
for w in str_words
]
# Skip characters that are not in the training set
chars = [[char_to_id[c] for c in w if c in char_to_id]
for w in str_words]
caps = [cap_feature(w) for w in str_words]
gazetteer = [gazetteer_feature(w, gazetteer_list) for w in str_words]
brown = [brown_feature(w, brown_dict) for w in str_words]
sent = {
'str_words': str_words,
'words': words,
'chars': chars,
'caps': caps,
'gazetteer': gazetteer,
'brown': brown,
}
if l1_model is not None:
input = create_input(sent, l1_model.parameters, False)
try:
if l1_model.parameters['crf']:
y_preds = np.array(f_eval(*input))[1:-1]
else:
y_preds = f_eval(*input).argmax(axis=1)
y_preds = [l1_model.id_to_tag[y_pred] for y_pred in y_preds]
except Exception as e:
y_preds = ["O"] * len(str_words)
sent['pred'] = [0 if y_pred == "O" else 1 for y_pred in y_preds]
tags = [tag_to_id[w[-1]] for w in s]
sent['tags'] = tags
data.append(sent)
return data
def tag():
if request.method == 'POST':
data = request.get_json()
text = data['text']
if data['split_sentences']:
sentences = split_sentences(text)
else:
sentences = text
if data['tokenize'] or data['split_sentences']:
tokenized_sentences = [tokenize(s) for s in sentences]
else:
tokenized_sentences = text
count = 0
output = []
for words in tokenized_sentences:
if len(words) == 0:
continue
# Lowercase sentence
if model.parameters['lower']:
line = line.lower()
# Replace all digits with zeros
if model.parameters['zeros']:
line = zero_digits(line)
# Prepare input
sentence = prepare_sentence(words,
word_to_id,
char_to_id,
lower=model.parameters['lower'])
input = create_input(sentence, model.parameters, False)
# Decoding
if model.parameters['crf']:
y_preds = np.array(f_eval(*input))[1:-1]
else:
y_preds = f_eval(*input).argmax(axis=1)
y_preds = [model.id_to_tag[y_pred] for y_pred in y_preds]
# Output tags in the IOB2 format
if model.parameters['tag_scheme'] == 'iobes':
y_preds = iobes_iob(y_preds)
# Write tags
assert len(y_preds) == len(
words
), "Predictions have different length than sentence. Something went wrong."
output.append(list(zip(words, y_preds)))
count += 1
if count % 100 == 0:
logging.info(count)
return jsonify(output)
def create_input(self):
imgs = []
wides = []
for i,box in enumerate(self.all_after_row_connect):
result = Box_Cell(box.bbox,self.img,self.types[i])
if result.img_wide>self.max_wide:
self.max_wide = result.img_wide
wides.append(result.img_wide)
imgs.append(result.normal_img)
self.results.append(result)
inputs,wides = utils.create_input(imgs,self.max_wide,wides)
return inputs,wides
def tag(model, line):
# Load existing model
print("Loading model...")
model = Model(model_path=model)
parameters = model.parameters
# Load reverse mappings
word_to_id, char_to_id, tag_to_id = [{
v: k
for k, v in x.items()
} for x in [model.id_to_word, model.id_to_char, model.id_to_tag]]
# Load the model
_, f_eval = model.build(training=False, **parameters)
model.reload()
start = time.time()
print('Tagging...')
words_ini = line.rstrip().split()
# Replace all digits with zeros
if parameters['zeros']:
line = zero_digits(line)
words = line.rstrip().split()
# Prepare input
sentence = prepare_sentence(words,
word_to_id,
char_to_id,
lower=parameters['lower'])
input = create_input(sentence, parameters, False)
# Decoding
if parameters['crf']:
y_preds = np.array(f_eval(*input))[1:-1]
else:
y_preds = f_eval(*input).argmax(axis=1)
y_preds = [model.id_to_tag[y_pred] for y_pred in y_preds]
# Output tags in the IOB2 format
if parameters['tag_scheme'] == 'iobes':
y_preds = iobes_iob(y_preds)
# Write tags
assert len(y_preds) == len(words)
print('---- sentence tagged in %.4fs ----' % (time.time() - start))
return ' '.join(w + '__' + str(y) for w, y in zip(words_ini, y_preds))
def create_input(self):
imgs = []
wides = []
def add_img(result):
if result.img_wide>self.max_wide:
self.max_wide = result.img_wide
wides.append(result.img_wide)
imgs.append(result.normal_img)
for i,result in enumerate(self.all_after_row_connect):
add_img(result)
inputs,wides = utils.create_input(imgs,self.max_wide,wides)
return inputs,wides
def tag_document(doc, parameters, model, f_eval, word_to_id, char_to_id):
count = 0
all_ypreds = list()
all_tokens = list()
for line in doc.sentences:
toks_text = [x.orth_ for x in line.tokens]
# line = ' '.join(toks_text)
if toks_text: # WL edit: used to be 'if line', was crashing on '\n' lines
# Lowercase sentence
if parameters['lower']:
toks_text = [line.lower() for line in toks_text]
# Replace all digits with zeros
if parameters['zeros']:
toks_text = [zero_digits(line) for line in toks_text]
# Prepare input
sentence = prepare_sentence(toks_text,
word_to_id,
char_to_id,
lower=parameters['lower'])
input = create_input(sentence, parameters, False)
# Decoding
if parameters['crf']:
y_preds = np.array(f_eval(*input))[1:-1]
else:
y_preds = f_eval(*input).argmax(axis=1)
y_preds = [model.id_to_tag[y_pred] for y_pred in y_preds]
# Output tags in the IOB2 format
if parameters['tag_scheme'] == 'iobes':
y_preds = iobes_iob(y_preds)
# Write tags
assert len(y_preds) == len(toks_text)
# strip IOB prefixes
y_preds = [x.split('-')[-1] for x in y_preds]
all_ypreds.append(y_preds)
all_tokens.append(toks_text)
count += 1
if count % 100 == 0:
print count
return (all_ypreds, all_tokens)
def prepare_sentence(str_words,
word_to_id,
char_to_id,
gazetteer_list={},
brown_dict={},
l1_model=None,
l1_f_eval=None,
lower=False):
"""
Prepare a sentence for evaluation.
"""
def f(x):
return x.lower() if lower else x
words = [
word_to_id[f(w) if f(w) in word_to_id else '<UNK>'] for w in str_words
]
chars = [[char_to_id[c] for c in w if c in char_to_id] for w in str_words]
caps = [cap_feature(w) for w in str_words]
gazetteer = [gazetteer_feature(w, gazetteer_list) for w in str_words]
brown = [brown_feature(w, brown_dict) for w in str_words]
sent = {
'str_words': str_words,
'words': words,
'chars': chars,
'caps': caps,
'gazetteer': gazetteer,
'brown': brown
}
if l1_model is not None:
input = create_input(sent, l1_model.parameters, False)
try:
if l1_model.parameters['crf']:
y_preds = np.array(f_eval(*input))[1:-1]
else:
y_preds = f_eval(*input).argmax(axis=1)
y_preds = [l1_model.id_to_tag[y_pred] for y_pred in y_preds]
except Exception as e:
y_preds = ["O"] * len(str_words)
sent['pred'] = [0 if y_pred == "O" else 1 for y_pred in y_preds]
return sent
def predicts(self, line):
if line:
# Save original bigrams
bigram_sent = self.to_bigram(line, 0).strip().split()
# Replave all digits with zeros
line = zero_digits(line)
input_seq = self.to_bigram(line, 0).strip().split()
# Prepare input
sentence = prepare_sentence(input_seq,
self.word_to_id,
self.char_to_id,
lower=self.parameters['lower'])
input = create_input(sentence, self.parameters, False)
if self.parameters['crf']:
y_preds = np.array(self.f_eval(*input))[1:-1]
else:
y_preds = self.f_eval(*input).argmax(axis=1)
tags = [self.id_to_tag[y_pred] for y_pred in y_preds]
# Output tags in the IOB2 format
if self.parameters['tag_scheme'] == 'iobes':
tags = iobes_iob(tags)
print(tags)
# Make output form
out_form = ""
unigram_sent = self.bigrams_to_unigrams(bigram_sent)
for i in range(len(tags)):
if tags[i].startswith('B'):
out_form += '<' + unigram_sent[i]
elif tags[i].startswith('I'):
if i == len(tags) - 1:
out_form += unigram_sent[i] + ':' + tags[i][2:] + '>'
elif tags[i + 1] == 'O':
out_form += unigram_sent[i] + ':' + tags[i][2:] + '>'
else:
out_form += unigram_sent[i]
else:
out_form += unigram_sent[i]
return out_form
def parseString(self, string):
#TO DO
#To be consumed by web-service
test_file = "test_file"
file = open(test_file, 'w')
file.write('\n'.join(string.encode('utf-8').split()))
file.close()
test_sentences = load_sentences(test_file, self.lower, self.zeros)
data = self.prepare_dataset(test_sentences)
result = ''
for citation in data:
input = create_input(citation, self.model.parameters, False)
y_pred = np.array(self.f[1](*input))[1:-1]
tags = []
for i in xrange(len(y_pred)):
tags.append(self.model.id_to_tag[y_pred[i]])
for num, word in enumerate(string.encode('utf-8').split()):
#print word.decode('utf-8')+'\t'+tags[num]
result += word.decode('utf-8')+'\t'+tags[num]+'\n'
return result
def evaluate(ctx, model, env, rounds=5, print_action=False, save=None):
env.reset_env()
for epoch in range(rounds):
env.reset_env()
done = 0
step = 0
while not done:
step += 1
data = create_input([translate_state(env.map.state())])
data = [torch.FloatTensor(i).to(ctx) for i in data]
pred = model.forward(data)
action = int(torch.argmax(pred).cpu().numpy())
old, new, reward, done = env.step(action)
if print_action:
print(pred, reward, env.map.battery)
if save is not None:
img = Image.fromarray(env.map.render(), 'RGB')
pred = [str(x)[0:5] for x in pred.detach().numpy().tolist()[0]]
filename = "torch-" + str(epoch) + "-" + str(step) + "-" + str(
reward) + "-" + "_".join(pred) + ".jpg"
img.save(save + "/" + filename)
return env.detect_rate
def evaluate(ctx, model, env, rounds=5, print_action=False, save=None):
for epoch in range(rounds):
env.reset_env()
done = 0
step = 0
while not done:
step += 1
data = create_input([translate_state(env.map.state())])
data = [nd.array(i, ctx=ctx) for i in data]
pred = model(data)
action = int(nd.argmax(pred, axis=1).asnumpy()[0])
old, new, reward, done = env.step(action)
if print_action:
print(pred, reward, env.map.battery)
if save is not None:
img = Image.fromarray(
env.map.grid.render(20, env.map.agent_pos,
env.map.agent_dir), 'RGB')
pred = [str(x)[0:5] for x in pred.asnumpy().tolist()[0]]
filename = str(epoch) + "-" + str(step) + "-" + str(
reward) + "-" + "_".join(pred) + ".jpg"
img.save(save + "/" + filename)
return env.detect_rate
def evaluate(parameters, f_eval, raw_sentences, parsed_sentences,
id_to_tag, input_file_path):
"""
Evaluate current model using CoNLL script.
"""
predictions = []
for raw_sentence, data in zip(raw_sentences, parsed_sentences):
input = create_input(data, parameters, False)
if parameters['crf']:
y_preds = np.array(f_eval(*input))[1:-1]
else:
y_preds = f_eval(*input).argmax(axis=1)
p_tags = [id_to_tag[y_pred] for y_pred in y_preds]
if parameters['tag_scheme'] == 'iobes':
p_tags = iobes_iob(p_tags)
for i, y_pred in enumerate(y_preds):
new_line = "%s %s" % (raw_sentence[i][0], p_tags[i])
predictions.append(new_line)
predictions.append("")
output_path = os.path.join(opts.output, os.path.basename(input_file_path[:-4] + "_Tagged.txt"))
with codecs.open(output_path, 'w', 'utf8') as f:
f.write("\n".join(predictions))
if tags[i][0] == 'O':
if len(preTag):
res.append("</" + preTag + ">")
preTag = ""
res.append(sentence[i])
if len(preTag):
res.append("</" + preTag + ">")
return res
#}}}
print 'Tagging...'
for line in test_data:
# Prepare input
input = create_input(line,
parameters,
False,
useAttend=parameters['useAttend'])
words = line['str_words']
# Decoding
if parameters['crf']:
y_preds = np.array(f_eval(*input))
else:
y_preds = f_eval(*input).argmax(axis=1)
y_preds = [model.id_to_tag[y_pred] for y_pred in y_preds]
# Output tags in the IOB2 format
if parameters['tag_scheme'] == 'iobes':
y_preds = iobes_iob(y_preds)
# Write tags
assert len(y_preds) == len(words)
# print words
with codecs.open(opts.input, 'r', 'utf-8') as f_input:
count = 0
for line in f_input:
words_ini = line.rstrip().split()
if line:
# Lowercase sentence
if parameters['lower']:
line = line.lower()
# Replace all digits with zeros
if parameters['zeros']:
line = zero_digits(line)
words = line.rstrip().split()
# Prepare input
sentence = prepare_sentence(words, word_to_id, char_to_id,
lower=parameters['lower'])
input = create_input(sentence, parameters, False)
# Decoding
if parameters['crf']:
y_preds = np.array(f_eval(*input))[1:-1]
else:
y_preds = f_eval(*input).argmax(axis=1)
y_preds = [model.id_to_tag[y_pred] for y_pred in y_preds]
# Output tags in the IOB2 format
if parameters['tag_scheme'] == 'iobes':
y_preds = iobes_iob(y_preds)
# Write tags
assert len(y_preds) == len(words)
if opts.outputFormat == 'json':
f_output.write(json.dumps({ "text": ' '.join(words), "ranges": iob_ranges(y_preds) }))
else:
#
# Train network
#
singletons = set([word_to_id[k] for k, v
in dico_words_train.items() if v == 1])
n_epochs = 100 # number of epochs over the training set
freq_eval = 1000 # evaluate on dev every freq_eval steps
best_dev = -np.inf
best_test = -np.inf
count = 0
for epoch in xrange(n_epochs):
epoch_costs = []
print "Starting epoch %i..." % epoch
for i, index in enumerate(np.random.permutation(len(train_data))):
count += 1
input = create_input(train_data[index], parameters, True, singletons)
new_cost = f_train(*input)
epoch_costs.append(new_cost)
if i % 50 == 0 and i > 0 == 0:
print "%i, cost average: %f" % (i, np.mean(epoch_costs[-50:]))
if count % freq_eval == 0:
dev_score = evaluate(parameters, f_eval, dev_sentences,
dev_data, id_to_tag, dico_tags)
test_score = evaluate(parameters, f_eval, test_sentences,
test_data, id_to_tag, dico_tags)
print "Score on dev: %.5f" % dev_score
print "Score on test: %.5f" % test_score
if dev_score > best_dev:
best_dev = dev_score
print "New best score on dev."
print "Saving model to disk..."
for line in f_input:
words_ini = line.rstrip().split()
if line:
# Lowercase sentence
if parameters['lower']:
line = line.lower()
# Replace all digits with zeros
if parameters['zeros']:
line = zero_digits(line)
words = line.rstrip().split()
# Prepare input
sentence = prepare_sentence(words,
word_to_id,
char_to_id,
lower=parameters['lower'])
input = create_input(sentence, parameters, False)
# Decoding
if parameters['crf']:
y_preds = np.array(f_eval(*input))[1:-1]
else:
y_preds = f_eval(*input).argmax(axis=1)
y_preds = [model.id_to_tag[y_pred] for y_pred in y_preds]
# Output tags in the IOB2 format
if parameters['tag_scheme'] == 'iobes':
y_preds = iobes_iob(y_preds)
# Write tags
assert len(y_preds) == len(words)
if opts.outputFormat == 'json':
f_output.write(
json.dumps({
def run_tagging(model,
f_eval,
parameters,
word_to_id,
char_to_id,
tag_to_id,
opts_input="",
opts_output="",
opts_delimiter="__",
opts_outputFormat=""):
# Check parameters validity
assert opts_delimiter
assert os.path.isfile(opts_input)
#set environment to use gpu
f_output = codecs.open(opts_output, 'w', 'utf-8')
start = time.time()
logger.info('Tagging...')
with codecs.open(opts_input, 'r', 'utf-8') as f_input:
count = 0
for line in f_input:
words_ini = line.rstrip().split()
if line:
# Lowercase sentence
if parameters['lower']:
line = line.lower()
# Replace all digits with zeros
if parameters['zeros']:
line = zero_digits(line)
words = line.rstrip().split()
# Prepare input
sentence = prepare_sentence(words,
word_to_id,
char_to_id,
lower=parameters['lower'])
input = create_input(sentence, parameters, False)
# Decoding
if parameters['crf']:
y_preds = np.array(f_eval(*input))[1:-1]
else:
y_preds = f_eval(*input).argmax(axis=1)
y_preds = [model.id_to_tag[y_pred] for y_pred in y_preds]
# Output tags in the IOB2 format
if parameters['tag_scheme'] == 'iobes':
y_preds = iobes_iob(y_preds)
# Write tags
assert len(y_preds) == len(words)
if opts_outputFormat == 'json':
f_output.write(
json.dumps({
"text": ' '.join(words),
"ranges": iob_ranges(y_preds)
}))
else:
#logger.info( "write out tags..."
f_output.write(
'%s\n' % ' '.join('%s%s%s' % (w, opts_delimiter, y)
for w, y in zip(words_ini, y_preds)))
else:
f_output.write('\n')
count += 1
# if count % 100 == 0:
# logger.info( count
logger.info('---- %i lines tagged in %.4fs ----' %
(count, time.time() - start))
f_output.close()
logger.info(opts_output)
logger.info("")
return opts_output + " has been tagged!"
# def main():
# logger.info( "executed"
# if __name__ == '__main__':
# main()
# Train network
#
count = 0
fvs = []
words = []
for i, index in enumerate(np.random.permutation(len(train_data))):
fv = []
if (train_data[index]['str_words'][0] in words):
continue
count += 1
words.append(train_data[index]['str_words'][0])
input = create_input(train_data[index], parameters, False)
# Get gradients vector from model
grads = f_eval(*input)
grads_rev = f_eval_rev(*input)
* Concatenate all gradients
for grad in grads_rev:
for g in grad:
try:
for s in g:
fv.append(s)
except:
fv.append(g)
for grad in grads:
string = raw_input("Enter the citation string: ")
strings = [string]
test_file = "test_file"
if os.path.exists(test_file):
os.remove(test_file)
file = open(test_file, 'a')
for string in strings:
file.write('\n'.join(string.split()) + '\n')
file.close()
test_sentences = load_sentences(test_file, lower, zeros)
data = prepare_dataset(test_sentences, word_to_id, char_to_id, {}, lower,
True)
for citation in data:
inputs = create_input(citation, model.parameters, False)
y_pred = np.array(f[1](*inputs))[1:-1]
tags = [model.id_to_tag[y_pred[i]] for i in range(len(y_pred))]
output = [
w + '\t' + tags[i] for i, w in enumerate(citation['str_words'])
]
if opts.run == 'file':
with closing(open(output_file, 'w')) as fh:
fh.write('\n'.join(output))
else:
print('\n'.join(output))
if opts.run == 'file':
best_test = -np.inf
count = 0
# costfile = './cost_vec_' + str(parameters['word_dim']) + '_' + str(parameters['word_hidden_dim']) + str(parameters['L2_reg']) + '.txt'
# fw = codecs.open(costfile, 'w', 'utf-8')
# fw.write('epoch\t\ttrain_loss\t\tdev_los\t\ttest_loss\t\tdev_F1\t\ttest_F1\n')
F1_file = './vec_' + opts.tagger + "gaze_" + opts.dictionary + str(parameters['use_gaze']) + "char" + str(parameters['char_dim']) + "_" + str(parameters['char_hidden_dim']) \
+ "_word" + str(parameters['word_dim']) + "_" + str(parameters['word_hidden_dim']) + 'taggerhidden' + str(parameters['tagger_hidden_dim']) + '.txt'
fw = codecs.open(F1_file, 'w', 'utf-8')
fw.write("epoch\t\tdev_F1\t\ttest_F1\n")
for epoch in xrange(n_epochs):
epoch_costs = []
print "Starting epoch %i..." % epoch
for i, index in enumerate(np.random.permutation(len(train_data))):
count += 1
input = create_input(train_data[index], parameters, True, use_gaze, pos, singletons)
new_cost = f_train(*input)
epoch_costs.append(new_cost)
if i % 50 == 0 and i > 0 == 0:
print "%i, cost average: %f" % (i, np.mean(epoch_costs[-50:]))
if count % freq_eval == 0:
#train_score = evaluate(parameters, f_eval, train_sentences,
# train_data, id_to_tag, dico_tags)
# train_cost = []
# for i, data in enumerate(train_data):
# input = create_input(data, parameters, True, use_gaze, pos, singletons)
# train_cost.append(f_plot_cost(*input))
# dev_cost = []
# for i, data in enumerate(dev_data):
# input = create_input(data, parameters, True, use_gaze, pos, singletons)
hidden_layers=parameters['hidden_layer'],
padding=parameters['padding'],
max_seq_len=max_seq_len,
train_size=len(train_data))
if parameters['reload']:
gramcnn.load(models_path, model_name)
for epoch in xrange(n_epochs):
epoch_costs = []
print "Starting epoch %i..." % epoch
for i, index in enumerate(np.random.permutation(len(train_data))):
inputs, word_len = create_input(train_data[index],
parameters,
True,
singletons,
padding=parameters['padding'],
max_seq_len=max_seq_len,
use_pts=parameters['pts'])
assert inputs['char_for']
assert inputs['word']
assert inputs['label']
# break
if len(inputs['label']) == 1:
continue
train_loss = []
temp = []
temp.append(word_len)
batch_loss = gramcnn.train(inputs, temp)
def runModelInLoop(dropout,char_dim,char_lstm_dim,word_dim,word_lstm_dim):
#results File
resultsPath = "/Users/Ehsan/Documents/Ehsan_General/HMQ/HMQ_Projects/DNR2/COLING-2016-Code/i2b2-2010/results/"
for u_dropout in dropout:
for v_char_dim in char_dim:
for w_char_lstm_dim in char_lstm_dim:
for x_word_dim in word_dim:
for y_word_lstm_dim in word_lstm_dim:
for dataset in datasets:
print "+++++++++++++++"
print u_dropout,v_char_dim,w_char_lstm_dim,x_word_dim,y_word_lstm_dim,dataset
parameters['dropout'] = u_dropout
parameters['char_dim'] = v_char_dim
parameters['char_lstm_dim'] =w_char_lstm_dim
parameters['word_dim'] = x_word_dim
parameters['word_lstm_dim'] = y_word_lstm_dim
# If dataset is DrugBank assign predefined path
if(dataset == "i2b2-2010"):
opts.train = i2b2BasePath+"train.txt"
opts.dev = i2b2BasePath+ "dev.txt"
opts.test = i2b2BasePath+ "test.txt"
resultsFile = resultsPath +"i2b2_2010_Results.txt"
# Initialize model
model = Model(parameters=parameters, models_path=models_path)
print "Model location: %s" % model.model_path
# Data parameters
lower = parameters['lower']
zeros = parameters['zeros']
tag_scheme = parameters['tag_scheme']
# Load sentences
train_sentences = loader.load_sentences(opts.train, lower, zeros)
dev_sentences = loader.load_sentences(opts.dev, lower, zeros)
test_sentences = loader.load_sentences(opts.test, lower, zeros)
# Use selected tagging scheme (IOB / IOBES)
update_tag_scheme(train_sentences, tag_scheme)
update_tag_scheme(dev_sentences, tag_scheme)
update_tag_scheme(test_sentences, tag_scheme)
# Create a dictionary / mapping of words
# If we use pretrained embeddings, we add them to the dictionary.
if parameters['pre_emb']:
dico_words_train = word_mapping(train_sentences, lower)[0]
dico_words, word_to_id, id_to_word = augment_with_pretrained(
dico_words_train.copy(),
parameters['pre_emb'],
list(itertools.chain.from_iterable(
[[w[0] for w in s] for s in dev_sentences + test_sentences])
) if not parameters['all_emb'] else None
)
else:
dico_words, word_to_id, id_to_word = word_mapping(train_sentences, lower)
dico_words_train = dico_words
# Create a dictionary and a mapping for words / POS tags / tags
dico_chars, char_to_id, id_to_char = char_mapping(train_sentences)
dico_tags, tag_to_id, id_to_tag = tag_mapping(train_sentences)
print "Calling the prepare_dataset :--"
# Index data
train_data = prepare_dataset(
train_sentences, word_to_id, char_to_id, tag_to_id, lower
)
dev_data = prepare_dataset(
dev_sentences, word_to_id, char_to_id, tag_to_id, lower
)
test_data = prepare_dataset(
test_sentences, word_to_id, char_to_id, tag_to_id, lower
)
print "%i / %i / %i sentences in train / dev / test." % (
len(train_data), len(dev_data), len(test_data))
# Save the mappings to disk
print 'Saving the mappings to disk...'
model.save_mappings(id_to_word, id_to_char, id_to_tag)
# Build the model
f_train, f_eval = model.build(**parameters)
# Reload previous model values
if opts.reload:
print 'Reloading previous model...'
model.reload()
# Train network
#
singletons = set([word_to_id[k] for k, v
in dico_words_train.items() if v == 1])
n_epochs = 2 # number of epochs over the training set
freq_eval = 1000 # evaluate on dev every freq_eval steps
best_dev = -np.inf
best_test = -np.inf
count = 0
for epoch in xrange(n_epochs):
epoch_costs = []
print "Starting epoch %i..." % epoch
for i, index in enumerate(np.random.permutation(len(train_data))):
count += 1
input = create_input(train_data[index], parameters, True, singletons)
new_cost = f_train(*input)
epoch_costs.append(new_cost)
#if i % 50 == 0 and i > 0 == 0:
# print "%i, cost average: %f" % (i, np.mean(epoch_costs[-50:]))
if count % freq_eval == 0:
dev_score = evaluate(parameters, f_eval, dev_sentences,
dev_data, id_to_tag, dico_tags)
test_score = evaluate(parameters, f_eval, test_sentences,
test_data, id_to_tag, dico_tags)
print "Score on dev: %.5f" % dev_score
print "Score on test: %.5f" % test_score
if dev_score > best_dev:
best_dev = dev_score
print "New best score on dev."+str(best_dev)
# print "Saving model to disk..."
# model.save()
if test_score > best_test:
best_test = test_score
print "New best score on test."+str(best_test)
# print "Config values used are : "
print "Epoch %i done. Average cost: %f" % (epoch, np.mean(epoch_costs))
# Write the best dev and test scores to the file
del model
with open(resultsFile, 'a') as f:
f.write("dropout: "+ str(parameters['dropout'] ) +"| char_dim: |"+str(parameters['char_dim'])+ "| char_lstm_dim: "+str(parameters['char_lstm_dim']) +" word_dim: "+ str(parameters['word_dim']) +" |word_lstm_dim: "+ str( parameters['word_lstm_dim'] )+" | Best Dev Score: "+str(best_dev) + " | Best Test Score: "+str(best_test) +"\n")
return
def train(self,
n_epochs=100,
freq_eval=1000,
verbose=True,
eval_test_set=False):
"""
:param n_epochs: number of epochs over the training set
:param freq_eval: evaluate on dev every freq_eval steps
:return: Saves the model with the best F1-Score, evaluated on the dev set
"""
# Initialize model
model = Model(parameters=self.parameters, models_path=models_path)
print("Model location: %s" % model.model_path)
# Data parameters
lower = self.parameters['lower']
zeros = self.parameters['zeros']
tag_scheme = self.parameters['tag_scheme']
# Load sentences
train_sentences = loader.load_sentences(self.parameters['train'],
lower, zeros)
dev_sentences = loader.load_sentences(self.parameters['dev'], lower,
zeros)
test_sentences = loader.load_sentences(self.parameters['test'], lower,
zeros)
# Use selected tagging scheme (IOB / IOBES)
update_tag_scheme(train_sentences, tag_scheme)
update_tag_scheme(dev_sentences, tag_scheme)
update_tag_scheme(test_sentences, tag_scheme)
# Create a dictionary / mapping of words
# If we use pretrained embeddings, we add them to the dictionary.
if self.parameters['pre_emb']:
dico_words_train = word_mapping(train_sentences, lower)[0]
dico_words, word_to_id, id_to_word = augment_with_pretrained(
dico_words_train.copy(), self.parameters['pre_emb'],
list(
itertools.chain.from_iterable([[w[0] for w in s]
for s in dev_sentences +
test_sentences]))
if not self.parameters['all_emb'] else None)
else:
dico_words, word_to_id, id_to_word = word_mapping(
train_sentences, lower)
dico_words_train = dico_words
# Create a dictionary and a mapping for words / POS tags / tags
dico_chars, char_to_id, id_to_char = char_mapping(train_sentences)
dico_tags, tag_to_id, id_to_tag = tag_mapping(train_sentences)
# Index data
train_data = prepare_dataset(train_sentences, word_to_id, char_to_id,
tag_to_id, lower)
dev_data = prepare_dataset(dev_sentences, word_to_id, char_to_id,
tag_to_id, lower)
test_data = prepare_dataset(test_sentences, word_to_id, char_to_id,
tag_to_id, lower)
print("%i / %i / %i sentences in train / dev / test." %
(len(train_data), len(dev_data), len(test_data)))
# Save the mappings to disk
print('Saving the mappings to disk...')
model.save_mappings(id_to_word, id_to_char, id_to_tag)
# Build the model
f_train, f_eval = model.build(**self.parameters)
# Reload previous model values
if self.parameters['reload']:
print('Reloading previous model...')
model.reload()
#
# Train network
#
singletons = set(
[word_to_id[k] for k, v in dico_words_train.items() if v == 1])
best_dev = -np.inf
best_test = -np.inf
count = 0
for epoch in range(n_epochs):
epoch_costs = []
print("Starting epoch %i at..." % epoch, time.ctime())
for i, index in enumerate(np.random.permutation(len(train_data))):
count += 1
input = create_input(train_data[index], self.parameters, True,
singletons)
new_cost = f_train(*input)
epoch_costs.append(new_cost)
if i % 50 == 0 and i > 0 == 0 and verbose:
print("%i, cost average: %f" %
(i, np.mean(epoch_costs[-50:])))
if count % freq_eval == 0:
dev_score = evaluate(self.parameters,
f_eval,
dev_sentences,
dev_data,
id_to_tag,
verbose=verbose)
if eval_test_set:
test_score = evaluate(self.parameters,
f_eval,
test_sentences,
test_data,
id_to_tag,
verbose=verbose)
print("Score on dev: %.5f" % dev_score)
if eval_test_set:
print("Score on test: %.5f" % test_score)
if dev_score > best_dev:
best_dev = dev_score
print("New best score on dev.")
print("Saving model to disk...")
model.save()
if eval_test_set:
if test_score > best_test:
best_test = test_score
print("New best score on test.")
print(
"Epoch %i done. Average cost: %f. Ended at..." %
(epoch, np.mean(epoch_costs)), time.ctime())
return best_dev
for phase in ['train', 'dev', 'test'][:]:
if phase == 'train':
optimizer = exp_lr_scheduler(optimizer_ft, epoch,
**lr_method_parameters)
model.train(True) # Set model to training mode
random.shuffle(dataset[phase])
else:
model.train(False) # Set model to evaluate mode
epoch_loss = []
# Iterate over data.
preds = []
for i in range(0, len(dataset[phase]), batch_size):
inputs, seq_index_mapping, char_index_mapping, seq_len, char_len = \
create_input(dataset[phase][i:i+batch_size], parameters)
# forward
outputs, loss = model.forward(inputs, seq_len, char_len,
char_index_mapping)
try:
epoch_loss.append(loss.data[0])
except AttributeError:
pass
# backward + optimize only if in training phase
if phase == 'train':
# zero the parameter gradients
optimizer.zero_grad()
loss.backward()
#
# Train network
#
singletons = set([word_to_id[k] for k, v
in dico_words_train.items() if v == 1])
n_epochs = 100 # number of epochs over the training set
freq_eval = 1000 # evaluate on dev every freq_eval steps
best_dev = -np.inf
best_test = -np.inf
count = 0
for epoch in xrange(n_epochs):
epoch_costs = []
print "Starting epoch %i..." % epoch
for i, index in enumerate(np.random.permutation(len(train_data))):
count += 1
input = create_input(train_data[index], parameters, True, singletons)
new_cost = f_train(*input)
epoch_costs.append(new_cost)
if i % 50 == 0 and i > 0 == 0:
print "%i, cost average: %f" % (i, np.mean(epoch_costs[-50:]))
if count % freq_eval == 0:
dev_score = evaluate(parameters, f_eval, dev_sentences,
dev_data, id_to_tag, dico_tags)
test_score = evaluate(parameters, f_eval, test_sentences,
test_data, id_to_tag, dico_tags)
print "Score on dev: %.5f" % dev_score
print "Score on test: %.5f" % test_score
if dev_score > best_dev:
best_dev = dev_score
print "New best score on dev."
print "Saving model to disk..."
if (valid_total_ph_pred != 0):
valid_precision = valid_true_pos / valid_total_ph_pred
if (valid_total_phrases_truth != 0):
valid_recall = valid_true_pos / valid_total_phrases_truth
if ((valid_precision + valid_recall) != 0):
valid_F1score = 2 * valid_precision * \
valid_recall/(valid_precision+valid_recall)
print("Precision : {} | Recall : {} | F1 Score : {}".format(
valid_precision, valid_recall, valid_F1score))
# Main Training Loop
with torch.no_grad():
precheck_sent = utils.create_input(taskB_in[0][0].split(),
tokenizer).to(device)
precheck_tags = torch.tensor(
[4] + [biluo_code[t] for t in taskB_label[0][0].split()] + [4],
dtype=torch.long).to(device)
print("Checkpoint reached! Starting model training......")
for epoch in range(4):
start = time.time()
model.train()
for i, file in enumerate(taskB_in):
try:
if (i % 10 == 0):
print(f"done with {i} of {len(taskB_in)}")
for j, sent in enumerate(file):
#
# Train network
#
n_epochs = 50 # number of epochs over the training set
freq_eval = 500 # evaluate on dev every freq_eval steps
best_dev = -np.inf
best_test = -np.inf
count = 0
for epoch in xrange(n_epochs):
epoch_costs = []
print "Starting epoch %i..." % epoch
for i, index in enumerate(np.random.permutation(len(train_data))):
count += 1
input = create_input(train_data[index], parameters, True, False if pos_tag==0 else True)
new_cost = f_train(*input)
epoch_costs.append(new_cost)
if i % 50 == 0 and i > 0 == 0:
print "%i, cost average: %f" % (i, np.mean(epoch_costs[-50:]))
if count % freq_eval == 0:
dev_score, pred_dev = evaluate_scope(parameters, model.model_path, f_eval, dev_data, id_to_y, False if pos_tag==0 else True)
if dev_score > best_dev:
best_dev = dev_score
print "New best score on dev."
print "Saving model to disk..."
model.save()
# Store predictions to disk
output_predDEV = os.path.join(model.model_path, "best_dev.output")
with codecs.open(output_predDEV, 'w', 'utf8') as f:
f.write("\n".join(pred_dev))
def ner():
global model
global f_eval
global parameters
global word_to_id
global char_to_id
global tag_to_id
model_name = request.json["model"]
words = request.json["words"]
begin_end = request.json["begin_end"]
if model is None:
## Model loading
print "Loading model " + model_name + ".."
model = Model(model_path="models/" + models[model_name])
parameters = model.parameters
# Load reverse mappings
word_to_id, char_to_id, tag_to_id = [{
v: k
for k, v in x.items()
} for x in [model.id_to_word, model.id_to_char, model.id_to_tag]]
# Load the model
_, f_eval = model.build(training=False, **parameters)
model.reload()
# else:
# parameters = model.parameters
# word_to_id, char_to_id, tag_to_id = [
# {v: k for k, v in x.items()}
# for x in [model.id_to_word, model.id_to_char, model.id_to_tag]
# ]
# Lowercase sentence
if parameters['lower']:
words = [w.lower() for w in words]
# Replace all digits with zeros
if parameters['zeros']:
words = [zero_digits(w) for w in words]
words = [w if not w.isupper() else w.title() for w in words]
# Prepare input
sentence = prepare_sentence(words,
word_to_id,
char_to_id,
lower=parameters['lower'])
input = create_input(sentence, parameters, False)
# Decoding
if parameters['crf']:
y_preds = np.array(f_eval(*input))[1:-1]
else:
y_preds = f_eval(*input).argmax(axis=1)
y_preds = [model.id_to_tag[y_pred] for y_pred in y_preds]
# Output tags in the IOB2 format
if parameters['tag_scheme'] == 'iobes':
y_preds = iobes_iob(y_preds)
# Write tags
assert len(y_preds) == len(words) # TODO:remove assert?
ents = [{
"start_char": b,
"end_char": e,
"label": label
} for (b, e), label in zip(begin_end, y_preds) if label != "O"]
return json.dumps({"ents": ents})
def test_inference_performance():
from sklearn.metrics import f1_score
from torchtext.datasets import SequenceTaggingDataset
from torchtext.data import Field, NestedField
WORD = Field(init_token='<bos>', eos_token='<eos>')
CHAR_NESTING = Field(tokenize=list, init_token='<bos>', eos_token='<eos>')
CHAR = NestedField(CHAR_NESTING, init_token='<bos>', eos_token='<eos>')
ENTITY = Field(init_token='<bos>', eos_token='<eos>')
data_file = tempfile.NamedTemporaryFile(delete=True)
# TODO Need to be decoded in Python 3
data_file.write(requests.get(CORA_URL).content)
fields = [(('text', 'char'),
(WORD, CHAR))] + [(None, None)] * 22 + [('entity', ENTITY)]
dataset = SequenceTaggingDataset(data_file.name, fields, separator=" ")
model = Model(model_path='models/neuralParsCit')
model.parameters['pre_emb'] = os.path.join(os.getcwd(),
'vectors_with_unk.kv')
f = model.build(training=False, **model.parameters)
model.reload()
word_to_id = {v: i for i, v in model.id_to_word.items()}
char_to_id = {v: i for i, v in model.id_to_char.items()}
tag_to_id = {tag: i for i, tag in model.id_to_tag.items()}
tf = tempfile.NamedTemporaryFile(delete=False)
tf.write("\n\n".join(
["\n".join(example.text) for example in dataset.examples]))
tf.close()
train_sentences = load_sentences(tf.name, model.parameters['lower'],
model.parameters['zeros'])
train_inputs = prepare_dataset(train_sentences, word_to_id, char_to_id,
model.parameters['lower'], True)
preds = []
for citation in train_inputs:
inputs = create_input(citation, model.parameters, False)
y_pred = np.array(f[1](*inputs))[1:-1]
preds.append([(w, y_pred[i])
for i, w in enumerate(citation['str_words'])])
assert len(preds) == len(dataset.examples)
results = []
for P, T in zip(preds, dataset.examples):
for p, t in zip(P, zip(T.text, T.entity)):
results.append((p[1], tag_to_id[t[1]]))
pred, true = zip(*results)
eval_metrics = {
'micro_f1': f1_score(true, pred, average='micro'),
'macro_f1': f1_score(true, pred, average='macro')
}
data_file.close()
assert eval_metrics == pytest.approx({
'macro_f1': 0.984,
'micro_f1': 0.993
},
abs=0.001)
