def main():
model_metrics = metrics.BinaryMetricsRecorder(
domains=riskofbias.CORE_DOMAINS)
stupid_metrics = metrics.BinaryMetricsRecorder(
domains=riskofbias.CORE_DOMAINS)
# parse the risk of bias data from Cochrane
data = riskofbias.RoBData(test_mode=False)
data.generate_data(doc_level_only=False)
docs = riskofbias.MultiTaskSentFilter(data)
uids = np.array(docs.get_ids())
no_studies = len(uids)
kf = KFold(no_studies, n_folds=5, shuffle=False)
tuned_parameters = {"alpha": np.logspace(-4, -1, 5)}
vec = modhashvec.ModularVectorizer(
norm=None,
non_negative=True,
binary=True,
ngram_range=(1, 2),
n_features=2**26) # since multitask + bigrams = huge feature space
for train, test in kf:
y_train = docs.y(uids[train])
vec.builder_clear()
vec.builder_add_interaction_features(docs.X(uids[train]),
low=7) # add base features
vec.builder_add_interaction_features(docs.X_i(uids[train]),
low=2) # then add interactions
X_train = vec.builder_fit_transform()
clf = GridSearchCV(SGDClassifier(loss="hinge", penalty="L2"),
tuned_parameters,
scoring='recall')
# import pdb; pdb.set_trace()
clf.fit(X_train, y_train)
del X_train, y_train
clf = clf.best_estimator_ # and we only need the best performing, discard the rest
# Test on each domain in turn
# filtered_data = riskofbias.SentFilter(data)
for domain in riskofbias.CORE_DOMAINS:
print "Testing on %s" % domain
vec.builder_clear()
vec.builder_add_interaction_features(
docs.X(uids[test], domain=domain)) # add base features
vec.builder_add_interaction_features(
docs.X_i(uids[test], domain=domain)) # then add interactions
X_test = vec.builder_transform()
y_test = docs.y(uids[test], domain=domain)
y_preds = clf.predict(X_test)
model_metrics.add_preds_test(y_preds, y_test, domain=domain)
stupid_metrics.add_preds_test([-1] * len(y_test),
y_test,
domain=domain)
del X_test, y_test, y_preds
del clf
model_metrics.save_csv(
os.path.join('results', outputnames.filename(label="model")))
stupid_metrics.save_csv(
os.path.join('results', outputnames.filename(label="stupid-baseline")))
def main(out_dir="results"):
model_metrics = metrics.BinaryMetricsRecorder(domains=skip_domains)
stupid_metrics = metrics.BinaryMetricsRecorder(domains=skip_domains)
human_metrics = metrics.BinaryMetricsRecorder(domains=skip_domains)
# parse the risk of bias data from Cochrane
print "risk of bias data!"
data = riskofbias.RoBData(test_mode=False)
data.generate_data(doc_level_only=False)
# filter the data by Document
filtered_data = riskofbias.DocFilter(data)
# get the uids of the desired training set
# (for this experiment those which appear in only one review)
uids_all = filtered_data.get_ids(
pmid_instance=0) # those with 1 or more assessment (i.e. all)
uids_double_assessed = filtered_data.get_ids(
pmid_instance=1
) # those with 2 (or more) assessments (to hide for training)
uids_train = np.setdiff1d(uids_all, uids_double_assessed)
########################
# sentence prediction #
########################
# The first stage is to make the sentence prediction model using the
# training data set
#
print "First, making sentence prediction model"
sent_docs = riskofbias.MultiTaskSentFilter(data)
uids = np.array(sent_docs.get_ids())
no_studies = len(uids)
# sentence tokenization
sent_vec = modhashvec.ModularVectorizer(
norm=None,
non_negative=True,
binary=True,
ngram_range=(1, 2),
n_features=2**26) # since multitask + bigrams = huge feature space
sent_vec.builder_clear()
# add base features; this effectively generates the shared feature
# space (i.e., features for all domains)
sent_vec.builder_add_interaction_features(sent_docs.X(uids_train,
domain=skip_domains),
low=7)
# now we add interaction features, which cross the domain with the
# tokens. specifically, the X_i method returns token tuples crossing
# every term with every domain, and the vectorizer (an instance of
# ModularVectorizer) deals with inserting the actual interaction tokens
# that cross domains with tokens.
domain_interaction_tuples = sent_docs.X_i(uids_train, domain=skip_domains)
sent_vec.builder_add_interaction_features(domain_interaction_tuples, low=2)
# setup sentence classifier
tuned_parameters = {
"alpha": np.logspace(-4, -1, 5),
"class_weight": [{
1: i,
-1: 1
} for i in np.logspace(0, 2, 5)]
}
# bcw: are we sure we want to do 'recall' here, and not (e.g.) F1?
sent_clf = GridSearchCV(SGDClassifier(loss="hinge", penalty="L2"),
tuned_parameters,
scoring='recall')
X_train = sent_vec.builder_fit_transform()
y_train = sent_docs.y(uids_train, domain=skip_domains)
sent_clf.fit(X_train, y_train)
del X_train, y_train
# we only need the best performing
sent_clf = sent_clf.best_estimator_
# now we have our multi-task sentence prediction model,
# which we'll use to make sentence-level predictions for
# documents.
########################
# document prediction #
########################
# we need different test ids for each domain
# (since we're testing on studies with more than one RoB assessment for *each domain*)
docs = riskofbias.MultiTaskDocFilter(data)
X_train_d = docs.Xyi(uids_train, domain=skip_domains)
tuned_parameters = {"alpha": np.logspace(-4, -1, 5)}
clf = GridSearchCV(SGDClassifier(loss="hinge", penalty="L2"),
tuned_parameters,
scoring='f1')
# bcw: note that I've amended the y method to
# return interactions as well (i.e., domain strs)
y_train = docs.y(uids_train, domain=skip_domains)
# add interaction features (here both domain + high prob sentences)
interactions = {domain: [] for domain in skip_domains}
high_prob_sents = []
interaction_domains = []
for doc_index, (doc_text, doc_domain) in enumerate(X_train_d):
doc_sents = sent_tokenizer.tokenize(doc_text)
doc_domains = [doc_domain] * len(doc_sents)
# interactions
doc_X_i = izip(doc_sents, doc_domains)
# sent_vec is from above.
sent_vec.builder_clear()
sent_vec.builder_add_interaction_features(
doc_sents) # add base features
sent_vec.builder_add_interaction_features(
doc_X_i) # then add interactions
doc_sents_X = sent_vec.builder_transform()
## bcw -- shouldn't we use the *true* sentence labels
# here, rather than predictions????
# sent_clf was trained above
doc_sents_preds = sent_clf.predict(doc_sents_X)
high_prob_sents.append(" ".join([
sent for sent, sent_pred in zip(doc_sents, doc_sents_preds)
if sent_pred == 1
]))
interaction_domains.append("-s-" + doc_domain)
if doc_index % 10 == 0:
print doc_index
# from collections import Counter
# prob_count = Counter(list(doc_sents_preds))
# print prob_count
# for domain in riskofbias.CORE_DOMAINS:
# if domain == doc_domain:
# interactions[domain].append(True)
# else:
# interactions[domain].append(False)
vec = modhashvec.ModularVectorizer(
norm=None,
non_negative=True,
binary=True,
ngram_range=(1, 2),
n_features=2**26) # since multitask + bigrams = huge feature space
vec.builder_clear()
vec.builder_add_docs(docs.X(uids_train, domain=skip_domains),
low=7) # add base features
vec.builder_add_docs(docs.Xyi(uids_train, domain=skip_domains),
low=2) # add domain interactions
# removed X_train_d since already been through the generator! (needed reset)
vec.builder_add_docs(izip(high_prob_sents, interaction_domains),
low=2) # then add sentence interaction terms
X_train = vec.builder_fit_transform()
clf.fit(X_train, y_train)
############
# testing #
############
# Test on each domain in turn
for domain in skip_domains:
uids_domain_all = filtered_data.get_ids(pmid_instance=0,
filter_domain=domain)
uids_domain_double_assessed = filtered_data.get_ids(
pmid_instance=1, filter_domain=domain)
uids_test_domain = np.intersect1d(uids_domain_all,
uids_domain_double_assessed)
X_test_d, y_test = filtered_data.Xy(uids_test_domain,
domain=domain,
pmid_instance=0)
X_ignore, y_human = filtered_data.Xy(uids_test_domain,
domain=domain,
pmid_instance=1)
X_ignore = None # don't need this bit
#
# get high prob sents from test data
#
high_prob_sents = []
for doc_text in X_test_d:
doc_sents = sent_tokenizer.tokenize(doc_text)
# bcw -- I think this (using doc_domain and not
# domain) was the bug before!
#doc_domains = [doc_domain] * len(doc_sents)
doc_domains = [domain] * len(doc_sents)
doc_X_i = izip(doc_sents, doc_domains)
sent_vec.builder_clear()
sent_vec.builder_add_interaction_features(
doc_sents) # add base features
sent_vec.builder_add_interaction_features(
doc_X_i) # then add interactions
doc_sents_X = sent_vec.builder_transform()
doc_sents_preds = sent_clf.predict(doc_sents_X)
high_prob_sents.append(" ".join([
sent for sent, sent_pred in zip(doc_sents, doc_sents_preds)
if sent_pred == 1
]))
sent_domain_interactions = ["-s-" + domain] * len(high_prob_sents)
domain_interactions = [domain] * len(high_prob_sents)
print
print "domain: %s" % domain
print "High prob sents:"
print '\n'.join(high_prob_sents)
# build up test vector
vec.builder_clear()
vec.builder_add_docs(X_test_d) # add base features
vec.builder_add_docs(izip(X_test_d,
domain_interactions)) # add interactions
vec.builder_add_docs(
izip(high_prob_sents,
sent_domain_interactions)) # sentence interactions
X_test = vec.builder_transform()
y_preds = clf.predict(X_test)
model_metrics.add_preds_test(y_preds, y_test, domain=domain)
human_metrics.add_preds_test(y_human, y_test, domain=domain)
stupid_metrics.add_preds_test([1] * len(y_test), y_test, domain=domain)
model_metrics.save_csv(
os.path.join(out_dir, outputnames.filename(label="model")))
stupid_metrics.save_csv(
os.path.join(out_dir, outputnames.filename(label="stupid-baseline")))
human_metrics.save_csv(
os.path.join(out_dir, outputnames.filename(label="human-performance")))
def main():
model_metrics = metrics.BinaryMetricsRecorder(domains=riskofbias.CORE_DOMAINS)
stupid_metrics = metrics.BinaryMetricsRecorder(domains=riskofbias.CORE_DOMAINS)
f = open('test_data.csv','wb')
w = csv.DictWriter(f, ["pmid", "domain", "sent_text", "random", "human", "algorithm", "top3", "top1"], escapechar="\\")
w.writeheader()
# parse the risk of bias data from Cochrane
data = riskofbias.RoBData(test_mode=False)
data.generate_data(doc_level_only=False)
docs = riskofbias.MultiTaskSentFilter(data)
uids = np.array(docs.get_ids())
no_studies = len(uids)
kf = KFold(no_studies, n_folds=5, shuffle=False)
tuned_parameters = {"alpha": np.logspace(-4, -1, 5), "class_weight": [{1: i, -1: 1} for i in np.logspace(0, 2, 5)]}
vec = modhashvec.ModularVectorizer(norm=None, non_negative=True, binary=True, ngram_range=(1, 2), n_features=2**26) # since multitask + bigrams = huge feature space
for k_i, (train, test) in enumerate(kf):
if k_i == 1:
break
y_train = docs.y(uids[train])
vec.builder_clear()
vec.builder_add_interaction_features(docs.X(uids[train]), low=7) # add base features
vec.builder_add_interaction_features(docs.X_i(uids[train]), low=2) # then add interactions
X_train = vec.builder_fit_transform()
clf = GridSearchCV(SGDClassifier(loss="hinge", penalty="L2"), tuned_parameters, scoring='recall', n_jobs=16)
# import pdb; pdb.set_trace()
clf.fit(X_train, y_train)
del X_train, y_train
clf = clf.best_estimator_ # and we only need the best performing, discard the rest
# Test on each domain in turn
# filtered_data = riskofbias.SentFilter(data)
for domain in riskofbias.CORE_DOMAINS:
print "Testing on %s" % domain
vec.builder_clear()
vec.builder_add_interaction_features(docs.X(uids[test], domain=domain)) # add base features
vec.builder_add_interaction_features(docs.X_i(uids[test], domain=domain)) # then add interactions
X_test = vec.builder_transform()
y_test = docs.y(uids[test], domain=domain)
y_preds = clf.predict(X_test)
y_df = clf.decision_function(X_test) # get distances from the decision boundary
# positive distances = more likely to be relevant sentences
r_len = len(y_preds)
y_top3 = []
y_top1 = []
y_rand = []
y_uids = np.array(docs.y_uids(uids[test], domain=domain))
# import pdb; pdb.set_trace()
for y_uid in np.unique(y_uids):
mask = np.where(y_uids == y_uid)[0]
doc_df = y_df[mask]
doc_top3 = np.argpartition(doc_df, -3)[-3:]
y_top3.extend(list(mask[doc_top3]))
doc_top1 = np.argmax(doc_df)
y_top1.append(mask[doc_top1])
doc_rand = np.random.randint(0, len(doc_df))
y_rand.append(mask[doc_rand])
human_sent_indices = np.where(y_test==1)[0]
algorithm_sent_indices = np.where(y_preds==1)[0]
model_metrics.add_preds_test(y_preds, y_test, domain=domain)
stupid_metrics.add_preds_test([-1] * len(y_test), y_test, domain=domain)
# import pdb; pdb.set_trace()
for doc_i, (doc, pmid) in enumerate(izip(docs.X(uids[test], domain=domain), docs.iter_pmid(uids[test], domain=domain))):
row = {"domain": domain,
"sent_text": doc,
"random": doc_i in y_rand,
"human": doc_i in human_sent_indices,
"algorithm": doc_i in algorithm_sent_indices,
"top3": doc_i in y_top3,
"top1": doc_i in y_top1,
"pmid": pmid}
if row["random"] or row["human"] or row["top3"] or row["top1"]:
# please note, the sentences will only be included in the analysis if
# in the top1 or top3
# we do have data on whether the raw classifier has predicted yes/no
#
# this in effect means where the classifier picks <= 3 sentences
# we use all raw classifier data
# where >3 sentences are predicted by raw classifier, only the
# top 3 are used; the rest are discarded
w.writerow(row)
del X_test, y_test, y_preds
del clf
model_metrics.save_csv(os.path.join('results', outputnames.filename(label="model")))
stupid_metrics.save_csv(os.path.join('results', outputnames.filename(label="stupid-baseline")))
f.close()
def _load_domain_map(filename=os.path.join(cochranenlp.PATH, "data", "domain_names.txt")):
with codecs.open(filename, 'rb', 'utf-8') as f:
raw_data = yaml.load(f)
mapping = {}
for key, value in raw_data.iteritems():
for synonym in value:
mapping[synonym] = key
return mapping
data = riskofbias.RoBData(test_mode=False)
data.generate_data(doc_level_only=False)
mapper = _load_domain_map()
all_pmids = Counter()
domains_present = defaultdict(Counter)
b = biviewer.BiViewer()
print "getting all pubmed ids in CDSR..."
p = ProgressBar(len(b))
for doc in b:
p.tap()
def main():
model_metrics = metrics.BinaryMetricsRecorder(domains=target_domains)
stupid_metrics = metrics.BinaryMetricsRecorder(domains=target_domains)
human_metrics = metrics.BinaryMetricsRecorder(domains=target_domains)
# parse the risk of bias data from Cochrane
data = riskofbias.RoBData(test_mode=False)
data.generate_data(doc_level_only=True)
# filter the data by Document
filtered_data = riskofbias.DocFilter(data)
# get the uids of the desired training set
# (for this experiment those which appear in only one review)
uids_all = filtered_data.get_ids(
pmid_instance=0) # those with 1 or more assessment (i.e. all)
uids_double_assessed = filtered_data.get_ids(
pmid_instance=1
) # those with 2 (or more) assessments (to hide for training)
uids_train = np.setdiff1d(uids_all, uids_double_assessed)
# we need different test ids for each domain
# (since we're testing on studies with more than one RoB assessment for *each domain*)
docs = riskofbias.MultiTaskDocFilter(data)
tuned_parameters = {"alpha": np.logspace(-4, -1, 10)}
clf = GridSearchCV(SGDClassifier(loss="hinge", penalty="L2"),
tuned_parameters,
scoring='accuracy')
X_train_d, y_train, i_train = docs.Xyi(uids_train, pmid_instance=0)
interactions = {domain: [] for domain in target_domains}
for doc_text, doc_domain in zip(X_train_d, i_train):
for domain in target_domains:
if domain == doc_domain:
interactions[domain].append(True)
else:
interactions[domain].append(False)
vec = modhashvec.ModularVectorizer(
norm=None,
non_negative=True,
binary=True,
ngram_range=(1, 2),
n_features=2**26) # since multitask + bigrams = huge feature space
vec.builder_clear()
vec.builder_add_docs(X_train_d, low=10) # add base features
for domain in target_domains:
print np.sum(interactions[domain]), "/", len(
interactions[domain]), "added for", domain
vec.builder_add_interaction_features(X_train_d,
interactions=interactions[domain],
prefix=domain + "_I_",
low=2) # then add interactions
X_train = vec.builder_fit_transform()
clf.fit(X_train, y_train)
# free some memory now, only need the model
del X_train_d # remove references to these
del X_train
del y_train
clf = clf.best_estimator_ # and we only need the best performing, discard the rest
# Test on each domain in turn
for domain in target_domains:
uids_domain_all = filtered_data.get_ids(pmid_instance=0,
filter_domain=domain)
uids_domain_double_assessed = filtered_data.get_ids(
pmid_instance=1, filter_domain=domain)
uids_test_domain = np.intersect1d(uids_domain_all,
uids_domain_double_assessed)
X_test_d, y_test = filtered_data.Xy(uids_test_domain,
domain=domain,
pmid_instance=0)
X_ignore, y_human = filtered_data.Xy(uids_test_domain,
domain=domain,
pmid_instance=1)
X_ignore = None # don't need this bit
# build up test vector
vec.builder_clear()
vec.builder_add_docs(X_test_d) # add base features
vec.builder_add_docs(X_test_d,
prefix=domain + '_I_') # add interactions
X_test = vec.builder_transform()
y_preds = clf.predict(X_test)
model_metrics.add_preds_test(y_preds, y_test, domain=domain)
human_metrics.add_preds_test(y_human, y_test, domain=domain)
stupid_metrics.add_preds_test([1] * len(y_test), y_test, domain=domain)
model_metrics.save_csv(
os.path.join('results', outputnames.filename(label="model")))
stupid_metrics.save_csv(
os.path.join('results', outputnames.filename(label="stupid-baseline")))
human_metrics.save_csv(
os.path.join('results',
outputnames.filename(label="human-performance")))
def main():
model_metrics = metrics.BinaryMetricsRecorder(
domains=riskofbias.CORE_DOMAINS)
stupid_metrics = metrics.BinaryMetricsRecorder(
domains=riskofbias.CORE_DOMAINS)
human_metrics = metrics.BinaryMetricsRecorder(
domains=riskofbias.CORE_DOMAINS)
# parse the risk of bias data from Cochrane
data = riskofbias.RoBData(test_mode=False)
data.generate_data(doc_level_only=False)
# filter the data by Document
filtered_data = riskofbias.DocFilter(data)
# get the uids of the desired training set
# (for this experiment those which appear in only one review)
uids_all = filtered_data.get_ids(
pmid_instance=0) # those with 1 or more assessment (i.e. all)
uids_double_assessed = filtered_data.get_ids(
pmid_instance=1
) # those with 2 (or more) assessments (to hide for training)
uids_train = np.setdiff1d(uids_all, uids_double_assessed)
###
### sentence prediction
###
# The first stage is to make the sentence prediction model using the
# training data set
#
print "First, making sentence prediction model"
sent_docs = riskofbias.SentFilter(data)
sent_models = {} #where the key is the domain name
sent_vec = modhashvec.InteractionHashingVectorizer(
norm=None,
non_negative=True,
binary=True,
ngram_range=(1, 2),
n_features=2**24
) # hashing vectorizer so doesn't change per domain in terms of feature space
for domain in riskofbias.CORE_DOMAINS:
sent_uids = np.intersect1d(
uids_train, np.array(sent_docs.get_ids(filter_domain=domain)))
no_studies = len(sent_uids)
print "%d docs obtained for domain: %s" % (no_studies, domain)
tuned_parameters = {
"alpha": np.logspace(-4, -1, 5),
"class_weight": [{
1: i,
-1: 1
} for i in np.logspace(-1, 1, 5)]
}
clf = GridSearchCV(SGDClassifier(loss="hinge", penalty="L2"),
tuned_parameters,
scoring='recall')
X_train_d, y_train = sent_docs.Xy(sent_uids, domain=domain)
X_train = sent_vec.fit_transform(X_train_d, low=2)
clf.fit(X_train, y_train)
sent_models[domain] = clf.best_estimator_
# import pdb; pdb.set_trace()
# we need different test ids for each domain
# (since we're testing on studies with more than one RoB assessment for *each domain*)
docs = riskofbias.MultiTaskDocFilter(data)
tuned_parameters = {"alpha": np.logspace(-4, -1, 10)}
clf = GridSearchCV(SGDClassifier(loss="hinge", penalty="L2"),
tuned_parameters,
scoring='accuracy')
# X_train_d, y_train, i_train = docs.Xyi(uids_train, pmid_instance=0)
# add interaction features (here both domain + high prob sentences)
# interactions = {domain:[] for domain in riskofbias.CORE_DOMAINS}
high_prob_sents = []
for doc_text, doc_domain in zip(X_train_d, i_train):
doc_sents = sent_tokenizer.tokenize(doc_text)
doc_sents_X = sent_vec.transform(doc_sents)
doc_sents_preds = sent_models[doc_domain].predict(doc_sents_X)
high_prob_sents.append(" ".join([
sent for sent, sent_pred in zip(doc_sents, doc_sents_preds)
if sent_pred == 1
]))
print "high prob sents:"
from collections import Counter
prob_count = Counter(list(doc_sents_preds))
print prob_count
for domain in riskofbias.CORE_DOMAINS:
if domain == doc_domain:
interactions[domain].append(True)
else:
interactions[domain].append(False)
vec = modhashvec.ModularVectorizer(
norm=None,
non_negative=True,
binary=True,
ngram_range=(1, 2),
n_features=2**26) # since multitask + bigrams = huge feature space
vec.builder_clear()
vec.builder_add_docs(X_train_d, low=10) # add base features
# print high_prob_sents
for domain in riskofbias.CORE_DOMAINS:
print np.sum(interactions[domain]), "/", len(
interactions[domain]), "added for", domain
vec.builder_add_docs(X_train_d,
interactions=interactions[domain],
prefix=domain + "-i-",
low=2) # then add interactions
vec.builder_add_docs(high_prob_sents, prefix="-s-", low=2)
X_train = vec.builder_fit_transform()
clf.fit(X_train, y_train)
# Test on each domain in turn
for domain in riskofbias.CORE_DOMAINS:
uids_domain_all = filtered_data.get_ids(pmid_instance=0,
filter_domain=domain)
uids_domain_double_assessed = filtered_data.get_ids(
pmid_instance=1, filter_domain=domain)
uids_test_domain = np.intersect1d(uids_domain_all,
uids_domain_double_assessed)
X_test_d, y_test = filtered_data.Xy(uids_test_domain,
domain=domain,
pmid_instance=0)
X_ignore, y_human = filtered_data.Xy(uids_test_domain,
domain=domain,
pmid_instance=1)
X_ignore = None # don't need this bit
#
# get high prob sents from test data
#
high_prob_sents = []
for doc_text in X_test_d:
doc_sents = sent_tokenizer.tokenize(doc_text)
doc_sents_X = sent_vec.transform(doc_sents)
doc_sents_preds = sent_models[domain].predict(doc_sents_X)
high_prob_sents.append(" ".join([
sent for sent, sent_pred in zip(doc_sents, doc_sents_preds)
if sent_pred == 1
]))
# build up test vector
vec.builder_clear()
vec.builder_add_docs(X_test_d) # add base features
vec.builder_add_docs(X_test_d,
prefix=domain + '-i-') # add interactions
vec.builder_add_docs(high_prob_sents, prefix="-s-")
X_test = vec.builder_transform()
y_preds = clf.predict(X_test)
model_metrics.add_preds_test(y_preds, y_test, domain=domain)
human_metrics.add_preds_test(y_human, y_test, domain=domain)
stupid_metrics.add_preds_test([1] * len(y_test), y_test, domain=domain)
model_metrics.save_csv(
os.path.join('results', outputnames.filename(label="model")))
stupid_metrics.save_csv(
os.path.join('results', outputnames.filename(label="stupid-baseline")))
human_metrics.save_csv(
os.path.join('results',
outputnames.filename(label="human-performance")))
def main(out_dir="results"):
# parse the risk of bias data from Cochrane
print "risk of bias data!"
data = riskofbias.RoBData(test_mode=False)
data.generate_data(doc_level_only=False, skip_small_files=True)
# filter the data by Document
filtered_data = riskofbias.DocFilter(data)
# get the uids of the desired training set
# (for this experiment those which appear in only one review)
uids_all = filtered_data.get_ids(
pmid_instance=0) # those with 1 or more assessment (i.e. all)
########################
# sentence prediction #
########################
# The first stage is to make the sentence prediction model using the
# training data set
#
print "First, making sentence prediction model"
sent_docs = riskofbias.MultiTaskSentFilter(data)
uids = np.array(sent_docs.get_ids())
no_studies = len(uids)
# sentence tokenization
sent_vec = modhashvec.ModularVectorizer(
norm=None,
non_negative=True,
binary=True,
ngram_range=(1, 2),
n_features=2**26) # since multitask + bigrams = huge feature space
sent_vec.builder_clear()
# add base features; this effectively generates the shared feature
# space (i.e., features for all domains)
sent_vec.builder_add_interaction_features(sent_docs.X(uids_all,
domain=skip_domains),
low=7)
# now we add interaction features, which cross the domain with the
# tokens. specifically, the X_i method returns token tuples crossing
# every term with every domain, and the vectorizer (an instance of
# ModularVectorizer) deals with inserting the actual interaction tokens
# that cross domains with tokens.
domain_interaction_tuples = sent_docs.X_i(uids_all, domain=skip_domains)
sent_vec.builder_add_interaction_features(domain_interaction_tuples, low=2)
# setup sentence classifier
tuned_parameters = {
"alpha": np.logspace(-4, -1, 5),
"class_weight": [{
1: i,
-1: 1
} for i in np.logspace(-1, 1, 5)]
}
# bcw: are we sure we want to do 'recall' here, and not (e.g.) F1?
sent_clf = GridSearchCV(SGDClassifier(loss="hinge",
penalty="L2",
shuffle=True),
tuned_parameters,
scoring='recall')
X_train = sent_vec.builder_fit_transform()
y_train = sent_docs.y(uids_all, domain=skip_domains)
sent_clf.fit(X_train, y_train)
del X_train, y_train
# we only need the best performing
sent_clf = sent_clf.best_estimator_
# now we have our multi-task sentence prediction model,
# which we'll use to make sentence-level predictions for
# documents.
########################
# document prediction #
########################
# we need different test ids for each domain
# (since we're testing on studies with more than one RoB assessment for *each domain*)
docs = riskofbias.MultiTaskDocFilter(data)
X_train_d = docs.Xyi(uids_all, domain=skip_domains)
tuned_parameters = {"alpha": np.logspace(-2, 2, 10)}
clf = GridSearchCV(SGDClassifier(loss="hinge", penalty="L2", shuffle=True),
tuned_parameters,
scoring='f1')
# bcw: note that I've amended the y method to
# return interactions as well (i.e., domain strs)
y_train = docs.y(uids_all, domain=skip_domains)
# add interaction features (here both domain + high prob sentences)
interactions = {domain: [] for domain in skip_domains}
high_prob_sents = []
interaction_domains = []
for doc_index, (doc_text, doc_domain) in enumerate(X_train_d):
doc_sents = sent_tokenizer.tokenize(doc_text)
doc_domains = [doc_domain] * len(doc_sents)
# interactions
doc_X_i = izip(doc_sents, doc_domains)
# sent_vec is from above.
sent_vec.builder_clear()
sent_vec.builder_add_interaction_features(
doc_sents) # add base features
sent_vec.builder_add_interaction_features(
doc_X_i) # then add interactions
doc_sents_X = sent_vec.builder_transform()
## bcw -- shouldn't we use the *true* sentence labels
# here, rather than predictions????
# sent_clf was trained above
doc_sents_preds = sent_clf.predict(doc_sents_X)
high_prob_sents.append(" ".join([
sent for sent, sent_pred in zip(doc_sents, doc_sents_preds)
if sent_pred == 1
]))
interaction_domains.append("-s-" + doc_domain)
if doc_index % 10 == 0:
print doc_index
# from collections import Counter
# prob_count = Counter(list(doc_sents_preds))
# print prob_count
# for domain in riskofbias.CORE_DOMAINS:
# if domain == doc_domain:
# interactions[domain].append(True)
# else:
# interactions[domain].append(False)
vec = modhashvec.ModularVectorizer(
norm=None,
non_negative=True,
binary=True,
ngram_range=(1, 2),
n_features=2**26) # since multitask + bigrams = huge feature space
vec.builder_clear()
vec.builder_add_docs(docs.X(uids_all, domain=skip_domains),
low=7) # add base features
vec.builder_add_docs(docs.Xyi(uids_all, domain=skip_domains),
low=2) # add domain interactions
# removed X_train_d since already been through the generator! (needed reset)
vec.builder_add_docs(izip(high_prob_sents, interaction_domains),
low=2) # then add sentence interaction terms
X_train = vec.builder_fit_transform()
clf.fit(X_train, y_train)
clf = clf.best_estimator_
with open('mt_mt_production_models.pck', 'wb') as f:
pickle.dump((sent_clf, clf), f)