summaryrefslogtreecommitdiff
path: root/corpus
diff options
context:
space:
mode:
authormjdenkowski <michael.j.denkowski@gmail.com>2014-12-20 03:01:42 -0500
committermjdenkowski <michael.j.denkowski@gmail.com>2014-12-20 03:01:42 -0500
commitd8e9f8c4df61a8162dcb7ac0e53c416eeeb36d26 (patch)
treef0db1494a6cd191c2291ccab541cd56e8a29f5eb /corpus
parent2723909270005be4f5cb7ede296b53d5bacf57dc (diff)
Generalize to sample any number of dev sets
Diffstat (limited to 'corpus')
-rwxr-xr-xcorpus/sample-dev-sets.py74
-rwxr-xr-xcorpus/sample-dev-test.py65
2 files changed, 74 insertions, 65 deletions
diff --git a/corpus/sample-dev-sets.py b/corpus/sample-dev-sets.py
new file mode 100755
index 00000000..3c969bbe
--- /dev/null
+++ b/corpus/sample-dev-sets.py
@@ -0,0 +1,74 @@
+#!/usr/bin/env python
+
+import gzip
+import os
+import sys
+
+HELP = '''Process an input corpus by dividing it into pseudo-documents and uniformly
+sampling train and dev sets (simulate uniform sampling at the document level
+when document boundaries are unknown)
+
+usage: {} in_file out_prefix doc_size docs_per_dev_set dev_sets [-lc]
+recommended: doc_size=20, docs_per_dev_set=100, dev_sets=2 (dev and test)
+'''
+
+def gzopen(f):
+ return gzip.open(f, 'rb') if f.endswith('.gz') else open(f, 'r')
+
+def wc(f):
+ return sum(1 for _ in gzopen(f))
+
+def main(argv):
+
+ if len(argv[1:]) < 5:
+ sys.stderr.write(HELP.format(os.path.basename(argv[0])))
+ sys.exit(2)
+
+ # Args
+ in_file = os.path.abspath(argv[1])
+ out_prefix = os.path.abspath(argv[2])
+ doc_size = int(argv[3])
+ docs_per_dev_set = int(argv[4])
+ dev_sets = int(argv[5])
+ lc = (len(argv[1:]) == 6 and argv[6] == '-lc')
+
+ # Compute sizes
+ corpus_size = wc(in_file)
+ total_docs = corpus_size / doc_size
+ leftover = corpus_size % doc_size
+ train_docs = total_docs - (dev_sets * docs_per_dev_set)
+ train_batch_size = (train_docs / docs_per_dev_set)
+
+ # Report
+ sys.stderr.write('Splitting {} lines ({} documents)\n'.format(corpus_size, total_docs + (1 if leftover else 0)))
+ sys.stderr.write('Train: {} ({})\n'.format((train_docs * doc_size) + leftover, train_docs + (1 if leftover else 0)))
+ sys.stderr.write('Dev: {} x {} ({})\n'.format(dev_sets, docs_per_dev_set * doc_size, docs_per_dev_set))
+
+ inp = gzopen(in_file)
+ train_out = open('{}.train'.format(out_prefix), 'w')
+ dev_out = [open('{}.dev.{}'.format(out_prefix, i + 1), 'w') for i in range(dev_sets)]
+ i = 0
+
+ # For each set of documents
+ for _ in range(docs_per_dev_set):
+ # Write several documents to train
+ for _ in range(train_batch_size):
+ for _ in range(doc_size):
+ i += 1
+ train_out.write('{} ||| {}'.format(i, inp.readline()) if lc else inp.readline())
+ # Write a document to each dev
+ for out in dev_out:
+ for _ in range(doc_size):
+ i += 1
+ out.write('{} ||| {}'.format(i, inp.readline()) if lc else inp.readline())
+ # Write leftover lines to train
+ for line in inp:
+ i += 1
+ train_out.write('{} ||| {}'.format(i, line) if lc else line)
+
+ train_out.close()
+ for out in dev_out:
+ out.close()
+
+if __name__ == '__main__':
+ main(sys.argv)
diff --git a/corpus/sample-dev-test.py b/corpus/sample-dev-test.py
deleted file mode 100755
index 0c0514ee..00000000
--- a/corpus/sample-dev-test.py
+++ /dev/null
@@ -1,65 +0,0 @@
-#!/usr/bin/env python
-
-import gzip
-import os
-import sys
-
-HELP = '''Process an input corpus by dividing it into pseudo-documents and uniformly
-sampling train, dev, and test sets (simulate uniform sampling at the document
-level when document boundaries are unknown)
-
-usage: {} in_file out_prefix doc_size dev_test_docs [-lc]
-recommended: doc_size=20, dev_test_docs=100
-'''
-
-def gzopen(f):
- return gzip.open(f, 'rb') if f.endswith('.gz') else open(f, 'r')
-
-def wc(f):
- return sum(1 for _ in gzopen(f))
-
-def main(argv):
-
- if len(argv[1:]) < 4:
- sys.stderr.write(HELP.format(os.path.basename(argv[0])))
- sys.exit(2)
-
- in_file = os.path.abspath(argv[1])
- out_prefix = os.path.abspath(argv[2])
- doc_size = int(argv[3])
- dev_test_docs = int(argv[4])
- lc = (len(argv[1:]) == 5 and argv[5] == '-lc')
-
- corpus_size = wc(in_file)
- total_docs = corpus_size / doc_size
- leftover = corpus_size % doc_size
- train_docs = total_docs - (2 * dev_test_docs)
- train_batch_size = (train_docs / dev_test_docs) - 2
-
- sys.stderr.write('Splitting {} lines ({} documents)\n'.format(corpus_size, total_docs + (1 if leftover else 0)))
- sys.stderr.write('Train: {} ({})\n'.format((train_docs * doc_size) + leftover, train_docs + (1 if leftover else 0)))
- sys.stderr.write('Dev: {} ({})\n'.format(dev_test_docs * doc_size, dev_test_docs))
- sys.stderr.write('Test: {} ({})\n'.format(dev_test_docs * doc_size, dev_test_docs))
-
- with gzopen(in_file) as inp, \
- open('{}.train'.format(out_prefix), 'w') as train_out, \
- open('{}.dev'.format(out_prefix), 'w') as dev_out, \
- open('{}.test'.format(out_prefix), 'w') as test_out:
- i = 0
- for _ in range(dev_test_docs):
- for _ in range(train_batch_size):
- for _ in range(doc_size):
- i += 1
- train_out.write('{} ||| {}'.format(i, inp.readline()) if lc else inp.readline())
- for _ in range(doc_size):
- i += 1
- dev_out.write('{} ||| {}'.format(i, inp.readline()) if lc else inp.readline())
- for _ in range(doc_size):
- i += 1
- test_out.write('{} ||| {}'.format(i, inp.readline()) if lc else inp.readline())
- for line in inp:
- i += 1
- train_out.write('{} ||| {}'.format(i, line) if lc else line)
-
-if __name__ == '__main__':
- main(sys.argv)