Use word2bits to create a Hamming dataset

ann_benchmarks/datasets.py

@@ -40,8 +40,8 @@ def write_output(train, test, fn, distance, count=100):
     f.attrs['distance'] = distance
     print('train size: %9d * %4d' % train.shape)
     print('test size:  %9d * %4d' % test.shape)
-    f.create_dataset('train', (len(train), len(train[0])), dtype='f')[:] = train
-    f.create_dataset('test', (len(test), len(test[0])), dtype='f')[:] = test
+    f.create_dataset('train', (len(train), len(train[0])), dtype=train.dtype)[:] = train
+    f.create_dataset('test', (len(test), len(test[0])), dtype=test.dtype)[:] = test
     neighbors = f.create_dataset('neighbors', (len(test), count), dtype='i')
     distances = f.create_dataset('distances', (len(test), count), dtype='f')
     bf = BruteForceBLAS(distance, precision=numpy.float32)
@@ -57,8 +57,13 @@ def write_output(train, test, fn, distance, count=100):
     f.close()
 
 
-def glove(out_fn, d):
+def train_test_split(X, test_size=10000):
     import sklearn.model_selection
+    print('Splitting %d*%d into train/test' % X.shape)
+    return sklearn.model_selection.train_test_split(X, test_size=test_size, random_state=1)
+
+
+def glove(out_fn, d):
     import zipfile
 
     url = 'http://nlp.stanford.edu/data/glove.twitter.27B.zip'
@@ -71,9 +76,7 @@ def glove(out_fn, d):
         for line in z.open(z_fn):
             v = [float(x) for x in line.strip().split()[1:]]
             X.append(numpy.array(v))
-        print('splitting output...')
-        X_train, X_test = sklearn.model_selection.train_test_split(X, test_size=10000, random_state=1)
-        print('writing output...')
+        X_train, X_test = train_test_split(X)
         write_output(numpy.array(X_train), numpy.array(X_test), out_fn, 'angular')
 
 
@@ -170,7 +173,6 @@ def fashion_mnist(out_fn):
 
 def transform_bag_of_words(filename, n_dimensions, out_fn):
     import gzip
-    import sklearn.model_selection
     from scipy.sparse import lil_matrix
     from sklearn.feature_extraction.text import TfidfTransformer
     from sklearn import random_projection
@@ -188,24 +190,41 @@ def transform_bag_of_words(filename, n_dimensions, out_fn):
         B = TfidfTransformer().fit_transform(A)
         print("reducing dimensionality...")
         C = random_projection.GaussianRandomProjection(n_components = n_dimensions).fit_transform(B)
-        X_train, X_test = sklearn.model_selection.train_test_split(C, test_size=10000, random_state=1)
-        print('writing output...')
+        X_train, X_test = train_test_split(C)
         write_output(numpy.array(X_train), numpy.array(X_test), out_fn, 'angular')
 
 
 def nytimes(out_fn, n_dimensions):
-    download('https://archive.ics.uci.edu/ml/machine-learning-databases/bag-of-words/docword.nytimes.txt.gz', 'nytimes.txt.gz')
-    transform_bag_of_words('nytimes.txt.gz', n_dimensions, out_fn)
+    fn = 'nytimes_%s.txt.gz' % n_dimensions
+    download('https://archive.ics.uci.edu/ml/machine-learning-databases/bag-of-words/docword.nytimes.txt.gz', fn)
+    transform_bag_of_words(fn, n_dimensions, out_fn)
 
 
 def random(out_fn, n_dims, n_samples, centers, distance):
-    import sklearn.model_selection
     import sklearn.datasets
 
     X, _ = sklearn.datasets.make_blobs(n_samples=n_samples, n_features=n_dims, centers=centers, random_state=1)
-    X_train, X_test = sklearn.model_selection.train_test_split(X, test_size=0.1, random_state=1)
+    X_train, X_test = train_test_split(X, test_size=0.1)
     write_output(X_train, X_test, out_fn, distance)
 
+
+def word2bits(out_fn, path, fn):
+    import tarfile
+    local_fn = fn + '.tar.gz'
+    url = 'http://web.stanford.edu/~maxlam/word_vectors/compressed/%s/%s.tar.gz' % (path, fn)
+    download(url, local_fn)
+    print('parsing vectors in %s...' % local_fn)
+    with tarfile.open(local_fn, 'r:gz') as t:
+        f = t.extractfile(fn)
+        n_words, k = [int(z) for z in next(f).strip().split()]
+        X = numpy.zeros((n_words, k), dtype=numpy.bool)
+        for i in range(n_words):
+            X[i] = [float(z) > 0 for z in next(f).strip().split()[1:]]
+
+        X_train, X_test = train_test_split(X)
+        write_output(X_train, X_test, out_fn, 'euclidean')  # TODO: use hamming
+
+
 DATASETS = {
     'fashion-mnist-784-euclidean': fashion_mnist,
     'gist-960-euclidean': gist,
@@ -221,4 +240,5 @@ def random(out_fn, n_dims, n_samples, centers, distance):
     'sift-128-euclidean': sift,
     'nytimes-256-angular': lambda out_fn: nytimes(out_fn, 256),
     'nytimes-16-angular': lambda out_fn: nytimes(out_fn, 16),
+    'word2bits-800-hamming': lambda out_fn: word2bits(out_fn, '400K', 'w2b_bitlevel1_size800_vocab400K'),
 }