diff options
Diffstat (limited to 'Wasserstein_Distance.py')
-rw-r--r-- | Wasserstein_Distance.py | 136 |
1 files changed, 72 insertions, 64 deletions
diff --git a/Wasserstein_Distance.py b/Wasserstein_Distance.py index 78bf9cf..60991b9 100644 --- a/Wasserstein_Distance.py +++ b/Wasserstein_Distance.py | |||
@@ -11,17 +11,20 @@ from sklearn.utils import check_array | |||
11 | 11 | ||
12 | class WassersteinMatcher(KNeighborsClassifier): | 12 | class WassersteinMatcher(KNeighborsClassifier): |
13 | """ | 13 | """ |
14 | Implements a nearest neighbors classifier for input distributions using the Wasserstein distance as metric. | 14 | Source and target distributions are l_1 normalized before computing the Wasserstein |
15 | Source and target distributions are l_1 normalized before computing the Wasserstein distance. | 15 | distance. Wasserstein is parametrized by the distances between the individual |
16 | Wasserstein is parametrized by the distances between the individual points of the distributions. | 16 | points of the distributions. |
17 | """ | 17 | """ |
18 | def __init__(self, | 18 | |
19 | W_embed, | 19 | def __init__( |
20 | n_neighbors=1, | 20 | self, |
21 | n_jobs=1, | 21 | W_embed, |
22 | verbose=False, | 22 | n_neighbors=1, |
23 | sinkhorn=False, | 23 | n_jobs=1, |
24 | sinkhorn_reg=0.1): | 24 | verbose=False, |
25 | sinkhorn=False, | ||
26 | sinkhorn_reg=0.1, | ||
27 | ): | ||
25 | """ | 28 | """ |
26 | Initialization of the class. | 29 | Initialization of the class. |
27 | Arguments | 30 | Arguments |
@@ -33,10 +36,12 @@ class WassersteinMatcher(KNeighborsClassifier): | |||
33 | self.sinkhorn_reg = sinkhorn_reg | 36 | self.sinkhorn_reg = sinkhorn_reg |
34 | self.W_embed = W_embed | 37 | self.W_embed = W_embed |
35 | self.verbose = verbose | 38 | self.verbose = verbose |
36 | super(WassersteinMatcher, self).__init__(n_neighbors=n_neighbors, | 39 | super(WassersteinMatcher, self).__init__( |
37 | n_jobs=n_jobs, | 40 | n_neighbors=n_neighbors, |
38 | metric='precomputed', | 41 | n_jobs=n_jobs, |
39 | algorithm='brute') | 42 | metric="precomputed", |
43 | algorithm="brute", | ||
44 | ) | ||
40 | 45 | ||
41 | def _wmd(self, i, row, X_train): | 46 | def _wmd(self, i, row, X_train): |
42 | union_idx = np.union1d(X_train[i].indices, row.indices) | 47 | union_idx = np.union1d(X_train[i].indices, row.indices) |
@@ -51,7 +56,7 @@ class WassersteinMatcher(KNeighborsClassifier): | |||
51 | W_dist, | 56 | W_dist, |
52 | self.sinkhorn_reg, | 57 | self.sinkhorn_reg, |
53 | numItermax=50, | 58 | numItermax=50, |
54 | method='sinkhorn_stabilized', | 59 | method="sinkhorn_stabilized", |
55 | )[0] | 60 | )[0] |
56 | else: | 61 | else: |
57 | return ot.emd2(bow_i, bow_j, W_dist) | 62 | return ot.emd2(bow_i, bow_j, W_dist) |
@@ -66,27 +71,27 @@ class WassersteinMatcher(KNeighborsClassifier): | |||
66 | 71 | ||
67 | if X_train is None: | 72 | if X_train is None: |
68 | X_train = self._fit_X | 73 | X_train = self._fit_X |
69 | pool = Pool(nodes=self.n_jobs | 74 | pool = Pool( |
70 | ) # Parallelization of the calculation of the distances | 75 | nodes=self.n_jobs |
76 | ) # Parallelization of the calculation of the distances | ||
71 | dist = pool.map(self._wmd_row, X_test) | 77 | dist = pool.map(self._wmd_row, X_test) |
72 | return np.array(dist) | 78 | return np.array(dist) |
73 | 79 | ||
74 | def fit(self, X, y): # X_train_idf | 80 | def fit(self, X, y): # X_train_idf |
75 | X = check_array(X, accept_sparse='csr', | 81 | X = check_array(X, accept_sparse="csr", copy=True) # check if array is sparse |
76 | copy=True) # check if array is sparse | 82 | X = normalize(X, norm="l1", copy=False) |
77 | X = normalize(X, norm='l1', copy=False) | ||
78 | return super(WassersteinMatcher, self).fit(X, y) | 83 | return super(WassersteinMatcher, self).fit(X, y) |
79 | 84 | ||
80 | def predict(self, X): | 85 | def predict(self, X): |
81 | X = check_array(X, accept_sparse='csr', copy=True) | 86 | X = check_array(X, accept_sparse="csr", copy=True) |
82 | X = normalize(X, norm='l1', copy=False) | 87 | X = normalize(X, norm="l1", copy=False) |
83 | dist = self._pairwise_wmd(X) | 88 | dist = self._pairwise_wmd(X) |
84 | dist = dist * 1000 # for lapjv, small floating point numbers are evil | 89 | dist = dist * 1000 # for lapjv, small floating point numbers are evil |
85 | return super(WassersteinMatcher, self).predict(dist) | 90 | return super(WassersteinMatcher, self).predict(dist) |
86 | 91 | ||
87 | def kneighbors(self, X, n_neighbors=1): | 92 | def kneighbors(self, X, n_neighbors=1): |
88 | X = check_array(X, accept_sparse='csr', copy=True) | 93 | X = check_array(X, accept_sparse="csr", copy=True) |
89 | X = normalize(X, norm='l1', copy=False) | 94 | X = normalize(X, norm="l1", copy=False) |
90 | dist = self._pairwise_wmd(X) | 95 | dist = self._pairwise_wmd(X) |
91 | dist = dist * 1000 # for lapjv, small floating point numbers are evil | 96 | dist = dist * 1000 # for lapjv, small floating point numbers are evil |
92 | return lapjv(dist) | 97 | return lapjv(dist) |
@@ -102,19 +107,24 @@ class WassersteinMatcher(KNeighborsClassifier): | |||
102 | percentage = p_at_one / n_neighbors * 100 | 107 | percentage = p_at_one / n_neighbors * 100 |
103 | return p_at_one, percentage | 108 | return p_at_one, percentage |
104 | 109 | ||
110 | |||
105 | class WassersteinRetriever(KNeighborsClassifier): | 111 | class WassersteinRetriever(KNeighborsClassifier): |
106 | """ | 112 | """ |
107 | Implements a nearest neighbors classifier for input distributions using the Wasserstein distance as metric. | 113 | Implements a nearest neighbors classifier for input distributions using |
108 | Source and target distributions are l_1 normalized before computing the Wasserstein distance. | 114 | the Wasserstein distance as metric. Source and target distributions |
109 | Wasserstein is parametrized by the distances between the individual points of the distributions. | 115 | are l_1 normalized before computing the Wasserstein distance. Wasserstein is |
116 | parametrized by the distances between the individual points of the distributions. | ||
110 | """ | 117 | """ |
111 | def __init__(self, | 118 | |
112 | W_embed, | 119 | def __init__( |
113 | n_neighbors=1, | 120 | self, |
114 | n_jobs=1, | 121 | W_embed, |
115 | verbose=False, | 122 | n_neighbors=1, |
116 | sinkhorn=False, | 123 | n_jobs=1, |
117 | sinkhorn_reg=0.1): | 124 | verbose=False, |
125 | sinkhorn=False, | ||
126 | sinkhorn_reg=0.1, | ||
127 | ): | ||
118 | """ | 128 | """ |
119 | Initialization of the class. | 129 | Initialization of the class. |
120 | Arguments | 130 | Arguments |
@@ -126,10 +136,12 @@ class WassersteinRetriever(KNeighborsClassifier): | |||
126 | self.sinkhorn_reg = sinkhorn_reg | 136 | self.sinkhorn_reg = sinkhorn_reg |
127 | self.W_embed = W_embed | 137 | self.W_embed = W_embed |
128 | self.verbose = verbose | 138 | self.verbose = verbose |
129 | super(WassersteinRetriever, self).__init__(n_neighbors=n_neighbors, | 139 | super(WassersteinRetriever, self).__init__( |
130 | n_jobs=n_jobs, | 140 | n_neighbors=n_neighbors, |
131 | metric='precomputed', | 141 | n_jobs=n_jobs, |
132 | algorithm='brute') | 142 | metric="precomputed", |
143 | algorithm="brute", | ||
144 | ) | ||
133 | 145 | ||
134 | def _wmd(self, i, row, X_train): | 146 | def _wmd(self, i, row, X_train): |
135 | union_idx = np.union1d(X_train[i].indices, row.indices) | 147 | union_idx = np.union1d(X_train[i].indices, row.indices) |
@@ -144,7 +156,7 @@ class WassersteinRetriever(KNeighborsClassifier): | |||
144 | W_dist, | 156 | W_dist, |
145 | self.sinkhorn_reg, | 157 | self.sinkhorn_reg, |
146 | numItermax=50, | 158 | numItermax=50, |
147 | method='sinkhorn_stabilized', | 159 | method="sinkhorn_stabilized", |
148 | )[0] | 160 | )[0] |
149 | else: | 161 | else: |
150 | return ot.emd2(bow_i, bow_j, W_dist) | 162 | return ot.emd2(bow_i, bow_j, W_dist) |
@@ -164,19 +176,19 @@ class WassersteinRetriever(KNeighborsClassifier): | |||
164 | return np.array(dist) | 176 | return np.array(dist) |
165 | 177 | ||
166 | def fit(self, X, y): | 178 | def fit(self, X, y): |
167 | X = check_array(X, accept_sparse='csr', copy=True) | 179 | X = check_array(X, accept_sparse="csr", copy=True) |
168 | X = normalize(X, norm='l1', copy=False) | 180 | X = normalize(X, norm="l1", copy=False) |
169 | return super(WassersteinRetriever, self).fit(X, y) | 181 | return super(WassersteinRetriever, self).fit(X, y) |
170 | 182 | ||
171 | def predict(self, X): | 183 | def predict(self, X): |
172 | X = check_array(X, accept_sparse='csr', copy=True) | 184 | X = check_array(X, accept_sparse="csr", copy=True) |
173 | X = normalize(X, norm='l1', copy=False) | 185 | X = normalize(X, norm="l1", copy=False) |
174 | dist = self._pairwise_wmd(X) | 186 | dist = self._pairwise_wmd(X) |
175 | return super(WassersteinRetriever, self).predict(dist) | 187 | return super(WassersteinRetriever, self).predict(dist) |
176 | 188 | ||
177 | def kneighbors(self, X, n_neighbors=1): | 189 | def kneighbors(self, X, n_neighbors=1): |
178 | X = check_array(X, accept_sparse='csr', copy=True) | 190 | X = check_array(X, accept_sparse="csr", copy=True) |
179 | X = normalize(X, norm='l1', copy=False) | 191 | X = normalize(X, norm="l1", copy=False) |
180 | dist = self._pairwise_wmd(X) | 192 | dist = self._pairwise_wmd(X) |
181 | return super(WassersteinRetriever, self).kneighbors(dist, n_neighbors) | 193 | return super(WassersteinRetriever, self).kneighbors(dist, n_neighbors) |
182 | 194 | ||
@@ -199,9 +211,9 @@ def load_embeddings(path, dimension=300): | |||
199 | The first line may or may not include the word count and dimension | 211 | The first line may or may not include the word count and dimension |
200 | """ | 212 | """ |
201 | vectors = {} | 213 | vectors = {} |
202 | with open(path, mode='r', encoding='utf8') as fp: | 214 | with open(path, mode="r", encoding="utf8") as fp: |
203 | first_line = fp.readline().rstrip('\n') | 215 | first_line = fp.readline().rstrip("\n") |
204 | if first_line.count(' ') == 1: | 216 | if first_line.count(" ") == 1: |
205 | # includes the "word_count dimension" information | 217 | # includes the "word_count dimension" information |
206 | (_, dimension) = map(int, first_line.split()) | 218 | (_, dimension) = map(int, first_line.split()) |
207 | else: | 219 | else: |
@@ -209,22 +221,19 @@ def load_embeddings(path, dimension=300): | |||
209 | fp.seek(0) | 221 | fp.seek(0) |
210 | for line in fp: | 222 | for line in fp: |
211 | elems = line.split() | 223 | elems = line.split() |
212 | vectors[" ".join(elems[:-dimension])] = " ".join( | 224 | vectors[" ".join(elems[:-dimension])] = " ".join(elems[-dimension:]) |
213 | elems[-dimension:]) | ||
214 | return vectors | 225 | return vectors |
215 | 226 | ||
216 | 227 | ||
217 | def clean_corpus_using_embeddings_vocabulary( | 228 | def clean_corpus_using_embeddings_vocabulary( |
218 | embeddings_dictionary, | 229 | embeddings_dictionary, corpus, vectors, language |
219 | corpus, | ||
220 | vectors, | ||
221 | language, | ||
222 | ): | 230 | ): |
223 | ''' | 231 | """ |
224 | Cleans corpus using the dictionary of embeddings. | 232 | Cleans corpus using the dictionary of embeddings. |
225 | Any word without an associated embedding in the dictionary is ignored. | 233 | Any word without an associated embedding in the dictionary is ignored. |
226 | Adds '__target-language' and '__source-language' at the end of the words according to their language. | 234 | Adds '__target-language' and '__source-language' at the end |
227 | ''' | 235 | of the words according to their language. |
236 | """ | ||
228 | clean_corpus, clean_vectors, keys = [], {}, [] | 237 | clean_corpus, clean_vectors, keys = [], {}, [] |
229 | words_we_want = set(embeddings_dictionary) | 238 | words_we_want = set(embeddings_dictionary) |
230 | tokenize = MosesTokenizer(language) | 239 | tokenize = MosesTokenizer(language) |
@@ -233,19 +242,18 @@ def clean_corpus_using_embeddings_vocabulary( | |||
233 | words = tokenize(doc) | 242 | words = tokenize(doc) |
234 | for word in words: | 243 | for word in words: |
235 | if word in words_we_want: | 244 | if word in words_we_want: |
236 | clean_doc.append(word + '__%s' % language) | 245 | clean_doc.append(word + "__%s" % language) |
237 | clean_vectors[word + '__%s' % language] = np.array( | 246 | clean_vectors[word + "__%s" % language] = np.array( |
238 | vectors[word].split()).astype(np.float) | 247 | vectors[word].split() |
248 | ).astype(np.float) | ||
239 | if len(clean_doc) > 3 and len(clean_doc) < 25: | 249 | if len(clean_doc) > 3 and len(clean_doc) < 25: |
240 | keys.append(key) | 250 | keys.append(key) |
241 | clean_corpus.append(' '.join(clean_doc)) | 251 | clean_corpus.append(" ".join(clean_doc)) |
242 | tokenize.close() | 252 | tokenize.close() |
243 | return np.array(clean_corpus), clean_vectors, keys | 253 | return np.array(clean_corpus), clean_vectors, keys |
244 | 254 | ||
245 | 255 | ||
246 | def mrr_precision_at_k(golden, preds, k_list=[ | 256 | def mrr_precision_at_k(golden, preds, k_list=[1]): |
247 | 1, | ||
248 | ]): | ||
249 | """ | 257 | """ |
250 | Calculates Mean Reciprocal Error and Hits@1 == Precision@1 | 258 | Calculates Mean Reciprocal Error and Hits@1 == Precision@1 |
251 | """ | 259 | """ |