Add DPR model (#5279)

* beginning of dpr modeling * wip * implement forward * remove biencoder + better init weights * export dpr model to embed model for nlp lib * add new api * remove old code * make style * fix dumb typo * don't load bert weights * docs * docs * style * move the `k` parameter * fix init_weights * add pretrained configs * minor * update config names * style * better config * style * clean code based on PR comments * change Dpr to DPR * fix config * switch encoder config to a dict * style * inheritance -> composition * add messages in assert startements * add dpr reader tokenizer * one tokenizer per model * fix base_model_prefix * fix imports * typo * add convert script * docs * change tokenizers conf names * style * change tokenizers conf names * minor * minor * fix wrong names * minor * remove unused convert functions * rename convert script * use return_tensors in tokenizers * remove n_questions dim * move generate logic to tokenizer * style * add docs * docs * quality * docs * add tests * style * add tokenization tests * DPR full tests * Stay true to the attention mask building * update docs * missing param in bert input docs * docs * style Co-authored-by: Lysandre <lysandre.debut@reseau.eseo.fr>
2020-07-07 14:56:12 +02:00
parent d2a9399115
commit fbd8792195
11 changed files with 1531 additions and 3 deletions
--- a/tests/test_tokenization_dpr.py
+++ b/tests/test_tokenization_dpr.py
@@ -0,0 +1,89 @@
+# coding=utf-8
+# Copyright 2020 Huggingface
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from transformers.tokenization_dpr import (
+    DPRContextEncoderTokenizer,
+    DPRContextEncoderTokenizerFast,
+    DPRQuestionEncoderTokenizer,
+    DPRQuestionEncoderTokenizerFast,
+    DPRReaderOutput,
+    DPRReaderTokenizer,
+    DPRReaderTokenizerFast,
+)
+from transformers.tokenization_utils_base import BatchEncoding
+
+from .test_tokenization_bert import BertTokenizationTest
+from .utils import slow
+
+
+class DPRContextEncoderTokenizationTest(BertTokenizationTest):
+
+    tokenizer_class = DPRContextEncoderTokenizer
+
+    def get_rust_tokenizer(self, **kwargs):
+        return DPRContextEncoderTokenizerFast.from_pretrained(self.tmpdirname, **kwargs)
+
+
+class DPRQuestionEncoderTokenizationTest(BertTokenizationTest):
+
+    tokenizer_class = DPRQuestionEncoderTokenizer
+
+    def get_rust_tokenizer(self, **kwargs):
+        return DPRQuestionEncoderTokenizerFast.from_pretrained(self.tmpdirname, **kwargs)
+
+
+class DPRReaderTokenizationTest(BertTokenizationTest):
+
+    tokenizer_class = DPRReaderTokenizer
+
+    def get_rust_tokenizer(self, **kwargs):
+        return DPRReaderTokenizerFast.from_pretrained(self.tmpdirname, **kwargs)
+
+    @slow
+    def test_decode_best_spans(self):
+        tokenizer = self.tokenizer_class.from_pretrained("bert-base-uncased")
+
+        text_1 = tokenizer.encode("question sequence", add_special_tokens=False)
+        text_2 = tokenizer.encode("title sequence", add_special_tokens=False)
+        text_3 = tokenizer.encode("text sequence " * 4, add_special_tokens=False)
+        input_ids = [[101] + text_1 + [102] + text_2 + [102] + text_3]
+        reader_input = BatchEncoding({"input_ids": input_ids})
+
+        start_logits = [[0] * len(input_ids[0])]
+        end_logits = [[0] * len(input_ids[0])]
+        relevance_logits = [0]
+        reader_output = DPRReaderOutput(start_logits, end_logits, relevance_logits)
+
+        start_index, end_index = 8, 9
+        start_logits[0][start_index] = 10
+        end_logits[0][end_index] = 10
+        predicted_spans = tokenizer.decode_best_spans(reader_input, reader_output)
+        self.assertEqual(predicted_spans[0].start_index, start_index)
+        self.assertEqual(predicted_spans[0].end_index, end_index)
+        self.assertEqual(predicted_spans[0].doc_id, 0)
+
+    @slow
+    def test_call(self):
+        tokenizer = self.tokenizer_class.from_pretrained("bert-base-uncased")
+
+        text_1 = tokenizer.encode("question sequence", add_special_tokens=False)
+        text_2 = tokenizer.encode("title sequence", add_special_tokens=False)
+        text_3 = tokenizer.encode("text sequence", add_special_tokens=False)
+        expected_input_ids = [101] + text_1 + [102] + text_2 + [102] + text_3
+        encoded_input = tokenizer(questions=["question sequence"], titles=["title sequence"], texts=["text sequence"])
+        self.assertIn("input_ids", encoded_input)
+        self.assertIn("attention_mask", encoded_input)
+        self.assertListEqual(encoded_input["input_ids"][0], expected_input_ids)