Documentation and example monobert

experimaestro · Jul 24, 2023 · 90126b1 · 90126b1
1 parent 7750213
commit 90126b1
Show file tree

Hide file tree

Showing 8 changed files with 41 additions and 15 deletions.
diff --git a/docs/source/letor/samplers.rst b/docs/source/letor/samplers.rst
@@ -1,35 +1,38 @@
 Samplers
 --------
 
+.. currentmodule:: xpmir.letor.samplers
+
 Samplers provide samples in the form of *records*. They all inherit from:
 
-.. autoxpmconfig:: xpmir.letor.samplers.Sampler
-.. autoclass:: xpmir.letor.samplers.SerializableIterator
+.. autoxpmconfig:: Sampler
+.. autoclass:: SerializableIterator
 
 
 Pointwise
 =========
 
-.. autoxpmconfig:: xpmir.letor.samplers.PointwiseSampler
+.. autoxpmconfig:: PointwiseSampler
     :members: pointwise_iter
 
-.. autoxpmconfig:: xpmir.letor.samplers.PointwiseModelBasedSampler
+.. autoxpmconfig:: PointwiseModelBasedSampler
 
 Pairwise
 =========
 
-.. autoxpmconfig:: xpmir.letor.samplers.PairwiseSampler
-.. autoxpmconfig:: xpmir.letor.samplers.PairwiseModelBasedSampler
+.. autoxpmconfig:: PairwiseSampler
+.. autoxpmconfig:: BatchwiseSampler
+.. autoxpmconfig:: PairwiseModelBasedSampler
 .. autoxpmconfig:: xpmir.documents.samplers.BatchwiseRandomSpanSampler
 
-.. autoxpmconfig:: xpmir.letor.samplers.TripletBasedSampler
-.. autoxpmconfig:: xpmir.letor.samplers.PairwiseDatasetTripletBasedSampler
+.. autoxpmconfig:: TripletBasedSampler
+.. autoxpmconfig:: PairwiseDatasetTripletBasedSampler
 
 Hard Negatives Sampling (Tasks)
-============
+===============================
 
-.. autoxpmconfig:: xpmir.letor.samplers.ModelBasedHardNegativeSampler
-.. autoxpmconfig:: xpmir.letor.samplers.TeacherModelBasedHardNegativesTripletSampler
+.. autoxpmconfig:: ModelBasedHardNegativeSampler
+.. autoxpmconfig:: TeacherModelBasedHardNegativesTripletSampler
 
 Distillation
 ============
@@ -45,3 +48,14 @@ Records for training
 
 .. automodule:: xpmir.letor.records
     :members: PointwiseRecord, PairwiseRecord
+
+
+Document samplers
+=================
+
+Useful for pre-training or when learning index parameters (e.g. for FAISS).
+
+.. currentmodule:: xpmir.documents.samplers
+.. autoxpmconfig:: DocumentSampler
+.. autoxpmconfig:: HeadDocumentSampler
+.. autoxpmconfig:: RandomDocumentSampler
diff --git a/docs/source/text/huggingface.rst b/docs/source/text/huggingface.rst
@@ -12,6 +12,13 @@ Encoders
 .. autoxpmconfig:: SentenceTransformerTextEncoder
 .. autoxpmconfig:: OneHotHuggingFaceEncoder
 
+
+Tokenizers
+==========
+
+.. autoxpmconfig:: OneHotHuggingFaceEncoder
+.. autoxpmconfig:: HuggingfaceTokenizer
+
 Hooks
 =====
 

diff --git a/requirements.txt b/requirements.txt
@@ -1,6 +1,6 @@
 # Experimaestro
 
-experimaestro>=0.29.1
+experimaestro>=1.0.0
 datamaestro>=0.8.13
 datamaestro_text>=2023.3.23
 ir_datasets

diff --git a/src/xpmir/letor/samplers.py b/src/xpmir/letor/samplers.py
@@ -61,7 +61,8 @@ def pairwise_iter(self) -> SerializableIterator[PairwiseRecord]:
 
 
 class BatchwiseSampler(Sampler):
-    """Batchwise samplers provide for each question a set of documents"""
+    """Base class for batchwise samplers, that provide for each question a list
+    of documents"""
 
     def batchwise_iter(self, batch_size: int) -> SerializableIterator[BatchwiseRecords]:
         """Iterate over batches of size (# of queries) batch_size

diff --git a/src/xpmir/papers/monobert/configuration.py b/src/xpmir/papers/monobert/configuration.py
@@ -46,3 +46,6 @@ class Monobert(RerankerMSMarcoV1Configuration):
 
     dev_test_size: int = 0
     """Development test size (0 to leave it like this)"""
+
+    base: str = "bert-base-uncased"
+    """Identifier for the base model"""
diff --git a/src/xpmir/papers/monobert/experiment.py b/src/xpmir/papers/monobert/experiment.py
@@ -115,7 +115,7 @@ def run(
 
     monobert_scorer: CrossScorer = CrossScorer(
         encoder=DualTransformerEncoder(
-            model_id="bert-base-uncased", trainable=True, maxlen=512, dropout=0.1
+            model_id=cfg.base, trainable=True, maxlen=512, dropout=0.1
         )
     ).tag("scorer", "monobert")
 

diff --git a/src/xpmir/papers/monobert/small.yaml b/src/xpmir/papers/monobert/small.yaml
@@ -6,6 +6,8 @@ description: |
 
     This model has been trained on MsMarco v1 but only a few iterations (debug)
 
+gpu: true
+base: "microsoft/MiniLM-L12-H384-uncased"
 dev_test_size: 50
 
 validation:

diff --git a/src/xpmir/text/huggingface.py b/src/xpmir/text/huggingface.py
@@ -260,7 +260,6 @@ def static(self):
 @deprecate
 class HuggingfaceTokenizer(OneHotHuggingFaceEncoder):
     """The old encoder for one hot"""
-
     pass
-Original file line number
+Diff line change
@@ Expand Up / @@ -260,7 +260,6 @@ def static(self): @@
     @deprecate
     class HuggingfaceTokenizer(OneHotHuggingFaceEncoder):
         """The old encoder for one hot"""
         pass
@@ Expand Down @@