Rename preset IDs for consistency (#612)

And start to add rules we would like to follow to our style guide.
keras-team · Dec 27, 2022 · 729815b · 729815b
1 parent f7d816f
commit 729815b
Show file tree

Hide file tree

Showing 26 changed files with 140 additions and 111 deletions.
diff --git a/STYLE_GUIDE.md b/STYLE_GUIDE.md
@@ -18,6 +18,29 @@ When a specific abbreviation is very common and is pronounceable (acronym),
 consider it as a standalone word, e.g. Bert, Deberta, etc. In this case, "Bert"
 is considered as a common noun and not an abbreviation anymore.
 
+## Naming of Models and Presets
+
+Naming of models and presets is a difficult and important element of our
+library usability. In general we try to to follow the branding of "upstream"
+model naming, subject to the consistency constraints laid out here.
+
+- The model and preset names should be recognizable to users familiar with the
+  original release. E.g. the model that goes with the "DeBERTaV3" paper should
+  be called `DebertaV3`. A release of a [toxic-bert](https://huggingface.co/unitary/toxic-bert)
+  checkpoint for `keras_nlp.models.Bert`, should include the string
+  `"toxic_bert"`.
+- All preset names should include the language of the pretraining data. If three
+  or more language are supported, the preset name should include `"multi"` (not
+  the single letter "m").
+- If a preset lowercases input for cased-based languages, the preset name should
+  be marked with `"uncased"`.
+- Don't abbreviate size names. E.g. "xsmall" or "XL" in an original checkpoint
+  releases should map to `"extra_small"` or `"extra_large"` in a preset names.
+- No configuration in names. E.g. use "bert_base" instead of
+  "bert_L-12_H-768_A-12".
+
+When in doubt, readability should win out!
+
 ## File names
 
 When possible, keep publicly documented classes in their own files, and make

diff --git a/keras_nlp/models/bert/bert_presets.py b/keras_nlp/models/bert/bert_presets.py
@@ -108,7 +108,7 @@
         "vocabulary_url": "https://storage.googleapis.com/keras-nlp/models/bert_base_en_uncased/v1/vocab.txt",
         "vocabulary_hash": "64800d5d8528ce344256daf115d4965e",
     },
-    "bert_base_en_cased": {
+    "bert_base_en": {
         "config": {
             "vocabulary_size": 28996,
             "num_layers": 12,
@@ -126,9 +126,9 @@
             "Base size of BERT where case is maintained. "
             "Trained on English Wikipedia + BooksCorpus."
         ),
-        "weights_url": "https://storage.googleapis.com/keras-nlp/models/bert_base_en_cased/v1/model.h5",
+        "weights_url": "https://storage.googleapis.com/keras-nlp/models/bert_base_en/v1/model.h5",
         "weights_hash": "f94a6cb012e18f4fb8ec92abb91864e9",
-        "vocabulary_url": "https://storage.googleapis.com/keras-nlp/models/bert_base_en_cased/v1/vocab.txt",
+        "vocabulary_url": "https://storage.googleapis.com/keras-nlp/models/bert_base_en/v1/vocab.txt",
         "vocabulary_hash": "bb6ca9b42e790e5cd986bbb16444d0e0",
     },
     "bert_base_zh": {
@@ -151,7 +151,7 @@
         "vocabulary_url": "https://storage.googleapis.com/keras-nlp/models/bert_base_zh/v1/vocab.txt",
         "vocabulary_hash": "3b5b76c4aef48ecf8cb3abaafe960f09",
     },
-    "bert_base_multi_cased": {
+    "bert_base_multi": {
         "config": {
             "vocabulary_size": 119547,
             "num_layers": 12,
@@ -169,9 +169,9 @@
             "Base size of BERT. Trained on trained on Wikipedias of 104 "
             "languages."
         ),
-        "weights_url": "https://storage.googleapis.com/keras-nlp/models/bert_base_multi_cased/v1/model.h5",
+        "weights_url": "https://storage.googleapis.com/keras-nlp/models/bert_base_multi/v1/model.h5",
         "weights_hash": "b0631cec0a1f2513c6cfd75ba29c33aa",
-        "vocabulary_url": "https://storage.googleapis.com/keras-nlp/models/bert_base_multi_cased/v1/vocab.txt",
+        "vocabulary_url": "https://storage.googleapis.com/keras-nlp/models/bert_base_multi/v1/vocab.txt",
         "vocabulary_hash": "d9d865138d17f1958502ed060ecfeeb6",
     },
     "bert_large_en_uncased": {
@@ -197,7 +197,7 @@
         "vocabulary_url": "https://storage.googleapis.com/keras-nlp/models/bert_large_en_uncased/v1/vocab.txt",
         "vocabulary_hash": "64800d5d8528ce344256daf115d4965e",
     },
-    "bert_large_en_cased": {
+    "bert_large_en": {
         "config": {
             "vocabulary_size": 28996,
             "num_layers": 24,
@@ -215,9 +215,9 @@
             "Base size of BERT where case is maintained. "
             "Trained on English Wikipedia + BooksCorpus."
         ),
-        "weights_url": "https://storage.googleapis.com/keras-nlp/models/bert_large_en_cased/v1/model.h5",
+        "weights_url": "https://storage.googleapis.com/keras-nlp/models/bert_large_en/v1/model.h5",
         "weights_hash": "8b8ab82290bbf4f8db87d4f100648890",
-        "vocabulary_url": "https://storage.googleapis.com/keras-nlp/models/bert_large_en_cased/v1/vocab.txt",
+        "vocabulary_url": "https://storage.googleapis.com/keras-nlp/models/bert_large_en/v1/vocab.txt",
         "vocabulary_hash": "bb6ca9b42e790e5cd986bbb16444d0e0",
     },
 }

diff --git a/keras_nlp/models/deberta_v3/deberta_v3_backbone.py b/keras_nlp/models/deberta_v3/deberta_v3_backbone.py
@@ -227,12 +227,14 @@ def from_preset(
         }
 
         # Load architecture and weights from preset
-        model = keras_nlp.models.DebertaV3Backbone.from_preset("deberta_base")
+        model = keras_nlp.models.DebertaV3Backbone.from_preset(
+            "deberta_base_en",
+        )
         output = model(input_data)
 
         # Load randomly initialized model from preset architecture
         model = keras_nlp.models.DebertaV3Backbone.from_preset(
-            "deberta_base", load_weights=False
+            "deberta_base_en", load_weights=False
         )
         output = model(input_data)
         ```

diff --git a/keras_nlp/models/deberta_v3/deberta_v3_classifier.py b/keras_nlp/models/deberta_v3/deberta_v3_classifier.py
@@ -203,7 +203,7 @@ def from_preset(
 
         # Create a DebertaV3Classifier and fit your data.
         classifier = keras_nlp.models.DebertaV3Classifier.from_preset(
-            "deberta_base",
+            "deberta_v3_base_en",
             num_classes=4,
         )
         classifier.compile(
@@ -220,13 +220,13 @@ def from_preset(
 
         # Use a shorter sequence length.
         preprocessor = keras_nlp.models.DebertaV3Preprocessor.from_preset(
-            "deberta_base",
+            "deberta_v3_base_en",
             sequence_length=128,
         )
 
         # Create a DebertaV3Classifier and fit your data.
         classifier = keras_nlp.models.DebertaV3Classifier.from_preset(
-            "deberta_base",
+            "deberta_v3_base_en",
             num_classes=4,
             preprocessor=preprocessor,
         )
@@ -249,7 +249,7 @@ def from_preset(
 
         # Create a DebertaV3Classifier and fit your data.
         classifier = keras_nlp.models.DebertaV3Classifier.from_preset(
-            "deberta_base",
+            "deberta_v3_base_en",
             num_classes=4,
             preprocessor=None,
         )

diff --git a/keras_nlp/models/deberta_v3/deberta_v3_preprocessor.py b/keras_nlp/models/deberta_v3/deberta_v3_preprocessor.py
@@ -226,13 +226,13 @@ def from_preset(
         ```python
         # Load preprocessor from preset
         preprocessor = keras_nlp.models.DebertaV3Preprocessor.from_preset(
-            "deberta_base",
+            "deberta_v3_base_en",
         )
         preprocessor("The quick brown fox jumped.")
 
         # Override sequence_length
         preprocessor = keras_nlp.models.DebertaV3Preprocessor.from_preset(
-            "deberta_base",
+            "deberta_v3_base_en",
             sequence_length=64
         )
         preprocessor("The quick brown fox jumped.")

diff --git a/keras_nlp/models/deberta_v3/deberta_v3_presets.py b/keras_nlp/models/deberta_v3/deberta_v3_presets.py
@@ -14,7 +14,7 @@
 """DeBERTa model preset configurations."""
 
 backbone_presets = {
-    "deberta_v3_extra_small": {
+    "deberta_v3_extra_small_en": {
         "config": {
             "vocabulary_size": 128100,
             "num_layers": 12,
@@ -30,12 +30,12 @@
             "Extra small size of DeBERTaV3. "
             "Trained on English Wikipedia, BookCorpus and OpenWebText."
         ),
-        "weights_url": "https://storage.googleapis.com/keras-nlp/models/deberta_v3_extra_small/v1/model.h5",
+        "weights_url": "https://storage.googleapis.com/keras-nlp/models/deberta_v3_extra_small_en/v1/model.h5",
         "weights_hash": "d8e10327107e5c5e20b45548a5028619",
-        "spm_proto_url": "https://storage.googleapis.com/keras-nlp/models/deberta_v3_extra_small/v1/vocab.spm",
+        "spm_proto_url": "https://storage.googleapis.com/keras-nlp/models/deberta_v3_extra_small_en/v1/vocab.spm",
         "spm_proto_hash": "1613fcbf3b82999c187b09c9db79b568",
     },
-    "deberta_v3_small": {
+    "deberta_v3_small_en": {
         "config": {
             "vocabulary_size": 128100,
             "num_layers": 6,
@@ -51,12 +51,12 @@
             "Small size of DeBERTaV3. "
             "Trained on English Wikipedia, BookCorpus and OpenWebText."
         ),
-        "weights_url": "https://storage.googleapis.com/keras-nlp/models/deberta_v3_small/v1/model.h5",
+        "weights_url": "https://storage.googleapis.com/keras-nlp/models/deberta_v3_small_en/v1/model.h5",
         "weights_hash": "84118eb7c5a735f2061ecccaf71bb888",
-        "spm_proto_url": "https://storage.googleapis.com/keras-nlp/models/deberta_v3_small/v1/vocab.spm",
+        "spm_proto_url": "https://storage.googleapis.com/keras-nlp/models/deberta_v3_small_en/v1/vocab.spm",
         "spm_proto_hash": "1613fcbf3b82999c187b09c9db79b568",
     },
-    "deberta_v3_base": {
+    "deberta_v3_base_en": {
         "config": {
             "vocabulary_size": 128100,
             "num_layers": 12,
@@ -72,12 +72,12 @@
             "Base size of DeBERTaV3. "
             "Trained on English Wikipedia, BookCorpus and OpenWebText."
         ),
-        "weights_url": "https://storage.googleapis.com/keras-nlp/models/deberta_v3_base/v1/model.h5",
+        "weights_url": "https://storage.googleapis.com/keras-nlp/models/deberta_v3_base_en/v1/model.h5",
         "weights_hash": "cebce044aeed36aec9b94e3b8a255430",
-        "spm_proto_url": "https://storage.googleapis.com/keras-nlp/models/deberta_v3_base/v1/vocab.spm",
+        "spm_proto_url": "https://storage.googleapis.com/keras-nlp/models/deberta_v3_base_en/v1/vocab.spm",
         "spm_proto_hash": "1613fcbf3b82999c187b09c9db79b568",
     },
-    "deberta_v3_large": {
+    "deberta_v3_large_en": {
         "config": {
             "vocabulary_size": 128100,
             "num_layers": 24,
@@ -93,9 +93,9 @@
             "Base size of DeBERTaV3. "
             "Trained on English Wikipedia, BookCorpus and OpenWebText."
         ),
-        "weights_url": "https://storage.googleapis.com/keras-nlp/models/deberta_v3_large/v1/model.h5",
+        "weights_url": "https://storage.googleapis.com/keras-nlp/models/deberta_v3_large_en/v1/model.h5",
         "weights_hash": "bce7690f358a9e39304f8c0ebc71a745",
-        "spm_proto_url": "https://storage.googleapis.com/keras-nlp/models/deberta_v3_large/v1/vocab.spm",
+        "spm_proto_url": "https://storage.googleapis.com/keras-nlp/models/deberta_v3_large_en/v1/vocab.spm",
         "spm_proto_hash": "1613fcbf3b82999c187b09c9db79b568",
     },
     "deberta_v3_base_multi": {

diff --git a/keras_nlp/models/deberta_v3/deberta_v3_presets_test.py b/keras_nlp/models/deberta_v3/deberta_v3_presets_test.py
@@ -38,15 +38,15 @@ class DebertaV3PresetSmokeTest(tf.test.TestCase, parameterized.TestCase):
 
     def test_tokenizer_output(self):
         tokenizer = DebertaV3Tokenizer.from_preset(
-            "deberta_v3_extra_small",
+            "deberta_v3_extra_small_en",
         )
         outputs = tokenizer("The quick brown fox.")
         expected_outputs = [279, 1538, 3258, 16123, 260]
         self.assertAllEqual(outputs, expected_outputs)
 
     def test_preprocessor_output(self):
         preprocessor = DebertaV3Preprocessor.from_preset(
-            "deberta_v3_extra_small",
+            "deberta_v3_extra_small_en",
             sequence_length=4,
         )
         outputs = preprocessor("The quick brown fox.")["token_ids"]
@@ -62,7 +62,7 @@ def test_backbone_output(self, load_weights):
             "padding_mask": tf.constant([[1, 1, 1, 1]]),
         }
         model = DebertaV3Backbone.from_preset(
-            "deberta_v3_extra_small", load_weights=load_weights
+            "deberta_v3_extra_small_en", load_weights=load_weights
         )
         outputs = model(input_data)
         if load_weights:
@@ -76,7 +76,7 @@ def test_backbone_output(self, load_weights):
     def test_classifier_output(self, load_weights):
         input_data = tf.constant(["The quick brown fox."])
         model = DebertaV3Classifier.from_preset(
-            "deberta_v3_extra_small", load_weights=load_weights
+            "deberta_v3_extra_small_en", load_weights=load_weights
         )
         # Never assert output values, as the head weights are random.
         model.predict(input_data)
@@ -90,7 +90,7 @@ def test_classifier_output_without_preprocessing(self, load_weights):
             "padding_mask": tf.constant([[1, 1, 1, 1]]),
         }
         model = DebertaV3Classifier.from_preset(
-            "deberta_v3_extra_small",
+            "deberta_v3_extra_small_en",
             load_weights=load_weights,
             preprocessor=None,
         )
@@ -117,7 +117,7 @@ def test_preset_docstring(self, cls):
     def test_unknown_preset_error(self, cls):
         # Not a preset name
         with self.assertRaises(ValueError):
-            cls.from_preset("deberta_v3_extra_small_clowntown")
+            cls.from_preset("deberta_v3_extra_small_en_clowntown")
 
 
 @pytest.mark.extra_large

diff --git a/keras_nlp/models/deberta_v3/deberta_v3_tokenizer.py b/keras_nlp/models/deberta_v3/deberta_v3_tokenizer.py
@@ -106,7 +106,7 @@ def from_preset(
         ```python
         # Load a preset tokenizer.
         tokenizer = keras_nlp.models.DebertaV3Tokenizer.from_preset(
-            "deberta_base",
+            "deberta_v3_base_en",
         )
 
         # Tokenize some input.

diff --git a/keras_nlp/models/distil_bert/distil_bert_classifier.py b/keras_nlp/models/distil_bert/distil_bert_classifier.py
@@ -219,12 +219,12 @@ def from_preset(
 
         # Use a shorter sequence length.
         preprocessor = keras_nlp.models.DistilBertBackbone.from_preset(
-            "bert_base_en_uncased",
+            "distil_bert_base_en_uncased",
             sequence_length=128,
         )
         # Create a DistilBertClassifier and fit your data.
         classifier = keras_nlp.models.DistilBertClassifier.from_preset(
-            "bert_base_en_uncased",
+            "distil_bert_base_en_uncased",
             num_classes=4,
             preprocessor=preprocessor,
         )
@@ -250,7 +250,7 @@ def from_preset(
 
         # Create a DistilBERT classifier and fit your data.
         classifier = keras_nlp.models.DistilBertClassifier.from_preset(
-            "bert_base_en_uncased",
+            "distil_bert_base_en_uncased",
             num_classes=4,
             preprocessor=None,
         )

diff --git a/keras_nlp/models/distil_bert/distil_bert_presets.py b/keras_nlp/models/distil_bert/distil_bert_presets.py
@@ -37,7 +37,7 @@
         "vocabulary_url": "https://storage.googleapis.com/keras-nlp/models/distil_bert_base_en_uncased/v1/vocab.txt",
         "vocabulary_hash": "64800d5d8528ce344256daf115d4965e",
     },
-    "distil_bert_base_en_cased": {
+    "distil_bert_base_en": {
         "config": {
             "vocabulary_size": 28996,
             "num_layers": 6,
@@ -55,12 +55,12 @@
             "Trained on English Wikipedia + BooksCorpus using BERT as the "
             "teacher model."
         ),
-        "weights_url": "https://storage.googleapis.com/keras-nlp/models/distil_bert_base_en_cased/v1/model.h5",
+        "weights_url": "https://storage.googleapis.com/keras-nlp/models/distil_bert_base_en/v1/model.h5",
         "weights_hash": "fa36aa6865978efbf85a5c8264e5eb57",
-        "vocabulary_url": "https://storage.googleapis.com/keras-nlp/models/distil_bert_base_en_cased/v1/vocab.txt",
+        "vocabulary_url": "https://storage.googleapis.com/keras-nlp/models/distil_bert_base_en/v1/vocab.txt",
         "vocabulary_hash": "bb6ca9b42e790e5cd986bbb16444d0e0",
     },
-    "distil_bert_base_multi_cased": {
+    "distil_bert_base_multi": {
         "config": {
             "vocabulary_size": 119547,
             "num_layers": 6,
@@ -77,9 +77,9 @@
             "Base size of DistilBERT. Trained on Wikipedias of 104 languages "
             "using BERT the teacher model."
         ),
-        "weights_url": "https://storage.googleapis.com/keras-nlp/models/distil_bert_base_multi_cased/v1/model.h5",
+        "weights_url": "https://storage.googleapis.com/keras-nlp/models/distil_bert_base_multi/v1/model.h5",
         "weights_hash": "c0f11095e2a6455bd3b1a6d14800a7fa",
-        "vocabulary_url": "https://storage.googleapis.com/keras-nlp/models/distil_bert_base_multi_cased/v1/vocab.txt",
+        "vocabulary_url": "https://storage.googleapis.com/keras-nlp/models/distil_bert_base_multi/v1/vocab.txt",
         "vocabulary_hash": "d9d865138d17f1958502ed060ecfeeb6",
     },
 }
diff --git a/keras_nlp/models/gpt2/gpt2_backbone.py b/keras_nlp/models/gpt2/gpt2_backbone.py
@@ -215,7 +215,7 @@ def from_preset(
         }
 
         # Load architecture and weights from preset
-        model = GPT2Backbone.from_preset("gpt2_base")
+        model = GPT2Backbone.from_preset("gpt2_base_en")
         output = model(input_data)
 
         # Load randomly initialized model from preset architecture