Regression test for pegasus bugfix (#6606)

src/transformers/configuration_pegasus.py

@@ -22,6 +22,7 @@ from .file_utils import add_start_docstrings_to_callable
 
 logger = logging.getLogger(__name__)
 
+# These config values do not vary between checkpoints
 DEFAULTS = dict(
     vocab_size=96103,
     max_position_embeddings=512,
@@ -46,6 +47,47 @@ DEFAULTS = dict(
     num_beams=8,
     activation_function="relu",
 )
+# Config values that vary between checkpoints: for testing and conversion
+max_gen_length = {
+    # See appendix C of paper
+    "xsum": 64,
+    "cnn_dailymail": 128,
+    "newsroom": 128,
+    "wikihow": 256,
+    "multi_news": 256,
+    "reddit_tifu": 128,
+    "big_patent": 256,
+    "arxiv": 256,
+    "pubmed": 256,
+    "gigaword": 32,
+    "aeslc": 32,
+    "billsum": 256,
+    "large": 256,  # @sshleifer chose arbitrarily
+}
+max_model_length = {
+    "xsum": 512,
+    "cnn_dailymail": 1024,
+    "newsroom": 512,
+    "wikihow": 512,
+    "multi_news": 1024,
+    "reddit_tifu": 512,
+    "big_patent": 1024,
+    "arxiv": 1024,
+    "pubmed": 1024,
+    "gigaword": 128,
+    "aeslc": 512,
+    "billsum": 1024,
+    "large": 1024,
+}
+expected_alpha = {
+    "multinews": 0.9,
+    "wikihow": 0.6,
+    "reddit_tifu": 0.6,
+    "big_patent": 0.7,
+    "gigaword": 0.6,
+    "aeslc": 0.6,
+    "billsum": 0.6,
+}  # otherwise 0.8
 
 
 @add_start_docstrings_to_callable(BART_CONFIG_ARGS_DOC)
@@ -56,7 +98,3 @@ class PegasusConfig(BartConfig):
     """
     model_type = "pegasus"
     # The implementation of the config object is in BartConfig
-
-    @property
-    def default_config_parameters(self):
-        return DEFAULTS

src/transformers/convert_pegasus_tf_to_pytorch.py

@@ -22,7 +22,7 @@ import torch
 from tqdm import tqdm
 
 from transformers import PegasusConfig, PegasusForConditionalGeneration, PegasusTokenizer
-from transformers.configuration_pegasus import DEFAULTS
+from transformers.configuration_pegasus import DEFAULTS, expected_alpha, max_gen_length, max_model_length
 
 
 PATTERNS = [
@@ -52,47 +52,7 @@ def rename_state_dict_key(k):
 
 
 # See appendix C of paper for all hyperparams
-max_gen_length = {
-    # See appendix C of paper
-    "xsum": 64,
-    "cnn_dailymail": 128,
-    "newsroom": 128,
-    "wikihow": 256,
-    "multi_news": 256,
-    "reddit_tifu": 128,
-    "big_patent": 256,
-    "arxiv": 256,
-    "pubmed": 256,
-    "gigaword": 32,
-    "aeslc": 32,
-    "billsum": 256,
-    "large": 256,  # @sshleifer chose arbitrarily
-}
-max_model_length = {
-    "xsum": 512,
-    "cnn_dailymail": 1024,
-    "newsroom": 512,
-    "wikihow": 512,
-    "multi_news": 1024,
-    "reddit_tifu": 512,
-    "big_patent": 1024,
-    "arxiv": 1024,
-    "pubmed": 1024,
-    "gigaword": 128,
-    "aeslc": 512,
-    "billsum": 1024,
-    "large": 1024,
-}
 
-expected_alpha = {
-    "multinews": 0.9,
-    "wikihow": 0.6,
-    "reddit_tifu": 0.6,
-    "big_patent": 0.7,
-    "gigaword": 0.6,
-    "aeslc": 0.6,
-    "billsum": 0.6,
-}  # otherwise 0.8
 # TODO(SS): one constant
 
 
@@ -151,7 +111,11 @@ def convert_pegasus_ckpt_to_pytorch(ckpt_path, save_dir):
 
     # convert model
     tf_weights = get_tf_weights_as_numpy(ckpt_path)
-    cfg_updates = dict(max_length=max_gen_length[dataset], length_penalty=expected_alpha.get(dataset, 0.8))
+    cfg_updates = dict(
+        max_length=max_gen_length[dataset],
+        length_penalty=expected_alpha.get(dataset, 0.8),
+        max_position_embeddings=desired_max_model_length,
+    )
     torch_model = convert_pegasus_to_bart(tf_weights, cfg_updates)
     torch_model.save_pretrained(save_dir)
 

src/transformers/modeling_pegasus.py

@@ -23,6 +23,13 @@ from .modeling_bart import BART_START_DOCSTRING, BartForConditionalGeneration
 @add_start_docstrings("The Pegasus Model for summarization ", BART_START_DOCSTRING)
 class PegasusForConditionalGeneration(BartForConditionalGeneration):
     config_class = PegasusConfig
+    authorized_missing_keys = [
+        r"final_logits_bias",
+        r"encoder\.version",
+        r"decoder\.version",
+        r"model.encoder.embed_positions",
+        "model.decoder.embed_positions",
+    ]
     r"""
     Pytorch version of google's pegasus model for summarization.
     Model API is identical to BartForConditionalGeneration.

tests/test_modeling_pegasus.py

@@ -1,6 +1,7 @@
 import unittest
 
-from transformers import AutoConfig, is_torch_available
+from transformers import AutoConfig, AutoTokenizer, is_torch_available
+from transformers.configuration_pegasus import max_gen_length, max_model_length
 from transformers.file_utils import cached_property
 from transformers.testing_utils import require_torch, slow, torch_device
 
@@ -50,28 +51,28 @@ class PegasusXSUMIntegrationTest(AbstractSeq2SeqIntegrationTest):
 
 class PegasusConfigTests(unittest.TestCase):
     def test_all_config_max_lengths(self):
-        expected_max_length = {
-            # See appendix C of paper
-            "xsum": 64,
-            "cnn_dailymail": 128,
-            "newsroom": 128,
-            "wikihow": 256,
-            "multi_news": 256,
-            "reddit_tifu": 128,
-            "big_patent": 256,
-            "arxiv": 256,
-            "pubmed": 256,
-            "gigaword": 32,
-            "aeslc": 32,
-            "billsum": 256,
-        }
         failures = []
         pegasus_prefix = "google/pegasus"
-        for dataset, max_len in expected_max_length.items():
+        for dataset, max_len in max_gen_length.items():
             mname = f"{pegasus_prefix}-{dataset}"
             cfg = AutoConfig.from_pretrained(mname)
+
             if cfg.max_length != max_len:
                 failures.append(f"config for {mname} had max_length: {cfg.max_length}, expected {max_len}")
+
+            if cfg.max_position_embeddings < max_model_length[dataset]:
+                # otherwise you get IndexError for e.g. position 513
+                # see https://github.com/huggingface/transformers/issues/6599
+                failures.append(
+                    f"config for {mname} had max_position_embeddings: {cfg.max_position_embeddings}, expected {max_model_length[dataset]}"
+                )
+
+            tokenizer = AutoTokenizer.from_pretrained(mname)
+            if max_model_length[dataset] != tokenizer.model_max_length:
+                failures.append(
+                    f"tokenizer.model_max_length {tokenizer.model_max_length} expected {max_model_length[dataset]}"
+                )
+
         if failures == []:
             return
         # error