Avoid invalid escape sequences, use raw strings (huggingface#22936)

* Avoid invalid escape sequences, use raw strings

* Integrate PR feedback

examples/research_projects/quantization-qdqbert/quant_trainer.py

@@ -41,8 +41,8 @@ def add_arguments(parser):
     group.add_argument("--quant-disable", action="store_true", help="disable all quantizers")
     group.add_argument("--quant-disable-embeddings", action="store_true", help="disable all embeddings quantizers")
     group.add_argument("--quant-disable-keyword", type=str, nargs="+", help="disable quantizers by keyword")
-    group.add_argument("--quant-disable-layer-module", type=str, help="disable quantizers by keyword under layer.\d+.")
-    group.add_argument("--quant-enable-layer-module", type=str, help="enable quantizers by keyword under layer.\d+.")
+    group.add_argument("--quant-disable-layer-module", type=str, help="disable quantizers by keyword under layer.")
+    group.add_argument("--quant-enable-layer-module", type=str, help="enable quantizers by keyword under layer")
     group.add_argument("--calibrator", default="max", help="which quantization range calibrator to use")
     group.add_argument("--percentile", default=None, type=float, help="percentile for PercentileCalibrator")
     group.add_argument("--fuse-qkv", action="store_true", help="use the same scale factor for qkv")
@@ -94,10 +94,10 @@ def configure_model(model, args, calib=False, eval=False):
             set_quantizer_by_name(model, args.quant_disable_keyword, _disabled=True)
 
         if args.quant_disable_layer_module:
-            set_quantizer_by_name(model, ["layer.\d+." + args.quant_disable_layer_module], _disabled=True)
+            set_quantizer_by_name(model, [r"layer.\d+." + args.quant_disable_layer_module], _disabled=True)
 
         if args.quant_enable_layer_module:
-            set_quantizer_by_name(model, ["layer.\d+." + args.quant_enable_layer_module], _disabled=False)
+            set_quantizer_by_name(model, [r"layer.\d+." + args.quant_enable_layer_module], _disabled=False)
 
         if args.recalibrate_weights:
             recalibrate_weights(model)

examples/research_projects/wav2vec2/run_asr.py

@@ -365,7 +365,7 @@ def main():
     target_sr = processor.feature_extractor.sampling_rate if data_args.target_feature_extractor_sampling_rate else None
     vocabulary_chars_str = "".join(t for t in processor.tokenizer.get_vocab().keys() if len(t) == 1)
     vocabulary_text_cleaner = re.compile(  # remove characters not in vocabulary
-        f"[^\s{re.escape(vocabulary_chars_str)}]",  # allow space in addition to chars in vocabulary
+        rf"[^\s{re.escape(vocabulary_chars_str)}]",  # allow space in addition to chars in vocabulary
         flags=re.IGNORECASE if processor.tokenizer.do_lower_case else 0,
     )
     text_updates = []

pyproject.toml

@@ -4,7 +4,7 @@ target-version = ['py37']
 
 [tool.ruff]
 # Never enforce `E501` (line length violations).
-ignore = ["C901", "E501", "E741", "W605"]
+ignore = ["C901", "E501", "E741"]
 select = ["C", "E", "F", "I", "W"]
 line-length = 119
 

src/transformers/commands/add_new_model_like.py

@@ -127,7 +127,7 @@ def find_indent(line: str) -> int:
     """
     Returns the number of spaces that start a line indent.
     """
-    search = re.search("^(\s*)(?:\S|$)", line)
+    search = re.search(r"^(\s*)(?:\S|$)", line)
     if search is None:
         return 0
     return len(search.groups()[0])
@@ -519,7 +519,7 @@ def duplicate_module(
     with open(module_file, "r", encoding="utf-8") as f:
         content = f.read()
 
-    content = re.sub("# Copyright (\d+)\s", f"# Copyright {CURRENT_YEAR} ", content)
+    content = re.sub(r"# Copyright (\d+)\s", f"# Copyright {CURRENT_YEAR} ", content)
     objects = parse_module_content(content)
 
     # Loop and treat all objects
@@ -568,7 +568,7 @@ def duplicate_module(
         # Regular classes functions
         old_obj = obj
         obj, replacement = replace_model_patterns(obj, old_model_patterns, new_model_patterns)
-        has_copied_from = re.search("^#\s+Copied from", obj, flags=re.MULTILINE) is not None
+        has_copied_from = re.search(r"^#\s+Copied from", obj, flags=re.MULTILINE) is not None
         if add_copied_from and not has_copied_from and _re_class_func.search(obj) is not None and len(replacement) > 0:
             # Copied from statement must be added just before the class/function definition, which may not be the
             # first line because of decorators.
@@ -667,7 +667,7 @@ def get_model_files(model_type: str, frameworks: Optional[List[str]] = None) ->
     return {"doc_file": doc_file, "model_files": model_files, "module_name": module_name, "test_files": test_files}
 
 
-_re_checkpoint_for_doc = re.compile("^_CHECKPOINT_FOR_DOC\s+=\s+(\S*)\s*$", flags=re.MULTILINE)
+_re_checkpoint_for_doc = re.compile(r"^_CHECKPOINT_FOR_DOC\s+=\s+(\S*)\s*$", flags=re.MULTILINE)
 
 
 def find_base_model_checkpoint(
@@ -913,8 +913,8 @@ def clean_frameworks_in_init(
             idx += 1
         # Otherwise we keep the line, except if it's a tokenizer import and we don't want to keep it.
         elif keep_processing or (
-            re.search('^\s*"(tokenization|processing|feature_extraction|image_processing)', lines[idx]) is None
-            and re.search("^\s*from .(tokenization|processing|feature_extraction|image_processing)", lines[idx])
+            re.search(r'^\s*"(tokenization|processing|feature_extraction|image_processing)', lines[idx]) is None
+            and re.search(r"^\s*from .(tokenization|processing|feature_extraction|image_processing)", lines[idx])
             is None
         ):
             new_lines.append(lines[idx])
@@ -1192,7 +1192,7 @@ def duplicate_doc_file(
     with open(doc_file, "r", encoding="utf-8") as f:
         content = f.read()
 
-    content = re.sub("<!--\s*Copyright (\d+)\s", f"<!--Copyright {CURRENT_YEAR} ", content)
+    content = re.sub(r"<!--\s*Copyright (\d+)\s", f"<!--Copyright {CURRENT_YEAR} ", content)
     if frameworks is None:
         frameworks = get_default_frameworks()
     if dest_file is None:
@@ -1218,7 +1218,7 @@ def duplicate_doc_file(
         if not block.startswith("#"):
             new_blocks.append(block)
         # Main title
-        elif re.search("^#\s+\S+", block) is not None:
+        elif re.search(r"^#\s+\S+", block) is not None:
             new_blocks.append(f"# {new_model_patterns.model_name}\n")
         # The config starts the part of the doc with the classes.
         elif not in_classes and old_model_patterns.config_class in block.split("\n")[0]:
@@ -1230,7 +1230,7 @@ def duplicate_doc_file(
         elif in_classes:
             in_classes = True
             block_title = block.split("\n")[0]
-            block_class = re.search("^#+\s+(\S.*)$", block_title).groups()[0]
+            block_class = re.search(r"^#+\s+(\S.*)$", block_title).groups()[0]
             new_block, _ = replace_model_patterns(block, old_model_patterns, new_model_patterns)
 
             if "Tokenizer" in block_class:

src/transformers/modeling_utils.py

@@ -1829,7 +1829,7 @@ def save_pretrained(
 
             # make sure that file to be deleted matches format of sharded file, e.g. pytorch_model-00001-of-00005
             filename_no_suffix = filename.replace(".bin", "").replace(".safetensors", "")
-            reg = re.compile("(.*?)-\d{5}-of-\d{5}")
+            reg = re.compile(r"(.*?)-\d{5}-of-\d{5}")
 
             if (
                 filename.startswith(weights_no_suffix)

src/transformers/models/bloom/convert_bloom_original_checkpoint_to_pytorch.py

@@ -71,7 +71,7 @@ def layer_name_mapping(key, file):
 def get_dtype_size(dtype):
     if dtype == torch.bool:
         return 1 / 8
-    bit_search = re.search("[^\d](\d+)$", str(dtype))
+    bit_search = re.search(r"[^\d](\d+)$", str(dtype))
     if bit_search is None:
         raise ValueError(f"`dtype` is not a valid dtype: {dtype}.")
     bit_size = int(bit_search.groups()[0])

src/transformers/models/cvt/convert_cvt_original_pytorch_checkpoint_to_pytorch.py

@@ -350,7 +350,7 @@ def convert_cvt_checkpoint(cvt_model, image_size, cvt_file_name, pytorch_dump_fo
     )
     parser.add_argument(
         "--cvt_file_name",
-        default="cvtmodels\CvT-w24-384x384-IN-22k.pth",
+        default=r"cvtmodels\CvT-w24-384x384-IN-22k.pth",
         type=str,
         help="Input Image Size",
     )

src/transformers/models/deformable_detr/modeling_deformable_detr.py

@@ -1823,7 +1823,7 @@ def forward(
 )
 class DeformableDetrForObjectDetection(DeformableDetrPreTrainedModel):
     # When using clones, all layers > 0 will be clones, but layer 0 *is* required
-    _keys_to_ignore_on_load_missing = ["bbox_embed\.[1-9]\d*", "class_embed\.[1-9]\d*"]
+    _keys_to_ignore_on_load_missing = [r"bbox_embed\.[1-9]\d*", r"class_embed\.[1-9]\d*"]
 
     def __init__(self, config: DeformableDetrConfig):
         super().__init__(config)

src/transformers/models/deta/modeling_deta.py

@@ -1775,7 +1775,7 @@ def forward(
 )
 class DetaForObjectDetection(DetaPreTrainedModel):
     # When using clones, all layers > 0 will be clones, but layer 0 *is* required
-    _keys_to_ignore_on_load_missing = ["bbox_embed\.[1-9]\d*", "class_embed\.[1-9]\d*"]
+    _keys_to_ignore_on_load_missing = [r"bbox_embed\.[1-9]\d*", r"class_embed\.[1-9]\d*"]
 
     # Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrForObjectDetection.__init__ with DeformableDetr->Deta
     def __init__(self, config: DetaConfig):

src/transformers/models/efficientformer/convert_efficientformer_original_pytorch_checkpoint_to_pytorch.py

@@ -50,12 +50,12 @@ def rename_key(old_name, num_meta4D_last_stage):
         else:
             new_name = old_name.replace("4", "batchnorm_after")
 
-    if "network" in old_name and re.search("\d\.\d", old_name):
+    if "network" in old_name and re.search(r"\d\.\d", old_name):
         two_digit_num = r"\b\d{2}\b"
         if bool(re.search(two_digit_num, old_name)):
-            match = re.search("\d\.\d\d.", old_name).group()
+            match = re.search(r"\d\.\d\d.", old_name).group()
         else:
-            match = re.search("\d\.\d.", old_name).group()
+            match = re.search(r"\d\.\d.", old_name).group()
         if int(match[0]) < 6:
             trimmed_name = old_name.replace(match, "")
             trimmed_name = trimmed_name.replace("network", match[0] + ".meta4D_layers.blocks." + match[2:-1])
@@ -78,7 +78,7 @@ def rename_key(old_name, num_meta4D_last_stage):
 
             new_name = "last_stage." + trimmed_name
 
-    elif "network" in old_name and re.search(".\d.", old_name):
+    elif "network" in old_name and re.search(r".\d.", old_name):
         new_name = old_name.replace("network", "intermediate_stages")
 
     if "fc" in new_name:

src/transformers/models/glpn/modeling_glpn.py

@@ -632,7 +632,7 @@ def forward(self, hidden_states: List[torch.Tensor]) -> List[torch.Tensor]:
 
 
 class SiLogLoss(nn.Module):
-    """
+    r"""
     Implements the Scale-invariant log scale loss [Eigen et al., 2014](https://arxiv.org/abs/1406.2283).
 
     $$L=\frac{1}{n} \sum_{i} d_{i}^{2}-\frac{1}{2 n^{2}}\left(\sum_{i} d_{i}^{2}\right)$$ where $d_{i}=\log y_{i}-\log

src/transformers/models/jukebox/convert_jukebox.py

@@ -97,23 +97,23 @@ def fix_jukebox_keys(state_dict, model_state_dict, key_prefix, mapping):
     new_dict = {}
     import re
 
-    re_encoder_block_conv_in = re.compile("encoders.(\d*).level_blocks.(\d*).model.(\d*).(\d).(bias|weight)")
+    re_encoder_block_conv_in = re.compile(r"encoders.(\d*).level_blocks.(\d*).model.(\d*).(\d).(bias|weight)")
     re_encoder_block_resnet = re.compile(
-        "encoders.(\d*).level_blocks.(\d*).model.(\d*).(\d).model.(\d*).model.(\d*).(bias|weight)"
+        r"encoders.(\d*).level_blocks.(\d*).model.(\d*).(\d).model.(\d*).model.(\d*).(bias|weight)"
     )
-    re_encoder_block_proj_out = re.compile("encoders.(\d*).level_blocks.(\d*).model.(\d*).(bias|weight)")
+    re_encoder_block_proj_out = re.compile(r"encoders.(\d*).level_blocks.(\d*).model.(\d*).(bias|weight)")
 
-    re_decoder_block_conv_out = re.compile("decoders.(\d*).level_blocks.(\d*).model.(\d*).(\d).(bias|weight)")
+    re_decoder_block_conv_out = re.compile(r"decoders.(\d*).level_blocks.(\d*).model.(\d*).(\d).(bias|weight)")
     re_decoder_block_resnet = re.compile(
-        "decoders.(\d*).level_blocks.(\d*).model.(\d*).(\d).model.(\d*).model.(\d*).(bias|weight)"
+        r"decoders.(\d*).level_blocks.(\d*).model.(\d*).(\d).model.(\d*).model.(\d*).(bias|weight)"
     )
-    re_decoder_block_proj_in = re.compile("decoders.(\d*).level_blocks.(\d*).model.(\d*).(bias|weight)")
+    re_decoder_block_proj_in = re.compile(r"decoders.(\d*).level_blocks.(\d*).model.(\d*).(bias|weight)")
 
-    re_prior_cond_conv_out = re.compile("conditioner_blocks.(\d*).cond.model.(\d*).(\d).(bias|weight)")
+    re_prior_cond_conv_out = re.compile(r"conditioner_blocks.(\d*).cond.model.(\d*).(\d).(bias|weight)")
     re_prior_cond_resnet = re.compile(
-        "conditioner_blocks.(\d*).cond.model.(\d*).(\d).model.(\d*).model.(\d*).(bias|weight)"
+        r"conditioner_blocks.(\d*).cond.model.(\d*).(\d).model.(\d*).model.(\d*).(bias|weight)"
     )
-    re_prior_cond_proj_in = re.compile("conditioner_blocks.(\d*).cond.model.(\d*).(bias|weight)")
+    re_prior_cond_proj_in = re.compile(r"conditioner_blocks.(\d*).cond.model.(\d*).(bias|weight)")
 
     for original_key, value in state_dict.items():
         # rename vqvae.encoder keys

src/transformers/models/jukebox/tokenization_jukebox.py

@@ -148,10 +148,10 @@ def __init__(
         with open(lyrics_file, encoding="utf-8") as vocab_handle:
             self.lyrics_encoder = json.load(vocab_handle)
 
-        oov = "[^A-Za-z0-9.,:;!?\-'\"()\[\] \t\n]+"
+        oov = r"[^A-Za-z0-9.,:;!?\-'\"()\[\] \t\n]+"
         # In v2, we had a n_vocab=80 and in v3 we missed + and so n_vocab=79 of characters.
         if len(self.lyrics_encoder) == 79:
-            oov = oov.replace("\-'", "\-+'")
+            oov = oov.replace(r"\-'", r"\-+'")
 
         self.out_of_vocab = regex.compile(oov)
         self.artists_decoder = {v: k for k, v in self.artists_encoder.items()}
@@ -230,7 +230,7 @@ def prepare_for_tokenization(
                 ]  # split is for the full dictionary with combined genres
 
         if self.version[0] == "v2":
-            self.out_of_vocab = regex.compile("[^A-Za-z0-9.,:;!?\-'\"()\[\] \t\n]+")
+            self.out_of_vocab = regex.compile(r"[^A-Za-z0-9.,:;!?\-'\"()\[\] \t\n]+")
             vocab = "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789.,:;!?-+'\"()[] \t\n"
             self.vocab = {vocab[index]: index + 1 for index in range(len(vocab))}
             self.vocab["<unk>"] = 0
@@ -239,7 +239,7 @@ def prepare_for_tokenization(
             self.lyrics_decoder = {v: k for k, v in self.vocab.items()}
             self.lyrics_decoder[0] = ""
         else:
-            self.out_of_vocab = regex.compile("[^A-Za-z0-9.,:;!?\-+'\"()\[\] \t\n]+")
+            self.out_of_vocab = regex.compile(r"[^A-Za-z0-9.,:;!?\-+'\"()\[\] \t\n]+")
 
         lyrics = self._run_strip_accents(lyrics)
         lyrics = lyrics.replace("\\", "\n")

src/transformers/models/maskformer/modeling_maskformer.py

@@ -1241,7 +1241,7 @@ def forward(self, features: List[Tensor]) -> List[Tensor]:
 
 class MaskFormerPixelDecoder(nn.Module):
     def __init__(self, *args, feature_size: int = 256, mask_feature_size: int = 256, **kwargs):
-        """
+        r"""
         Pixel Decoder Module proposed in [Per-Pixel Classification is Not All You Need for Semantic
         Segmentation](https://arxiv.org/abs/2107.06278). It first runs the backbone's features into a Feature Pyramid
         Network creating a list of feature maps. Then, it projects the last one to the correct `mask_size`.
@@ -1250,7 +1250,7 @@ def __init__(self, *args, feature_size: int = 256, mask_feature_size: int = 256,
             feature_size (`int`, *optional*, defaults to 256):
                 The feature size (channel dimension) of the FPN feature maps.
             mask_feature_size (`int`, *optional*, defaults to 256):
-                The features (channels) of the target masks size \\C_{\epsilon}\\ in the paper.
+                The features (channels) of the target masks size \\(C_{\epsilon}\\) in the paper.
         """
         super().__init__()
 

src/transformers/models/megatron_bert/convert_megatron_bert_checkpoint.py

@@ -150,7 +150,7 @@ def convert_megatron_checkpoint(args, input_state_dict, config):
     transformer = lm["transformer"] if "transformer" in lm.keys() else lm["encoder"]
 
     # The regex to extract layer names.
-    layer_re = re.compile("layers\.(\d+)\.([a-z0-9_.]+)\.([a-z]+)")
+    layer_re = re.compile(r"layers\.(\d+)\.([a-z0-9_.]+)\.([a-z]+)")
 
     # The simple map of names for "automated" rules.
     megatron_to_transformers = {

src/transformers/models/megatron_gpt2/checkpoint_reshaping_and_interoperability.py

@@ -394,7 +394,7 @@ def convert_checkpoint_from_megatron_to_transformers(args):
     pp_size = megatron_args.pipeline_model_parallel_size
     dtype = torch.float32
     # The regex to extract layer names.
-    layer_re = re.compile("layers\.(\d+)\.([a-z0-9_.]+)\.([a-z]+)")
+    layer_re = re.compile(r"layers\.(\d+)\.([a-z0-9_.]+)\.([a-z]+)")
 
     # Convert.
     print("Converting")
@@ -746,7 +746,7 @@ def convert_checkpoint_from_transformers_to_megatron(args):
         )
     num_layers = config.num_hidden_layers // args.target_pipeline_model_parallel_size
 
-    layer_re = re.compile("transformer.h\.(\d+)\.([a-z0-9_.]+)\.([a-z]+)")
+    layer_re = re.compile(r"transformer.h\.(\d+)\.([a-z0-9_.]+)\.([a-z]+)")
     # The number of heads.
     heads = config.n_head
     # The hidden_size per head.

src/transformers/models/megatron_gpt2/convert_megatron_gpt2_checkpoint.py

@@ -148,7 +148,7 @@ def convert_megatron_checkpoint(args, input_state_dict, config):
     transformer = lm["transformer"] if "transformer" in lm.keys() else lm["encoder"]
 
     # The regex to extract layer names.
-    layer_re = re.compile("layers\.(\d+)\.([a-z0-9_.]+)\.([a-z]+)")
+    layer_re = re.compile(r"layers\.(\d+)\.([a-z0-9_.]+)\.([a-z]+)")
 
     # The simple map of names for "automated" rules.
     megatron_to_transformers = {

tests/sagemaker/conftest.py

@@ -32,15 +32,15 @@ class SageMakerTestEnvironment:
     def metric_definitions(self) -> str:
         if self.framework == "pytorch":
             return [
-                {"Name": "train_runtime", "Regex": "train_runtime.*=\D*(.*?)$"},
-                {"Name": "eval_accuracy", "Regex": "eval_accuracy.*=\D*(.*?)$"},
-                {"Name": "eval_loss", "Regex": "eval_loss.*=\D*(.*?)$"},
+                {"Name": "train_runtime", "Regex": r"train_runtime.*=\D*(.*?)$"},
+                {"Name": "eval_accuracy", "Regex": r"eval_accuracy.*=\D*(.*?)$"},
+                {"Name": "eval_loss", "Regex": r"eval_loss.*=\D*(.*?)$"},
             ]
         else:
             return [
-                {"Name": "train_runtime", "Regex": "train_runtime.*=\D*(.*?)$"},
-                {"Name": "eval_accuracy", "Regex": "loss.*=\D*(.*?)]?$"},
-                {"Name": "eval_loss", "Regex": "sparse_categorical_accuracy.*=\D*(.*?)]?$"},
+                {"Name": "train_runtime", "Regex": r"train_runtime.*=\D*(.*?)$"},
+                {"Name": "eval_accuracy", "Regex": r"loss.*=\D*(.*?)]?$"},
+                {"Name": "eval_loss", "Regex": r"sparse_categorical_accuracy.*=\D*(.*?)]?$"},
             ]
 
     @property

tests/utils/test_add_new_model_like.py

@@ -157,14 +157,14 @@ def test_add_content_to_text(self):
         self.assertEqual(
             add_content_to_text(test_text, line, add_before='    "bert": "BertConfig",', exact_match=True), expected
         )
-        self.assertEqual(add_content_to_text(test_text, line, add_before=re.compile('^\s*"bert":')), expected)
+        self.assertEqual(add_content_to_text(test_text, line, add_before=re.compile(r'^\s*"bert":')), expected)
 
         self.assertEqual(add_content_to_text(test_text, line, add_after="gpt"), expected)
         self.assertEqual(add_content_to_text(test_text, line, add_after="gpt", exact_match=True), test_text)
         self.assertEqual(
             add_content_to_text(test_text, line, add_after='    "gpt": "GPTConfig",', exact_match=True), expected
         )
-        self.assertEqual(add_content_to_text(test_text, line, add_after=re.compile('^\s*"gpt":')), expected)
+        self.assertEqual(add_content_to_text(test_text, line, add_after=re.compile(r'^\s*"gpt":')), expected)
 
     def test_add_content_to_file(self):
         test_text = """all_configs = {
@@ -197,7 +197,7 @@ def test_add_content_to_file(self):
             self.check_result(file_name, expected)
 
             self.init_file(file_name, test_text)
-            add_content_to_file(file_name, line, add_before=re.compile('^\s*"bert":'))
+            add_content_to_file(file_name, line, add_before=re.compile(r'^\s*"bert":'))
             self.check_result(file_name, expected)
 
             self.init_file(file_name, test_text)
@@ -213,7 +213,7 @@ def test_add_content_to_file(self):
             self.check_result(file_name, expected)
 
             self.init_file(file_name, test_text)
-            add_content_to_file(file_name, line, add_after=re.compile('^\s*"gpt":'))
+            add_content_to_file(file_name, line, add_after=re.compile(r'^\s*"gpt":'))
             self.check_result(file_name, expected)
 
     def test_simplify_replacements(self):