Merge pull request #1390 from BossPi/formers_tokenizer

Formers tokenizer

Author: lugimzzz

data_processor/image_preprocessor/image_preprocessor_adaptive.py

@@ -241,7 +241,7 @@ def get_smarted_resize(self, height, width, min_pixels=None, max_pixels=None):
 
     def to_dict(self):
         encoder_dict = super().to_dict()
-        encoder_dict.pop("image_processor_type", None)
+        encoder_dict["image_processor_type"] = "Ernie4_5_VLImageProcessor"
         return encoder_dict
 
     def _preprocess(

data_processor/steps/input_ids_messaging/data_process/process_video.py

@@ -33,6 +33,7 @@
     SFT_IMAGE_END_TOKEN,
     SFT_IMAGE_START_TOKEN,
 )
+from paddleformers.transformers.legacy.tokenizer_utils_base import BatchEncoding
 
 
 class VideoProcess(Process):
@@ -335,7 +336,11 @@ def calculate_ratios_with_min_one(numbers):
         # calculate the ratio of each video
         text_token_count = 0
         for item in meta["text_info"]:
-            text_token_count += len(self.tokenizer.encode(item["text"])["input_ids"])
+            tokens = self.tokenizer.encode(item["text"])
+            if isinstance(tokens, BatchEncoding):
+                text_token_count += len(tokens["input_ids"])
+            else:
+                text_token_count += len(tokens)
         text_token_count += 1  # for eos token
 
         if not self.is_training:

data_processor/steps/input_ids_messaging/data_utils.py

@@ -22,11 +22,16 @@
 
 import xxhash
 from PIL import Image
+from paddleformers.transformers.legacy.tokenizer_utils_base import BatchEncoding
 
 
 def get_text_token_num(tokenizer, text: str):
     """text tokenize and count"""
-    return len(tokenizer.encode(text)["input_ids"])
+    tokens = tokenizer.encode(text)
+    if isinstance(tokens, BatchEncoding):
+        return len(tokens["input_ids"])
+    else:
+        return len(tokens)
 
 
 def get_uniq_id(text):

data_processor/steps/input_ids_messaging/example_to_feature.py

@@ -51,6 +51,7 @@
     SFT_VIDEO_END_TOKEN,
     SFT_VIDEO_START_TOKEN,
 )
+from paddleformers.transformers.legacy.tokenizer_utils_base import BatchEncoding
 
 
 class SlidingWindowsContextManager:
@@ -539,7 +540,9 @@ def _text_tokenization(self, sample, dataset_name, data_type):
             else:
                 cur_tokens = self.tokenizer.encode(
                     item["text"], add_special_tokens=False, return_attention_mask=False
-                )["input_ids"]
+                )
+                if isinstance(cur_tokens, BatchEncoding):
+                    cur_tokens = cur_tokens["input_ids"]
             input_ids.append(cur_tokens)
 
             mask_flag = item.get("tag", "no_mask")

ernie/tokenizer.py

@@ -24,7 +24,10 @@
 import paddle
 import sentencepiece as spm
 from paddleformers.transformers import PretrainedTokenizer
-from paddleformers.transformers.tokenizer_utils_base import PaddingStrategy, TextInput
+from paddleformers.transformers.legacy.tokenizer_utils_base import (
+    PaddingStrategy,
+    TextInput,
+)
 from paddleformers.utils.log import logger
 
 

ernie/tokenizer_vl.py

@@ -29,7 +29,7 @@
 
 from paddleformers.utils.log import logger
 from paddleformers.transformers import PretrainedTokenizer
-from paddleformers.transformers.tokenizer_utils_base import (
+from paddleformers.transformers.legacy.tokenizer_utils_base import (
     PaddingStrategy,
     TextInput,
 )

erniekit/train/vl_sft/workflow.py

@@ -20,9 +20,7 @@
 
 import numpy as np
 import paddle
-from paddleformers.transformers import (
-    AutoConfig,
-)
+from paddleformers.transformers import AutoConfig, AutoTokenizer
 from paddle.distributed import fleet
 from paddleformers.datasets import IterDataset
 from paddleformers.trainer import get_last_checkpoint
@@ -238,11 +236,18 @@ def run_vl_sft(
     )
     print("data_processor_args:\n", preprocess_args)
 
-    tokenizer = Ernie4_5_VLTokenizer.from_pretrained(
-        model_args.model_name_or_path,
-        padding_side="right",
-        model_max_length=data_args.max_seq_len,
-    )
+    if convert_from_hf:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.model_name_or_path,
+            padding_side="right",
+            model_max_length=data_args.max_seq_len,
+        )
+    else:
+        tokenizer = Ernie4_5_VLTokenizer.from_pretrained(
+            model_args.model_name_or_path,
+            padding_side="right",
+            model_max_length=data_args.max_seq_len,
+        )
     data_processor = End2EndProcessor(preprocess_args, tokenizer, image_preprocess_save)
     data_processor.train().sft()
     logger.info(f"[DEBUG] data_processor_args: {preprocess_args}")