update

finetune/dataset.py

@@ -14,6 +14,9 @@ from PIL import Image
 from torch.nn.utils.rnn import pad_sequence
 from torch.utils.data import Dataset
 from transformers import AutoProcessor, AutoTokenizer
+import logging
+
+logger = logging.getLogger(__name__)
 
 llama3_chat_template = "{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}"
 
@@ -30,6 +33,7 @@ class SupervisedDataset(Dataset):
         patch_size=14,
         query_nums=64,
         batch_vision=False,
+        max_length=None,
     ):
         super(SupervisedDataset, self).__init__()
         self.raw_data = raw_data
@@ -40,13 +44,13 @@ class SupervisedDataset(Dataset):
         self.patch_size = patch_size
         self.query_nums=query_nums
         self.batch_vision = batch_vision
+        self.max_length = max_length
 
     def __len__(self):
         return len(self.raw_data)
 
     def __getitem__(self, i) -> Dict[str, torch.Tensor]:
-        # try:
-        if 1:
+        try:
             if isinstance(self.raw_data[i]["image"], str):
                 images_dict = { "<image>" : Image.open(self.raw_data[i]["image"]).convert("RGB") }
             elif isinstance(self.raw_data[i]["image"], Dict):
@@ -63,6 +67,7 @@ class SupervisedDataset(Dataset):
                 llm_type=self.llm_type,
                 patch_size=self.patch_size,
                 batch_vision=self.batch_vision,
+                max_length=self.max_length
             )
             ret = dict(
                 input_ids=ret["input_ids"],
@@ -73,13 +78,12 @@ class SupervisedDataset(Dataset):
                 tgt_sizes=ret["tgt_sizes"],
                 image_bound=ret["image_bound"],
             )
-        # except:
-        #     print(f"data fetch error")
-        #     return self.__getitem__(random.randint(0, len(self)))
+        except:
+            logger.error(f"data fetch error")
+            return self.__getitem__(random.randint(0, len(self)))
         return ret
 
         
-
 def data_collator(examples, padding_value=0, max_length=2048):
     def trim_and_pad(seq, batch_first, padding_value):
         return pad_sequence([s[:max_length] for s in seq], batch_first=True, padding_value=padding_value)
@@ -118,7 +122,7 @@ def data_collator(examples, padding_value=0, max_length=2048):
     }
 
 
-def conversation_to_ids(conversation, tokenizer, llm_type=None, new_schema=False):
+def conversation_to_ids(conversation, tokenizer, llm_type=None, new_schema=False, max_length=None):
     """
     for single image multi-turn conversation
     conversation: [{'role': 'user', 'content': 'Describe this image'},
@@ -139,6 +143,14 @@ def conversation_to_ids(conversation, tokenizer, llm_type=None, new_schema=False
 
     ids = torch.from_numpy(np.hstack(input_ids, dtype=np.int32))
     context = torch.from_numpy(np.hstack(context, dtype=np.int8))
+    if input_ids.shape[-1] > max_length:
+        ids =ids[:max_length]
+        context = context[:max_length]
+        logger.warning(f"The input length ({input_ids.shape[-1]}) exceeds the model's maximum length ({max_length}), so it has been truncated")
+    
+    if torch.all(context):
+        logger.error("No tokens available to compute loss.")
+        raise Exception("No tokens available to compute loss.")
 
     # build target
     target = torch.full_like(ids, -100, dtype=torch.int32)
@@ -164,7 +176,8 @@ def conversation_to_ids(conversation, tokenizer, llm_type=None, new_schema=False
         image_start_tokens += 1
         image_end_tokens = torch.where(ids == tokenizer.im_end_id)[0]
     if len(image_start_tokens) != len(image_end_tokens):
-        print("image start token != image end tokens")
+        logger.error("image start token != image end tokens")
+        raise Exception("image start token != image end tokens")
     
     if len(image_start_tokens) > 0:
         image_bound = torch.hstack(
@@ -294,7 +307,6 @@ def conversation_to_ids_qwen2(conversation, tokenizer):
     return input_ids, context, raw_msg
 
 
-
 def preprocess(
     images_dict,
     conversations,
@@ -305,6 +317,7 @@ def preprocess(
     llm_type=None,
     patch_size=14,
     batch_vision=False,
+    max_length=None,
 ):
     """
     single(multi) image(s) preprocess, the image(s) will be placed at the top of the conversation
@@ -369,7 +382,7 @@ def preprocess(
             conversations[0]["content"] = (
                 image_placeholder + "\n" + conversation[0]["content"]
             )
-        input_dict = conversation_to_ids(conversations, tokenizer, llm_type, new_schema)
+        input_dict = conversation_to_ids(conversations, tokenizer, llm_type, new_schema, max_length)
     else:
         pattern = r'<image_\d+>'
         new_conversations = []
@@ -377,16 +390,19 @@ def preprocess(
             content = conversation['content']
             parts = re.split(f'({pattern})', content)
             for i, part in enumerate(parts):
+                if not part.strip():
+                    continue
                 if re.match(pattern, part):  
                     if part in image_placeholder_dict:
                         parts[i] = image_placeholder_dict[part] 
                     else:
-                        print(f'Unreplaced image tag: {part}')
-            conversation['content'] = ''.join(parts)
+                        raise Exception(f"not found {part} in image dict")
+            conversation['content'] = '\n'.join(parts)
             new_conversations.append(conversation)
         conversations = new_conversations
         
-        input_dict = conversation_to_ids(conversations, tokenizer, llm_type, new_schema)
+        input_dict = conversation_to_ids(conversations, tokenizer, llm_type, new_schema, max_length)
+
     if batch_vision:
         tgt_sizes = []
         reshape_images = []

finetune/finetune.py

@@ -108,6 +108,7 @@ def make_supervised_data_module(
         patch_size=patch_size,
         query_nums=query_nums,
         batch_vision=batch_vision,
+        max_length=max_length,
     )
 
     if data_args.eval_data_path:
@@ -121,6 +122,7 @@ def make_supervised_data_module(
             patch_size=patch_size,
             query_nums=query_nums,
             batch_vision=batch_vision,
+            max_length=max_length,
         )
     else:
         eval_dataset = None
@@ -205,7 +207,6 @@ def train():
     tokenizer = AutoTokenizer.from_pretrained(
         model_args.model_name_or_path, trust_remote_code=True
     )
-    tokenizer.model_max_length = training_args.model_max_length
 
     if not training_args.tune_vision:
         model.vpm.requires_grad_(False)

finetune/finetune_lora.sh

@@ -15,7 +15,7 @@ LLM_TYPE="qwen2"
 # if use openbmb/MiniCPM-V-2, please set LLM_TYPE=minicpm
 #if use openbmb/MiniCPM-Llama3-V-2_5, please set LLM_TYPE=llama3
 
-MODEL_MAX_Length=4096 # if use openbmb/MiniCPM-V-2 or openbmb/MiniCPM-Llama3-V-2_5, please set MODEL_MAX_Length=2048
+MODEL_MAX_Length=2048 # if conduct multi-images sft, please set MODEL_MAX_Length=4096
 
 DISTRIBUTED_ARGS="
     --nproc_per_node $GPUS_PER_NODE \

finetune/readme.md

@@ -53,7 +53,9 @@ If your input consists of a single image, you can use a single placeholder **\<i
 </details>
 
 #### Multiple Images Example
-For inputs with multiple images, you should use a dictionary where each key represents a unique placeholder (e.g., **\<image_00\>**, **\<image_01\>**), and the corresponding value is the image path. You can then use these placeholders in the conversation to insert the images at specific positions.
+For inputs containing multiple images, utilize a dictionary where each key represents a unique placeholder (e.g., \textbf{\textbackslash image\_00}, \textbf{\textbackslash image\_01}) with the corresponding image path as its value. These placeholders can then be used within the conversation to seamlessly insert images at specific positions.
+
+Additionally, to optimize resource management, especially when dealing with large batches of images during training or inference, consider reducing \texttt{max\_slice\_nums}. If you are performing multi-image supervised fine-tuning (SFT), it's recommended to set \texttt{MODEL\_MAX\_LENGTH=4096} in your script for better performance.
 
 <details>
   <summary>