Bump version to v1.2.1 (#46)

1. Support dynamic sequence length during training
2. Update README.md
3. Update evaluation code

Author: czczup

BLOG.md

@@ -58,7 +58,6 @@ We released [InternVL-Chat-V1.1](https://huggingface.co/OpenGVLab/InternVL-Chat-
 
 <img width="650" alt="image" src="https://github.com/OpenGVLab/InternVL/assets/8529570/0e60912e-c52b-46fa-bd61-5f94a221d1fc">
 
-
 ## InternVL
 
 > Date: 2023/12/12<br>

README.md

@@ -332,7 +332,7 @@ InternVL scales up the ViT to _**6B parameters**_ and aligns it with LLM.
   <summary>Multimodal Dialogue (click to expand)</summary>
 
 - Compared with SOTA VLLMs
-  
+
   | name               | image size | MMMU<br>(val) | MMMU<br>(test) | MathVista<br>(testmini) | MMB<br>(test) | MMB−CN<br>(test) | MMVP | MME      | ScienceQA<br>(image) | POPE | TextVQA | SEEDv1<br>(image) | VizWiz<br>(test) | GQA<br>(test) |
   | ------------------ | ---------- | ------------- | -------------- | ----------------------- | ------------- | ---------------- | ---- | -------- | -------------------- | ---- | ------- | ----------------- | ---------------- | ------------- |
   | GPT-4V\*           | unknown    | 56.8          | 55.7           | 49.9                    | 77.0          | 74.4             | 38.7 | 1409/517 | -                    | -    | 78.0    | 71.6              | -                | -             |
@@ -343,7 +343,7 @@ InternVL scales up the ViT to _**6B parameters**_ and aligns it with LLM.
   |                    |            |               |                |                         |               |                  |      |          |                      |      |         |                   |                  |               |
   | LLaVA-NEXT-34B     | 672x672    | 51.1          | 44.7           | 46.5                    | 79.3          | 79.0             | -    | 1631/397 | 81.8                 | 87.7 | 69.5    | 75.9              | 63.8             | 67.1          |
   | InternVL-Chat-V1.2 | 448x448    | 51.6          | 46.2           | 47.7                    | 82.2          | 81.2             | 56.7 | 1672/509 | 83.3                 | 88.0 | 69.7    | 75.6              | 60.0             | 64.0          |
-  
+
   \* denotes proprietary models. MMBench results are collected from the [leaderboard](https://mmbench.opencompass.org.cn/leaderboard). In most benchmarks, InternVL-Chat-V1.2 achieves better performance than LLaVA-NeXT-34B.
 
 - Zero-Shot Image Captioning [\[see details\]](./internvl_g#zero-shot-image-captioning)

internvl_chat/README.md

@@ -141,6 +141,8 @@ The hyperparameters used for finetuning are listed in the following table.
 
 ## 📊 Evaluation
 
+\* Training set observed.
+
 **MultiModal Benchmark**
 
 | model                                                                             | MME            | MMB<sub>dev/test</sub> | MMB-CN<sub>dev/test</sub> | POPE | MMVP | MathVista |
@@ -151,14 +153,14 @@ The hyperparameters used for finetuning are listed in the following table.
 | model                                                                             | MMMU<sub>val/test</sub>                                                            | CMMMU<sub>val/test</sub> | Tiny<sub>LVLM</sub> | LLaVA<sub>bench</sub> | MM-Vet |
 | --------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------- | ------------------------ | ------------------- | --------------------- | ------ |
 | [InternVL-Chat-V1.1](https://huggingface.co/OpenGVLab/InternVL-Chat-Chinese-V1-1) | 39.1 / 35.3                                                                        | 34.8 / 34.0              | 344.5               | 76.3                  | 45.0   |
-| [InternVL-Chat-V1.2](https://huggingface.co/OpenGVLab/InternVL-Chat-Chinese-V1-2) | 51.6 / [46.2](https://eval.ai/web/challenges/challenge-page/2179/leaderboard/5377) | TODO                     | 350.3               | -                     | 48.9   |
+| [InternVL-Chat-V1.2](https://huggingface.co/OpenGVLab/InternVL-Chat-Chinese-V1-2) | 51.6 / [46.2](https://eval.ai/web/challenges/challenge-page/2179/leaderboard/5377) | -                        | 350.3               | -                     | 48.9   |
 
 **Visual Question Answering**
 
 | model                                                                             | VQAv2<sub>test</sub> | OKVQA<sub>val</sub> | TextVQA<sub>val</sub> | VizWiz<sub>val/test</sub> | AI2D<sub>test</sub> | GQA<sub>test</sub> | SQA<sub>test</sub> |
 | --------------------------------------------------------------------------------- | -------------------- | ------------------- | --------------------- | ------------------------- | ------------------- | ------------------ | ------------------ |
-| [InternVL-Chat-V1.1](https://huggingface.co/OpenGVLab/InternVL-Chat-Chinese-V1-1) | 80.9                 | 64.2                | 65.8                  | 58.3 / 57.3               | 70.2                | 62.4               | 91.2               |
-| [InternVL-Chat-V1.2](https://huggingface.co/OpenGVLab/InternVL-Chat-Chinese-V1-2) | -                    | 62.5                | 69.7                  | 61.9 / 60.0               | 71.6                | 64.0               | 83.3               |
+| [InternVL-Chat-V1.1](https://huggingface.co/OpenGVLab/InternVL-Chat-Chinese-V1-1) | 80.9\*               | 64.2\*              | 65.8                  | 58.3 / 57.3               | 70.2\*              | 62.4\*             | 91.2\*             |
+| [InternVL-Chat-V1.2](https://huggingface.co/OpenGVLab/InternVL-Chat-Chinese-V1-2) | -                    | 62.5\*              | 69.7                  | 61.9 / 60.0               | 71.6\*              | 64.0\*             | 83.3               |
 
 **Image Captioning**
 

internvl_chat/eval/scienceqa/evaluate_scienceqa.py

@@ -114,7 +114,7 @@ def post_process(pred, option):
             if v in pred:
                 return k
 
-    return random.choice(option_candidate)
+    return pred
 
 
 def evaluate_chat_model():

internvl_chat/internvl/patch/__init__.py

@@ -2,7 +2,11 @@
 from .llama_flash_attn_monkey_patch import replace_llama_attn_with_flash_attn
 from .llama_rmsnorm_monkey_patch import \
     replace_llama_rmsnorm_with_fused_rmsnorm
+from .pad_data_collator import pad_data_collator
+from .train_sampler_patch import replace_train_sampler
 
 __all__ = ['replace_llama_attn_with_flash_attn',
            'replace_llama_rmsnorm_with_fused_rmsnorm',
-           'replace_llama2_attn_with_flash_attn']
+           'replace_llama2_attn_with_flash_attn',
+           'replace_train_sampler',
+           'pad_data_collator']

internvl_chat/internvl/patch/pad_data_collator.py

@@ -0,0 +1,49 @@
+import numpy as np
+import torch
+
+IGNORE_INDEX = -100
+
+
+def pad_data_collator(features, pad_id=0):
+
+    first = features[0]
+    batch = {}
+
+    batch_lens = [feat['input_ids'].shape for feat in features]
+    max_item_length = max(batch_lens)[0]
+    for idx in range(len(features)):
+        feat = features[idx]
+        temp_input_ids = torch.LongTensor([pad_id] * max_item_length)
+        temp_input_ids[:feat['input_ids'].shape[0]] = feat['input_ids']
+        feat['input_ids'] = temp_input_ids
+        temp_labels = torch.LongTensor([IGNORE_INDEX] * max_item_length)
+        temp_labels[:feat['labels'].shape[0]] = feat['labels']
+        feat['labels'] = temp_labels
+        feat['attention_mask'] = feat['input_ids'].ne(pad_id)
+
+    # Special handling for labels.
+    # Ensure that tensor is created with the correct type
+    # (it should be automatically the case, but let's make sure of it.)
+    if 'label' in first and first['label'] is not None:
+        label = first['label'].item() if isinstance(first['label'], torch.Tensor) else first['label']
+        dtype = torch.long if isinstance(label, int) else torch.float
+        batch['labels'] = torch.tensor([f['label'] for f in features], dtype=dtype)
+    elif 'label_ids' in first and first['label_ids'] is not None:
+        if isinstance(first['label_ids'], torch.Tensor):
+            batch['labels'] = torch.stack([f['label_ids'] for f in features])
+        else:
+            dtype = torch.long if isinstance(first['label_ids'][0], int) else torch.float
+            batch['labels'] = torch.tensor([f['label_ids'] for f in features], dtype=dtype)
+
+    # Handling of all other possible keys.
+    # Again, we will use the first element to figure out which key/values are not None for this model.
+    for k, v in first.items():
+        if k not in ('label', 'label_ids') and v is not None and not isinstance(v, str):
+            if isinstance(v, torch.Tensor):
+                batch[k] = torch.stack([f[k] for f in features])
+            elif isinstance(v, np.ndarray):
+                batch[k] = torch.tensor(np.stack([f[k] for f in features]))
+            else:
+                batch[k] = torch.tensor([f[k] for f in features])
+
+    return batch

internvl_chat/internvl/patch/train_sampler_patch.py

@@ -0,0 +1,31 @@
+from typing import Optional
+
+import torch
+import transformers
+from transformers.trainer import (LengthGroupedSampler, RandomSampler,
+                                  has_length)
+
+
+# patch trainer
+def _get_train_sampler(self) -> Optional[torch.utils.data.Sampler]:
+    if self.train_dataset is None or not has_length(self.train_dataset):
+        return None
+    # Build the sampler.
+    if self.args.group_by_length:
+        lengths = []
+        for dataset in self.train_dataset.datasets:
+            lengths = lengths + dataset.length
+        model_input_name = self.tokenizer.model_input_names[0] if self.tokenizer is not None else None
+        return LengthGroupedSampler(
+            self.args.train_batch_size * self.args.gradient_accumulation_steps,
+            dataset=self.train_dataset,
+            lengths=lengths,
+            model_input_name=model_input_name,
+        )
+    else:
+        return RandomSampler(self.train_dataset)
+
+
+def replace_train_sampler():
+    transformers.Trainer._get_train_sampler = _get_train_sampler
+    print('Replace train sampler!!')

internvl_chat/internvl/train/internvl_chat_finetune.py

@@ -23,8 +23,10 @@
                                           InternVisionModel,
                                           InternVLChatConfig,
                                           InternVLChatModel)
-from internvl.patch import (replace_llama2_attn_with_flash_attn,
-                            replace_llama_rmsnorm_with_fused_rmsnorm)
+from internvl.patch import (pad_data_collator,
+                            replace_llama2_attn_with_flash_attn,
+                            replace_llama_rmsnorm_with_fused_rmsnorm,
+                            replace_train_sampler)
 from internvl.train.dataset import (TCSLoader, WeightedConcatDataset,
                                     build_transform)
 from PIL import Image, ImageFile, PngImagePlugin
@@ -39,6 +41,7 @@
 # Upgrade transformers to v4.36.2, we don't need it anymore
 # replace_llama2_attn_with_flash_attn()
 replace_llama_rmsnorm_with_fused_rmsnorm()
+replace_train_sampler()
 
 try:
     from petrel_client.client import Client
@@ -182,6 +185,7 @@ def preprocess(
         tokenizer: transformers.PreTrainedTokenizer,
         num_image_token: int,
         text_only: bool = False,
+        group_by_length: bool = False,
 ) -> Dict:
     conv = get_conv_template(template_name)
     roles = {'human': conv.roles[0], 'gpt': conv.roles[1]}
@@ -213,7 +217,7 @@ def preprocess(
     input_ids = tokenizer(
         conversations,
         return_tensors='pt',
-        padding='max_length',
+        padding=False if group_by_length else 'max_length',
         max_length=tokenizer.model_max_length,
         truncation=True,
     ).input_ids
@@ -283,6 +287,7 @@ def preprocess_mpt(
         tokenizer: transformers.PreTrainedTokenizer,
         num_image_token: int,
         text_only: bool = False,
+        group_by_length: bool = False,
 ) -> Dict:
     conv = get_conv_template(template_name)
     roles = {'human': conv.roles[0], 'gpt': conv.roles[1]}
@@ -314,7 +319,7 @@ def preprocess_mpt(
     input_ids = tokenizer(
         conversations,
         return_tensors='pt',
-        padding='max_length',
+        padding=False if group_by_length else 'max_length',
         max_length=tokenizer.model_max_length,
         truncation=True,
     ).input_ids
@@ -368,7 +373,7 @@ class LazySupervisedDataset(Dataset):
     """Dataset for supervised fine-tuning."""
 
     def __init__(self, template_name, meta, tokenizer, tcs_loader, num_image_token,
-                 image_size=224, is_train=True, pad2square=False):
+                 image_size=224, is_train=True, pad2square=False, group_by_length=False):
         super(LazySupervisedDataset, self).__init__()
         self.tokenizer = tokenizer
         self.template_name = template_name
@@ -384,6 +389,21 @@ def __init__(self, template_name, meta, tokenizer, tcs_loader, num_image_token,
         self.root = meta['root']
         self.cached_data_dict = {}
         self.tcs_loader = tcs_loader
+        self.group_by_length = group_by_length
+        if self.group_by_length:
+            self.conv2length = {}
+            self.length = []
+            for data_item in self.raw_data:
+                conversations = ''.join(data_item.split('conversations')[1:])
+                str_length = len(conversations)
+                if str_length not in self.conv2length:
+                    token_length = tokenizer(
+                        conversations, return_tensors='pt', padding=False, truncation=False,
+                    ).input_ids.size(1)
+                    self.conv2length[str_length] = token_length
+                else:
+                    token_length = self.conv2length[str_length]
+                self.length.append(token_length)
 
     def __len__(self):
         return len(self.raw_data)
@@ -405,7 +425,7 @@ def multi_modal_get_item(self, data_item):
         else:
             preprocess_function = preprocess
         ret = preprocess_function(self.template_name, [deepcopy(data_item['conversations'])],
-                                  self.tokenizer, self.num_image_token)
+                                  self.tokenizer, self.num_image_token, group_by_length=self.group_by_length)
         ret = dict(
             input_ids=ret['input_ids'][0],
             labels=ret['labels'][0],
@@ -425,7 +445,8 @@ def pure_text_get_item(self, data_item):
         else:
             preprocess_function = preprocess
         ret = preprocess_function(self.template_name, [deepcopy(data_item['conversations'])],
-                                  self.tokenizer, self.num_image_token, text_only=True)
+                                  self.tokenizer, self.num_image_token, text_only=True,
+                                  group_by_length=self.group_by_length)
         ret = dict(
             input_ids=ret['input_ids'][0],
             labels=ret['labels'][0],
@@ -455,7 +476,7 @@ def __getitem__(self, i) -> Dict[str, torch.Tensor]:
         return ret
 
 
-def build_datasets(data_args, tokenizer, tcs_loader, model):
+def build_datasets(data_args, tokenizer, tcs_loader, model, group_by_length=False):
     datasets = []
     lengths = []
     ds_collections = json.loads(open(data_args.meta_path).read())
@@ -469,7 +490,8 @@ def build_datasets(data_args, tokenizer, tcs_loader, model):
                 num_image_token=model.num_image_token,
                 image_size=data_args.force_image_size,
                 is_train=ds_collections[ds_name]['data_augment'],
-                pad2square=data_args.pad2square
+                pad2square=data_args.pad2square,
+                group_by_length=group_by_length
             )
         except Exception:
             logger.info(f'Error in loading dataset: {ds_name}')
@@ -623,7 +645,8 @@ def main():
     if model_args.grad_checkpoint:
         model.language_model._set_gradient_checkpointing()
 
-    train_dataset = build_datasets(data_args, tokenizer, tcs_loader, model)
+    train_dataset = build_datasets(data_args, tokenizer, tcs_loader, model,
+                                   group_by_length=training_args.group_by_length)
 
     def _freeze_params(module):
         for param in module.parameters():
@@ -672,7 +695,7 @@ def _freeze_params(module):
         train_dataset=train_dataset if training_args.do_train else None,
         eval_dataset=None,
         tokenizer=tokenizer,
-        data_collator=default_data_collator,
+        data_collator=default_data_collator if not training_args.group_by_length else pad_data_collator,
     )
 
     # Training

internvl_chat/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "internvl_chat"
-version = "1.2.0"
+version = "1.2.1"
 description = "Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks."
 readme = "README.md"
 requires-python = ">=3.8"

internvl_chat/tools/json2jsonl.py

@@ -11,10 +11,11 @@
 
 data = json.load(open(args.path))
 writer = open(args.path.replace('.json', '.jsonl'), 'w')
-for item in data:
+for idx, item in enumerate(data):
     conversations = item['conversations']
     if conversations[0]['from'] == 'system':
         item['conversations'] = item['conversations'][1:]
+    item['id'] = idx
     writer.write(json.dumps(item, ensure_ascii=False) + '\n')
 
 writer.close()