[Model] Support for PaliGemma model (open-compass#202)

* Support for PaliGemma model

* Fix message_to_promptimg

* update

README.md

@@ -25,6 +25,7 @@ English | [<a href="README_zh-CN.md">简体中文</a>]
 
 ## 🆕 News
 
+- **[2024-05-15]** We have supported [**PaliGemma-3B**](https://huggingface.co/google/paligemma-3b-pt-448), a versatile and lightweight vision-language model released by Google 🔥🔥🔥
 - **[2024-05-14]** We have supported [**GPT-4o**](https://openai.com/index/hello-gpt-4o/) 🔥🔥🔥
 - **[2024-05-07]** We have supported [**XVERSE-V-13B**](https://github.com/xverse-ai/XVERSE-V-13B/blob/main/vxverse/models/vxverse.py), thanks to [**YJY123**](https://github.com/YJY123) 🔥🔥🔥
 - **[2024-05-06]** We have launched a discord channel for VLMEvalKit users: https://discord.gg/evDT4GZmxN. Latest updates and discussion will be posted here
@@ -34,7 +35,6 @@ English | [<a href="README_zh-CN.md">简体中文</a>]
 - **[2024-04-25]** We have supported [**Reka API**](https://www.reka.ai), the API model ranked first in [**Vision-Arena**](https://huggingface.co/spaces/WildVision/vision-arena) 🔥🔥🔥
 - **[2024-04-21]** We have noticed a minor issue with the MathVista evaluation script (which may negatively affect the performance). We have fixed it and updated the leaderboard accordingly
 - **[2024-04-17]** We have supported [**InternVL-Chat-V1.5**](https://github.com/OpenGVLab/InternVL/) 🔥🔥🔥
-- **[2024-04-15]** We have supported [**RealWorldQA**](https://x.ai/blog/grok-1.5v), a multimodal benchmark for real-world spatial understanding 🔥🔥🔥
 
 ## 📊 Datasets, Models, and Evaluation Results
 
@@ -84,6 +84,7 @@ VLMEvalKit will use an **judge LLM** to extract answer from the output if you se
 | [**Monkey**](https://github.com/Yuliang-Liu/Monkey)🚅         | [**EMU2-Chat**](https://github.com/baaivision/Emu)🚅🎞️         | [**Yi-VL-[6B/34B]**](https://huggingface.co/01-ai/Yi-VL-6B)  | [**MMAlaya**](https://huggingface.co/DataCanvas/MMAlaya)🚅    |
 | [**InternLM-XComposer2-[1.8B/7B]**](https://huggingface.co/internlm/internlm-xcomposer2-vl-7b)🚅🎞️ | [**MiniCPM-[V1/V2]**](https://huggingface.co/openbmb/MiniCPM-V)🚅 | [**OmniLMM-12B**](https://huggingface.co/openbmb/OmniLMM-12B) | [**InternVL-Chat Series**](https://github.com/OpenGVLab/InternVL)🚅 |
 | [**DeepSeek-VL**](https://github.com/deepseek-ai/DeepSeek-VL/tree/main)🎞️ | [**LLaVA-NeXT**](https://llava-vl.github.io/blog/2024-01-30-llava-next/)🚅 | [**Bunny-Llama3**](https://huggingface.co/BAAI/Bunny-Llama-3-8B-V)🚅 | [**XVERSE-V-13B**](https://github.com/xverse-ai/XVERSE-V-13B/blob/main/vxverse/models/vxverse.py) |
+| [**PaliGemma-3B**](https://huggingface.co/google/paligemma-3b-pt-448) 🚅 | | |
 
 🎞️: Support multiple images as inputs.
 

README_zh-CN.md

@@ -23,7 +23,8 @@
 
 ## 🆕 更新
 
-- **[2024-05-07]** 支持了 [**GPT-4o**](https://openai.com/index/hello-gpt-4o/) 🔥🔥🔥
+- **[2024-05-15]** 支持了 [**PaliGemma-3B**](https://huggingface.co/google/paligemma-3b-pt-448), 一个谷歌开源的 3B 多模态模型。 🔥🔥🔥
+- **[2024-05-14]** 支持了 [**GPT-4o**](https://openai.com/index/hello-gpt-4o/) 🔥🔥🔥
 - **[2024-05-07]** 支持了 [**XVERSE-V-13B**](https://github.com/xverse-ai/XVERSE-V-13B/blob/main/vxverse/models/vxverse.py), 感谢 [**YJY123**](https://github.com/YJY123) 🔥🔥🔥
 - **[2024-05-06]** 成立了 VLMEvalKit 用户群组的 Discord 频道: https://discord.gg/evDT4GZmxN，将在这里分享关于 VLMEvalKit 的更新并进行讨论
 - **[2024-05-06]** 支持了两个基于 Llama3 的 VLM 🔥🔥🔥: Bunny-llama3-8B (SigLIP, 输入图像大小 384) 和 llava-llama-3-8b (CLIP-L, 输入图像大小 336), 用户可在我们支持的数十个测试基准上测试这两个模型
@@ -32,7 +33,6 @@
 - **[2024-04-25]** 支持了 [**Reka**](https://www.reka.ai), 这个 API 模型在 [**Vision-Arena**](https://huggingface.co/spaces/WildVision/vision-arena) 排名第一 🔥🔥🔥
 - **[2024-04-21]** 修复了 MathVista 评估脚本的一个小问题（可能会对性能产生较小的负面影响），并相应更新了排行榜
 - **[2024-04-17]** 支持 [**InternVL-Chat-V1.5**](https://github.com/OpenGVLab/InternVL/) 🔥🔥🔥
-- **[2024-04-15]** 支持 [**RealWorldQA**](https://x.ai/blog/grok-1.5v)， 这是一个用于真实世界空间理解的多模态基准测试  🔥🔥🔥
 
 ## 📊 评测结果，支持的数据集和模型 <a id="data-model-results"></a>
 ### 评测结果
@@ -82,6 +82,7 @@
 | [**Monkey**](https://github.com/Yuliang-Liu/Monkey)🚅         | [**EMU2-Chat**](https://github.com/baaivision/Emu)🚅🎞️         | [**Yi-VL-[6B/34B]**](https://huggingface.co/01-ai/Yi-VL-6B)  | [**MMAlaya**](https://huggingface.co/DataCanvas/MMAlaya)🚅    |
 | [**InternLM-XComposer2-[1.8B/7B]**](https://huggingface.co/internlm/internlm-xcomposer2-vl-7b)🚅🎞️ | [**MiniCPM-[V1/V2]**](https://huggingface.co/openbmb/MiniCPM-V)🚅 | [**OmniLMM-12B**](https://huggingface.co/openbmb/OmniLMM-12B) | [**InternVL-Chat Series**](https://github.com/OpenGVLab/InternVL)🚅 |
 | [**DeepSeek-VL**](https://github.com/deepseek-ai/DeepSeek-VL/tree/main)🎞️ | [**LLaVA-NeXT**](https://llava-vl.github.io/blog/2024-01-30-llava-next/)🚅 | [**Bunny-Llama3**](https://huggingface.co/BAAI/Bunny-Llama-3-8B-V)🚅 | [**XVERSE-V-13B**](https://github.com/xverse-ai/XVERSE-V-13B/blob/main/vxverse/models/vxverse.py) |
+| [**PaliGemma-3B**](https://huggingface.co/google/paligemma-3b-pt-448) 🚅 | | |
 
 🎞️ 表示支持多图片输入。
 

vlmeval/config.py

@@ -27,6 +27,7 @@
     'MGM_7B':partial(Mini_Gemini, model_path='YanweiLi/MGM-7B-HD', root=Mini_Gemini_ROOT),
     'Bunny-llama3-8B': partial(BunnyLLama3, model_path='BAAI/Bunny-Llama-3-8B-V'),
     'VXVERSE':partial(VXVERSE, model_name='XVERSE-V-13B', root=VXVERSE_ROOT),
+    'paligemma-3b-mix-448': partial(PaliGemma, model_path='google/paligemma-3b-mix-448'),
 }
 
 api_models = {

vlmeval/vlm/__init__.py

@@ -26,3 +26,4 @@
 from .mgm import Mini_Gemini
 from .bunnyllama3 import BunnyLLama3
 from .vxverse import VXVERSE
+from .paligemma import PaliGemma

vlmeval/vlm/base.py

@@ -144,10 +144,7 @@ def message_to_promptimg(self, message):
         if num_images == 0:
             prompt = '\n'.join([x['value'] for x in message if x['type'] == 'text'])
             image = None
-        elif num_images == 1:
-            prompt = '\n'.join([x['value'] for x in message if x['type'] == 'text'])
-            image = [x['value'] for x in message if x['type'] == 'image'][0]
         else:
-            prompt = '\n'.join([x['value'] if x['type'] == 'text' else '<image>' for x in message])
+            prompt = '\n'.join([x['value'] for x in message if x['type'] == 'text'])
             image = [x['value'] for x in message if x['type'] == 'image'][0]
         return prompt, image

vlmeval/vlm/paligemma.py

@@ -0,0 +1,38 @@
+from PIL import Image
+import torch
+
+from .base import BaseModel
+from ..smp import *
+
+
+class PaliGemma(BaseModel):
+    INSTALL_REQ = False
+    INTERLEAVE = False
+
+    def __init__(self, model_path='google/paligemma-3b-mix-448', **kwargs):
+        try:
+            from transformers import AutoProcessor, PaliGemmaForConditionalGeneration
+        except:
+            warnings.warn('Please install the latest version transformers.')
+            sys.exit(-1)
+        self.model = PaliGemmaForConditionalGeneration.from_pretrained(
+            model_path,
+            torch_dtype=torch.bfloat16,
+            device_map='auto',
+            revision='bfloat16',
+        ).eval()
+        self.processor = AutoProcessor.from_pretrained(model_path)
+        self.kwargs = kwargs
+
+    def generate_inner(self, message, dataset=None):
+        prompt, image_path = self.message_to_promptimg(message)
+        image = Image.open(image_path).convert('RGB')
+
+        model_inputs = self.processor(text=prompt, images=image, return_tensors='pt').to(self.model.device)
+        input_len = model_inputs['input_ids'].shape[-1]
+
+        with torch.inference_mode():
+            generation = self.model.generate(**model_inputs, max_new_tokens=512, do_sample=False)
+            generation = generation[0][input_len:]
+            res = self.processor.decode(generation, skip_special_tokens=True)
+        return res