[Refactor] Refactor instructblip (#227)

* refactor instructblip

* add post processor

* add forward

* fix lint

* update

* update

configs/multimodal/instructblip/README.md

@@ -7,3 +7,43 @@ git clone https://github.com/salesforce/LAVIS.git
 cd ./LAVIS
 pip install -e .
 ```
+
+### Modify the config
+
+Modify the config of InstructBlip, like model path of LLM and Qformer.
+
+Then update `tasks.py` like the following code snippet.
+
+```python
+from mmengine.config import read_base
+
+with read_base():
+    from .instructblip.instructblip_mmbench import (instruct_blip_dataloader,
+                                                    instruct_blip_evaluator,
+                                                    instruct_blip_load_from,
+                                                    instruct_blip_model)
+
+models = [instruct_blip_model]
+datasets = [instruct_blip_dataloader]
+evaluators = [instruct_blip_evaluator]
+load_froms = [instruct_blip_load_from]
+num_gpus = 8
+num_procs = 8
+launcher = 'pytorch'  # or 'slurm'
+```
+
+### Start evaluation
+
+#### Slurm
+
+```sh
+cd $root
+python run.py configs/multimodal/tasks.py --mm-eval --slurm -p $PARTITION
+```
+
+#### PyTorch
+
+```sh
+cd $root
+python run.py configs/multimodal/tasks.py --mm-eval 
+```

configs/multimodal/instructblip/instructblip_mmbench.py

@@ -1,3 +1,6 @@
+from opencompass.multimodal.models.instructblip import (
+    InstructBlipMMBenchPromptConstructor, InstructBlipMMBenchPostProcessor)
+
 # dataloader settings
 val_pipeline = [
     dict(type='mmpretrain.torchvision/Resize',
@@ -9,24 +12,27 @@ val_pipeline = [
          std=(0.26862954, 0.26130258, 0.27577711)),
     dict(type='mmpretrain.PackInputs',
          algorithm_keys=[
-             'question', 'category', 'l2-category', 'context',
+             'question', 'category', 'l2-category', 'context', 'index',
-             'index', 'options_dict', 'options', 'split'
+             'options_dict', 'options', 'split'
          ])
 ]
 
-dataset = dict(type='opencompass.MMBench',
+dataset = dict(type='opencompass.MMBenchDataset',
                data_file='data/mmbench/mmbench_test_20230712.tsv',
                pipeline=val_pipeline)
 
-dataloader = dict(batch_size=1,
+instruct_blip_dataloader = dict(batch_size=1,
-                  num_workers=4,
+                                num_workers=4,
-                  dataset=dataset,
+                                dataset=dataset,
-                  collate_fn=dict(type='pseudo_collate'),
+                                collate_fn=dict(type='pseudo_collate'),
-                  sampler=dict(type='DefaultSampler', shuffle=False))
+                                sampler=dict(type='DefaultSampler',
+                                             shuffle=False))
 
 # model settings
-model = dict(
+instruct_blip_model = dict(
-    type='blip2-vicuna-instruct-mmbench',
+    type='blip2-vicuna-instruct',
+    prompt_constructor=dict(type=InstructBlipMMBenchPromptConstructor),
+    post_processor=dict(type=InstructBlipMMBenchPostProcessor),
     freeze_vit=True,
     low_resource=False,
     llm_model='/path/to/vicuna-7b/',
@@ -35,11 +41,11 @@ model = dict(
 )
 
 # evaluation settings
-evaluator = [
+instruct_blip_evaluator = [
     dict(
         type='opencompass.DumpResults',
         save_path=  # noqa: E251
         'work_dirs/instructblip_vicuna7b/instructblipvicuna_mmbench.xlsx')
 ]
 
-load_from = '/path/to/instruct_blip_vicuna7b_trimmed.pth'  # noqa
+instruct_blip_load_from = '/path/to/instruct_blip_vicuna7b_trimmed'

opencompass/multimodal/models/instructblip/__init__.py

@@ -1,3 +1,8 @@
-from .blip2_vicuna_instruct import Blip2VicunaInstructMMBench
+from .blip2_vicuna_instruct import InstructBlipInferencer
+from .post_processor import InstructBlipMMBenchPostProcessor
+from .prompt_constructor import InstructBlipMMBenchPromptConstructor
 
-__all__ = ['Blip2VicunaInstructMMBench']
+__all__ = [
+    'InstructBlipInferencer', 'InstructBlipMMBenchPromptConstructor',
+    'InstructBlipMMBenchPostProcessor'
+]

opencompass/multimodal/models/instructblip/blip2_vicuna_instruct.py

@@ -1,8 +1,8 @@
 """Requires Transformer 4.28 and above, implementation may change according the
 Llama implementation."""
 import logging
-import re
 
+import mmengine
 import torch
 import torch.nn as nn
 from lavis.models.blip2_models.blip2 import Blip2Base, disabled_train
@@ -12,27 +12,36 @@ from transformers import LlamaForCausalLM, LlamaTokenizer
 from opencompass.registry import MM_MODELS
 
 
-@MM_MODELS.register_module('blip2-vicuna-instruct-mmbench')
+@MM_MODELS.register_module('blip2-vicuna-instruct')
-class Blip2VicunaInstructMMBench(Blip2Base):
+class InstructBlipInferencer(Blip2Base):
 
     def __init__(
         self,
-        vit_model='eva_clip_g',
+        prompt_constructor: dict,
-        img_size=224,
+        post_processor: dict,
-        drop_path_rate=0,
+        vit_model: str = 'eva_clip_g',
-        use_grad_checkpoint=False,
+        img_size: int = 224,
-        vit_precision='fp16',
+        drop_path_rate: float = 0,
-        freeze_vit=True,
+        use_grad_checkpoint: bool = False,
-        num_query_token=32,
+        vit_precision: str = 'fp16',
-        llm_model='',
+        freeze_vit: bool = True,
-        sys_prompt='',
+        num_query_token: int = 32,
-        prompt='',
+        llm_model: str = '',
-        max_txt_len=128,
+        sys_prompt: str = '',
-        max_output_txt_len=256,
+        prompt: str = '',
-        qformer_text_input=True,
+        max_txt_len: int = 128,
-        low_resource=False,
+        max_output_txt_len: int = 256,
+        qformer_text_input: bool = True,
+        low_resource: bool = False,
+        mode: str = 'generation',
     ):
         super().__init__()
+        self.mode = mode
+        self.prompt_constructor = mmengine.registry.build_from_cfg(
+            prompt_constructor, MM_MODELS)
+        self.post_processor = mmengine.registry.build_from_cfg(
+            post_processor, MM_MODELS)
+
         self.tokenizer = self.init_tokenizer(truncation_side='left')
 
         self.visual_encoder, self.ln_vision = self.init_vision_encoder(
@@ -92,6 +101,12 @@ class Blip2VicunaInstructMMBench(Blip2Base):
 
         self.qformer_text_input = qformer_text_input
 
+    def forward(self, batch):
+        if self.mode == 'generation':
+            return self.generate(batch)
+        else:
+            raise RuntimeError(f'Invalid mode "{self.mode}".')
+
     def concat_text_input_output(self, input_ids, input_atts, output_ids,
                                  output_atts):
         input_part_targets_len = []
@@ -136,31 +151,13 @@ class Blip2VicunaInstructMMBench(Blip2Base):
         temperature=1,
     ):
         inputs = self.pack_inputs(batch)
-        image = inputs.pop('image')
+        inputs = self.prompt_constructor(inputs)
+        image = inputs['image']
+        prompt = inputs['prompt']
         data_samples = inputs['data_samples']
-        samples = {'image': image}
-        questions = [
-            data_sample.get('question') for data_sample in data_samples
-        ]
-        options = [data_sample.get('options') for data_sample in data_samples]
-        if data_samples[0].get('context') is not None:
-            contexts = [
-                data_sample.get('context') for data_sample in data_samples
-            ]
-            prompt = [
-                context + ' ' + question + ' ' + option for context, question,
-                option in zip(contexts, questions, options)
-            ]
-        else:
-            prompt = [
-                question + ' ' + option
-                for question, option in zip(questions, options)
-            ]
 
         self.llm_tokenizer.padding_side = 'left'
 
-        image = samples['image']
-
         bs = image.size(0)
 
         if isinstance(prompt, str):
@@ -237,24 +234,10 @@ class Blip2VicunaInstructMMBench(Blip2Base):
                 length_penalty=length_penalty,
                 num_return_sequences=num_captions,
             )
-        outputs[outputs == 0] = 2  # convert output id 0 to 2 (eos_token_id)
-        output_text = self.llm_tokenizer.batch_decode(outputs,
-                                                      skip_special_tokens=True)
-        output_text = [text.strip() for text in output_text]
-        output_text = self.post_process(output_text[0])
-        data_sample = data_samples[0]
-        data_sample.pred_answer = output_text
 
-        return data_sample
+        for i, data_sample in enumerate(data_samples):
-
+            output_token = outputs[i]
-    def post_process(self, output_text):
+            output_text = self.post_processor(output_token, self.llm_tokenizer)
-        output_text = output_text.split('###')[0]
+            data_sample.pred_answer = output_text
-        output_text = output_text.split('Assistant:')[-1].strip()
+            data_samples[i] = data_sample
-        output_text = output_text.strip('</s><s>')
+        return data_samples
-        output_text = output_text.strip('</Img>')
-        output_text = output_text.strip()
-        pattern = re.compile(r'([A-Z]\.)')
-        res = pattern.findall(output_text)
-        if len(res) > 0:
-            output_text = res[0][:-1]
-        return output_text

opencompass/multimodal/models/instructblip/post_processor.py

@@ -0,0 +1,31 @@
+import re
+
+import torch
+
+
+class InstructBlipMMBenchPostProcessor:
+    """"Post processor for MiniGPT-4 on MMBench."""
+
+    def __init__(self) -> None:
+        pass
+
+    def __call__(self, output_token: torch.tensor, tokenizer) -> str:
+        # convert output id 0 to 2 (eos_token_id)
+        output_token[output_token == 0] = 2
+        output_text = tokenizer.decode(output_token,
+                                       add_special_tokens=False)  # noqa
+        output_text = self._extract_key_words(output_text.strip())
+        return output_text
+
+    def _extract_key_words(self, output_text: str) -> str:
+
+        output_text = output_text.split('###')[0]
+        output_text = output_text.split('Assistant:')[-1].strip()
+        output_text = output_text.strip('</s><s>')
+        output_text = output_text.strip('</Img>')
+        output_text = output_text.strip()
+        pattern = re.compile(r'([A-Z]\.)')
+        res = pattern.findall(output_text)
+        if len(res) > 0:
+            output_text = res[0][:-1]
+        return output_text

opencompass/multimodal/models/instructblip/prompt_constructor.py

@@ -0,0 +1,55 @@
+from typing import List
+
+from mmpretrain.structures import DataSample
+
+
+class InstructBlipMMBenchPromptConstructor:
+    """Prompt constructor for InstructBlip on MMBench.
+
+    Args:
+        image_prompt (str): Image prompt.
+        reply_prompt (str): Reply prompt.
+    """
+
+    def __init__(self, image_prompt: str = '', reply_prompt: str = '') -> None:
+        self.image_prompt = image_prompt
+        self.reply_prompt = reply_prompt
+
+    def __call__(self, inputs: dict) -> dict:
+        """Construct prompt.
+
+        Args:
+            inputs (dict): Input data containing image and data_samples.
+
+        Returns:
+            dict: A dict containing prompt, images and data_samples.
+        """
+        data_samples = inputs['data_samples']
+        prompt = self._process(data_samples)
+        inputs.update({'prompt': prompt})
+
+        return inputs
+
+    def _process(self, data_samples: List[DataSample]) -> str:
+        """Process data sample to prompt.
+
+        Args:
+            data_samples (List[DataSample]): A list of data_samples.
+
+        Returns:
+            str: Prompt.
+        """
+        assert len(data_samples) == 1, 'Only support batch size 1.'
+        questions = [
+            data_sample.get('question') for data_sample in data_samples
+        ]
+        options = [data_sample.get('options') for data_sample in data_samples]
+        contexts = [data_sample.get('context') for data_sample in data_samples]
+        question = questions[0]
+        option = options[0]
+        context = contexts[0]
+        if context is not None:
+            prompt = self.image_prompt + ' ' + context + ' ' + question + ' ' + option + ' ' + self.reply_prompt  # noqa
+        else:
+            prompt = self.image_prompt + ' ' + question + ' ' + option + ' ' + self.reply_prompt  # noqa
+        return prompt