[Feature] Support otter (#1651)

* [Feature] Support Otter

* Update docs

README.md

@@ -256,6 +256,7 @@ Results and models are available in the [model zoo](https://mmpretrain.readthedo
         <li><a href="configs/flamingo">Flamingo (NeurIPS'2022)</a></li>
         <li><a href="configs/chinese_clip">Chinese CLIP (arxiv'2022)</a></li>
         <li><a href="configs/minigpt4">MiniGPT-4 (arxiv'2023)</a></li>
+        <li><a href="configs/otter">Otter (arxiv'2023)</a></li>
         </ul>
       </td>
       <td>

README_zh-CN.md

@@ -252,6 +252,7 @@ mim install -e ".[multimodal]"
         <li><a href="configs/flamingo">Flamingo (NeurIPS'2022)</a></li>
         <li><a href="configs/chinese_clip">Chinese CLIP (arxiv'2022)</a></li>
         <li><a href="configs/minigpt4">MiniGPT-4 (arxiv'2023)</a></li>
+        <li><a href="configs/otter">Otter (arxiv'2023)</a></li>
         </ul>
       </td>
       <td>

configs/otter/README.md

@@ -0,0 +1,78 @@
+# Otter
+
+> [Otter: A Multi-Modal Model with In-Context Instruction Tuning](https://arxiv.org/abs/2305.03726)
+
+<!-- [ALGORITHM] -->
+
+## Abstract
+
+Large language models (LLMs) have demonstrated significant universal capabilities as few/zero-shot learners in various tasks due to their pre-training on vast amounts of text data, as exemplified by GPT-3, which boosted to InstrctGPT and ChatGPT, effectively following natural language instructions to accomplish real-world tasks. In this paper, we propose to introduce instruction tuning into multi-modal models, motivated by the Flamingo model's upstream interleaved format pretraining dataset. We adopt a similar approach to construct our MultI-Modal In-Context Instruction Tuning (MIMIC-IT) dataset. We then introduce Otter, a multi-modal model based on OpenFlamingo (open-sourced version of DeepMind's Flamingo), trained on MIMIC-IT and showcasing improved instruction-following ability and in-context learning. We also optimize OpenFlamingo's implementation for researchers, democratizing the required training resources from 1$\times$ A100 GPU to 4$\times$ RTX-3090 GPUs, and integrate both OpenFlamingo and Otter into Huggingface Transformers for more researchers to incorporate the models into their customized training and inference pipelines.
+
+<div align=center>
+<img src="https://camo.githubusercontent.com/70613ab882a7827808148a2c577029d544371e707b0832a0b01151c54ce553c3/68747470733a2f2f692e706f7374696d672e63632f5477315a304243572f6f7474657276302d322d64656d6f2e706e67" width="80%"/>
+</div>
+
+## How to use it?
+
+<!-- [TABS-BEGIN] -->
+
+**Use the model**
+
+```python
+import torch
+from mmpretrain import get_model, inference_model
+
+model = get_model('otter-9b_3rdparty_caption', pretrained=True, device='cuda')
+out = inference_model(model, 'demo/cat-dog.png')
+print(out)
+```
+
+**Test Command**
+
+Prepare your dataset according to the [docs](https://mmpretrain.readthedocs.io/en/latest/user_guides/dataset_prepare.html#prepare-dataset).
+
+Test:
+
+```shell
+python tools/test.py configs/otter/otter-9b_caption.py https://download.openmmlab.com/mmclassification/v1/otter/otter-9b-adapter_20230613-51c5be8d.pth
+```
+
+<!-- [TABS-END] -->
+
+## Models and results
+
+### Image Caption on COCO
+
+| Model                         |   Pretrain   | Params (M) |  BLEU-4  |  CIDER   |            Config             |                                                 Download                                                 |
+| :---------------------------- | :----------: | :--------: | :------: | :------: | :---------------------------: | :------------------------------------------------------------------------------------------------------: |
+| `otter-9b_3rdparty_caption`\* | From scratch |  8220.45   | Upcoming | Upcoming | [config](otter-9b_caption.py) | [model](https://download.openmmlab.com/mmclassification/v1/otter/otter-9b-adapter_20230613-51c5be8d.pth) |
+
+*Models with * are converted from the [official repo](https://github.com/Luodian/Otter/tree/main). The config files of these models are only for inference. We haven't reprodcue the training results.*
+
+### Visual Question Answering on VQAv2
+
+| Model                     |   Pretrain   | Params (M) | Accuracy |          Config           |                                                 Download                                                 |
+| :------------------------ | :----------: | :--------: | :------: | :-----------------------: | :------------------------------------------------------------------------------------------------------: |
+| `otter-9b_3rdparty_vqa`\* | From scratch |  8220.45   | Upcoming | [config](otter-9b_vqa.py) | [model](https://download.openmmlab.com/mmclassification/v1/otter/otter-9b-adapter_20230613-51c5be8d.pth) |
+
+*Models with * are converted from the [official repo](https://github.com/Luodian/Otter/tree/main). The config files of these models are only for inference. We haven't reprodcue the training results.*
+
+## Citation
+
+```bibtex
+@article{li2023otter,
+  title={Otter: A Multi-Modal Model with In-Context Instruction Tuning},
+  author={Li, Bo and Zhang, Yuanhan and Chen, Liangyu and Wang, Jinghao and Yang, Jingkang and Liu, Ziwei},
+  journal={arXiv preprint arXiv:2305.03726},
+  year={2023}
+}
+
+@article{li2023mimicit,
+    title={MIMIC-IT: Multi-Modal In-Context Instruction Tuning},
+    author={Bo Li and Yuanhan Zhang and Liangyu Chen and Jinghao Wang and Fanyi Pu and Jingkang Yang and Chunyuan Li and Ziwei Liu},
+    year={2023},
+    eprint={2306.05425},
+    archivePrefix={arXiv},
+    primaryClass={cs.CV}
+}
+```

configs/otter/metafile.yml

@@ -0,0 +1,43 @@
+Collections:
+  - Name: Otter
+    Metadata:
+      Architecture:
+        - Transformer
+        - Gated Cross-Attention Dense
+    Paper:
+      Title: 'Otter: A Multi-Modal Model with In-Context Instruction Tuning'
+      URL: https://arxiv.org/abs/2305.03726
+    README: configs/otter/README.md
+
+Models:
+  - Name: otter-9b_3rdparty_caption
+    Metadata:
+      FLOPs: null
+      Parameters: 8220452880
+    In Collection: Otter
+    Results:
+      - Task: Image Caption
+        Dataset: COCO
+        Metrics:
+          BLEU-4: null
+          CIDER: null
+    Weights: https://download.openmmlab.com/mmclassification/v1/otter/otter-9b-adapter_20230613-51c5be8d.pth
+    Config: configs/otter/otter-9b_caption.py
+    Converted From:
+      Weights: https://huggingface.co/luodian/otter-9b-hf
+      Code: https://github.com/Luodian/Otter/tree/main
+  - Name: otter-9b_3rdparty_vqa
+    Metadata:
+      FLOPs: null
+      Parameters: 8220452880
+    In Collection: Otter
+    Results:
+      - Task: Visual Question Answering
+        Dataset: VQAv2
+        Metrics:
+          Accuracy: null
+    Weights: https://download.openmmlab.com/mmclassification/v1/otter/otter-9b-adapter_20230613-51c5be8d.pth
+    Config: configs/otter/otter-9b_vqa.py
+    Converted From:
+      Weights: https://huggingface.co/luodian/otter-9b-hf
+      Code: https://github.com/Luodian/Otter/tree/main

configs/otter/otter-9b_caption.py

@@ -0,0 +1,91 @@
+_base_ = [
+    '../_base_/default_runtime.py',
+]
+
+# model settings
+model = dict(
+    type='Otter',
+    tokenizer=dict(type='LlamaTokenizer', name_or_path='huggyllama/llama-7b'),
+    vision_encoder=dict(
+        type='VisionTransformer',
+        arch='l',
+        patch_size=14,
+        pre_norm=True,
+        norm_cfg=dict(type='LN', eps=1e-5),
+        layer_cfgs=dict(act_cfg=dict(type='mmpretrain.QuickGELU')),
+        final_norm=False,
+        out_type='raw',
+        pretrained=(
+            'https://download.openmmlab.com/mmclassification/v0/clip/'
+            'vit-large-p14_clip-openai-pre_3rdparty_20230517-95e2af0b.pth'),
+    ),
+    lang_encoder=dict(
+        base=dict(
+            type='AutoModelForCausalLM',
+            name_or_path='huggyllama/llama-7b',
+            local_files_only=True),
+        adapter=dict(
+            type='FlamingoLMAdapter',
+            vis_hidden_size=1024,
+            cross_attn_every_n_layers=4,
+            use_media_placement_augmentation=False,
+            only_attend_previous=True,
+        ),
+    ),
+    task='caption',
+    final_prompt_tmpl='<image>User:Please describe the image. GPT:<answer>',
+    generation_cfg=dict(
+        num_beams=3, max_new_tokens=24, no_repeat_ngram_size=3),
+)
+
+# data settings
+data_preprocessor = dict(
+    type='MultiModalDataPreprocessor',
+    mean=[122.770938, 116.7460125, 104.09373615],
+    std=[68.5005327, 66.6321579, 70.32316305],
+    to_rgb=True,
+)
+
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='ResizeEdge',
+        scale=224,
+        interpolation='bicubic',
+        backend='pillow'),
+    dict(type='CenterCrop', crop_size=(224, 224)),
+    dict(
+        type='PackInputs',
+        algorithm_keys=['gt_caption'],
+        meta_keys=['image_id'],
+    ),
+]
+
+val_dataloader = dict(
+    batch_size=8,
+    num_workers=8,
+    dataset=dict(
+        type='FlamingoEvalCOCOCaption',
+        data_root='data/coco',
+        ann_file='annotations/captions_train2014.json',
+        data_prefix=dict(img_path='train2014'),
+        pipeline=test_pipeline,
+        num_shots=0,
+        num_support_examples=2048,
+        num_query_examples=5000,
+    ),
+    sampler=dict(type='DefaultSampler', shuffle=False),
+    persistent_workers=True,
+)
+
+val_evaluator = dict(
+    type='COCOCaption',
+    ann_file='data/coco/annotations/captions_train2014.json')
+
+# If you want standard test, please manually configure the test dataset
+test_dataloader = val_dataloader
+test_evaluator = val_evaluator
+
+# schedule settings
+val_cfg = dict()
+test_cfg = dict()

configs/otter/otter-9b_vqa.py

@@ -0,0 +1,104 @@
+_base_ = [
+    '../_base_/default_runtime.py',
+]
+
+# model settings
+model = dict(
+    type='Otter',
+    tokenizer=dict(type='LlamaTokenizer', name_or_path='huggyllama/llama-7b'),
+    vision_encoder=dict(
+        type='VisionTransformer',
+        arch='l',
+        patch_size=14,
+        pre_norm=True,
+        norm_cfg=dict(type='LN', eps=1e-5),
+        layer_cfgs=dict(act_cfg=dict(type='QuickGELU')),
+        final_norm=False,
+        out_type='raw',
+        pretrained=(
+            'https://download.openmmlab.com/mmclassification/v0/clip/'
+            'vit-large-p14_clip-openai-pre_3rdparty_20230517-95e2af0b.pth'),
+    ),
+    lang_encoder=dict(
+        base=dict(
+            type='AutoModelForCausalLM',
+            name_or_path='huggyllama/llama-7b',
+            local_files_only=True),
+        adapter=dict(
+            type='FlamingoLMAdapter',
+            vis_hidden_size=1024,
+            cross_attn_every_n_layers=4,
+            use_media_placement_augmentation=False,
+            only_attend_previous=True,
+        ),
+    ),
+    task='vqa',
+    final_prompt_tmpl='<image>User:{question} GPT:<answer>',
+    generation_cfg=dict(
+        num_beams=3, max_new_tokens=24, no_repeat_ngram_size=3),
+)
+
+# data settings
+data_preprocessor = dict(
+    type='MultiModalDataPreprocessor',
+    mean=[122.770938, 116.7460125, 104.09373615],
+    std=[68.5005327, 66.6321579, 70.32316305],
+    to_rgb=True,
+)
+
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='ResizeEdge',
+        scale=224,
+        interpolation='bicubic',
+        backend='pillow'),
+    dict(type='CenterCrop', crop_size=(224, 224)),
+    dict(
+        type='PackInputs',
+        algorithm_keys=['question', 'gt_answer', 'gt_answer_weight', 'shots'],
+        meta_keys=['image_id'],
+    ),
+]
+
+val_dataloader = dict(
+    batch_size=8,
+    num_workers=8,
+    dataset=dict(
+        type='FlamingoEvalCOCOVQA',
+        data_root='data/coco',
+        data_prefix='val2014',
+        question_file='annotations/v2_OpenEnded_mscoco_val2014_questions.json',
+        ann_file='annotations/v2_mscoco_val2014_annotations.json',
+        pipeline=test_pipeline,
+        num_shots=0,
+        num_support_examples=2048,
+        num_query_examples=5000,
+    ),
+    sampler=dict(type='DefaultSampler', shuffle=False),
+    persistent_workers=True,
+)
+val_evaluator = dict(type='VQAAcc')
+
+test_dataloader = dict(
+    batch_size=8,
+    num_workers=8,
+    dataset=dict(
+        type='FlamingoEvalCOCOVQA',
+        data_root='data/coco',
+        data_prefix='test2015',
+        question_file=
+        'annotations/v2_OpenEnded_mscoco_test-dev2015_questions.json',
+        pipeline=test_pipeline,
+        num_shots=0,
+        num_support_examples=2048,
+        num_query_examples=5000,
+    ),
+    sampler=dict(type='DefaultSampler', shuffle=False),
+    persistent_workers=True,
+)
+test_evaluator = dict(type='ReportVQA', file_path='vqa_test-dev.json')
+
+# schedule settings
+val_cfg = dict()
+test_cfg = dict()

docs/en/api/models.rst

@@ -144,6 +144,7 @@ Multi-Modality Algorithms
    Flamingo
    OFA
    MiniGPT4
+   Otter
 
 .. module:: mmpretrain.models.backbones
 

mmpretrain/models/multimodal/__init__.py

@@ -8,12 +8,13 @@
     from .flamingo import *  # noqa: F401, F403
     from .minigpt4 import *  # noqa: F401, F403
     from .ofa import *  # noqa: F401, F403
+    from .otter import *  # noqa: F401, F403
 else:
     from mmpretrain.registry import MODELS
     from mmpretrain.utils.dependency import register_multimodal_placeholder
 
     register_multimodal_placeholder([
         'Blip2Caption', 'Blip2Retrieval', 'Blip2VQA', 'BlipCaption',
         'BlipNLVR', 'BlipRetrieval', 'BlipGrounding', 'BlipVQA', 'Flamingo',
-        'OFA', 'ChineseCLIP', 'MiniGPT4'
+        'OFA', 'ChineseCLIP', 'MiniGPT4', 'Otter'
     ], MODELS)

mmpretrain/models/multimodal/flamingo/adapter.py

@@ -19,6 +19,7 @@ def extend_init(
         vis_hidden_size: int,
         cross_attn_every_n_layers: int,
         use_media_placement_augmentation: bool,
+        only_attend_previous: bool = False,
     ):
         """Initialize Flamingo by adding a new gated cross attn to the decoder.
 
@@ -48,6 +49,7 @@ def extend_init(
             ]))
         base.use_media_placement_augmentation = use_media_placement_augmentation  # noqa
         base.initialized_flamingo = True
+        base.only_attend_previous = only_attend_previous
         return base
 
     def set_decoder_layers_attr_name(self, decoder_layers_attr_name):
@@ -67,8 +69,12 @@ def forward(self, *input, **kwargs):
         function."""
         input_ids = kwargs['input_ids'] if 'input_ids' in kwargs else input[0]
         media_locations = input_ids == self.media_token_id
-        attend_previous = ((random.random() < 0.5)
-                           if self.use_media_placement_augmentation else False)
+        if self.only_attend_previous:
+            attend_previous = True
+        elif self.use_media_placement_augmentation:
+            attend_previous = (random.random() < 0.5)
+        else:
+            attend_previous = False
 
         for layer in self.get_decoder().layers:
             layer.condition_media_locations(media_locations)

mmpretrain/models/multimodal/otter/__init__.py

@@ -0,0 +1,4 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from .otter import Otter
+
+__all__ = ['Otter']