From ae3c1869dacc6afc56076ab516edec6b8f648696 Mon Sep 17 00:00:00 2001
From: Yuan Liu <30762564+YuanLiuuuuuu@users.noreply.github.com>
Date: Thu, 17 Aug 2023 11:11:26 +0800
Subject: [PATCH] [Feature]: Add other public datasets config (#214)

* [Feature]: Add flickr30k

* [Feature]: Add GQA

* [Feature]: Add OCR VQA

* [Feature]: Add OK VQA

* [Feature]: Add text vqa

* [Feature]: Add other vqa
---
 .../minigpt_4/minigpt_4_7b_flickr30k.py       | 53 ++++++++++++++++++
 .../multimodal/minigpt_4/minigpt_4_7b_gqa.py  | 52 ++++++++++++++++++
 .../minigpt_4/minigpt_4_7b_ocr-vqa.py         | 53 ++++++++++++++++++
 .../minigpt_4/minigpt_4_7b_ok-vqa.py          | 55 +++++++++++++++++++
 .../minigpt_4/minigpt_4_7b_textvqa.py         | 55 +++++++++++++++++++
 .../minigpt_4/minigpt_4_7b_vizwiz.py          | 52 ++++++++++++++++++
 6 files changed, 320 insertions(+)
 create mode 100644 configs/multimodal/minigpt_4/minigpt_4_7b_flickr30k.py
 create mode 100644 configs/multimodal/minigpt_4/minigpt_4_7b_gqa.py
 create mode 100644 configs/multimodal/minigpt_4/minigpt_4_7b_ocr-vqa.py
 create mode 100644 configs/multimodal/minigpt_4/minigpt_4_7b_ok-vqa.py
 create mode 100644 configs/multimodal/minigpt_4/minigpt_4_7b_textvqa.py
 create mode 100644 configs/multimodal/minigpt_4/minigpt_4_7b_vizwiz.py

diff --git a/configs/multimodal/minigpt_4/minigpt_4_7b_flickr30k.py b/configs/multimodal/minigpt_4/minigpt_4_7b_flickr30k.py
new file mode 100644
index 00000000..69bd2afc
--- /dev/null
+++ b/configs/multimodal/minigpt_4/minigpt_4_7b_flickr30k.py
@@ -0,0 +1,53 @@
+from opencompass.multimodal.models.minigpt_4 import (
+    MiniGPT4COCOCaotionPromptConstructor,
+    MiniGPT4COCOCaptionPostProcessor,
+)
+
+# dataloader settings
+val_pipeline = [
+    dict(type='mmpretrain.LoadImageFromFile'),
+    dict(type='mmpretrain.ToPIL', to_rgb=True),
+    dict(type='mmpretrain.torchvision/Resize',
+         size=(384, 384),
+         interpolation=3),
+    dict(type='mmpretrain.torchvision/ToTensor'),
+    dict(type='mmpretrain.torchvision/Normalize',
+         mean=(0.48145466, 0.4578275, 0.40821073),
+         std=(0.26862954, 0.26130258, 0.27577711)),
+    dict(type='mmpretrain.PackInputs', algorithm_keys=['image_id'])
+]
+
+dataset = dict(type='mmpretrain.Flickr30kCaption',
+               data_root='data/flickr30k',
+               ann_file='annotations/dataset_flickr30k.json',
+               data_prefix='images',
+               split='val',
+               pipeline=val_pipeline)
+
+minigpt_4_flickr30k_dataloader = dict(
+    batch_size=1,
+    num_workers=4,
+    dataset=dataset,
+    collate_fn=dict(type='pseudo_collate'),
+    sampler=dict(type='DefaultSampler', shuffle=False))
+
+# model settings
+minigpt_4_flickr30k_model = dict(
+    type='minigpt-4',
+    low_resource=False,
+    img_size=384,
+    llama_model='/path/to/vicuna-7b/',
+    prompt_constructor=dict(type=MiniGPT4COCOCaotionPromptConstructor,
+                            image_prompt='###Human: <Img><ImageHere></Img>',
+                            reply_prompt='###Assistant:'),
+    post_processor=dict(type=MiniGPT4COCOCaptionPostProcessor))
+
+# evaluation settings
+minigpt_4_flickr30k_evaluator = [
+    dict(
+        type='mmpretrain.COCOCaption',
+        ann_file='data/coco/annotations/coco_karpathy_val_gt.json',
+    )  # noqa
+]
+
+minigpt_4_flickr30k_load_from = '/path/to/prerained_minigpt4_7b.pth'  # noqa
diff --git a/configs/multimodal/minigpt_4/minigpt_4_7b_gqa.py b/configs/multimodal/minigpt_4/minigpt_4_7b_gqa.py
new file mode 100644
index 00000000..f0cec5a6
--- /dev/null
+++ b/configs/multimodal/minigpt_4/minigpt_4_7b_gqa.py
@@ -0,0 +1,52 @@
+from opencompass.multimodal.models.minigpt_4 import (
+    MiniGPT4VQAPromptConstructor,
+    MiniGPT4VQAPostProcessor,
+)
+
+# dataloader settings
+val_pipeline = [
+    dict(type='mmpretrain.LoadImageFromFile'),
+    dict(type='mmpretrain.ToPIL', to_rgb=True),
+    dict(type='mmpretrain.torchvision/Resize',
+         size=(224, 224),
+         interpolation=3),
+    dict(type='mmpretrain.torchvision/ToTensor'),
+    dict(type='mmpretrain.torchvision/Normalize',
+         mean=(0.48145466, 0.4578275, 0.40821073),
+         std=(0.26862954, 0.26130258, 0.27577711)),
+    dict(
+        type='mmpretrain.PackInputs',
+        algorithm_keys=['question', 'gt_answer', 'gt_answer_weight'],
+        meta_keys=['question_id', 'image_id'],
+    )
+]
+
+dataset = dict(type='mmpretrain.GQA',
+               data_root='data/gqa',
+               data_prefix='images',
+               ann_file='annotations/testdev_balanced_questions.json',
+               pipeline=val_pipeline)
+
+minigpt_4_gqa_dataloader = dict(batch_size=1,
+                                num_workers=4,
+                                dataset=dataset,
+                                collate_fn=dict(type='pseudo_collate'),
+                                sampler=dict(type='DefaultSampler',
+                                             shuffle=False))
+
+# model settings
+minigpt_4_gqa_model = dict(type='minigpt-4',
+                           low_resource=False,
+                           img_size=224,
+                           max_length=10,
+                           llama_model='/path/to/vicuna-7b/',
+                           prompt_constructor=dict(
+                               type=MiniGPT4VQAPromptConstructor,
+                               image_prompt='###Human: <Img><ImageHere></Img>',
+                               reply_prompt='###Assistant:'),
+                           post_processor=dict(type=MiniGPT4VQAPostProcessor))
+
+# evaluation settings
+minigpt_4_gqa_evaluator = [dict(type='mmpretrain.GQAAcc')]
+
+minigpt_4_gqa_load_from = '/path/to/prerained_minigpt4_7b.pth'  # noqa
diff --git a/configs/multimodal/minigpt_4/minigpt_4_7b_ocr-vqa.py b/configs/multimodal/minigpt_4/minigpt_4_7b_ocr-vqa.py
new file mode 100644
index 00000000..f95b2f85
--- /dev/null
+++ b/configs/multimodal/minigpt_4/minigpt_4_7b_ocr-vqa.py
@@ -0,0 +1,53 @@
+from opencompass.multimodal.models.minigpt_4 import (
+    MiniGPT4VQAPromptConstructor,
+    MiniGPT4VQAPostProcessor,
+)
+
+# dataloader settings
+val_pipeline = [
+    dict(type='mmpretrain.LoadImageFromFile'),
+    dict(type='mmpretrain.ToPIL', to_rgb=True),
+    dict(type='mmpretrain.torchvision/Resize',
+         size=(224, 224),
+         interpolation=3),
+    dict(type='mmpretrain.torchvision/ToTensor'),
+    dict(type='mmpretrain.torchvision/Normalize',
+         mean=(0.48145466, 0.4578275, 0.40821073),
+         std=(0.26862954, 0.26130258, 0.27577711)),
+    dict(
+        type='mmpretrain.PackInputs',
+        algorithm_keys=['question', 'gt_answer', 'gt_answer_weight'],
+        meta_keys=['question_id', 'image_id'],
+    )
+]
+
+dataset = dict(type='mmpretrain.OCRVQA',
+               data_root='data/ocrvqa',
+               ann_file='annotations/dataset.json',
+               split='test',
+               data_prefix='images',
+               pipeline=val_pipeline)
+
+minigpt_4_ocr_vqa_dataloader = dict(batch_size=1,
+                                    num_workers=4,
+                                    dataset=dataset,
+                                    collate_fn=dict(type='pseudo_collate'),
+                                    sampler=dict(type='DefaultSampler',
+                                                 shuffle=False))
+
+# model settings
+minigpt_4_ocr_vqa_model = dict(
+    type='minigpt-4',
+    low_resource=False,
+    img_size=224,
+    max_length=10,
+    llama_model='/path/to/vicuna-7b/',
+    prompt_constructor=dict(type=MiniGPT4VQAPromptConstructor,
+                            image_prompt='###Human: <Img><ImageHere></Img>',
+                            reply_prompt='###Assistant:'),
+    post_processor=dict(type=MiniGPT4VQAPostProcessor))
+
+# evaluation settings
+minigpt_4_ocr_vqa_evaluator = [dict(type='mmpretrain.VQAAcc')]
+
+minigpt_4_ocr_vqa_load_from = '/path/to/prerained_minigpt4_7b.pth'  # noqa
diff --git a/configs/multimodal/minigpt_4/minigpt_4_7b_ok-vqa.py b/configs/multimodal/minigpt_4/minigpt_4_7b_ok-vqa.py
new file mode 100644
index 00000000..249223f5
--- /dev/null
+++ b/configs/multimodal/minigpt_4/minigpt_4_7b_ok-vqa.py
@@ -0,0 +1,55 @@
+from opencompass.multimodal.models.minigpt_4 import (
+    MiniGPT4VQAPromptConstructor,
+    MiniGPT4VQAPostProcessor,
+)
+
+# dataloader settings
+val_pipeline = [
+    dict(type='mmpretrain.LoadImageFromFile'),
+    dict(type='mmpretrain.ToPIL', to_rgb=True),
+    dict(type='mmpretrain.torchvision/Resize',
+         size=(224, 224),
+         interpolation=3),
+    dict(type='mmpretrain.torchvision/ToTensor'),
+    dict(type='mmpretrain.torchvision/Normalize',
+         mean=(0.48145466, 0.4578275, 0.40821073),
+         std=(0.26862954, 0.26130258, 0.27577711)),
+    dict(
+        type='mmpretrain.PackInputs',
+        algorithm_keys=['question', 'gt_answer', 'gt_answer_weight'],
+        meta_keys=['question_id', 'image_id'],
+    )
+]
+
+dataset = dict(
+    type='mmpretrain.COCOVQA',
+    data_root='data/okvqa',
+    question_file='annotations/OpenEnded_mscoco_val2014_questions.json',
+    ann_file='annotations/mscoco_val2014_annotations.json',
+    pipeline=val_pipeline,
+    data_prefix='images/val2014',
+)
+
+minigpt_4_ok_vqa_dataloader = dict(batch_size=1,
+                                  num_workers=4,
+                                  dataset=dataset,
+                                  collate_fn=dict(type='pseudo_collate'),
+                                  sampler=dict(type='DefaultSampler',
+                                               shuffle=False))
+
+# model settings
+minigpt_4_ok_vqa_model = dict(
+    type='minigpt-4',
+    low_resource=False,
+    img_size=224,
+    max_length=10,
+    llama_model='/path/to/vicuna-7b/',
+    prompt_constructor=dict(type=MiniGPT4VQAPromptConstructor,
+                            image_prompt='###Human: <Img><ImageHere></Img>',
+                            reply_prompt='###Assistant:'),
+    post_processor=dict(type=MiniGPT4VQAPostProcessor))
+
+# evaluation settings
+minigpt_4_ok_vqa_evaluator = [dict(type='mmpretrain.VQAAcc')]
+
+minigpt_4_ok_vqa_load_from = '/path/to/prerained_minigpt4_7b.pth'  # noqa
diff --git a/configs/multimodal/minigpt_4/minigpt_4_7b_textvqa.py b/configs/multimodal/minigpt_4/minigpt_4_7b_textvqa.py
new file mode 100644
index 00000000..3ad2b7ef
--- /dev/null
+++ b/configs/multimodal/minigpt_4/minigpt_4_7b_textvqa.py
@@ -0,0 +1,55 @@
+from opencompass.multimodal.models.minigpt_4 import (
+    MiniGPT4VQAPromptConstructor,
+    MiniGPT4VQAPostProcessor,
+)
+
+
+# dataloader settings
+val_pipeline = [
+    dict(type='mmpretrain.LoadImageFromFile'),
+    dict(type='mmpretrain.ToPIL', to_rgb=True),
+    dict(type='mmpretrain.torchvision/Resize',
+         size=(224, 224),
+         interpolation=3),
+    dict(type='mmpretrain.torchvision/ToTensor'),
+    dict(type='mmpretrain.torchvision/Normalize',
+         mean=(0.48145466, 0.4578275, 0.40821073),
+         std=(0.26862954, 0.26130258, 0.27577711)),
+    dict(
+        type='mmpretrain.PackInputs',
+        algorithm_keys=['question', 'gt_answer', 'gt_answer_weight'],
+        meta_keys=['question_id', 'image_id'],
+    )
+]
+
+dataset = dict(
+    type='mmpretrain.TextVQA',
+    data_root='data/textvqa',
+    ann_file='annotations/TextVQA_0.5.1_val.json',
+    pipeline=val_pipeline,
+    data_prefix='images/train_images',
+)
+
+minigpt_4_textvqa_dataloader = dict(batch_size=1,
+                                  num_workers=4,
+                                  dataset=dataset,
+                                  collate_fn=dict(type='pseudo_collate'),
+                                  sampler=dict(type='DefaultSampler',
+                                               shuffle=False))
+
+# model settings
+minigpt_4_textvqa_model = dict(
+    type='minigpt-4',
+    low_resource=False,
+    img_size=224,
+    max_length=10,
+    llama_model='/path/to/vicuna-7b/',
+    prompt_constructor=dict(type=MiniGPT4VQAPromptConstructor,
+                            image_prompt='###Human: <Img><ImageHere></Img>',
+                            reply_prompt='###Assistant:'),
+    post_processor=dict(type=MiniGPT4VQAPostProcessor))
+
+# evaluation settings
+minigpt_4_textvqa_evaluator = [dict(type='mmpretrain.VQAAcc')]
+
+minigpt_4_textvqa_load_from = '/path/to/prerained_minigpt4_7b.pth'  # noqa
diff --git a/configs/multimodal/minigpt_4/minigpt_4_7b_vizwiz.py b/configs/multimodal/minigpt_4/minigpt_4_7b_vizwiz.py
new file mode 100644
index 00000000..0a371aed
--- /dev/null
+++ b/configs/multimodal/minigpt_4/minigpt_4_7b_vizwiz.py
@@ -0,0 +1,52 @@
+from opencompass.multimodal.models.minigpt_4 import (
+    MiniGPT4VQAPromptConstructor,
+    MiniGPT4VQAPostProcessor,
+)
+
+# dataloader settings
+val_pipeline = [
+    dict(type='mmpretrain.LoadImageFromFile'),
+    dict(type='mmpretrain.ToPIL', to_rgb=True),
+    dict(type='mmpretrain.torchvision/Resize',
+         size=(224, 224),
+         interpolation=3),
+    dict(type='mmpretrain.torchvision/ToTensor'),
+    dict(type='mmpretrain.torchvision/Normalize',
+         mean=(0.48145466, 0.4578275, 0.40821073),
+         std=(0.26862954, 0.26130258, 0.27577711)),
+    dict(
+        type='mmpretrain.PackInputs',
+        algorithm_keys=['question', 'gt_answer', 'gt_answer_weight'],
+        meta_keys=['question_id', 'image_id'],
+    )
+]
+
+dataset = dict(type='mmpretrain.VizWiz',
+               data_root='data/vizwiz/',
+               data_prefix='Images/val',
+               ann_file='Annotations/val.json',
+               pipeline=val_pipeline)
+
+minigpt_4_vizwiz_dataloader = dict(batch_size=1,
+                                   num_workers=4,
+                                   dataset=dataset,
+                                   collate_fn=dict(type='pseudo_collate'),
+                                   sampler=dict(type='DefaultSampler',
+                                                shuffle=False))
+
+# model settings
+minigpt_4_vizwiz_model = dict(
+    type='minigpt-4',
+    low_resource=False,
+    img_size=224,
+    max_length=10,
+    llama_model='/path/to/vicuna-7b/',
+    prompt_constructor=dict(type=MiniGPT4VQAPromptConstructor,
+                            image_prompt='###Human: <Img><ImageHere></Img>',
+                            reply_prompt='###Assistant:'),
+    post_processor=dict(type=MiniGPT4VQAPostProcessor))
+
+# evaluation settings
+minigpt_4_vizwiz_evaluator = [dict(type='mmpretrain.VQAAcc')]
+
+minigpt_4_vizwiz_load_from = '/path/to/prerained_minigpt4_7b.pth'  # noqa