chris/MiniGPT-4
1
0
Fork 0
mirror of https://github.com/Vision-CAIR/MiniGPT-4.git synced 2025-04-05 02:20:47 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'main' of https://github.com/junchen14/MiniGPT-4_finetune into main

Browse Source
This commit is contained in:
junchen14 2023-10-23 23:56:58 +03:00
commit 3fdb6f35c9
10 changed files with 123 additions and 131 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
dataset/Evaluation.md Normal file
View File

@ -0,0 +1,26 @@
### OKVQA
### GQA
Images and question-answer pairs will be loaded during the evaluation.
``` python run_eval.py xxxx ```
### VSR
Images and question-answer pairs will be loaded during the evaluation.
``` python run_eval.py xxxx ```
### IconVQA
### VizWiz
1. Download [`test.json`](https://vizwiz.cs.colorado.edu/VizWiz_final/vqa_data/Annotations.zip) and extract [`test.zip`](https://vizwiz.cs.colorado.edu/VizWiz_final/images/test.zip) to `test`. Put them under `your_path/vizwiz`.
2. Single-GPU inference.
``` python run_eval.py xxxx ```
### HM

183
dataset/README_MINIGPTv2_FINETUNE.md
View File

@ -1,133 +1,94 @@
## Download the COCO captions, RefCOCO, RefCOCO+. RefCOCOg, visual genome, textcaps, LLaVA, gqa, AOK-VQA, OK-VQA, OCR-VQA, filtered Flickr-30k, multi-task conversation, and Unnatural instruction datasets
### COCO captions
- [train2017](http://images.cocodataset.org/zips/train2017.zip)
### RefCOCO, RefCOCO+, RefCOCOg
### Visual genome
- [part1](https://cs.stanford.edu/people/rak248/VG_100K_2/images.zip), [part2](https://cs.stanford.edu/people/rak248/VG_100K_2/images2.zip)
### textcaps
### TextCaps
- [TextCaps_0.1_train](https://dl.fbaipublicfiles.com/textvqa/data/textcaps/TextCaps_0.1_train.json)
- [Images](https://dl.fbaipublicfiles.com/textvqa/images/train_val_images.zip)
### RefCOCO, RefCOCO+, RefCOCOg
Make sure you have the COCO 2014 images first.
Then,
download RefCOCO, RefCOCO+, and RefCOCOg annotation files in the following links.
- https://bvisionweb1.cs.unc.edu/licheng/referit/data/refcoco.zip
- https://bvisionweb1.cs.unc.edu/licheng/referit/data/refcoco+.zip
- https://bvisionweb1.cs.unc.edu/licheng/referit/data/refcocog.zip
Unzip these files to the location you like. It should have the structure like the following
```
Location_you_like
├── refcoco
│ ├── instances.json
│ ├── refs(google).p
│ └── refs(unc).p
├── refcoco+
│ ├── instances.json
│ └── refs(unc).p
└── refcocog
├── instances.json
├── refs(google).p
└── refs(umd).p
```
Set **image_path** in all the following dataset configuration files to the COCO 2014 image folder.
Similarly, set **ann_path** in all the following configs to the above folder (Location_you_like) that contains refcoco, refcoco+, and refcocog.
- [minigpt4/configs/datasets/coco_bbox/refcoco.yaml](../minigpt4/configs/datasets/coco_bbox/refcoco.yaml)
- [minigpt4/configs/datasets/coco_bbox/refcocog.yaml](../minigpt4/configs/datasets/coco_bbox/refcocog.yaml)
- [minigpt4/configs/datasets/coco_bbox/refcocop.yaml](../minigpt4/configs/datasets/coco_bbox/refcocop.yaml)
- [minigpt4/configs/datasets/coco_bbox/invrefcoco.yaml](../minigpt4/configs/datasets/coco_bbox/invrefcoco.yaml)
- [minigpt4/configs/datasets/coco_bbox/invrefcocog.yaml](../minigpt4/configs/datasets/coco_bbox/invrefcocog.yaml)
- [minigpt4/configs/datasets/coco_bbox/invrefcocop.yaml](../minigpt4/configs/datasets/coco_bbox/invrefcocop.yaml)
### LLaVA
Makesure you have the COCO 2014 images first.
Download Llava annotation files in the following link to the place you like.
- https://huggingface.co/datasets/liuhaotian/LLaVA-Instruct-150K/resolve/main/conversation_58k.json
- https://huggingface.co/datasets/liuhaotian/LLaVA-Instruct-150K/resolve/main/detail_23k.json
- https://huggingface.co/datasets/liuhaotian/LLaVA-Instruct-150K/resolve/main/complex_reasoning_77k.json
Set **image_path** in all the following dataset configuration files to the COCO 2014 image folder.
Similarly, set **ann_path** to the location of the previous downloaded conversation_58k.json,
detail_23k.json, and complex_reasoning_77k.json in conversation.yaml, detail.yaml, and reason.yaml, respectively.
- [minigpt4/configs/datasets/llava/conversation.yaml](../minigpt4/configs/datasets/llava/conversation.yaml)
- [minigpt4/configs/datasets/llava/detail.yaml](../minigpt4/configs/datasets/llava/detail.yaml)
- [minigpt4/configs/datasets/llava/reason.yaml](../minigpt4/configs/datasets/llava/reason.yaml)
### gqa
### OKVQA
- [OK-VQA Input Questions](https://okvqa.allenai.org/static/data/OpenEnded_mscoco_train2014_questions.json.zip)
- [OK-VQA Annotations](https://okvqa.allenai.org/static/data/mscoco_train2014_annotations.json.zip)
- [okvqa_train](https://storage.googleapis.com/sfr-vision-language-research/LAVIS/datasets/okvqa/okvqa_train.json)
- Images are from COCO
### AOK-VQA
```
export AOKVQA_DIR=YOUR_DATASET_PATH
mkdir -p ${AOKVQA_DIR}
curl -fsSL https://prior-datasets.s3.us-east-2.amazonaws.com/aokvqa/aokvqa_v1p0.tar.gz | tar xvz -C ${AOKVQA_DIR}
```
### OCR-VQA
- [download script](https://drive.google.com/drive/folders/1_GYPY5UkUy7HIcR0zq3ZCFgeZN7BAfm_?usp=sharing), **we save all files as `.jpg`**
### filtered Flickr-30k
### Multi-task conversation
### Unnatural instruction
### Pre-training datasets download:
We use the filtered synthetic captions prepared by BLIP. For more details about the dataset, please refer to [BLIP](https://github.com/salesforce/BLIP).
It requires ~2.3T to store LAION and CC3M+CC12M+SBU datasets
Image source | Filtered synthetic caption by ViT-L
--- | :---:
CC3M+CC12M+SBU | <a href="https://storage.googleapis.com/sfr-vision-language-research/BLIP/datasets/ccs_synthetic_filtered_large.json">Download</a>
LAION115M | <a href="https://storage.googleapis.com/sfr-vision-language-research/BLIP/datasets/laion_synthetic_filtered_large.json">Download</a>
This will download two json files
```
ccs_synthetic_filtered_large.json
laion_synthetic_filtered_large.json
```
## prepare the data step-by-step
### setup the dataset folder and move the annotation file to the data storage folder
```
export MINIGPT4_DATASET=/YOUR/PATH/FOR/LARGE/DATASET/
mkdir ${MINIGPT4_DATASET}/cc_sbu
mkdir ${MINIGPT4_DATASET}/laion
mv ccs_synthetic_filtered_large.json ${MINIGPT4_DATASET}/cc_sbu
mv laion_synthetic_filtered_large.json ${MINIGPT4_DATASET}/laion
```
### Convert the scripts to data storate folder
```
cp convert_cc_sbu.py ${MINIGPT4_DATASET}/cc_sbu
cp download_cc_sbu.sh ${MINIGPT4_DATASET}/cc_sbu
cp convert_laion.py ${MINIGPT4_DATASET}/laion
cp download_laion.sh ${MINIGPT4_DATASET}/laion
```
### Convert the laion and cc_sbu annotation file format to be img2dataset format
```
cd ${MINIGPT4_DATASET}/cc_sbu
python convert_cc_sbu.py
cd ${MINIGPT4_DATASET}/laion
python convert_laion.py
```
### Download the datasets with img2dataset
```
cd ${MINIGPT4_DATASET}/cc_sbu
sh download_cc_sbu.sh
cd ${MINIGPT4_DATASET}/laion
sh download_laion.sh
```
The final dataset structure
```
.
├── ${MINIGPT4_DATASET}
│ ├── cc_sbu
│ ├── convert_cc_sbu.py
│ ├── download_cc_sbu.sh
│ ├── ccs_synthetic_filtered_large.json
│ ├── ccs_synthetic_filtered_large.tsv
│ └── cc_sbu_dataset
│ ├── 00000.tar
│ ├── 00000.parquet
│ ...
│ ├── laion
│ ├── convert_laion.py
│ ├── download_laion.sh
│ ├── laion_synthetic_filtered_large.json
│ ├── laion_synthetic_filtered_large.tsv
│ └── laion_dataset
│ ├── 00000.tar
│ ├── 00000.parquet
│ ...
...
```
## Set up the dataset configuration files
Then, set up the LAION dataset loading path in
[here](../minigpt4/configs/datasets/laion/defaults.yaml#L5) at Line 5 as
${MINIGPT4_DATASET}/laion/laion_dataset/{00000..10488}.tar
and the Conceptual Captoin and SBU datasets loading path in
[here](../minigpt4/configs/datasets/cc_sbu/defaults.yaml#L5) at Line 5 as
${MINIGPT4_DATASET}/cc_sbu/cc_sbu_dataset/{00000..01255}.tar

3
minigpt4/models/minigpt_base.py
View File

@ -178,7 +178,6 @@ class MiniGPTBase(BaseModel):
answers = [self.llama_tokenizer(a + self.end_sym,
return_tensors="pt",
add_special_tokens=False).to(self.device) for a in answers]
cur_id = []
cur_target = []
for i in range(len(questions)):
@ -226,8 +225,6 @@ class MiniGPTBase(BaseModel):
conv_q = [[self.prompt_template.format(item) for item in items] for items in conv_q]
cond_embeds, cond_atts = self.prompt_wrap(img_embeds, img_atts, [q[0] for q in conv_q])
regress_token_ids, regress_atts, part_targets = self.tokenize_conversation(conv_q, conv_a)

2
minigpt4/models/modeling_llama.py
View File

@ -75,7 +75,7 @@ class LlamaForCausalLM(LlamaForCausalLMOrig):
)
hidden_states = outputs[0]
if self.config.pretraining_tp > 1:
if hasattr(self.config, 'pretraining_tp') and self.config.pretraining_tp > 1:
lm_head_slices = self.lm_head.weight.split(self.vocab_size // self.config.pretraining_tp, dim=0)
logits = [F.linear(hidden_states, lm_head_slices[i]) for i in range(self.config.pretraining_tp)]
logits = torch.cat(logits, dim=-1)

14
train.py
View File

@ -12,6 +12,7 @@ import random
import numpy as np
import torch
import torch.backends.cudnn as cudnn
import wandb
import minigpt4.tasks as tasks
from minigpt4.common.config import Config
@ -30,7 +31,6 @@ from minigpt4.models import *
from minigpt4.processors import *
from minigpt4.runners import *
from minigpt4.tasks import *
import wandb
def parse_args():
@ -44,12 +44,10 @@ def parse_args():
"in xxx=yyy format will be merged into config file (deprecate), "
"change to --cfg-options instead.",
)
parser.add_argument("--wandb_log", default=False)
parser.add_argument("--job_name",default="minigpt_v2",type=str)
parser.add_argument("--job_name", default="minigpt_v2",type=str)
args = parser.parse_args()
return args
@ -80,16 +78,13 @@ def main():
# set before init_distributed_mode() to ensure the same job_id shared across all ranks.
job_id = now()
args = parse_args()
cfg = Config(parse_args())
cfg = Config(args)
init_distributed_mode(cfg.run_cfg)
setup_seeds(cfg)
# set after init_distributed_mode() to only log on master.
setup_logger()
cfg.pretty_print()
task = tasks.setup_task(cfg)
@ -98,10 +93,9 @@ def main():
if cfg.run_cfg.wandb_log:
wandb.login()
wandb.init(project="minigptv2",name=args.job_name)
wandb.init(project="minigptv", name=cfg.run_cfg.job_name)
wandb.watch(model)
runner = get_runner_class(cfg)(
cfg=cfg, job_id=job_id, task=task, model=model, datasets=datasets
)

3
train_configs/minigpt4_llama2_stage1_pretrain.yaml
View File

@ -53,3 +53,6 @@ run:
world_size: 1
dist_url: "env://"
distributed: True
wandb_log: True
job_name: minigpt4_llama2_pretrain

3
train_configs/minigpt4_llama2_stage2_finetune.yaml
View File

@ -47,3 +47,6 @@ run:
world_size: 1
dist_url: "env://"
distributed: True
wandb_log: True
job_name: minigpt4_llama2_finetune

3
train_configs/minigpt4_stage1_pretrain.yaml
View File

@ -53,3 +53,6 @@ run:
world_size: 1
dist_url: "env://"
distributed: True
wandb_log: True
job_name: minigpt4_pretrain

3
train_configs/minigpt4_stage2_finetune.yaml
View File

@ -47,3 +47,6 @@ run:
world_size: 1
dist_url: "env://"
distributed: True
wandb_log: True
job_name: minigpt4_finetune

4
train_configs/minigpt_v2_finetune.yaml
View File

@ -276,7 +276,6 @@ run:
init_lr: 1e-5
min_lr: 8e-5
warmup_lr: 1e-6
wandb_log: True
weight_decay: 0.05
max_epoch: 50
@ -297,3 +296,6 @@ run:
world_size: 1
dist_url: "env://"
distributed: True
wandb_log: True
job_name: minigptv2_finetune