Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[Feature] Support multi-image training and chatting for LLaVA #459

Open
wants to merge 11 commits into
base: main
Choose a base branch
from
Open

[Feature] Support multi-image training and chatting for LLaVA #459

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
6ca277c
per_image_token
LZHgrla Feb 5, 2024
e5dd387
update
LZHgrla Feb 5, 2024
b2222f0
Merge branch 'main' into lzh/per_image_token
LZHgrla Feb 29, 2024
97927c1
update
LZHgrla Feb 29, 2024
b9e0e3d
update
LZHgrla Feb 29, 2024
d12e1e7
update
LZHgrla Mar 5, 2024
61f1482
update
LZHgrla Mar 11, 2024
54ce706
update
LZHgrla Mar 11, 2024
26715ad
update
LZHgrla Mar 11, 2024
febc2d7
update
LZHgrla Mar 11, 2024
968641d
Merge branch 'main' into lzh/multi_image
LZHgrla Mar 11, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
4 changes: 3 additions & 1 deletion ...arge_p14_336/finetune/llava_internlm2_chat_20b_clip_vit_large_p14_336_e1_gpu8_finetune.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,7 +30,8 @@
data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
image_folder = data_root + 'llava_images'
prompt_template = PROMPT_TEMPLATE.internlm2_chat
max_length = int(2048 - (336 / 14)**2)
max_length = 2048
per_image_length = int(336 / 14)**2 # inferred from the visual encoder

# Scheduler & Optimizer
batch_size = 4 # per_device
Expand Down Expand Up @@ -96,6 +97,7 @@
template_map_fn=dict(
type=template_map_fn_factory, template=prompt_template),
max_length=max_length,
per_image_length=per_image_length,
pad_image_to_square=True)

train_dataloader = dict(
Expand Down
4 changes: 3 additions & 1 deletion ...6/finetune/llava_internlm2_chat_20b_qlora_clip_vit_large_p14_336_lora_e1_gpu8_finetune.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,7 +32,8 @@
data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
image_folder = data_root + 'llava_images'
prompt_template = PROMPT_TEMPLATE.internlm2_chat
max_length = int(2048 - (336 / 14)**2)
max_length = 2048
per_image_length = int(336 / 14)**2 # inferred from the visual encoder

# Scheduler & Optimizer
batch_size = 8 # per_device
Expand Down Expand Up @@ -115,6 +116,7 @@
template_map_fn=dict(
type=template_map_fn_factory, template=prompt_template),
max_length=max_length,
per_image_length=per_image_length,
pad_image_to_square=True)

train_dataloader = dict(
Expand Down
4 changes: 3 additions & 1 deletion ...arge_p14_336/pretrain/llava_internlm2_chat_20b_clip_vit_large_p14_336_e1_gpu8_pretrain.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,7 +29,8 @@
data_path = data_root + 'LLaVA-Pretrain/blip_laion_cc_sbu_558k.json'
image_folder = data_root + 'LLaVA-Pretrain/images'
prompt_template = PROMPT_TEMPLATE.internlm2_chat
max_length = int(2048 - (336 / 14)**2)
max_length = 2048
per_image_length = int(336 / 14)**2 # inferred from the visual encoder

# Scheduler & Optimizer
batch_size = 32 # per_device
Expand Down Expand Up @@ -102,6 +103,7 @@
template_map_fn=dict(
type=template_map_fn_factory, template=prompt_template),
max_length=max_length,
per_image_length=per_image_length,
pad_image_to_square=False)

train_dataloader = dict(
Expand Down
4 changes: 3 additions & 1 deletion ...large_p14_336/finetune/llava_internlm2_chat_7b_clip_vit_large_p14_336_e1_gpu8_finetune.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,7 +30,8 @@
data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
image_folder = data_root + 'llava_images'
prompt_template = PROMPT_TEMPLATE.internlm2_chat
max_length = int(2048 - (336 / 14)**2)
max_length = 2048
per_image_length = int(336 / 14)**2 # inferred from the visual encoder

# Scheduler & Optimizer
batch_size = 8 # per_device
Expand Down Expand Up @@ -95,6 +96,7 @@
template_map_fn=dict(
type=template_map_fn_factory, template=prompt_template),
max_length=max_length,
per_image_length=per_image_length,
pad_image_to_square=True)

train_dataloader = dict(
Expand Down
4 changes: 3 additions & 1 deletion ...36/finetune/llava_internlm2_chat_7b_qlora_clip_vit_large_p14_336_lora_e1_gpu8_finetune.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,7 +32,8 @@
data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
image_folder = data_root + 'llava_images'
prompt_template = PROMPT_TEMPLATE.internlm2_chat
max_length = int(2048 - (336 / 14)**2)
max_length = 2048
per_image_length = int(336 / 14)**2 # inferred from the visual encoder

# Scheduler & Optimizer
batch_size = 16 # per_device
Expand Down Expand Up @@ -115,6 +116,7 @@
template_map_fn=dict(
type=template_map_fn_factory, template=prompt_template),
max_length=max_length,
per_image_length=per_image_length,
pad_image_to_square=True)

train_dataloader = dict(
Expand Down
4 changes: 3 additions & 1 deletion ...large_p14_336/pretrain/llava_internlm2_chat_7b_clip_vit_large_p14_336_e1_gpu8_pretrain.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,7 +29,8 @@
data_path = data_root + 'LLaVA-Pretrain/blip_laion_cc_sbu_558k.json'
image_folder = data_root + 'LLaVA-Pretrain/images'
prompt_template = PROMPT_TEMPLATE.internlm2_chat
max_length = int(2048 - (336 / 14)**2)
max_length = 2048
per_image_length = int(336 / 14)**2 # inferred from the visual encoder

# Scheduler & Optimizer
batch_size = 32 # per_device
Expand Down Expand Up @@ -102,6 +103,7 @@
template_map_fn=dict(
type=template_map_fn_factory, template=prompt_template),
max_length=max_length,
per_image_length=per_image_length,
pad_image_to_square=False)

train_dataloader = dict(
Expand Down
4 changes: 3 additions & 1 deletion ..._large_p14_336/finetune/llava_internlm_chat_7b_clip_vit_large_p14_336_e1_gpu8_finetune.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,7 +30,8 @@
data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
image_folder = data_root + 'llava_images'
prompt_template = PROMPT_TEMPLATE.internlm_chat
max_length = int(2048 - (336 / 14)**2)
max_length = 2048
per_image_length = int(336 / 14)**2 # inferred from the visual encoder

# Scheduler & Optimizer
batch_size = 16 # per_device
Expand Down Expand Up @@ -95,6 +96,7 @@
template_map_fn=dict(
type=template_map_fn_factory, template=prompt_template),
max_length=max_length,
per_image_length=per_image_length,
pad_image_to_square=True)

train_dataloader = dict(
Expand Down
4 changes: 3 additions & 1 deletion ...336/finetune/llava_internlm_chat_7b_qlora_clip_vit_large_p14_336_lora_e1_gpu8_finetune.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,7 +32,8 @@
data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
image_folder = data_root + 'llava_images'
prompt_template = PROMPT_TEMPLATE.internlm_chat
max_length = int(2048 - (336 / 14)**2)
max_length = 2048
per_image_length = int(336 / 14)**2 # inferred from the visual encoder

# Scheduler & Optimizer
batch_size = 16 # per_device
Expand Down Expand Up @@ -115,6 +116,7 @@
template_map_fn=dict(
type=template_map_fn_factory, template=prompt_template),
max_length=max_length,
per_image_length=per_image_length,
pad_image_to_square=True)

train_dataloader = dict(
Expand Down
4 changes: 3 additions & 1 deletion ..._large_p14_336/pretrain/llava_internlm_chat_7b_clip_vit_large_p14_336_e1_gpu8_pretrain.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,7 +29,8 @@
data_path = data_root + 'LLaVA-Pretrain/blip_laion_cc_sbu_558k.json'
image_folder = data_root + 'LLaVA-Pretrain/images'
prompt_template = PROMPT_TEMPLATE.internlm_chat
max_length = int(2048 - (336 / 14)**2)
max_length = 2048
per_image_length = int(336 / 14)**2 # inferred from the visual encoder

# Scheduler & Optimizer
batch_size = 32 # per_device
Expand Down Expand Up @@ -102,6 +103,7 @@
template_map_fn=dict(
type=template_map_fn_factory, template=prompt_template),
max_length=max_length,
per_image_length=per_image_length,
pad_image_to_square=False)

train_dataloader = dict(
Expand Down
4 changes: 3 additions & 1 deletion ...it_large_p14_336/finetune/llava_vicuna_13b_v15_clip_vit_large_p14_336_e1_gpu8_finetune.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,7 +30,8 @@
data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
image_folder = data_root + 'llava_images'
prompt_template = PROMPT_TEMPLATE.vicuna
max_length = int(2048 - (336 / 14)**2)
max_length = 2048
per_image_length = int(336 / 14)**2 # inferred from the visual encoder

# Scheduler & Optimizer
batch_size = 16 # per_device
Expand Down Expand Up @@ -95,6 +96,7 @@
template_map_fn=dict(
type=template_map_fn_factory, template=prompt_template),
max_length=max_length,
per_image_length=per_image_length,
pad_image_to_square=True)

train_dataloader = dict(
Expand Down
4 changes: 3 additions & 1 deletion ...4_336/finetune/llava_vicuna_13b_v15_qlora_clip_vit_large_p14_336_lora_e1_gpu8_finetune.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,7 +32,8 @@
data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
image_folder = data_root + 'llava_images'
prompt_template = PROMPT_TEMPLATE.vicuna
max_length = int(2048 - (336 / 14)**2)
max_length = 2048
per_image_length = int(336 / 14)**2 # inferred from the visual encoder

# Scheduler & Optimizer
batch_size = 16 # per_device
Expand Down Expand Up @@ -115,6 +116,7 @@
template_map_fn=dict(
type=template_map_fn_factory, template=prompt_template),
max_length=max_length,
per_image_length=per_image_length,
pad_image_to_square=True)

train_dataloader = dict(
Expand Down
4 changes: 3 additions & 1 deletion ...it_large_p14_336/pretrain/llava_vicuna_13b_v15_clip_vit_large_p14_336_e1_gpu8_pretrain.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,7 +29,8 @@
data_path = data_root + 'LLaVA-Pretrain/blip_laion_cc_sbu_558k.json'
image_folder = data_root + 'LLaVA-Pretrain/images'
prompt_template = PROMPT_TEMPLATE.vicuna
max_length = int(2048 - (336 / 14)**2)
max_length = 2048
per_image_length = int(336 / 14)**2 # inferred from the visual encoder

# Scheduler & Optimizer
batch_size = 32 # per_device
Expand Down Expand Up @@ -102,6 +103,7 @@
template_map_fn=dict(
type=template_map_fn_factory, template=prompt_template),
max_length=max_length,
per_image_length=per_image_length,
pad_image_to_square=False)

train_dataloader = dict(
Expand Down
4 changes: 3 additions & 1 deletion ...vit_large_p14_336/finetune/llava_vicuna_7b_v15_clip_vit_large_p14_336_e1_gpu8_finetune.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,7 +30,8 @@
data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
image_folder = data_root + 'llava_images'
prompt_template = PROMPT_TEMPLATE.vicuna
max_length = int(2048 - (336 / 14)**2)
max_length = 2048
per_image_length = int(336 / 14)**2 # inferred from the visual encoder

# Scheduler & Optimizer
batch_size = 16 # per_device
Expand Down Expand Up @@ -95,6 +96,7 @@
template_map_fn=dict(
type=template_map_fn_factory, template=prompt_template),
max_length=max_length,
per_image_length=per_image_length,
pad_image_to_square=True)

train_dataloader = dict(
Expand Down
4 changes: 3 additions & 1 deletion ...14_336/finetune/llava_vicuna_7b_v15_qlora_clip_vit_large_p14_336_lora_e1_gpu8_finetune.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,7 +32,8 @@
data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
image_folder = data_root + 'llava_images'
prompt_template = PROMPT_TEMPLATE.vicuna
max_length = int(2048 - (336 / 14)**2)
max_length = 2048
per_image_length = int(336 / 14)**2 # inferred from the visual encoder

# Scheduler & Optimizer
batch_size = 16 # per_device
Expand Down Expand Up @@ -115,6 +116,7 @@
template_map_fn=dict(
type=template_map_fn_factory, template=prompt_template),
max_length=max_length,
per_image_length=per_image_length,
pad_image_to_square=True)

train_dataloader = dict(
Expand Down
4 changes: 3 additions & 1 deletion ...vit_large_p14_336/pretrain/llava_vicuna_7b_v15_clip_vit_large_p14_336_e1_gpu8_pretrain.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,7 +29,8 @@
data_path = data_root + 'LLaVA-Pretrain/blip_laion_cc_sbu_558k.json'
image_folder = data_root + 'LLaVA-Pretrain/images'
prompt_template = PROMPT_TEMPLATE.vicuna
max_length = int(2048 - (336 / 14)**2)
max_length = 2048
per_image_length = int(336 / 14)**2 # inferred from the visual encoder

# Scheduler & Optimizer
batch_size = 32 # per_device
Expand Down Expand Up @@ -102,6 +103,7 @@
template_map_fn=dict(
type=template_map_fn_factory, template=prompt_template),
max_length=max_length,
per_image_length=per_image_length,
pad_image_to_square=False)

train_dataloader = dict(
Expand Down
19 changes: 5 additions & 14 deletions xtuner/dataset/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,19 +17,10 @@
warnings.simplefilter(action='ignore', category=FutureWarning)

__all__ = [
'process_hf_dataset',
'ConcatDataset',
'MOSSSFTDataset',
'process_ms_dataset',
'LLaVADataset',
'expand2square',
'decode_base64_to_image',
'load_image',
'process_ms_dataset',
'process_hf_dataset', 'ConcatDataset', 'MOSSSFTDataset',
'process_ms_dataset', 'LLaVADataset', 'expand2square',
'decode_base64_to_image', 'load_image', 'process_ms_dataset',
'load_intern_repo_tokenized_dataset',
'load_intern_repo_untokenized_dataset',
'build_packed_dataset',
'RefCOCOJsonDataset',
'RefCOCOJsonEvalDataset',
'InvRefCOCOJsonDataset',
'load_intern_repo_untokenized_dataset', 'build_packed_dataset',
'RefCOCOJsonDataset', 'RefCOCOJsonEvalDataset', 'InvRefCOCOJsonDataset'
]
5 changes: 4 additions & 1 deletion xtuner/dataset/collate_fns/defalut_collate_fn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,7 +33,10 @@ def default_collate_fn(instances: Sequence[Dict],
indexes.append(torch.LongTensor(example['indexes']))

if has_image:
pixel_values.append(example['pixel_values'])
if isinstance(example['pixel_values'], list):
pixel_values.extend(example['pixel_values'])
else:
pixel_values.append(example['pixel_values'])

if len(instances) > 1:
input_ids = pad_sequence(
Expand Down