423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-08-02 03:32:50 +08:00
Code Issues Packages Projects Releases Wiki Activity

[data] update vlm args (#6976)

Browse Source 
Former-commit-id: 3da2cc2710c9b13ab450815a92fff14b03251984
This commit is contained in:
hoshi-hiyouga 2025-02-18 02:12:51 +08:00 committed by GitHub
parent b1d31ff0f9
commit 1f4a0b11ba
6 changed files with 33 additions and 33 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
examples/train_full/qwen2vl_full_sft.yaml
View File

@ -1,7 +1,7 @@
### model ### model
model_name_or_path: Qwen/Qwen2-VL-7B-Instruct model_name_or_path: Qwen/Qwen2-VL-7B-Instruct
image_max_resolution: 262144 image_max_pixels: 262144
video_max_resolution: 16384 video_max_pixels: 16384
trust_remote_code: true trust_remote_code: true
### method ### method

4
examples/train_lora/qwen2vl_lora_dpo.yaml
View File

@ -1,7 +1,7 @@
### model ### model
model_name_or_path: Qwen/Qwen2-VL-7B-Instruct model_name_or_path: Qwen/Qwen2-VL-7B-Instruct
image_max_resolution: 262144 image_max_pixels: 262144
video_max_resolution: 16384 video_max_pixels: 16384
trust_remote_code: true trust_remote_code: true
### method ### method

4
examples/train_lora/qwen2vl_lora_sft.yaml
View File

@ -1,7 +1,7 @@
### model ### model
model_name_or_path: Qwen/Qwen2-VL-7B-Instruct model_name_or_path: Qwen/Qwen2-VL-7B-Instruct
image_max_resolution: 262144 image_max_pixels: 262144
video_max_resolution: 16384 video_max_pixels: 16384
trust_remote_code: true trust_remote_code: true
### method ### method

38
src/llamafactory/data/mm_plugin.py
View File

@ -105,18 +105,18 @@ class MMPluginMixin:
) )
def _preprocess_image( def _preprocess_image(
self, image: "ImageObject", image_max_resolution: int, image_min_resolution: int, **kwargs self, image: "ImageObject", image_max_pixels: int, image_min_pixels: int, **kwargs
) -> "ImageObject": ) -> "ImageObject":
r""" r"""
Pre-processes a single image. Pre-processes a single image.
""" """
if (image.width * image.height) > image_max_resolution: if (image.width * image.height) > image_max_pixels:
resize_factor = math.sqrt(image_max_resolution / (image.width * image.height)) resize_factor = math.sqrt(image_max_pixels / (image.width * image.height))
width, height = int(image.width * resize_factor), int(image.height * resize_factor) width, height = int(image.width * resize_factor), int(image.height * resize_factor)
image = image.resize((width, height), resample=Image.Resampling.NEAREST) image = image.resize((width, height), resample=Image.Resampling.NEAREST)
if (image.width * image.height) < image_min_resolution: if (image.width * image.height) < image_min_pixels:
resize_factor = math.sqrt(image_min_resolution / (image.width * image.height)) resize_factor = math.sqrt(image_min_pixels / (image.width * image.height))
width, height = int(image.width * resize_factor), int(image.height * resize_factor) width, height = int(image.width * resize_factor), int(image.height * resize_factor)
image = image.resize((width, height), resample=Image.Resampling.NEAREST) image = image.resize((width, height), resample=Image.Resampling.NEAREST)
@ -225,16 +225,16 @@ class MMPluginMixin:
if len(images) != 0: if len(images) != 0:
images = self._regularize_images( images = self._regularize_images(
images, images,
image_max_resolution=getattr(processor, "image_max_resolution", 768 * 768), image_max_pixels=getattr(processor, "image_max_pixels", 768 * 768),
image_min_resolution=getattr(processor, "image_min_resolution", 32 * 32), image_min_pixels=getattr(processor, "image_min_pixels", 32 * 32),
) )
mm_inputs.update(image_processor(images, return_tensors="pt")) mm_inputs.update(image_processor(images, return_tensors="pt"))
if len(videos) != 0: if len(videos) != 0:
videos = self._regularize_videos( videos = self._regularize_videos(
videos, videos,
image_max_resolution=getattr(processor, "video_max_resolution", 256 * 256), image_max_pixels=getattr(processor, "video_max_pixels", 256 * 256),
image_min_resolution=getattr(processor, "video_min_resolution", 16 * 16), image_min_pixels=getattr(processor, "video_min_pixels", 16 * 16),
video_fps=getattr(processor, "video_fps", 2.0), video_fps=getattr(processor, "video_fps", 2.0),
video_maxlen=getattr(processor, "video_maxlen", 128), video_maxlen=getattr(processor, "video_maxlen", 128),
) )
@ -616,8 +616,8 @@ class MiniCPMVPlugin(BasePlugin):
if len(images) != 0: if len(images) != 0:
images = self._regularize_images( images = self._regularize_images(
images, images,
image_max_resolution=getattr(processor, "image_max_resolution", 768 * 768), image_max_pixels=getattr(processor, "image_max_pixels", 768 * 768),
image_min_resolution=getattr(processor, "image_min_resolution", 32 * 32), image_min_pixels=getattr(processor, "image_min_pixels", 32 * 32),
) )
if "valid_image_nums_ls" in kwargs: if "valid_image_nums_ls" in kwargs:
valid_image_nums_ls = kwargs["valid_image_nums_ls"] valid_image_nums_ls = kwargs["valid_image_nums_ls"]
@ -637,8 +637,8 @@ class MiniCPMVPlugin(BasePlugin):
if len(videos) != 0: if len(videos) != 0:
videos = self._regularize_videos( videos = self._regularize_videos(
videos, videos,
image_max_resolution=getattr(processor, "video_max_resolution", 256 * 256), image_max_pixels=getattr(processor, "video_max_pixels", 256 * 256),
image_min_resolution=getattr(processor, "video_min_resolution", 16 * 16), image_min_pixels=getattr(processor, "video_min_pixels", 16 * 16),
video_fps=getattr(processor, "video_fps", 2.0), video_fps=getattr(processor, "video_fps", 2.0),
video_maxlen=getattr(processor, "video_maxlen", 128), video_maxlen=getattr(processor, "video_maxlen", 128),
) )
@ -788,8 +788,8 @@ class MllamaPlugin(BasePlugin):
image_processor: "BaseImageProcessor" = getattr(processor, "image_processor") image_processor: "BaseImageProcessor" = getattr(processor, "image_processor")
images = self._regularize_images( images = self._regularize_images(
images, images,
image_max_resolution=getattr(processor, "image_max_resolution", 768 * 768), image_max_pixels=getattr(processor, "image_max_pixels", 768 * 768),
image_min_resolution=getattr(processor, "image_min_resolution", 32 * 32), image_min_pixels=getattr(processor, "image_min_pixels", 32 * 32),
) )
batch_images = [] batch_images = []
for image_length in imglens: for image_length in imglens:
@ -1082,16 +1082,16 @@ class Qwen2vlPlugin(BasePlugin):
if len(images) != 0: if len(images) != 0:
images = self._regularize_images( images = self._regularize_images(
images, images,
image_max_resolution=getattr(processor, "image_max_resolution", 768 * 768), image_max_pixels=getattr(processor, "image_max_pixels", 768 * 768),
image_min_resolution=getattr(processor, "image_min_resolution", 32 * 32), image_min_pixels=getattr(processor, "image_min_pixels", 32 * 32),
) )
mm_inputs.update(image_processor(images, return_tensors="pt")) mm_inputs.update(image_processor(images, return_tensors="pt"))
if len(videos) != 0: if len(videos) != 0:
videos, fps_per_video = self._regularize_videos( videos, fps_per_video = self._regularize_videos(
videos, videos,
image_max_resolution=getattr(processor, "video_max_resolution", 256 * 256), image_max_pixels=getattr(processor, "video_max_pixels", 256 * 256),
image_min_resolution=getattr(processor, "video_min_resolution", 16 * 16), image_min_pixels=getattr(processor, "video_min_pixels", 16 * 16),
video_fps=getattr(processor, "video_fps", 2.0), video_fps=getattr(processor, "video_fps", 2.0),
video_maxlen=getattr(processor, "video_maxlen", 128), video_maxlen=getattr(processor, "video_maxlen", 128),
) )

8
src/llamafactory/hparams/model_args.py
View File

@ -58,19 +58,19 @@ class ProcessorArguments:
Arguments pertaining to the image processor. Arguments pertaining to the image processor.
""" """
image_max_resolution: int = field( image_max_pixels: int = field(
default=768 * 768, default=768 * 768,
metadata={"help": "The maximum number of pixels of image inputs."}, metadata={"help": "The maximum number of pixels of image inputs."},
) )
image_min_resolution: int = field( image_min_pixels: int = field(
default=32 * 32, default=32 * 32,
metadata={"help": "The minimum number of pixels of image inputs."}, metadata={"help": "The minimum number of pixels of image inputs."},
) )
video_max_resolution: int = field( video_max_pixels: int = field(
default=256 * 256, default=256 * 256,
metadata={"help": "The maximum number of pixels of video inputs."}, metadata={"help": "The maximum number of pixels of video inputs."},
) )
video_min_resolution: int = field( video_min_pixels: int = field(
default=16 * 16, default=16 * 16,
metadata={"help": "The minimum number of pixels of video inputs."}, metadata={"help": "The minimum number of pixels of video inputs."},
) )

8
src/llamafactory/model/patcher.py
View File

@ -80,10 +80,10 @@ def patch_processor(
if getattr(config, "vision_config", None) is not None: # visual models if getattr(config, "vision_config", None) is not None: # visual models
setattr(processor, "image_seqlen", get_image_seqlen(config)) setattr(processor, "image_seqlen", get_image_seqlen(config))
setattr(processor, "patch_size", get_patch_size(config, processor)) setattr(processor, "patch_size", get_patch_size(config, processor))
setattr(processor, "image_max_resolution", model_args.image_max_resolution) setattr(processor, "image_max_pixels", model_args.image_max_pixels)
setattr(processor, "image_min_resolution", model_args.image_min_resolution) setattr(processor, "image_min_pixels", model_args.image_min_pixels)
setattr(processor, "video_max_resolution", model_args.video_max_resolution) setattr(processor, "video_max_pixels", model_args.video_max_pixels)
setattr(processor, "video_min_resolution", model_args.video_min_resolution) setattr(processor, "video_min_pixels", model_args.video_min_pixels)
setattr(processor, "video_fps", model_args.video_fps) setattr(processor, "video_fps", model_args.video_fps)
setattr(processor, "video_maxlen", model_args.video_maxlen) setattr(processor, "video_maxlen", model_args.video_maxlen)
setattr(processor, "vision_feature_select_strategy", get_vision_feature_select_strategy(config, processor)) setattr(processor, "vision_feature_select_strategy", get_vision_feature_select_strategy(config, processor))