LLaVa-NeXT-Video
Overview
The LLaVa-NeXT-Video model was proposed in LLaVA-NeXT: A Strong Zero-shot Video Understanding Model by Yuanhan Zhang, Bo Li, Haotian Liu, Yong Jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, Chunyuan Li. LLaVa-NeXT-Video improves upon LLaVa-NeXT by fine-tuning on a mix if video and image dataset thus increasing the model’s performance on videos.
LLaVA-NeXT surprisingly has strong performance in understanding video content in zero-shot fashion with the AnyRes technique that it uses. The AnyRes technique naturally represents a high-resolution image into multiple images. This technique is naturally generalizable to represent videos because videos can be considered as a set of frames (similar to a set of images in LLaVa-NeXT). The current version of LLaVA-NeXT makes use of AnyRes and trains with supervised fine-tuning (SFT) on top of LLaVA-Next on video data to achieves better video understanding capabilities.The model is a current SOTA among open-source models on VideoMME bench.
The introduction from the blog is the following:
On January 30, 2024, we released LLaVA-NeXT, an open-source Large Multimodal Model (LMM) that has been trained exclusively on text-image data. With the proposed AnyRes technique, it boosts capabilities in reasoning, OCR, and world knowledge, demonstrating remarkable performance across a spectrum of image-based multimodal understanding tasks, and even exceeding Gemini-Pro on several image benchmarks, e.g. MMMU and MathVista.
**In today’s exploration, we delve into the performance of LLaVA-NeXT within the realm of video understanding tasks. We reveal that LLaVA-NeXT surprisingly has strong performance in understanding video content. The current version of LLaVA-NeXT for videos has several improvements:
- Zero-shot video representation capabilities with AnyRes: The AnyRes technique naturally represents a high-resolution image into multiple images that a pre-trained VIT is able to digest, and forms them into a concantenated sequence. This technique is naturally generalizable to represent videos (consisting of multiple frames), allowing the image-only-trained LLaVA-Next model to perform surprisingly well on video tasks. Notably, this is the first time that LMMs show strong zero-shot modality transfer ability.
- Inference with length generalization improves on longer videos. The linear scaling technique enables length generalization, allowing LLaVA-NeXT to effectively handle long-video beyond the limitation of the “max_token_length” of the LLM.
- Strong video understanding ability. (1) LLaVA-Next-Image, which combines the above two techniques, yields superior zero-shot performance than open-source LMMs tuned on videos. (2) LLaVA-Next-Video, further supervised fine-tuning (SFT) LLaVA-Next-Image on video data, achieves better video understanding capabilities compared to LLaVA-Next-Image. (3) LLaVA-Next-Video-DPO, which aligns the model response with AI feedback using direct preference optimization (DPO), showing significant performance boost.
- Efficient deployment and inference with SGLang. It allows 5x faster inference on video tasks, allowing more scalable serving such as million-level video re-captioning. See instructions in our repo.**
This model was contributed by RaushanTurganbay. The original code can be found here.
Usage tips
- We advise users to use
padding_side="left"
when computing batched generation as it leads to more accurate results. Simply make sure to callprocessor.tokenizer.padding_side = "left"
before generating.
- Llava-Next uses different number of patches for images and thus has to pad the inputs inside modeling code, aside from the padding done when processing the inputs. The default setting is “left-padding” if model is in
eval()
mode, otherwise “right-padding”.
- Note that each checkpoint has been trained with a specific prompt format, depending on which large language model (LLM) was used. You can use tokenizer’s
apply_chat_template
to format your prompts correctly. Below is an example of how to do that.
We will use LLaVA-NeXT-Video-7B-hf and a conversation history of videos and images. Each content field has to be a list of dicts, as follows:
from transformers import LlavaNextVideoProcessor
processor = LlavaNextVideoProcessor.from_pretrained("llava-hf/LLaVA-NeXT-Video-7B-hf")
conversation = [
{
"role": "system",
"content": [
{"type": "text", "text": "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions."},
],
},
{
"role": "user",
"content": [
{"type": "text", "text": "What’s shown in this image?"},
{"type": "image"},
],
},
{
"role": "assistant",
"content": [{"type": "text", "text": "This image shows a red stop sign."},]
},
{
"role": "user",
"content": [
{"type": "text", "text": "Why is this video funny?"},
{"type": "video"},
],
},
]
text_prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
# Note that the template simply formats your prompt, you still have to tokenize it and obtain pixel values for your visuals
print(text_prompt)
Usage example
Single Media Mode
The model can accept both images and videos as input. Here’s an example code for inference in half-precision (torch.float16
):
import av
import torch
import numpy as np
from transformers import LlavaNextVideoForConditionalGeneration, LlavaNextVideoProcessor
def read_video_pyav(container, indices):
'''
Decode the video with PyAV decoder.
Args:
container (`av.container.input.InputContainer`): PyAV container.
indices (`List[int]`): List of frame indices to decode.
Returns:
result (np.ndarray): np array of decoded frames of shape (num_frames, height, width, 3).
'''
frames = []
container.seek(0)
start_index = indices[0]
end_index = indices[-1]
for i, frame in enumerate(container.decode(video=0)):
if i > end_index:
break
if i >= start_index and i in indices:
frames.append(frame)
return np.stack([x.to_ndarray(format="rgb24") for x in frames])
# Load the model in half-precision
model = LlavaNextVideoForConditionalGeneration.from_pretrained("llava-hf/LLaVA-NeXT-Video-7B-hf", torch_dtype=torch.float16, device_map="auto")
processor = LlavaNextVideoProcessor.from_pretrained("llava-hf/LLaVA-NeXT-Video-7B-hf")
# Load the video as an np.array, sampling uniformly 8 frames (can sample more for longer videos)
video_path = hf_hub_download(repo_id="raushan-testing-hf/videos-test", filename="sample_demo_1.mp4", repo_type="dataset")
container = av.open(video_path)
total_frames = container.streams.video[0].frames
indices = np.arange(0, total_frames, total_frames / 8).astype(int)
video = read_video_pyav(container, indices)
conversation = [
{
"role": "user",
"content": [
{"type": "text", "text": "Why is this video funny?"},
{"type": "video"},
],
},
]
prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
inputs = processor(text=prompt, videos=video, return_tensors="pt")
out = model.generate(**inputs, max_new_tokens=60)
processor.batch_decode(out, skip_special_tokens=True, clean_up_tokenization_spaces=True)
Mixed Media Mode
The model can also generate from an interleaved image-video inputs. However note, that it was not trained in interleaved image-video setting which might affect the performance. Below is an example usage for mixed media input, add the following lines to the above code snippet:
from PIL import Image
import requests
# Generate from image and video mixed inputs
# Load and image and write a new prompt
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw)
conversation = [
{
"role": "user",
"content": [
{"type": "text", "text": "How many cats are there in the image?"},
{"type": "image"},
],
},
{
"role": "assistant",
"content": [{"type": "text", "text": "There are two cats"}],
},
{
"role": "user",
"content": [
{"type": "text", "text": "Why is this video funny?"},
{"type": "video"},
],
},
]
prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
inputs = processor(text=prompt, images=image, videos=clip, padding=True, return_tensors="pt")
# Generate
generate_ids = model.generate(**inputs, max_length=50)
processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True)
Model optimization
Quantization using Bitsandbytes for memory efficiency
The model can be loaded in lower bits, significantly reducing memory burden while maintaining the performance of the original model. This allows for efficient deployment on resource-constrained cases.
First make sure to install bitsandbytes by running pip install bitsandbytes
and to have access to a CUDA compatible GPU device. Load the quantized model by simply adding BitsAndBytesConfig
as shown below:
from transformers import LlavaNextVideoForConditionalGeneration, LlavaNextVideoProcessor
# specify how to quantize the model
quantization_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.float16,
)
model = LlavaNextVideoForConditionalGeneration.from_pretrained("llava-hf/LLaVA-NeXT-Video-7B-hf", quantization_config=quantization_config, device_map="auto")
Flash-Attention 2 to speed-up generation
Additionally, we can greatly speed-up model inference by using Flash Attention, which is a faster implementation of the attention mechanism used inside the model.
First, make sure to install the latest version of Flash Attention 2:
pip install -U flash-attn --no-build-isolation
Also, you should have a hardware that is compatible with Flash-Attention 2. Read more about it in the official documentation of the flash attention repository. FlashAttention-2 can only be used when a model is loaded in torch.float16
or torch.bfloat16
.
To load and run a model using Flash Attention-2, simply add attn_implementation="flash_attention_2"
when loading the model as follows:
from transformers import LlavaNextVideoForConditionalGeneration
model = LlavaNextVideoForConditionalGeneration.from_pretrained(
"llava-hf/LLaVA-NeXT-Video-7B-hf",
torch_dtype=torch.float16,
attn_implementation="flash_attention_2",
).to(0)
LlavaNextVideoConfig
class transformers.LlavaNextVideoConfig
< source >( vision_config = None text_config = None ignore_index = -100 image_token_index = 32001 projector_hidden_act = 'gelu' vision_feature_select_strategy = 'default' vision_feature_layer = -2 image_grid_pinpoints = None tie_word_embeddings = False video_token_index = 32000 spatial_pool_mode = 'average' spatial_pool_stride = 2 **kwargs )
Parameters
- vision_config (
Union[AutoConfig, dict]
, optional, defaults toCLIPVisionConfig
) — The config object or dictionary of the vision backbone. - text_config (
Union[AutoConfig, dict]
, optional, defaults toLlamaConfig
) — The config object or dictionary of the text backbone. - ignore_index (
int
, optional, defaults to -100) — The ignore index for the loss function. - image_token_index (
int
, optional, defaults to 32001) — The image token index to encode the image prompt. - projector_hidden_act (
str
, optional, defaults to"gelu"
) — The activation function used by the multimodal projector. - vision_feature_select_strategy (
str
, optional, defaults to"default"
) — The feature selection strategy used to select the vision feature from the vision backbone. Can be one of"default"
or"full"
. If"default"
, the CLS token is removed from the vision features. If"full"
, the full vision features are used. - vision_feature_layer (
int
, optional, defaults to -2) — The index of the layer to select the vision feature. - image_grid_pinpoints (
List
, optional, defaults to[[336, 672], [672, 336], [672, 672], [1008, 336], [336, 1008]]
) — A list of possible resolutions to use for processing high resolution images. Each item in the list should be a tuple or list of the form(height, width)
. - tie_word_embeddings (
bool
, optional, defaults toFalse
) — Whether the model’s input and output word embeddings should be tied. - video_token_index (
int
, optional, defaults to 32000) — The video token index to encode the image prompt. - spatial_pool_mode (
str
, optional, defaults to"average"
) — Pooling mode to use for videos. Can be “average”, “max” or “conv”. - spatial_pool_stride (
int
, optional, defaults to 2) — Stride used in the pooling layer for videos.
This is the configuration class to store the configuration of a LlavaNextVideoForConditionalGeneration. It is used to instantiate an Llava-NeXT model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of the llava-hf/LLaVA-NeXT-Video-7B-hf model. Configuration objects inherit from PretrainedConfig and can be used to control the model outputs. Read the documentation from PretrainedConfig for more information.
Example:
>>> from transformers import LlavaNextVideoForConditionalGeneration, LlavaNextVideoConfig, CLIPVisionConfig, LlamaConfig
>>> # Initializing a CLIP-vision config
>>> vision_config = CLIPVisionConfig()
>>> # Initializing a Llama config
>>> text_config = LlamaConfig()
>>> configuration = LlavaNextVideoConfig(vision_config, text_config)
>>> model = LlavaNextVideoForConditionalGeneration(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
LlavaNextVideoProcessor
class transformers.LlavaNextVideoProcessor
< source >( video_processor = None image_processor = None tokenizer = None chat_template = None **kwargs )
Parameters
- video_processor (LlavaNextVideoImageProcessor, optional) — The video processor is a required input.
- image_processor (LlavaNextImageProcessor, optional) — The image processor is a required input.
- tokenizer (LlamaTokenizerFast, optional) — The tokenizer is a required input.
- chat_template (
str
, optional) — Jinja chat template that will be used in tokenizer’sapply_chat_template
Constructs a LLaVa-NeXT-Video processor which wraps a LLaVa-NeXT image processor, LLaVa-NeXT-Video video processor and a LLaMa tokenizer into a single processor.
LlavaNextVideoProcessor offers all the functionalities of LlavaNextImageProcessor, LlavaNextVideoImageProcessor and
LlamaTokenizerFast. See the __call__()
and decode() for more information.
This method forwards all its arguments to LlamaTokenizerFast’s batch_decode(). Please refer to the docstring of this method for more information.
This method forwards all its arguments to LlamaTokenizerFast’s decode(). Please refer to the docstring of this method for more information.
LlavaNextVideoImageProcessor
class transformers.LlavaNextVideoImageProcessor
< source >( do_resize: bool = True size: Dict = None image_grid_pinpoints: List = None resample: Resampling = <Resampling.BICUBIC: 3> do_center_crop: bool = True crop_size: Dict = None do_rescale: bool = True rescale_factor: Union = 0.00392156862745098 do_normalize: bool = True image_mean: Union = None image_std: Union = None do_convert_rgb: bool = True **kwargs )
Parameters
- do_resize (
bool
, optional, defaults toTrue
) — Whether to resize the image’s (height, width) dimensions to the specifiedsize
. Can be overridden bydo_resize
in thepreprocess
method. - size (
Dict[str, int]
optional, defaults to{"shortest_edge" -- 224}
): Size of the image after resizing. The shortest edge of the image is resized to size[“shortest_edge”], with the longest edge resized to keep the input aspect ratio. Can be overridden bysize
in thepreprocess
method. - image_grid_pinpoints (
List
optional, defaults to[[672, 336], [336, 672], [672, 672], [336, 1008], [1008, 336]]
) — A list of possible resolutions to use for processing high resolution images. The best resolution is selected based on the original size of the image. Can be overridden byimage_grid_pinpoints
in thepreprocess
method. Not used for processinf videos. - resample (
PILImageResampling
, optional, defaults toResampling.BICUBIC
) — Resampling filter to use if resizing the image. Can be overridden byresample
in thepreprocess
method. - do_center_crop (
bool
, optional, defaults toTrue
) — Whether to center crop the image to the specifiedcrop_size
. Can be overridden bydo_center_crop
in thepreprocess
method. - crop_size (
Dict[str, int]
optional, defaults to 224) — Size of the output image after applyingcenter_crop
. Can be overridden bycrop_size
in thepreprocess
method. - do_rescale (
bool
, optional, defaults toTrue
) — Whether to rescale the image by the specified scalerescale_factor
. Can be overridden bydo_rescale
in thepreprocess
method. - rescale_factor (
int
orfloat
, optional, defaults to1/255
) — Scale factor to use if rescaling the image. Can be overridden byrescale_factor
in thepreprocess
method. - do_normalize (
bool
, optional, defaults toTrue
) — Whether to normalize the image. Can be overridden bydo_normalize
in thepreprocess
method. - image_mean (
float
orList[float]
, optional, defaults to[0.48145466, 0.4578275, 0.40821073]
) — Mean to use if normalizing the image. This is a float or list of floats the length of the number of channels in the image. Can be overridden by theimage_mean
parameter in thepreprocess
method. - image_std (
float
orList[float]
, optional, defaults to[0.26862954, 0.26130258, 0.27577711]
) — Standard deviation to use if normalizing the image. This is a float or list of floats the length of the number of channels in the image. Can be overridden by theimage_std
parameter in thepreprocess
method. Can be overridden by theimage_std
parameter in thepreprocess
method. - do_convert_rgb (
bool
, optional, defaults toTrue
) — Whether to convert the image to RGB.
Constructs a LLaVa-NeXT-Video video processor. Based on CLIPImageProcessor with incorporation of processing each video frame.
preprocess
< source >( images: Union do_resize: bool = None size: Dict = None resample: Resampling = None do_center_crop: bool = None crop_size: int = None do_rescale: bool = None rescale_factor: float = None do_normalize: bool = None image_mean: Union = None image_std: Union = None do_convert_rgb: bool = None return_tensors: Union = None data_format: Optional = <ChannelDimension.FIRST: 'channels_first'> input_data_format: Union = None )
Parameters
- images (
VideoInput
) — Videos to preprocess. Expects a single or batch of videos with pixel values ranging from 0 to 255. If passing in images with pixel values between 0 and 1, setdo_rescale=False
. - do_resize (
bool
, optional, defaults toself.do_resize
) — Whether to resize the video. - size (
Dict[str, int]
, optional, defaults toself.size
) — Size of the video after resizing. Shortest edge of the video is resized to size[“shortest_edge”], with the longest edge resized to keep the input aspect ratio. - resample (
int
, optional, defaults toself.resample
) — Resampling filter to use if resizing the video. This can be one of the enumPILImageResampling
. Only has an effect ifdo_resize
is set toTrue
. - do_center_crop (
bool
, optional, defaults toself.do_center_crop
) — Whether to center crop the video. - crop_size (
Dict[str, int]
, optional, defaults toself.crop_size
) — Size of the center crop. Only has an effect ifdo_center_crop
is set toTrue
. - do_rescale (
bool
, optional, defaults toself.do_rescale
) — Whether to rescale the video. - rescale_factor (
float
, optional, defaults toself.rescale_factor
) — Rescale factor to rescale the video by ifdo_rescale
is set toTrue
. - do_normalize (
bool
, optional, defaults toself.do_normalize
) — Whether to normalize the video. - image_mean (
float
orList[float]
, optional, defaults toself.image_mean
) — Frame mean to use for normalization. Only has an effect ifdo_normalize
is set toTrue
. - image_std (
float
orList[float]
, optional, defaults toself.image_std
) — Frame standard deviation to use for normalization. Only has an effect ifdo_normalize
is set toTrue
. - do_convert_rgb (
bool
, optional, defaults toself.do_convert_rgb
) — Whether to convert the video to RGB. - return_tensors (
str
orTensorType
, optional) — The type of tensors to return. Can be one of:- Unset: Return a list of
np.ndarray
. TensorType.TENSORFLOW
or'tf'
: Return a batch of typetf.Tensor
.TensorType.PYTORCH
or'pt'
: Return a batch of typetorch.Tensor
.TensorType.NUMPY
or'np'
: Return a batch of typenp.ndarray
.TensorType.JAX
or'jax'
: Return a batch of typejax.numpy.ndarray
.
- Unset: Return a list of
- data_format (
ChannelDimension
orstr
, optional, defaults toChannelDimension.FIRST
) — The channel dimension format for the output image. Can be one of:"channels_first"
orChannelDimension.FIRST
: image in (num_channels, height, width) format."channels_last"
orChannelDimension.LAST
: image in (height, width, num_channels) format.- Unset: Use the channel dimension format of the input image.
- input_data_format (
ChannelDimension
orstr
, optional) — The channel dimension format for the input image. If unset, the channel dimension format is inferred from the input image. Can be one of:"channels_first"
orChannelDimension.FIRST
: image in (num_channels, height, width) format."channels_last"
orChannelDimension.LAST
: image in (height, width, num_channels) format."none"
orChannelDimension.NONE
: image in (height, width) format.
resize
< source >( image: ndarray size: Dict resample: Resampling = <Resampling.BICUBIC: 3> data_format: Union = None input_data_format: Union = None **kwargs )
Parameters
- image (
np.ndarray
) — Image to resize. - size (
Dict[str, int]
) — Size of the output image. - resample (
PILImageResampling
, optional, defaults toPILImageResampling.BICUBIC
) — Resampling filter to use when resiizing the image. - data_format (
str
orChannelDimension
, optional) — The channel dimension format of the image. If not provided, it will be the same as the input image. - input_data_format (
ChannelDimension
orstr
, optional) — The channel dimension format of the input image. If not provided, it will be inferred.
Resize an image. The shortest edge of the image is resized to size[“shortest_edge”], with the longest edge resized to keep the input aspect ratio.
LlavaNextVideoForConditionalGeneration
class transformers.LlavaNextVideoForConditionalGeneration
< source >( config: LlavaNextVideoConfig )
Parameters
- config (LlavaNextVideoConfig or
LlavaNextVideoVisionConfig
) — Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.
The LLAVA-NeXT model which consists of a vision backbone and a language model. This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)
This model is also a PyTorch torch.nn.Module subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.
forward
< source >( input_ids: LongTensor = None pixel_values: FloatTensor = None pixel_values_videos: FloatTensor = None image_sizes: Optional = None attention_mask: Optional = None position_ids: Optional = None past_key_values: Optional = None inputs_embeds: Optional = None vision_feature_layer: Optional = None vision_feature_select_strategy: Optional = None labels: Optional = None use_cache: Optional = None output_attentions: Optional = None output_hidden_states: Optional = None return_dict: Optional = None ) → transformers.models.llava_next_video.modeling_llava_next_video.LlavaNextVideoCausalLMOutputWithPast
or tuple(torch.FloatTensor)
Parameters
- input_ids (
torch.LongTensor
of shape(batch_size, sequence_length)
) — Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it.Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.
- pixel_values (
torch.FloatTensor
of shape `(batch_size, num_channels, image_size, image_size)) — The tensors corresponding to the input images. Pixel values can be obtained using AutoImageProcessor. See LlavaNextVideoImageProcessor.call() for details. LlavaProcessor uses LlavaNextVideoImageProcessor for processing images. - image_sizes (
torch.LongTensor
of shape(batch_size, 2)
, optional) — The sizes of the images in the batch, being (height, width) for each image. - attention_mask (
torch.Tensor
of shape(batch_size, sequence_length)
, optional) — Mask to avoid performing attention on padding token indices. Mask values selected in[0, 1]
:- 1 for tokens that are not masked,
- 0 for tokens that are masked.
Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.
If
past_key_values
is used, optionally only the lastdecoder_input_ids
have to be input (seepast_key_values
).If you want to change padding behavior, you should read
modeling_opt._prepare_decoder_attention_mask
and modify to your needs. See diagram 1 in the paper for more information on the default strategy.- 1 indicates the head is not masked,
- 0 indicates the head is masked.
- position_ids (
torch.LongTensor
of shape(batch_size, sequence_length)
, optional) — Indices of positions of each input sequence tokens in the position embeddings. Selected in the range[0, config.n_positions - 1]
. What are position IDs? - past_key_values (
tuple(tuple(torch.FloatTensor))
, optional, returned whenuse_cache=True
is passed or whenconfig.use_cache=True
) — Tuple oftuple(torch.FloatTensor)
of lengthconfig.n_layers
, with each tuple having 2 tensors of shape(batch_size, num_heads, sequence_length, embed_size_per_head)
) and 2 additional tensors of shape(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)
.Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention blocks) that can be used (see
past_key_values
input) to speed up sequential decoding.If
past_key_values
are used, the user can optionally input only the lastdecoder_input_ids
(those that don’t have their past key value states given to this model) of shape(batch_size, 1)
instead of alldecoder_input_ids
of shape(batch_size, sequence_length)
. - inputs_embeds (
torch.FloatTensor
of shape(batch_size, sequence_length, hidden_size)
, optional) — Optionally, instead of passinginput_ids
you can choose to directly pass an embedded representation. This is useful if you want more control over how to convertinput_ids
indices into associated vectors than the model’s internal embedding lookup matrix. - vision_feature_layer (
int
, optional, defaults to -2) — The index of the layer to select the vision feature. - vision_feature_select_strategy (
str
, optional, defaults to"default"
) — The feature selection strategy used to select the vision feature from the vision backbone. Can be one of"default"
or"full"
. If"default"
, the CLS token is removed from the vision features. If"full"
, the full vision features are used. - use_cache (
bool
, optional) — If set toTrue
,past_key_values
key value states are returned and can be used to speed up decoding (seepast_key_values
). - output_attentions (
bool
, optional) — Whether or not to return the attentions tensors of all attention layers. Seeattentions
under returned tensors for more detail. - output_hidden_states (
bool
, optional) — Whether or not to return the hidden states of all layers. Seehidden_states
under returned tensors for more detail. - return_dict (
bool
, optional) — Whether or not to return a ModelOutput instead of a plain tuple.Args — pixel_values_videos (
torch.FloatTensor
of shape(batch_size, num_frames, num_channels, image_size, image_size)): The tensors corresponding to the input videos. Pixel values can be obtained using [AutoImageProcessor](/docs/transformers/v4.44.2/en/model_doc/auto#transformers.AutoImageProcessor). See
LlavaNextVideoVideoProcessor.callfor details. [LlavaProcessor](/docs/transformers/v4.44.2/en/model_doc/llava#transformers.LlavaProcessor) uses
LlavaNextVideoVideoProcessorfor processing videos. labels (
torch.LongTensorof shape
(batch_size, sequence_length), *optional*): Labels for computing the masked language modeling loss. Indices should either be in
[0, …, config.vocab_size]or -100 (see
input_idsdocstring). Tokens with indices set to
-100are ignored (masked), the loss is only computed for the tokens with labels in
[0, …, config.vocab_size]`.
Returns
transformers.models.llava_next_video.modeling_llava_next_video.LlavaNextVideoCausalLMOutputWithPast
or tuple(torch.FloatTensor)
A transformers.models.llava_next_video.modeling_llava_next_video.LlavaNextVideoCausalLMOutputWithPast
or a tuple of
torch.FloatTensor
(if return_dict=False
is passed or when config.return_dict=False
) comprising various
elements depending on the configuration (LlavaNextVideoConfig) and inputs.
-
loss (
torch.FloatTensor
of shape(1,)
, optional, returned whenlabels
is provided) — Language modeling loss (for next-token prediction). -
logits (
torch.FloatTensor
of shape(batch_size, sequence_length, config.vocab_size)
) — Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax). -
past_key_values (
tuple(tuple(torch.FloatTensor))
, optional, returned whenuse_cache=True
is passed or whenconfig.use_cache=True
) — Tuple oftuple(torch.FloatTensor)
of lengthconfig.n_layers
, with each tuple having 2 tensors of shape(batch_size, num_heads, sequence_length, embed_size_per_head)
)Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
past_key_values
input) to speed up sequential decoding. -
hidden_states (
tuple(torch.FloatTensor)
, optional, returned whenoutput_hidden_states=True
is passed or whenconfig.output_hidden_states=True
) — Tuple oftorch.FloatTensor
(one for the output of the embeddings, if the model has an embedding layer, + one for the output of each layer) of shape(batch_size, sequence_length, hidden_size)
.Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
-
attentions (
tuple(torch.FloatTensor)
, optional, returned whenoutput_attentions=True
is passed or whenconfig.output_attentions=True
) — Tuple oftorch.FloatTensor
(one for each layer) of shape(batch_size, num_heads, sequence_length, sequence_length)
.Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
-
image_hidden_states (
tuple(torch.FloatTensor)
, optional) — Tuple oftorch.FloatTensor
(one for the output of the image embeddings,(batch_size, num_images, sequence_length, hidden_size)
.image_hidden_states of the model produced by the vision encoder, and optionally by the perceiver
The LlavaNextVideoForConditionalGeneration forward method, overrides the __call__
special method.
Although the recipe for forward pass needs to be defined within this function, one should call the Module
instance afterwards instead of this since the former takes care of running the pre and post processing steps while
the latter silently ignores them.
Example:
>>> from PIL import Image
>>> import requests
>>> import av
>>> from transformers import AutoProcessor, LlavaNextVideoForConditionalGeneration
>>> def read_video_pyav(container, indices):
... '''
... Decode the video with PyAV decoder.
... Args:
... container (`av.container.input.InputContainer`): PyAV container.
... indices (`List[int]`): List of frame indices to decode.
... Returns:
... result (np.ndarray): np array of decoded frames of shape (num_frames, height, width, 3).
... '''
... frames = []
... container.seek(0)
... start_index = indices[0]
... end_index = indices[-1]
... for i, frame in enumerate(container.decode(video=0)):
... if i > end_index:
... break
... if i >= start_index and i in indices:
... frames.append(frame)
... return np.stack([x.to_ndarray(format="rgb24") for x in frames])
>>> model = LlavaNextVideoForConditionalGeneration.from_pretrained("llava-hf/LLaVA-NeXT-Video-7B-hf", device_map="auto)
>>> processor = AutoProcessor.from_pretrained("llava-hf/LLaVA-NeXT-Video-7B-hf")
>>> prompt = "USER: <video>\nWhy is this video funny? ASSISTANT:"
>>> video_path = hf_hub_download(repo_id="raushan-testing-hf/videos-test", filename="sample_demo_1.mp4", repo_type="dataset")
>>> container = av.open(video_path)
>>> # sample uniformly 8 frames from the video (model was trained with 32 frames per video, but this video is short)
>>> total_frames = container.streams.video[0].frames
>>> indices = np.arange(0, total_frames, total_frames / 8).astype(int)
>>> clip = read_video_pyav(container, indices)
>>> inputs_video = processor(text=prompt, videos=clip, return_tensors="pt").to(model.device)
>>> # load an image to generate from an image
>>> prompt = "USER:<image>\nWhat is shown in this image? ASSISTANT:"
>>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> inputs_image = processor(text=prompt, images=image, return_tensors="pt").to(model.device)
>>> # Generate from video
>>> generate_ids = model.generate(**inputs_video, max_length=50)
>>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
"USER:\nWhy is this video funny? ASSISTANT: The humor in this video comes from the unexpected and endearing sight of a baby wearing glasses and (...)"
>>> # Generate from image
>>> generate_ids = model.generate(**inputs_image, max_length=30)
>>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
"USER: \nWhat's the content of the image? ASSISTANT: The image shows a red stop sign on a pole, with a traditional Chinese archway (...)"