---
library_name: transformers
tags:
- generated_from_trainer
model-index:
- name: siglip-video_16384
  results: []
---

<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->

# siglip-video_16384

This model was trained from scratch on an unknown dataset.

## Model description

More information needed

## Intended uses & limitations

More information needed

## Training and evaluation data

More information needed

## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 0.0001
- train_batch_size: 16
- eval_batch_size: 1
- seed: 42
- distributed_type: multi-GPU
- num_devices: 8
- total_train_batch_size: 128
- total_eval_batch_size: 8
- optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: cosine
- lr_scheduler_warmup_ratio: 0.03
- num_epochs: 1.0

### Training results



### Framework versions

- Transformers 4.47.1
- Pytorch 2.5.1+cu124
- Datasets 2.18.0
- Tokenizers 0.21.0


### Wandb
https://wandb.ai/dongfu/Mantis/runs/tatohkvl?nw=nwuserdongfu

### code
```python

from PIL import Image
import requests
from transformers import AutoProcessor, AutoModel
from mantis.models.siglip_video import SiglipVideoModel
import torch
import numpy as np
import av
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:
        np.ndarray: np array of decoded frames of shape (num_frames, height, width, 3).
    '''
    frames = []
    container.seek(0)
    if len(indices) == 0:
        # to debug
        indices = [0]
        print("No indices to decode, might be an empty video please check")
    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 = SiglipVideoModel.from_pretrained("google/siglip-so400m-patch14-384")
model = SiglipVideoModel.from_pretrained("Mantis-VL/siglip-video_16384_2fps_128").to("cuda:2")
processor = AutoProcessor.from_pretrained("google/siglip-so400m-patch14-384")


container = av.open("../mochi.mp4")
# container = av.open("/home/dongfu/WorkSpace/Mantis/data/llava-video/data/0_30_s_youtube_v0_1/videos/liwei_youtube_videos/videos/youtube_video_2024/ytb_F-FpE2GWW84.mp4")
total_frames = container.streams.video[0].frames
sample_fps = 2
ori_fps = container.streams.video[0].average_rate
indices = np.arange(0, total_frames, int(ori_fps/sample_fps))
frames = read_video_pyav(container, indices)

text = "Close-up of a chameleon's eye, with its scaly skin changing color. Ultra high resolution 4k."
# text = "The video showcases a group of individuals dressed in matching military-style uniforms, consisting of long, light-colored tunics and dark vests, marching in unison. They are carrying large, black, shoulder-mounted weapons, and the background appears to be an open area, possibly a parade ground or a military base, with a clear sky overhead. The text overlay in English reads, 'Talabani won victory over America with an impossible weapon,' suggesting a narrative of triumph using unconventional means. The individuals are seen marching in a coordinated manner, emphasizing discipline and uniformity. As the video progresses, the group continues their synchronized march, maintaining the same background setting. The text overlay, 'Talabani won victory over America with an impossible weapon,' reappears, reinforcing the narrative of triumph. One individual in the foreground is prominently holding a rifle, adding to the display of military prowess. The video emphasizes the themes of discipline, coordination, and military strength."

print(frames.shape)
inputs = processor(text=[text], images=frames, padding="max_length", return_tensors="pt")
inputs = {k: v.to(model.device) for k, v in inputs.items()}
inputs['pixel_values'] = [inputs['pixel_values']]
with torch.no_grad():
    outputs = model(**inputs)
logits_per_video = outputs.logits_per_video
print(logits_per_video)
probs = torch.sigmoid(logits_per_video) # these are the probabilities
print(f"{probs[0][0]:.1%} the video contains the text: '{text}'")

```