File size: 6,926 Bytes
35e4140 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 | # Copyright 2025-2026 The Moonshot AI Team and HuggingFace Inc. team. All rights reserved.
#
# Extracted from moonshotai/Kimi-K3 for standalone MoonViT-V2 use.
# Licensed under the Kimi K3 License (see LICENSE in this repository).
"""Image / video processor for MoonViT-V2 (NaViT-style native resolution)."""
from __future__ import annotations
import math
from typing import Optional, Sequence, Union
import numpy as np
import torch
from PIL import Image
from transformers.image_processing_utils import BaseImageProcessor, BatchFeature
from transformers.image_utils import ImageInput, make_list_of_images, valid_images
from transformers.utils import TensorType
def _navit_resize(
width: int,
height: int,
patch_size: int,
merge_kernel_size: int,
in_patch_limit: int,
patch_limit_on_one_side: int,
):
s1 = math.sqrt(
in_patch_limit
/ (max(1.0, width // patch_size) * max(1.0, height // patch_size))
)
s2 = patch_limit_on_one_side * patch_size / width
s3 = patch_limit_on_one_side * patch_size / height
scale = min(1.0, s1, s2, s3)
new_w = min(max(1, int(width * scale)), patch_limit_on_one_side * patch_size)
new_h = min(max(1, int(height * scale)), patch_limit_on_one_side * patch_size)
factor = merge_kernel_size * patch_size
pad_height = (factor - new_h % factor) % factor
pad_width = (factor - new_w % factor) % factor
token_height = (new_h + pad_height) // factor
token_width = (new_w + pad_width) // factor
return {
"new_width": new_w,
"new_height": new_h,
"pad_width": pad_width,
"pad_height": pad_height,
"num_tokens": token_height * token_width,
}
def _patchify(pixel_values: np.ndarray, patch_size: int) -> dict:
"""pixel_values: (t, h, w, c) -> patches + grid_thw."""
T, H, W, C = pixel_values.shape
assert C == 3
patches = pixel_values.reshape(
T, H // patch_size, patch_size, W // patch_size, patch_size, C
)
patches = patches.transpose(0, 1, 3, 5, 2, 4)
patches = patches.reshape(-1, C, patch_size, patch_size)
grid_thw = np.array([T, H // patch_size, W // patch_size], dtype=np.int64)
return {"pixel_values": patches, "grid_thw": grid_thw}
def _as_pil(image) -> Image.Image:
if isinstance(image, Image.Image):
return image.convert("RGB")
return Image.fromarray(np.asarray(image)).convert("RGB")
class MoonViTV2ImageProcessor(BaseImageProcessor):
model_type = "moonvit_v2"
def __init__(
self,
patch_size: int = 14,
merge_kernel_size: int = 2,
in_patch_limit: int = 65536,
patch_limit_on_one_side: int = 512,
max_num_frames: int = 4,
image_mean: tuple[float, float, float] = (0.5, 0.5, 0.5),
image_std: tuple[float, float, float] = (0.5, 0.5, 0.5),
**kwargs,
):
super().__init__(**kwargs)
self.patch_size = patch_size
self.merge_kernel_size = merge_kernel_size
self.in_patch_limit = in_patch_limit
self.patch_limit_on_one_side = patch_limit_on_one_side
# Matches MoonViT-V2 init_pos_emb_time / temporal pos-emb capacity.
self.max_num_frames = max_num_frames
self.image_mean = list(image_mean)
self.image_std = list(image_std)
def _resize_config(self, image: Image.Image) -> dict:
w, h = image.size
return _navit_resize(
w,
h,
self.patch_size,
self.merge_kernel_size,
self.in_patch_limit,
self.patch_limit_on_one_side,
)
def _normalize_frame(self, image: Image.Image, cfg: dict) -> np.ndarray:
image = image.resize(
(cfg["new_width"], cfg["new_height"]), resample=Image.Resampling.BICUBIC
)
arr = np.asarray(image)
if cfg["pad_height"] or cfg["pad_width"]:
arr = np.pad(
arr,
((0, cfg["pad_height"]), (0, cfg["pad_width"]), (0, 0)),
mode="constant",
constant_values=0,
)
mean = np.array(self.image_mean, dtype=np.float32)
std_inv = 1.0 / np.array(self.image_std, dtype=np.float32)
return (arr.astype(np.float32) / 255.0 - mean) * std_inv
def _preprocess_one(self, image: Image.Image) -> tuple[np.ndarray, np.ndarray]:
image = image.convert("RGB")
cfg = self._resize_config(image)
arr = self._normalize_frame(image, cfg)
packed = _patchify(np.expand_dims(arr, axis=0), self.patch_size)
return packed["pixel_values"], packed["grid_thw"]
def preprocess(
self,
images: ImageInput,
return_tensors: Optional[Union[str, TensorType]] = None,
) -> BatchFeature:
"""Preprocess one or more **images** (each becomes an independent sample with T=1)."""
images = make_list_of_images(images)
if not valid_images(images):
raise ValueError(
"Invalid image type. Must be PIL.Image.Image, numpy.ndarray, or torch.Tensor."
)
pixel_values, grid_thws = [], []
for image in images:
patches, grid_thw = self._preprocess_one(_as_pil(image))
pixel_values.append(torch.from_numpy(patches))
grid_thws.append(torch.from_numpy(grid_thw).unsqueeze(0))
data = {
"pixel_values": torch.cat(pixel_values, dim=0),
"grid_thws": torch.cat(grid_thws, dim=0),
}
return BatchFeature(data=data, tensor_type=return_tensors)
def preprocess_video(
self,
frames: Sequence[ImageInput],
return_tensors: Optional[Union[str, TensorType]] = None,
) -> BatchFeature:
"""Preprocess a **video** as an ordered list of frames (one sample with T=len(frames)).
All frames share the resize/pad config of the first frame so spatial grids align.
``T`` must be in ``[1, max_num_frames]`` (default 4, matching temporal pos-emb).
"""
if not isinstance(frames, (list, tuple)) or len(frames) == 0:
raise ValueError("`frames` must be a non-empty list/tuple of images.")
if len(frames) > self.max_num_frames:
raise ValueError(
f"Got {len(frames)} frames, but max_num_frames={self.max_num_frames} "
f"(MoonViT-V2 temporal pos-emb capacity)."
)
pil_frames = [_as_pil(f) for f in frames]
cfg = self._resize_config(pil_frames[0])
arrs = [self._normalize_frame(fr, cfg) for fr in pil_frames]
pixels = np.stack(arrs, axis=0) # (T, H, W, C)
packed = _patchify(pixels, self.patch_size)
data = {
"pixel_values": torch.from_numpy(packed["pixel_values"]),
"grid_thws": torch.from_numpy(packed["grid_thw"]).unsqueeze(0),
}
return BatchFeature(data=data, tensor_type=return_tensors)
|