github/ai-toolkit
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Change pixtral to crop based on number of pixels instead of largest dimension

Browse Source
This commit is contained in:
Jaret Burkett
2024-09-28 13:05:26 -06:00
parent 58537fc92b
commit a508caad1d
1 changed files with 7 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
toolkit/models/pixtral_vision.py
View File

@@ -1,3 +1,4 @@
import math
from typing import List, Optional, Tuple, Any, Union from typing import List, Optional, Tuple, Any, Union
import os import os
import torch import torch
@@ -461,7 +462,12 @@ class PixtralVisionImagePreprocessor:
w, h = img.shape[-1], img.shape[-2] w, h = img.shape[-1], img.shape[-2]
ratio = max(h / self.max_image_size, w / self.max_image_size) # originally, pixtral used the largest of the 2 dimensions, but we
# will use the base size of the image based on number of pixels.
# ratio = max(h / self.max_image_size, w / self.max_image_size) # original
base_size = int(math.sqrt(w * h))
ratio = base_size / self.max_image_size
if ratio > 1: if ratio > 1:
w = round(w / ratio) w = round(w / ratio)
h = round(h / ratio) h = round(h / ratio)