AdamCodd

/

vit-base-nsfw-detector

like 42

Hi @AdamCodd :
I have been trying the model and I found the results probabilities don't match when we use two methods listed in model_card file. Is there some other way the probability calculated from raw logits ?
This is the reproducible example :

from transformers import ViTImageProcessor, AutoModelForImageClassification, pipeline
from PIL import Image
import requests

url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
image = Image.open(requests.get(url, stream=True).raw)

# Direct Model
processor = ViTImageProcessor.from_pretrained('AdamCodd/vit-base-nsfw-detector')
model = AutoModelForImageClassification.from_pretrained('AdamCodd/vit-base-nsfw-detector')
inputs = processor(images=image, return_tensors="pt")
outputs = model(**inputs)
logits = outputs.logits

print(f"AutoModelForImageClassification Probs: {torch.nn.functional.softmax(logits)}")
# Predicted class: sfw


# Pipeline Model 
# Load image classification pipeline
pipeline_model = pipeline(task="image-classification", model="AdamCodd/vit-base-nsfw-detector", device=device)
print(f"Pipeline Probs: {pipeline_model(image)}")

OUTPUTS :

Device set to use cuda
AutoModelForImageClassification Probs: tensor([[0.9933, 0.0067]], grad_fn=<SoftmaxBackward0>)
Pipeline Probs: [{'label': 'sfw', 'score': 0.8909250497817993}, {'label': 'nsfw', 'score': 0.05239786580204964}]

Edit : The 2nd one also don't addup to 1, they are raw scores and not probs ?