import torch
from transformers import AutoFeatureExtractor, AutoModelForImageClassification
from einops import rearrange
import gradio
import call_labels
# define the feature extractor
extractor = AutoFeatureExtractor.from_pretrained("vincentclaes/mit-indoor-scenes")
# define the pretrained model
model = AutoModelForImageClassification.from_pretrained("vincentclaes/mit-indoor-scenes")
# retrieve the labels provided from MIT Indoor Scenes dataset (https://www.kaggle.com/itsahmad/indoor-scenes-cvpr-2019)
labels = call_labels.call_labels()
# call model.eval() to assert that we are evaluating the model and not updating the weights
model.eval()
# define the function used for model inference
def classify(image):
# disable gradient calculation
with torch.no_grad():
# extract features from the image input
inputs = extractor(images=image, return_tensors='pt')
# call the logits parameter only (object: SequenceClassifierOutput)
outputs = model(**inputs).logits
# remove the batch dimension
outputs = rearrange(outputs, '1 j->j')
# use the softmax function to convert the logits into probabilities
outputs = torch.nn.functional.softmax(outputs)
# convert the data type from tensor to a numpy array
outputs = outputs.cpu().numpy()
# returns a key-value pair of the id labels and its corresponding probabilities
return {labels[str(i)]: float(outputs[i]) for i in range(len(labels))}
# define the gradio interface
gradio.Interface(fn=classify,
inputs=gradio.inputs.Image(shape=(224,224),
image_mode='RGB',
source='upload',
tool='editor',
type='pil',
label=None,
optional=False),
outputs=gradio.outputs.Label(num_top_classes=5,
type='auto'),
theme='dark-huggingface',
examples=[['bedroom.jpg'],
['bathroom_AS.jpg'],
['samsung_room.jpg']],
live=True,
title='Indoor Scene Recognition',
description='An indoor scene classifier. Start by uploading an input image. The outputs are the top five indoor scene classes that best fit your input image.',
interpretation='default',
article='''Additional Information
This indoor scene classifier employs the google/vit-base-patch16-224-in21k, a Visual Transformer (ViT) model pre-trained on ImageNet-21k (14 million images, 21,843 classes) at resolution 224x224 introduced in the paper An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale by Dosovitskiy et al. The original GitHub repository of the Visual Transformer is found in this link. This model was fine-tuned on the MIT Indoor Scenes from Kaggle. The source model is found in this link.
''',
allow_flagging='never').launch()