import datasets
from transformers import AutoFeatureExtractor, AutoModelForImageClassification
import numpy as np
import gradio as gr
import torch

dataset = datasets.load_dataset("beans") # This should be the same as the first line of Python code in this Colab notebook

extractor = AutoFeatureExtractor.from_pretrained("andresgtn/vit-base-bean-health-classifier")
model = AutoModelForImageClassification.from_pretrained("andresgtn/vit-base-bean-health-classifier")

# add to cuda?
#model.eval()
#model.to(device)

labels = dataset['train'].features['labels'].names

def classify(im):
  features = extractor(im, return_tensors='pt')
  #features.to(device) # move to gpu as model, if available
  with torch.no_grad():
    logits = model(**features).logits
  probability = torch.nn.functional.softmax(logits, dim=-1)
  #probs = probability[0].to('cpu').detach().numpy()
  probs = probability[0].detach().numpy()
  confidences = {label: float(probs[i]) for i, label in enumerate(labels)}
  return confidences

interface = gr.Interface(classify, gr.Image(shape=(200, 200)), 'text')
#demo.launch()
interface.launch(debug=False)