import torch
from transformers import ViTImageProcessor, ViTForImageClassification
from PIL import Image
import torch.nn.functional as F
import gradio as gr
 
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
 
processor = ViTImageProcessor.from_pretrained("ViT_LCZs_v3",local_files_only=True)
model = ViTForImageClassification.from_pretrained("ViT_LCZs_v3",local_files_only=True).to(device)


def classify_image(image):

    with torch.no_grad():
        model.eval()

        inputs = processor(images=image, return_tensors="pt")
        outputs = model(**inputs)
    
    logits = outputs.logits
    prob = torch.nn.functional.softmax(logits, dim=1)

    top10_prob, top10_indices = torch.topk(prob, 10)
    top10_confidences = {}
    for i in range(10):
        top10_confidences[model.config.id2label[int(top10_indices[0][i])]] = float(top10_prob[0][i])

    return top10_confidences #confidences


with gr.Blocks(title="ViT LCZ Classification - ClassCat",
            css=".gradio-container {background:white;}"
        ) as demo:
    gr.HTML("""<div style="font-family:'Times New Roman', 'Serif'; font-size:16pt; font-weight:bold; text-align:center; color:royalblue;">LCZ Classification with ViT</div>""")

    with gr.Row(): 
        input_image = gr.Image(type="pil")        
        output_label=gr.Label(label="Probabilities", num_top_classes=3)

    send_btn = gr.Button("Classify")
    send_btn.click(fn=classify_image, inputs=input_image, outputs=output_label)
    
    with gr.Row():
        gr.Examples(['data/closed_highrise.png'], label='Closed highrise', inputs=input_image)
        gr.Examples(['data/open_lowrise.png'], label='Sparsey built', inputs=input_image)
        gr.Examples(['data/dense_trees.png'], label='Dense trees', inputs=input_image)
        gr.Examples(['data/large_lowrise.png'], label='Large lowrise', inputs=input_image)

demo.launch(debug=True)