import requests
import gradio as gr
import torch
from timm.data import resolve_data_config
from torchvision.models import alexnet
import torch.nn as nn
from torchvision import transforms

LABELS = {0:'Cat', 1:'Dog'}


model = alexnet(pretrained=True)
for param in model.parameters():
    param.requires_grad = False

# Add a avgpool here
avgpool = nn.AdaptiveAvgPool2d((7, 7))

# Replace the classifier layer
# to customise it according to our output
model.classifier = nn.Sequential(
    nn.Linear(256 * 7 * 7, 1024),
    nn.Linear(1024, 256),
    nn.Linear(256, 2),
)

checkpoint = torch.load(
    "CatVsDogsModel.pth", map_location=torch.device("cpu")
)

model.load_state_dict(checkpoint["state_dict"])
model = model.to('cpu')



transform = transforms.Compose(
    [transforms.Resize((128, 128)), transforms.ToTensor()]
)



def predict(img):
    img = transform(img).to('cpu')
    img = img.convert('RGB')
    img = transform(img).unsqueeze(0)
    with torch.no_grad():
        out=  model(img)
    probability = torch.nn.functional.softmax(out[0],dim=0)

    values, indices = torch.topk(probability,k=2)
    return {LABELS[i]: v.item() for i,v in zip(indices,values)}


transform = create_transform(**resolve_data_config({},model=model))


iface = gr.Interface(fn=predict, inputs=gr.inputs.Image(type='pil'), outputs="label").launch()
iface.launch()