app.py

from huggingface_hub import from_pretrained_fastai
import gradio as gr
from fastai.vision.all import *
from icevision.all import *
from icevision.models.checkpoint import *
import PIL

# class map > y con esto done;

class_map = ClassMap(['raccoon','banana'])

size = 384

model_type = models.ross.efficientdet

model_2 = model_type.model(
    backbone= model_type.backbones.tf_d0 (pretrained=True),
    num_classes=len(class_map),
    img_size = size
)

# load from model_repo:

from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="Alesteba/deep_model_02", filename="ross_racoon.pth")

state_dict = torch.load('./ross_racoon.pth', map_location=torch.device('cpu'))

model_2.load_state_dict(state_dict)

# use test img:

infer_tfms = tfms.A.Adapter([*tfms.A.resize_and_pad(size),tfms.A.Normalize()])

def predict(img):
    
    img = PIL.Image.fromarray(img, "RGB")
    
    pred_dict_2  = model_type.end2end_detect(
    
        img, 
        infer_tfms,
        model_2.to("cpu"), 
        class_map=class_map, 
        detection_threshold=0.5
    )
    
    return pred_dict_2["img"]
    
gr.Interface(
    fn=predict, 
    inputs=gr.inputs.Image(shape=(128, 128)), 
    outputs=[gr.outputs.Image(type="pil", label="VFNet Inference")],
    examples=['raccoon-test_1.jpg','raccoon-test_2.jpg']
).launch(share=False)