app.py

from PIL import Image, ImageDraw

import torch
from torchvision import transforms
import torch.nn.functional as F

import gradio as gr

# import sys
# sys.path.insert(0, './')
from test import create_letr, get_lines_and_draw
from models.preprocessing import *
from models.misc import nested_tensor_from_tensor_list


model = create_letr('resnet50/checkpoint0024.pth')
model101 = create_letr('resnet101/checkpoint0024.pth')
# PREPARE PREPROCESSING
# transform_test = transforms.Compose([
#     transforms.Resize((test_size)),
#     transforms.ToTensor(),
#     transforms.Normalize([0.538, 0.494, 0.453], [0.257, 0.263, 0.273]),
# ])
normalize = Compose([
        ToTensor(),
        Normalize([0.538, 0.494, 0.453], [0.257, 0.263, 0.273]),
        Resize([256]),
])
normalize_512 = Compose([
        ToTensor(),
        Normalize([0.538, 0.494, 0.453], [0.257, 0.263, 0.273]),
        Resize([512]),
])
normalize_1100 = Compose([
        ToTensor(),
        Normalize([0.538, 0.494, 0.453], [0.257, 0.263, 0.273]),
        Resize([1100]),
])


def predict(inp, size, model_name):
    image = Image.fromarray(inp.astype('uint8'), 'RGB')
    h, w = image.height, image.width
    orig_size = torch.as_tensor([int(h), int(w)])

    if size == '1100':
        img = normalize_1100(image)
    elif size == '512':
        img = normalize_512(image)
    else:
        img = normalize(image)
    inputs = nested_tensor_from_tensor_list([img])

    with torch.no_grad():
        if model_name == 'resnet101':
            outputs = model101(inputs)[0]
        else:
            outputs = model(inputs)[0]

    lines = get_lines_and_draw(image, outputs, orig_size)

    return image, str(lines)


inputs = [
    gr.inputs.Image(),
    gr.inputs.Radio(["256", "512", "1100"]),
    gr.inputs.Radio(["resnet50", "resnet101"]),
]
outputs = [
    gr.outputs.Image(label='Image with Lines', type='numpy'),
    gr.outputs.Textbox(label='Lines points List')
]
gr.Interface(
    fn=predict,
    inputs=inputs,
    outputs=outputs,
    examples=[
        ["demo.png", '256', "resnet50"], 
        ["tappeto-per-calibrazione.jpg", '256', "resnet50"]
    ],
    title="LETR: Line Segment Detection Using Transformers without Edges",
    description="It is an end-to-end line segment detection algorithm using Transformers [published on CVPR 2021](https://github.com/mlpc-ucsd/LETR)."
).launch()