app.py

import json
from pprint import pprint

import tensorflow as tf
import tensorflow.keras as keras
from gradio import Interface, inputs, outputs

RESNET50_SIZE = 256
RESNET101_SIZE = 360

DEVICE = "/cpu:0"


with open("./tags.json", "rt", encoding="utf-8") as f:
    tags = json.load(f)


with tf.device(DEVICE):
    model_resnet50 = keras.Sequential(
        [
            keras.applications.resnet.ResNet50(
                include_top=False,
                weights=None,
                input_shape=(RESNET50_SIZE, RESNET50_SIZE, 3),
            ),
            keras.layers.Conv2D(filters=len(tags), kernel_size=(1, 1), padding="same"),
            keras.layers.BatchNormalization(epsilon=1.001e-5),
            keras.layers.GlobalAveragePooling2D(name="avg_pool"),
            keras.layers.Activation("sigmoid"),
        ]
    )
    model_resnet50.load_weights("./tf_model_resnet50.h5")

with tf.device(DEVICE):
    model_resnet101 = keras.Sequential(
        [
            keras.applications.resnet.ResNet101(
                include_top=False,
                weights=None,
                input_shape=(RESNET101_SIZE, RESNET101_SIZE, 3),
            ),
            keras.layers.Conv2D(filters=len(tags), kernel_size=(1, 1), padding="same"),
            keras.layers.BatchNormalization(epsilon=1.001e-5),
            keras.layers.GlobalAveragePooling2D(name="avg_pool"),
            keras.layers.Activation("sigmoid"),
        ]
    )
    model_resnet101.load_weights("./tf_model_resnet101.h5")


def predict_resnet50(img):
    with tf.device(DEVICE):
        img = tf.image.resize_with_pad(img, RESNET50_SIZE, RESNET50_SIZE)
        img = tf.image.per_image_standardization(img)
        data = tf.expand_dims(img, 0)
        out, *_ = model_resnet50(data)
    return out


def predict_resnet101(img):
    with tf.device(DEVICE):
        img = tf.image.resize_with_pad(img, RESNET101_SIZE, RESNET101_SIZE)
        img = tf.image.per_image_standardization(img)
        data = tf.expand_dims(img, 0)
        out, *_ = model_resnet101(data)
    return out


def main(img, hide: float, model: str):
    if model.endswith("50"):
        out = predict_resnet50(img)
    elif model.endswith("101"):
        out = predict_resnet101(img)
    else:
        raise ValueError(f"Invalid model type: {model!r}")
    result = {
        tag: confidence
        for i, tag in enumerate(tags)
        if (confidence := float(out[i].numpy())) >= hide
    }
    pprint(result)
    return result


image = inputs.Image(label="Upload your image here!")
hide_threshold = inputs.Slider(
    label="Hide confidence lower than", default=0.5, maximum=1, minimum=0
)
select_model = inputs.Radio(
    ["ResNet50", "ResNet101"], label="Select model", type="value"
)

labels = outputs.Label(label="Tags", type="confidences")

interface = Interface(
    main, inputs=[image, hide_threshold, select_model], outputs=[labels]
)
interface.launch()