app.py

import csv
import os
import sys

import cv2
import gradio as gr
import matplotlib.pyplot as plt
import numpy as np
import onnxruntime as ort
from matplotlib import gridspec

ade_palette = []
labels_list = []

csv.field_size_limit(sys.maxsize)

with open(r"labels.txt", "r") as fp:
    for line in fp:
        labels_list.append(line[:-1])

with open(r"ade_palette.txt", "r") as fp:
    for line in fp:
        tmp_list = list(map(int, line[:-1].strip("][").split(", ")))
        ade_palette.append(tmp_list)

colormap = np.asarray(ade_palette)

model_filename = "segformer-b5-finetuned-ade-640-640.onnx"
sess_options = ort.SessionOptions()
sess_options.intra_op_num_threads = os.cpu_count()
sess = ort.InferenceSession(
    model_filename, sess_options, providers=["CPUExecutionProvider"]
)


def label_to_color_image(label):
    if label.ndim != 2:
        raise ValueError("Expect 2-D input label")

    if np.max(label) >= len(colormap):
        raise ValueError("label value too large.")

    return colormap[label]


def draw_plot(pred_img, seg):
    fig = plt.figure(figsize=(20, 15))

    grid_spec = gridspec.GridSpec(1, 2, width_ratios=[6, 1])

    plt.subplot(grid_spec[0])
    plt.imshow(pred_img)
    plt.axis("off")

    LABEL_NAMES = np.asarray(labels_list)
    FULL_LABEL_MAP = np.arange(len(LABEL_NAMES)).reshape(len(LABEL_NAMES), 1)
    FULL_COLOR_MAP = label_to_color_image(FULL_LABEL_MAP)

    unique_labels = np.unique(seg)
    ax = plt.subplot(grid_spec[1])
    plt.imshow(FULL_COLOR_MAP[unique_labels].astype(np.uint8), interpolation="nearest")
    ax.yaxis.tick_right()
    plt.yticks(range(len(unique_labels)), LABEL_NAMES[unique_labels])
    plt.xticks([], [])
    ax.tick_params(width=0.0, labelsize=25)
    return fig


def sepia(input_img):
    img = cv2.imread(input_img)
    img = cv2.resize(img, (640, 640)).astype(np.float32)
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    img_batch = np.expand_dims(img, axis=0)
    img_batch = np.transpose(img_batch, (0, 3, 1, 2))

    logits = sess.run(None, {"pixel_values": img_batch})[0]

    logits = np.transpose(logits, (0, 2, 3, 1))
    seg = np.argmax(logits, axis=-1)[0].astype("float32")
    seg = cv2.resize(seg, (640, 640)).astype("uint8")

    color_seg = np.zeros(
        (seg.shape[0], seg.shape[1], 3), dtype=np.uint8
    )  # height, width, 3

    for label, color in enumerate(colormap):
        color_seg[seg == label, :] = color

    # Convert to BGR
    color_seg = color_seg[..., ::-1]

    # Show image + mask
    pred_img = img * 0.5 + color_seg * 0.5
    pred_img = pred_img.astype(np.uint8)

    fig = draw_plot(pred_img, seg)
    return fig


title = "SegFormer(ADE20k) in TensorFlow"
description = """

This is demo TensorFlow SegFormer from 🤗 `transformers` official package. The pre-trained model was trained to segment scene specific images. We are **currently using ONNX model converted from the TensorFlow based SegFormer to improve the latency**. The average latency of an inference is **21** and **8** seconds for TensorFlow and ONNX converted models respectively (in [Colab](https://github.com/deep-diver/segformer-tf-transformers/blob/main/notebooks/TFSegFormer_ONNX.ipynb)). Check out the [repository](https://github.com/deep-diver/segformer-tf-transformers) to find out how to make inference, finetune the model with custom dataset, and further information.

"""

demo = gr.Interface(
    sepia,
    gr.inputs.Image(type="filepath"),
    outputs=["plot"],
    examples=["ADE_val_00000001.jpeg"],
    allow_flagging="never",
    title=title,
    description=description,
)

demo.launch()