app.py

import gradio as gr
import mathutils
import math
import numpy as np
import matplotlib.pyplot as plt
import matplotlib
import matplotlib.cm as cmx
import os.path as osp
import h5py
import random
import torch
import torch.nn as nn

from GDANet_cls import GDANET
from DGCNN import DGCNN

with open('shape_names.txt') as f:
    CLASS_NAME = f.read().splitlines()

model_gda = GDANET()
model_gda = nn.DataParallel(model_gda)
model_gda.load_state_dict(torch.load('./GDANet_WOLFMix.t7', map_location=torch.device('cpu')))
# model_gda.load_state_dict(torch.load('/Users/renjiawei/Downloads/pretrained_models/GDANet_WOLFMix.t7', map_location=torch.device('cpu')))
model_gda.eval()

model_dgcnn = DGCNN()
model_dgcnn = nn.DataParallel(model_dgcnn)
model_dgcnn.load_state_dict(torch.load('./dgcnn.t7', map_location=torch.device('cpu')))
# model_dgcnn.load_state_dict(torch.load('/Users/renjiawei/Downloads/pretrained_models/dgcnn.t7', map_location=torch.device('cpu')))
model_dgcnn.eval()

def pyplot_draw_point_cloud(points, corruption):
    rot1 = mathutils.Euler([-math.pi / 2, 0, 0]).to_matrix().to_3x3()
    rot2 = mathutils.Euler([0, 0, math.pi]).to_matrix().to_3x3()
    points = np.dot(points, rot1)
    points = np.dot(points, rot2)
    x, y, z = points[:, 0], points[:, 1], points[:, 2]
    colorsMap = 'winter'
    cs = y
    cm = plt.get_cmap(colorsMap)
    cNorm = matplotlib.colors.Normalize(vmin=-1, vmax=1)
    scalarMap = cmx.ScalarMappable(norm=cNorm, cmap=cm)
    fig = plt.figure(figsize=(5, 5))
    ax = fig.add_subplot(111, projection='3d')
    ax.scatter(x, y, z, c=scalarMap.to_rgba(cs))
    scalarMap.set_array(cs)
    ax.set_xlim(-1, 1)
    ax.set_ylim(-1, 1)
    ax.set_zlim(-1, 1)
    plt.axis('off')
    plt.title(corruption, fontsize=30)
    plt.tight_layout()
    plt.savefig('visualization.png', bbox_inches='tight', dpi=200)
    plt.close()



def load_dataset(corruption_idx, severity):
    corruptions = [
        'clean',
        'scale',
        'jitter',
        'rotate',
        'dropout_global',
        'dropout_local',
        'add_global',
        'add_local',
    ]
    corruption_type = corruptions[corruption_idx]
    if corruption_type == 'clean':
        f = h5py.File(osp.join('modelnet_c', corruption_type + '.h5'))
        # f = h5py.File(osp.join('/Users/renjiawei/Downloads/modelnet_c', corruption_type + '.h5'))
    else:
        f = h5py.File(osp.join('modelnet_c', corruption_type + '_{}'.format(severity-1) + '.h5'))
        # f = h5py.File(osp.join('/Users/renjiawei/Downloads/modelnet_c', corruption_type + '_{}'.format(severity - 1) + '.h5'))
    data = f['data'][:].astype('float32')
    label = f['label'][:].astype('int64')
    f.close()
    return data, label

def recognize_pcd(model, pcd):
    pcd = torch.tensor(pcd).unsqueeze(0)
    pcd = pcd.permute(0, 2, 1)
    output = model(pcd)
    prediction = output.softmax(-1).flatten()
    _, top5_idx = torch.topk(prediction, 5)
    return {CLASS_NAME[i]: float(prediction[i]) for i in top5_idx.tolist()}

def run(seed, corruption_idx, severity):
    data, label = load_dataset(corruption_idx, severity)
    random.seed(seed)
    sample_indx = random.randint(0, data.shape[0])
    pcd, cls = data[sample_indx], label[sample_indx]
    pyplot_draw_point_cloud(pcd, CLASS_NAME[cls[0]])
    output = 'visualization.png'
    return output, recognize_pcd(model_dgcnn, pcd), recognize_pcd(model_gda, pcd)


description = """

Welcome to the demo of PointCloud-C! [PointCloud-C](https://pointcloud-c.github.io/home.html) is a test-suite for point cloud robustness analysis under corruptions. In this demo, you may:

- __Visualize__ various types of corrupted point clouds in [ModelNet-C](https://github.com/jiawei-ren/ModelNet-C).

- __Compare__ our proposed techniques to the baseline in terms of prediction robustness.

For more details, checkout our paper [Benchmarking and Analyzing Point Cloud Classification under Corruptions, __ICML 2022__](https://arxiv.org/abs/2202.03377)!

📣 News: [The first PointCloud-C challenge](https://codalab.lisn.upsaclay.fr/competitions/6437) with Classification track and Part Segmentation track in [ECCV'22 SenseHuman workshop](https://sense-human.github.io/) is open for submission now!
"""


if __name__ == '__main__':
    iface = gr.Interface(
        fn=run,
        inputs=[
            gr.components.Number(label='Sample Seed', precision=0),
            gr.components.Radio(
                ['Clean', 'Scale', 'Jitter', 'Rotate', 'Drop Global', 'Drop Local', 'Add Global', 'Add Local'],
                value='Jitter', type="index", label='Corruption Type'),
            gr.components.Slider(1, 5, value=5, step=1, label='Corruption severity'),
        ],
        outputs=[
            gr.components.Image(type="file", label="Visualization"),
            gr.components.Label(num_top_classes=5, label="Baseline (DGCNN) Prediction"),
            gr.components.Label(num_top_classes=5, label="Ours (GDANet+WolfMix) Prediction")
        ],
        live=False,
        allow_flagging='never',
        title="PointCloud-C",
        description=description,
        examples=[
            [0, 'Jitter', 5],
            [999, 'Drop Local', 5],
        ],
        css=".output-image, .image-preview {height: 100px !important}",
        article="<p style='text-align: center'><a href='https://github.com/ldkong1205/PointCloud-C' target='_blank'>PointNet-C @ GitHub</a></p> "
    )
    iface.launch()