limb_wavefield.py

import os
import torch
import numpy as np
from torch.utils.data import Dataset, DataLoader
from collections import OrderedDict
import matplotlib.pyplot as plt
import torch.nn as nn
from torch import optim
import torch.nn.functional as F
# import sys
# import random
from matplotlib.patches import Rectangle
import argparse
torch.set_float32_matmul_precision("medium")
def main(model_name):
    
    if model_name == 'S2NO_big':
        from S2NO_pretrain import S2NO_pretrain
        model = S2NO_pretrain().cuda()
        PATH = './S2NO/big/600k.ckpt'
    if model_name == 'S2NO_small':
        from S2NO_pretrain import S2NO_pretrain
        model = S2NO_pretrain(width = 20).cuda()
        PATH = './S2NO/small/600k.ckpt'
        # PATH = '/gpfs/share/home/2401112587/neuralFWI/NCBSO_ckpt/NCBSO_20width_mq_lr01_final/600k_model-epoch=007-val_loss=0.0856-val_loss_breast=0.0772-val_loss_arm=0.0773-val_loss_limb=0.1391.ckpt'
    if model_name == 'FNO_big':
        from FNO_pretrain import FNO_pretrain
        model = FNO_pretrain(features_ = 40).cuda()
        PATH = './FNO/big/600k.ckpt'
    if model_name == 'FNO_small':
        from FNO_pretrain import FNO_pretrain
        model = FNO_pretrain(features_ = 20).cuda()
        PATH = './FNO/small/600k.ckpt'
    if model_name == 'UNet':
        from UNet_pretrain import UNet_pretrain
        model =UNet_pretrain().cuda()
        PATH = './UNet/600k.ckpt'
    checkpoint = torch.load(PATH, map_location=lambda storage, loc: storage)
    model.load_state_dict(checkpoint['state_dict'])

    homo = np.load('./homo/homo_600k.npy')[0:1,:,:]
    field_real = torch.tensor(np.real(homo))
    field_imag = torch.tensor(np.imag(homo))
    model.eval()

    def inference(data, field_real, field_imag):
        data = (1500/data - 1)*30
        data = torch.tensor(data, dtype=torch.float).cuda()
        batchsize = field_real.shape[0]
        sos = data.reshape(1,480, 480, 1).repeat(batchsize,1,1,1).cuda()
        field = torch.concat([field_real.unsqueeze(-1), field_imag.unsqueeze(-1)], dim=-1).cuda() * 2e-3
        src = field    
        pred = model(sos, src)
        pred = pred * 500
        pred = pred[...,0] + 1j*pred[...,1]
        return pred

    # 这里定义两块子图的行列范围（row_start:row_end, col_start:col_end）
    sub1_row_start, sub1_row_end = 323, 343
    sub1_col_start, sub1_col_end = 230, 250
    sub2_row_start, sub2_row_end = 234, 254
    sub2_col_start, sub2_col_end = 184, 204

    index = [36]
    for i in range(len(index)):
        path = f'./speed/test_{index[i]}.npy'
        data = np.load(path)
        # 推理得到 pred
        pred = inference(data, field_real, field_imag)
        pred_np = pred.detach().cpu().numpy()
        pred_real = np.real(pred_np)
        
        # 分别画出两个子图并保存
        # 子图1
        fig, ax = plt.subplots(figsize=(5,5), dpi=300)
        ax.imshow(np.squeeze(pred_real)[
                sub1_row_start:sub1_row_end,
                sub1_col_start:sub1_col_end],
                cmap='seismic',
                vmin=-2000, vmax=2000)
        ax.spines['top'].set_visible(False)
        ax.spines['right'].set_visible(False)
        ax.spines['bottom'].set_visible(False)
        ax.spines['left'].set_visible(False)
        ax.get_xaxis().set_visible(False)
        ax.get_yaxis().set_visible(False)
        plt.savefig(f'./result/{model_name}_limb_{index[i]}_small1.pdf',
                    bbox_inches='tight', pad_inches=0)
        plt.close()

        # 子图2
        fig, ax = plt.subplots(figsize=(5,5), dpi=300)
        ax.imshow(np.squeeze(pred_real)[
                sub2_row_start:sub2_row_end,
                sub2_col_start:sub2_col_end],
                cmap='seismic',
                vmin=-2000, vmax=2000)
        ax.spines['top'].set_visible(False)
        ax.spines['right'].set_visible(False)
        ax.spines['bottom'].set_visible(False)
        ax.spines['left'].set_visible(False)
        ax.get_xaxis().set_visible(False)
        ax.get_yaxis().set_visible(False)
        plt.savefig(f'./result/{model_name}_limb_{index[i]}_small2.pdf',
                    bbox_inches='tight', pad_inches=0)
        plt.close()

        # 在大图上两个子图相应位置标注出方框
        # 注意：matplotlib 中默认 (x, y) 是 (列索引, 行索引)，
        # 因此填入 Rectangle 的时候，需要 (col_start, row_start, width, height)
        # 画大图 + 标注方框
        fig, ax = plt.subplots(figsize=(5,5), dpi=300)
        ax.imshow(np.squeeze(pred_real),
                cmap='seismic',
                vmin=-2000, vmax=2000)

        # 第一个方框
        rect1 = Rectangle((sub1_col_start, sub1_row_start),
                      sub1_col_end - sub1_col_start,  # width
                      sub1_row_end - sub1_row_start,  # height
                      fill=False,
                      edgecolor='#8CA5D3',
                      linewidth=2)
        ax.add_patch(rect1)

        # 第二个方框
        rect2 = Rectangle((sub2_col_start, sub2_row_start),
                        sub2_col_end - sub2_col_start,
                        sub2_row_end - sub2_row_start,
                        fill=False,
                        edgecolor='#EDAD81',
                        linewidth=2)
        ax.add_patch(rect2)

        ax.spines['top'].set_visible(False)
        ax.spines['right'].set_visible(False)
        ax.spines['bottom'].set_visible(False)
        ax.spines['left'].set_visible(False)
        ax.get_xaxis().set_visible(False)
        ax.get_yaxis().set_visible(False)

        # 保存带方框的大图
        plt.savefig(f'./result/{model_name}_limb_{index[i]}_with_box.pdf',
                    bbox_inches='tight', pad_inches=0)
        plt.close()
if __name__ == '__main__':
    # 解析命令行参数
    parser = argparse.ArgumentParser(description='Run model inference with specified model.')
    parser.add_argument('--model_name', type=str, required=True, 
                        choices=['S2NO_big', 'S2NO_small', 'FNO_big', 'FNO_small','UNet'],
                        help='Name of the model to use (e.g., S2NO_small)')
    args = parser.parse_args()
    # 调用主函数
    main(args.model_name)