Paper-Repetition-EMFFN

import torch import torch.nn as nn from torchvision import transforms def extract_pc(image, pc=3): channel, height, width = image.shape # input float tensor image with CxHxW # data = image.view(channel, height*width) data = image.contiguous().view(channel, height * width) data_c = data - data.mean(dim=1).unsqueeze(1) u, s, vt = torch.svd(data_c.matmul(data_c.T)) sorted_data, indices = s.sort(descending=True) image_pc = u[:, indices[0:pc]].T.matmul(data) return image_pc.view(pc, height, width) def std_norm(image): image = image.permute(1, 2, 0).numpy() trans = transforms.Compose([ transforms.ToTensor(), transforms.Normalize(torch.tensor(image).mean(dim=[0, 1]), torch.tensor(image).std(dim=[0, 1])) ]) return trans(image) def one_zero_norm(image): channel, height, width = image.shape data = image.view(channel, height*width) data_max = data.max(dim=1)[0] data_min = data.min(dim=1)[0] data = (data-data_min.unsqueeze(1))/(data_max.unsqueeze(1) - data_min.unsqueeze(1)) return data.view(channel, height, width) def construct_sample(image, windows=27): _, height, width = image.shape half_window = int(windows/2) pad = nn.ReplicationPad2d(half_window) pad_image = pad(image.unsqueeze(0)).squeeze(0) batch_image_indices = torch.zeros((height*width, 4), dtype=torch.long) t = 0 for h in range(height): for w in range(width): batch_image_indices[t, :] = torch.tensor([h, h+windows, w, w+windows]) t += 1 return pad_image, batch_image_indices def label_transform(gt): label = torch.unique(gt) gt_new = torch.zeros_like(gt) for each in range(len(label)): indices = torch.where(gt == label[each]) if label[0] == 0: gt_new[indices] = each - 1 else: gt_new[indices] = each label_new = torch.unique(gt_new) return gt_new def label_inverse_transform(predict_result, gt): label_origin = torch.unique(gt) label_predict = torch.unique(predict_result) predict_result_origin = torch.zeros_like(predict_result) for each in range(len(label_predict)): indices = torch.where(predict_result == label_predict[each]) if len(label_predict) != len(label_origin): predict_result_origin[indices] = label_origin[each + 1] else: predict_result_origin[indices] = label_origin[each] return predict_result_origin def select_sample(gt, ntr): gt_vector = gt.reshape(-1, 1).squeeze(1) label = torch.unique(gt) for each in range(len(label)): indices_vector = torch.where(gt_vector == label[each]) # 1 tuple 返回一维的索引 indices = torch.where(gt == label[each]) # 2 tuple 返回二维的索引 indices_vector = indices_vector[0] indices_row = indices[0] indices_column = indices[1] # 背景 -1 # 10776 if label[each] == -1: no_gt_indices = torch.cat([indices_vector.unsqueeze(1), indices_row.unsqueeze(1), indices_column.unsqueeze(1)], dim=1 ) no_gt_num = torch.tensor(len(indices_vector)) else: class_num = torch.tensor(len(indices_vector)) rand_indices0 = torch.randperm(class_num) rand_indices = indices_vector[rand_indices0] # indian_pines # num_workers = [30, 150, 150, 100, 150, 150, 20, 150, 15, 150, 150, 150, 150, 150, 50, 50] # paviau # num_workers = [150, 150, 150, 150, 150, 150, 150, 150, 150] # KSC num_workers = [33, 23, 24, 24, 15, 22, 9, 38, 51, 39, 41, 49, 91] sel_num = num_workers[each - 1] sel_num = torch.tensor(sel_num) tr_ind0 = rand_indices0[0:sel_num] te_ind0 = rand_indices0[sel_num:] tr_ind = rand_indices[0:sel_num] te_ind = rand_indices[sel_num:] # 训练集:索引+坐标 sel_tr_ind = torch.cat([ tr_ind.unsqueeze(1), indices_row[tr_ind0].unsqueeze(1), indices_column[tr_ind0].unsqueeze(1) ], dim=1) # 测试集： sel_te_ind = torch.cat([ te_ind.unsqueeze(1), indices_row[te_ind0].unsqueeze(1), indices_column[te_ind0].unsqueeze(1) ], dim=1) if each == 1: train_indices = sel_tr_ind train_num = sel_num.unsqueeze(0) test_indices = sel_te_ind test_num = (class_num - sel_num).unsqueeze(0) else: train_indices = torch.cat([train_indices, sel_tr_ind], dim=0) train_num = torch.cat([train_num, sel_num.unsqueeze(0)]) test_indices = torch.cat([test_indices, sel_te_ind], dim=0) test_num = torch.cat([test_num, (class_num - sel_num).unsqueeze(0)]) # 训练集 rand_tr_ind = torch.randperm(train_num.sum()) train_indices = train_indices[rand_tr_ind, ] # 测试集 rand_te_ind = torch.randperm(test_num.sum()) # torch.Size([42506]) test_indices = test_indices[rand_te_ind, ] # 背景图 rand_no_gt_ind = torch.randperm(no_gt_num.sum()) no_gt_indices = no_gt_indices[rand_no_gt_ind, ] data_sample = {'train_indices': train_indices, 'train_num': train_num, 'test_indices': test_indices, 'test_num': test_num, 'no_gt_indices': no_gt_indices, 'no_gt_num': no_gt_num.unsqueeze(0)} return data_sample