def predict():
#加载模型
hsid = HSIDCNN()
#hsid = nn.DataParallel(hsid).to(DEVICE)
hsid.load_state_dict(torch.load('./PNMN_064SIGMA025.pth'))
#加载数据
test_data_dir = './data/test/'
test_set = HsiTrainDataset(test_data_dir)
test_dataloader = DataLoader(test_set, batch_size=1, shuffle=False)
#指定结果输出路径
test_result_output_path = './data/testresult/'
if not os.path.exists(test_result_output_path):
os.makedirs(test_result_output_path)
#逐个通道的去噪
"""
分配一个numpy数组，存储去噪后的结果
遍历所有通道，
对于每个通道，通过get_adjacent_spectral_bands获取其相邻的K个通道
调用hsid进行预测
将预测到的residual和输入的noise加起来，得到输出band
将去噪后的结果保存成mat结构
"""
for batch_idx, (noisy, label) in enumerate(test_dataloader):
noisy = noisy.type(torch.FloatTensor)
label = label.type(torch.FloatTensor)
batch_size, width, height, band_num = noisy.shape
denoised_hsi = np.zeros((width, height, band_num))
#noisy = noisy.to(DEVICE)
#label = label.to(DEVICE)
with torch.no_grad():
for i in range(band_num): #遍历每个band去处理
current_noisy_band = noisy[:,:,:,i]
current_noisy_band = current_noisy_band[:,None]
adj_spectral_bands = get_adjacent_spectral_bands(noisy, K, i)
#adj_spectral_bands = torch.transpose(adj_spectral_bands,3,1) #将通道数置换到第二维
adj_spectral_bands = adj_spectral_bands.permute(0, 3,1,2)
adj_spectral_bands_unsqueezed = adj_spectral_bands.unsqueeze(1)
denoised_band = hsid(current_noisy_band, adj_spectral_bands_unsqueezed)
denoised_band_numpy = denoised_band.cpu().numpy().astype(np.float32)
denoised_band_numpy = np.squeeze(denoised_band_numpy)
denoised_hsi[:,:,i] += denoised_band_numpy
#mdict是python字典类型，value值需要是一个numpy数组
scio.savemat(test_result_output_path + 'result.mat', {'denoised': denoised_hsi})
def predict_lowlight_residual():
#加载模型
#hsid = HSID(36)
hsid = MultiStageHSIDUpscale(36)
#hsid = nn.DataParallel(hsid).to(DEVICE)
#device = torch.device("cuda:1" if torch.cuda.is_available() else "cpu")
hsid = hsid.to(DEVICE)
hsid.load_state_dict(torch.load('./checkpoints/hsid_multistage_upscale_patchsize64_best.pth', map_location='cuda:0')['gen'])
#加载数据
mat_src_path = './data/test_lowlight/origin/soup_bigcorn_orange_1ms.mat'
test_label = scio.loadmat(mat_src_path)['label']
#test=test.transpose((2,0,1)) #将通道维放在最前面：191*1280*307
test_data_dir = './data/test_lowlight/origin/'
test_set = HsiLowlightTestDataset(test_data_dir)
test_dataloader = DataLoader(test_set, batch_size=1, shuffle=False)
#指定结果输出路径
test_result_output_path = './data/testresult/'
if not os.path.exists(test_result_output_path):
os.makedirs(test_result_output_path)
#逐个通道的去噪
"""
分配一个numpy数组，存储去噪后的结果
遍历所有通道，
对于每个通道，通过get_adjacent_spectral_bands获取其相邻的K个通道
调用hsid进行预测
将预测到的residual和输入的noise加起来，得到输出band
将去噪后的结果保存成mat结构
"""
for batch_idx, (noisy, label) in enumerate(test_dataloader):
noisy = noisy.type(torch.FloatTensor)
label = label.type(torch.FloatTensor)
batch_size, width, height, band_num = noisy.shape
denoised_hsi = np.zeros((width, height, band_num))
noisy = noisy.to(DEVICE)
label = label.to(DEVICE)
with torch.no_grad():
for i in range(band_num): #遍历每个band去处理
current_noisy_band = noisy[:,:,:,i]
current_noisy_band = current_noisy_band[:,None]
adj_spectral_bands = get_adjacent_spectral_bands(noisy, K, i)# shape: batch_size, width, height, band_num
adj_spectral_bands = adj_spectral_bands.permute(0, 3,1,2)#交换第一维和第三维 ，shape: batch_size, band_num, height, width
adj_spectral_bands = adj_spectral_bands.to(DEVICE)
residual = hsid(current_noisy_band, adj_spectral_bands)
denoised_band = current_noisy_band + residual[0]
denoised_band_numpy = denoised_band.cpu().numpy().astype(np.float32)
denoised_band_numpy = np.squeeze(denoised_band_numpy)
denoised_hsi[:,:,i] = denoised_band_numpy
#mdict是python字典类型，value值需要是一个numpy数组
scio.savemat(test_result_output_path + 'result.mat', {'denoised': denoised_hsi})
psnr = PSNR(denoised_hsi, test_label)
ssim = SSIM(denoised_hsi, test_label)
sam = SAM(denoised_hsi, test_label)
#计算pnsr和ssim
print("=====averPSNR:{:.3f}=====averSSIM:{:.4f}=====averSAM:{:.3f}".format(psnr, ssim, sam))
def predict_lowlight_hsid_origin():
#加载模型
#hsid = HSID(36)
hsid = HSIRDNECA_Denoise(K)
hsid = nn.DataParallel(hsid).to(DEVICE)
#device = torch.device("cuda:1" if torch.cuda.is_available() else "cpu")
save_model_path = './checkpoints/hsirnd_denoise_l1loss'
#hsid = hsid.to(DEVICE)
hsid.load_state_dict(
torch.load(save_model_path +
'/hsid_rdn_eca_l1_loss_600epoch_patchsize32_best.pth',
map_location='cuda:0')['gen'])
#加载数据
test_data_dir = './data/denoise/test/level25'
test_set = HsiTrainDataset(test_data_dir)
test_dataloader = DataLoader(test_set, batch_size=1, shuffle=False)
#指定结果输出路径
test_result_output_path = './data/denoise/testresult/'
if not os.path.exists(test_result_output_path):
os.makedirs(test_result_output_path)
#逐个通道的去噪
"""
分配一个numpy数组，存储去噪后的结果
遍历所有通道，
对于每个通道，通过get_adjacent_spectral_bands获取其相邻的K个通道
调用hsid进行预测
将预测到的residual和输入的noise加起来，得到输出band
将去噪后的结果保存成mat结构
"""
hsid.eval()
psnr_list = []
for batch_idx, (noisy, label) in enumerate(test_dataloader):
noisy = noisy.type(torch.FloatTensor)
label = label.type(torch.FloatTensor)
batch_size, width, height, band_num = noisy.shape
denoised_hsi = np.zeros((width, height, band_num))
noisy = noisy.to(DEVICE)
label = label.to(DEVICE)
with torch.no_grad():
for i in range(band_num): #遍历每个band去处理
current_noisy_band = noisy[:, :, :, i]
current_noisy_band = current_noisy_band[:, None]
adj_spectral_bands = get_adjacent_spectral_bands(noisy, K, i)
#adj_spectral_bands = torch.transpose(adj_spectral_bands,3,1) #将通道数置换到第二维
adj_spectral_bands = adj_spectral_bands.permute(0, 3, 1, 2)
adj_spectral_bands_unsqueezed = adj_spectral_bands.unsqueeze(1)
#print(current_noisy_band.shape, adj_spectral_bands.shape)
residual = hsid(current_noisy_band,
adj_spectral_bands_unsqueezed)
denoised_band = residual + current_noisy_band
denoised_band_numpy = denoised_band.cpu().numpy().astype(
np.float32)
denoised_band_numpy = np.squeeze(denoised_band_numpy)
denoised_hsi[:, :, i] += denoised_band_numpy
test_label_current_band = label[:, :, :, i]
label_band_numpy = test_label_current_band.cpu().numpy(
).astype(np.float32)
label_band_numpy = np.squeeze(label_band_numpy)
#print(denoised_band_numpy.shape, label_band_numpy.shape, label.shape)
psnr = PSNR(denoised_band_numpy, label_band_numpy)
psnr_list.append(psnr)
mpsnr = np.mean(psnr_list)
denoised_hsi_trans = denoised_hsi.transpose(2, 0, 1)
test_label_hsi_trans = np.squeeze(label.cpu().numpy().astype(
np.float32)).transpose(2, 0, 1)
mssim = SSIM(denoised_hsi_trans, test_label_hsi_trans)
sam = SAM(denoised_hsi_trans, test_label_hsi_trans)
#计算pnsr和ssim
print("=====averPSNR:{:.4f}=====averSSIM:{:.4f}=====averSAM:{:.3f}".
format(mpsnr, mssim, sam))
#mdict是python字典类型，value值需要是一个numpy数组
scio.savemat(test_result_output_path + 'result.mat',
{'denoised': denoised_hsi})
def train_model_residual_lowlight_rdn():
device = DEVICE
#准备数据
train = np.load('./data/denoise/train_washington8.npy')
train = train.transpose((2, 1, 0))
test = np.load('./data/denoise/train_washington8.npy')
#test=test.transpose((2,1,0))
test = test.transpose((2, 1, 0)) #将通道维放在最前面
save_model_path = './checkpoints/hsirnd_denoise_l1loss'
if not os.path.exists(save_model_path):
os.mkdir(save_model_path)
#创建模型
net = HSIRDNECA_Denoise(K)
init_params(net)
net = nn.DataParallel(net).to(device)
#net = net.to(device)
#创建优化器
#hsid_optimizer = optim.Adam(net.parameters(), lr=INIT_LEARNING_RATE, betas=(0.9, 0,999))
hsid_optimizer = optim.Adam(net.parameters(), lr=INIT_LEARNING_RATE)
scheduler = MultiStepLR(hsid_optimizer, milestones=[200, 400], gamma=0.5)
#定义loss 函数
#criterion = nn.MSELoss()
gen_epoch_loss_list = []
cur_step = 0
best_psnr = 0
best_epoch = 0
best_iter = 0
start_epoch = 1
num_epoch = 600
mpsnr_list = []
for epoch in range(start_epoch, num_epoch + 1):
epoch_start_time = time.time()
scheduler.step()
print('epoch = ', epoch, 'lr={:.6f}'.format(scheduler.get_lr()[0]))
print(scheduler.get_lr())
gen_epoch_loss = 0
net.train()
channels = 191 # 191 channels
data_patches, data_cubic_patches = datagenerator(train, channels)
data_patches = torch.from_numpy(data_patches.transpose((
0,
3,
1,
2,
)))
data_cubic_patches = torch.from_numpy(
data_cubic_patches.transpose((0, 4, 1, 2, 3)))
DDataset = DenoisingDataset(data_patches, data_cubic_patches, SIGMA)
print('yes')
DLoader = DataLoader(dataset=DDataset,
batch_size=BATCH_SIZE,
shuffle=True) # loader出问题了
epoch_loss = 0
start_time = time.time()
#for batch_idx, (noisy, label) in enumerate([first_batch] * 300):
for step, x_y in enumerate(DLoader):
#print('batch_idx=', batch_idx)
batch_x_noise, batch_y_noise, batch_x = x_y[0], x_y[1], x_y[2]
batch_x_noise = batch_x_noise.to(device)
batch_y_noise = batch_y_noise.to(device)
batch_x = batch_x.to(device)
hsid_optimizer.zero_grad()
#denoised_img = net(noisy, cubic)
#loss = loss_fuction(denoised_img, label)
residual = net(batch_x_noise, batch_y_noise)
alpha = 0.8
loss = recon_criterion(residual, batch_x - batch_x_noise)
#loss = alpha*recon_criterion(residual, label-noisy) + (1-alpha)*loss_function_mse(residual, label-noisy)
#loss = recon_criterion(residual, label-noisy)
loss.backward() # calcu gradient
hsid_optimizer.step() # update parameter
if step % 10 == 0:
print('%4d %4d / %4d loss = %2.8f' %
(epoch + 1, step, data_patches.size(0) // BATCH_SIZE,
loss.item() / BATCH_SIZE))
#scheduler.step()
#print("Decaying learning rate to %g" % scheduler.get_last_lr()[0])
torch.save(
{
'gen': net.state_dict(),
'gen_opt': hsid_optimizer.state_dict(),
},
f"{save_model_path}/hsid_rdn_eca_l1_loss_600epoch_patchsize32_{epoch}.pth"
)
#测试代码
net.eval()
"""
channel_s = 191 # 设置多少波段
data_patches, data_cubic_patches = datagenerator(test, channel_s)
data_patches = torch.from_numpy(data_patches.transpose((0, 3, 1, 2,)))
data_cubic_patches = torch.from_numpy(data_cubic_patches.transpose((0, 4, 1, 2, 3)))
DDataset = DenoisingDataset(data_patches, data_cubic_patches, SIGMA)
DLoader = DataLoader(dataset=DDataset, batch_size=BATCH_SIZE, shuffle=True)
epoch_loss = 0
for step, x_y in enumerate(DLoader):
batch_x_noise, batch_y_noise, batch_x = x_y[0], x_y[1], x_y[2]
batch_x_noise = batch_x_noise.to(DEVICE)
batch_y_noise = batch_y_noise.to(DEVICE)
batch_x = batch_x.to(DEVICE)
residual = net(batch_x_noise, batch_y_noise)
loss = loss_fuction(residual, batch_x-batch_x_noise)
epoch_loss += loss.item()
if step % 10 == 0:
print('%4d %4d / %4d test loss = %2.4f' % (
epoch + 1, step, data_patches.size(0) // BATCH_SIZE, loss.item() / BATCH_SIZE))
"""
#加载数据
test_data_dir = './data/denoise/test/'
test_set = HsiTrainDataset(test_data_dir)
test_dataloader = DataLoader(test_set, batch_size=1, shuffle=False)
#指定结果输出路径
test_result_output_path = './data/denoise/testresult/'
if not os.path.exists(test_result_output_path):
os.makedirs(test_result_output_path)
#逐个通道的去噪
"""
分配一个numpy数组，存储去噪后的结果
遍历所有通道，
对于每个通道，通过get_adjacent_spectral_bands获取其相邻的K个通道
调用hsid进行预测
将预测到的residual和输入的noise加起来，得到输出band
将去噪后的结果保存成mat结构
"""
psnr_list = []
for batch_idx, (noisy, label) in enumerate(test_dataloader):
noisy = noisy.type(torch.FloatTensor)
label = label.type(torch.FloatTensor)
batch_size, width, height, band_num = noisy.shape
denoised_hsi = np.zeros((width, height, band_num))
noisy = noisy.to(DEVICE)
label = label.to(DEVICE)
with torch.no_grad():
for i in range(band_num): #遍历每个band去处理
current_noisy_band = noisy[:, :, :, i]
current_noisy_band = current_noisy_band[:, None]
adj_spectral_bands = get_adjacent_spectral_bands(
noisy, K, i)
#adj_spectral_bands = torch.transpose(adj_spectral_bands,3,1) #将通道数置换到第二维
adj_spectral_bands = adj_spectral_bands.permute(0, 3, 1, 2)
adj_spectral_bands_unsqueezed = adj_spectral_bands.unsqueeze(
1)
#print(current_noisy_band.shape, adj_spectral_bands.shape)
residual = net(current_noisy_band,
adj_spectral_bands_unsqueezed)
denoised_band = residual + current_noisy_band
denoised_band_numpy = denoised_band.cpu().numpy().astype(
np.float32)
denoised_band_numpy = np.squeeze(denoised_band_numpy)
denoised_hsi[:, :, i] += denoised_band_numpy
test_label_current_band = label[:, :, :, i]
label_band_numpy = test_label_current_band.cpu().numpy(
).astype(np.float32)
label_band_numpy = np.squeeze(label_band_numpy)
#print(denoised_band_numpy.shape, label_band_numpy.shape, label.shape)
psnr = PSNR(denoised_band_numpy, label_band_numpy)
psnr_list.append(psnr)
mpsnr = np.mean(psnr_list)
mpsnr_list.append(mpsnr)
denoised_hsi_trans = denoised_hsi.transpose(2, 0, 1)
test_label_hsi_trans = np.squeeze(label.cpu().numpy().astype(
np.float32)).transpose(2, 0, 1)
mssim = SSIM(denoised_hsi_trans, test_label_hsi_trans)
sam = SAM(denoised_hsi_trans, test_label_hsi_trans)
#计算pnsr和ssim
print(
"=====averPSNR:{:.3f}=====averSSIM:{:.4f}=====averSAM:{:.3f}".
format(mpsnr, mssim, sam))
#保存best模型
if mpsnr > best_psnr:
best_psnr = mpsnr
best_epoch = epoch
best_iter = cur_step
torch.save(
{
'epoch': epoch,
'gen': net.state_dict(),
'gen_opt': hsid_optimizer.state_dict(),
},
f"{save_model_path}/hsid_rdn_eca_l1_loss_600epoch_patchsize32_best.pth"
)
print(
"[epoch %d it %d PSNR: %.4f --- best_epoch %d best_iter %d Best_PSNR %.4f]"
% (epoch, cur_step, mpsnr, best_epoch, best_iter, best_psnr))
print(
"------------------------------------------------------------------"
)
print("Epoch: {}\tTime: {:.4f}\tLoss: {:.4f}\tLearningRate {:.6f}".
format(epoch,
time.time() - epoch_start_time, gen_epoch_loss,
INIT_LEARNING_RATE))
print(
"------------------------------------------------------------------"
)
def predict_lowlight_residual():
#加载模型
encam = ENCAM()
#hsid = nn.DataParallel(hsid).to(DEVICE)
#device = torch.device("cuda:1" if torch.cuda.is_available() else "cpu")
encam = encam.to(DEVICE)
encam.eval()
encam.load_state_dict(
torch.load('./checkpoints/encam_best_08_27.pth',
map_location='cuda:0')['gen'])
#加载数据
mat_src_path = '../HSID/data/test_lowlight/origin/soup_bigcorn_orange_1ms.mat'
test_label = scio.loadmat(mat_src_path)['label']
#test=test.transpose((2,0,1)) #将通道维放在最前面：191*1280*307
test_data_dir = '../HSID/data/test_lowlight/origin/'
test_set = HsiLowlightTestDataset(test_data_dir)
test_dataloader = DataLoader(test_set, batch_size=1, shuffle=False)
#指定结果输出路径
test_result_output_path = './data/testresult/'
if not os.path.exists(test_result_output_path):
os.makedirs(test_result_output_path)
#逐个通道的去噪
"""
分配一个numpy数组，存储去噪后的结果
遍历所有通道，
对于每个通道，通过get_adjacent_spectral_bands获取其相邻的K个通道
调用hsid进行预测
将预测到的residual和输入的noise加起来，得到输出band
将去噪后的结果保存成mat结构
"""
for batch_idx, (noisy, label) in enumerate(test_dataloader):
noisy = noisy.type(torch.FloatTensor)
label = label.type(torch.FloatTensor)
batch_size, width, height, band_num = noisy.shape
denoised_hsi = np.zeros((width, height, band_num))
noisy = noisy.to(DEVICE)
label = label.to(DEVICE)
with torch.no_grad():
for i in range(band_num): #遍历每个band去处理
current_noisy_band = noisy[:, :, :, i]
current_noisy_band = current_noisy_band[:, None]
adj_spectral_bands = get_adjacent_spectral_bands(
noisy, K, i) # shape: batch_size, width, height, band_num
adj_spectral_bands = adj_spectral_bands.permute(
0, 3, 1,
2) #交换第一维和第三维 ，shape: batch_size, band_num, height, width
adj_spectral_bands = torch.unsqueeze(adj_spectral_bands, 1)
adj_spectral_bands = adj_spectral_bands.to(DEVICE)
print('adj_spectral_bands : ', adj_spectral_bands.shape)
print('adj_spectral_bands shape[4] =',
adj_spectral_bands.shape[4])
#这里需要将current_noisy_band和adj_spectral_bands拆分成4份，每份大小为batchsize，1， band_num , height/2, width/2
current_noisy_band_00 = current_noisy_band[:, :,
0:current_noisy_band
.shape[2] // 2,
0:current_noisy_band
.shape[3] // 2]
adj_spectral_bands_00 = adj_spectral_bands[:, :, :,
0:adj_spectral_bands
.shape[3] // 2,
0:adj_spectral_bands
.shape[4] // 2]
residual_00 = encam(current_noisy_band_00,
adj_spectral_bands_00)
denoised_band_00 = current_noisy_band_00 + residual_00
current_noisy_band_00 = current_noisy_band[:, :,
0:current_noisy_band
.shape[2] // 2,
0:current_noisy_band
.shape[3] // 2]
adj_spectral_bands_00 = adj_spectral_bands[:, :, :,
0:adj_spectral_bands
.shape[3] // 2,
0:adj_spectral_bands
.shape[4] // 2]
residual_00 = encam(current_noisy_band_00,
adj_spectral_bands_00)
denoised_band_01 = current_noisy_band_00 + residual_00
current_noisy_band_00 = current_noisy_band[:, :, 0:(
current_noisy_band.shape[2] //
2), 0:(current_noisy_band.shape[3] // 2)]
adj_spectral_bands_00 = adj_spectral_bands[:, :, :,
0:adj_spectral_bands
.shape[3] // 2,
0:adj_spectral_bands
.shape[4] // 2]
residual_00 = encam(current_noisy_band_00,
adj_spectral_bands_00)
denoised_band_10 = current_noisy_band_00 + residual_00
current_noisy_band_00 = current_noisy_band[:, :,
0:current_noisy_band
.shape[2] // 2,
0:current_noisy_band
.shape[3] // 2]
adj_spectral_bands_11 = adj_spectral_bands[:, :, :,
0:adj_spectral_bands
.shape[3] // 2,
0:adj_spectral_bands
.shape[4] // 2]
residual_00 = encam(current_noisy_band_00,
adj_spectral_bands_00)
denoised_band_11 = current_noisy_band_00 + residual_00
denoised_band_0 = torch.cat(
(denoised_band_00, denoised_band_01), dim=3)
denoised_band_1 = torch.cat(
(denoised_band_10, denoised_band_11), dim=3)
denoised_band = torch.cat((denoised_band_0, denoised_band_1),
dim=2)
denoised_band_numpy = denoised_band.cpu().numpy().astype(
np.float32)
denoised_band_numpy = np.squeeze(denoised_band_numpy)
denoised_hsi[:, :, i] = denoised_band_numpy
#mdict是python字典类型，value值需要是一个numpy数组
scio.savemat(test_result_output_path + 'result.mat',
{'denoised': denoised_hsi})
psnr = PSNR(denoised_hsi, test_label)
ssim = SSIM(denoised_hsi, test_label)
sam = SAM(denoised_hsi, test_label)
#计算pnsr和ssim
print("=====averPSNR:{:.3f}=====averSSIM:{:.4f}=====averSAM:{:.3f}".format(
psnr, ssim, sam))
def main():
device = DEVICE
#准备数据
train_set = HsiTrainDataset('./data/train/')
train_loader = DataLoader(dataset=train_set,
batch_size=BATCH_SIZE,
shuffle=True)
#创建模型
net = HSID_1x3(K)
init_params(net)
net = nn.DataParallel(net).to(device)
#创建优化器
#hsid_optimizer = optim.Adam(net.parameters(), lr=INIT_LEARNING_RATE, betas=(0.9, 0,999))
hsid_optimizer = optim.Adam(
net.parameters(),
lr=INIT_LEARNING_RATE) #betas default value 就是0.9和0.999
scheduler = MultiStepLR(hsid_optimizer,
milestones=[15, 30, 45],
gamma=0.25)
#定义loss 函数
#criterion = nn.MSELoss()
global tb_writer
tb_writer = get_summary_writer(log_dir='logs')
gen_minibatch_loss_list = []
gen_epoch_loss_list = []
cur_step = 0
first_batch = next(iter(train_loader))
for epoch in range(NUM_EPOCHS):
gen_epoch_loss = 0
net.train()
#for batch_idx, (noisy, label) in enumerate([first_batch] * 300):
for batch_idx, (noisy, label) in enumerate(train_loader):
noisy = noisy.to(device)
label = label.to(device)
batch_size, height, width, band_num = noisy.shape
""""
our method traverses all the bands through one-by-one mode,
which simultaneously employing spatial–spectral information
with spatial and spatial–spectral filters, respectively
"""
band_loss = 0
for i in range(band_num): #遍历每个band去处理
single_noisy_band = noisy[:, :, :, i]
single_noisy_band_cloned = single_noisy_band[:, None].clone()
single_label_band = label[:, :, :, i]
single_label_band_cloned = single_label_band[:, None].clone()
adj_spectral_bands = get_adjacent_spectral_bands(noisy, K, i)
#print('adj_spectral_bands.shape =', adj_spectral_bands.shape)
#print(type(adj_spectral_bands))
adj_spectral_bands_transposed = torch.transpose(
adj_spectral_bands, 3, 1).clone()
#print('transposed adj_spectral_bands.shape =', adj_spectral_bands.shape)
#print(type(adj_spectral_bands))
denoised_img = net(single_noisy_band_cloned,
adj_spectral_bands_transposed)
loss = loss_fuction(single_label_band_cloned, denoised_img)
hsid_optimizer.zero_grad()
loss.backward() # calcu gradient
hsid_optimizer.step() # update parameter
## Logging
band_loss += loss.item()
if i % 20 == 0:
print(
f"Epoch {epoch}: Step {cur_step}: bandnum {i}: band MSE loss: {loss.item()}"
)
gen_minibatch_loss_list.append(band_loss)
gen_epoch_loss += band_loss
if cur_step % display_step == 0:
if cur_step > 0:
print(
f"Epoch {epoch}: Step {cur_step}: MSE loss: {band_loss}"
)
else:
print("Pretrained initial state")
tb_writer.add_scalar("MSE loss", band_loss, cur_step)
#step ++,每一次循环，每一个batch的处理，叫做一个step
cur_step += 1
scheduler.step()
print("Decaying learning rate to %g" % scheduler.get_lr()[0])
gen_epoch_loss_list.append(gen_epoch_loss)
tb_writer.add_scalar("mse epoch loss", gen_epoch_loss, epoch)
torch.save(
{
'gen': net.state_dict(),
'gen_opt': hsid_optimizer.state_dict(),
}, f"checkpoints/hsid_1x3{epoch}.pth")
tb_writer.close()
def train_model():
device = DEVICE
#准备数据
train_set = HsiCubicTrainDataset('./data/train_cubic/')
train_loader = DataLoader(dataset=train_set,
batch_size=BATCH_SIZE,
shuffle=True)
#加载测试label数据
test_label_hsi = np.load('./data/origin/test_washington.npy')
#加载测试数据
test_data_dir = './data/test_level25/'
test_set = HsiTrainDataset(test_data_dir)
test_dataloader = DataLoader(test_set, batch_size=1, shuffle=False)
#创建模型
net = HSID_1x3(K)
init_params(net)
net = nn.DataParallel(net).to(device)
#创建优化器
#hsid_optimizer = optim.Adam(net.parameters(), lr=INIT_LEARNING_RATE, betas=(0.9, 0,999))
hsid_optimizer = optim.Adam(net.parameters(), lr=INIT_LEARNING_RATE)
scheduler = MultiStepLR(hsid_optimizer,
milestones=[15, 30, 45],
gamma=0.25)
#定义loss 函数
#criterion = nn.MSELoss()
global tb_writer
tb_writer = get_summary_writer(log_dir='logs')
gen_epoch_loss_list = []
cur_step = 0
first_batch = next(iter(train_loader))
for epoch in range(NUM_EPOCHS):
gen_epoch_loss = 0
net.train()
#for batch_idx, (noisy, label) in enumerate([first_batch] * 300):
for batch_idx, (noisy, cubic, label) in enumerate(train_loader):
noisy = noisy.to(device)
label = label.to(device)
cubic = cubic.to(device)
hsid_optimizer.zero_grad()
denoised_img = net(noisy, cubic)
loss = loss_fuction(denoised_img, label)
loss.backward() # calcu gradient
hsid_optimizer.step() # update parameter
gen_epoch_loss += loss.item()
if cur_step % display_step == 0:
if cur_step > 0:
print(
f"Epoch {epoch}: Step {cur_step}: MSE loss: {loss.item()}"
)
else:
print("Pretrained initial state")
tb_writer.add_scalar("MSE loss", loss.item(), cur_step)
#step ++,每一次循环，每一个batch的处理，叫做一个step
cur_step += 1
gen_epoch_loss_list.append(gen_epoch_loss)
tb_writer.add_scalar("mse epoch loss", gen_epoch_loss, epoch)
scheduler.step()
print("Decaying learning rate to %g" % scheduler.get_last_lr()[0])
torch.save(
{
'gen': net.state_dict(),
'gen_opt': hsid_optimizer.state_dict(),
}, f"checkpoints/hsid_{epoch}.pth")
#预测代码
net.eval()
for batch_idx, (noisy, label) in enumerate(test_dataloader):
noisy = noisy.type(torch.FloatTensor)
label = label.type(torch.FloatTensor)
batch_size, width, height, band_num = noisy.shape
denoised_hsi = np.zeros((width, height, band_num))
noisy = noisy.to(DEVICE)
label = label.to(DEVICE)
with torch.no_grad():
for i in range(band_num): #遍历每个band去处理
current_noisy_band = noisy[:, :, :, i]
current_noisy_band = current_noisy_band[:, None]
adj_spectral_bands = get_adjacent_spectral_bands(
noisy, K,
i) # shape: batch_size, width, height, band_num
adj_spectral_bands = torch.transpose(
adj_spectral_bands, 3, 1
) #交换第一维和第三维 ，shape: batch_size, band_num, height, width
denoised_band = net(current_noisy_band, adj_spectral_bands)
denoised_band_numpy = denoised_band.cpu().numpy().astype(
np.float32)
denoised_band_numpy = np.squeeze(denoised_band_numpy)
denoised_hsi[:, :, i] = denoised_band_numpy
psnr = PSNR(denoised_hsi, test_label_hsi)
ssim = SSIM(denoised_hsi, test_label_hsi)
sam = SAM(denoised_hsi, test_label_hsi)
#计算pnsr和ssim
print("=====averPSNR:{:.3f}=====averSSIM:{:.4f}=====averSAM:{:.3f}".
format(psnr, ssim, sam))
tb_writer.close()
def predict():
#加载模型
hsid = HSID(36)
hsid = nn.DataParallel(hsid).to(DEVICE)
hsid.load_state_dict(torch.load('./checkpoints/hsid_5.pth')['gen'])
#加载数据
test=np.load('./data/origin/test_washington.npy')
#test=test.transpose((2,0,1)) #将通道维放在最前面：191*1280*307
test_data_dir = './data/test_level25/'
test_set = HsiTrainDataset(test_data_dir)
test_dataloader = DataLoader(test_set, batch_size=1, shuffle=False)
#指定结果输出路径
test_result_output_path = './data/testresult/'
if not os.path.exists(test_result_output_path):
os.makedirs(test_result_output_path)
#逐个通道的去噪
"""
分配一个numpy数组，存储去噪后的结果
遍历所有通道，
对于每个通道，通过get_adjacent_spectral_bands获取其相邻的K个通道
调用hsid进行预测
将预测到的residual和输入的noise加起来，得到输出band
将去噪后的结果保存成mat结构
"""
for batch_idx, (noisy, label) in enumerate(test_dataloader):
noisy = noisy.type(torch.FloatTensor)
label = label.type(torch.FloatTensor)
batch_size, width, height, band_num = noisy.shape
denoised_hsi = np.zeros((width, height, band_num))
noisy = noisy.to(DEVICE)
label = label.to(DEVICE)
with torch.no_grad():
for i in range(band_num): #遍历每个band去处理
current_noisy_band = noisy[:,:,:,i]
current_noisy_band = current_noisy_band[:,None]
adj_spectral_bands = get_adjacent_spectral_bands(noisy, K, i)# shape: batch_size, width, height, band_num
adj_spectral_bands = torch.transpose(adj_spectral_bands,3,1)#交换第一维和第三维 ，shape: batch_size, band_num, height, width
denoised_band = hsid(current_noisy_band, adj_spectral_bands)
denoised_band_numpy = denoised_band.cpu().numpy().astype(np.float32)
denoised_band_numpy = np.squeeze(denoised_band_numpy)
denoised_hsi[:,:,i] = denoised_band_numpy
#mdict是python字典类型，value值需要是一个numpy数组
scio.savemat(test_result_output_path + 'result.mat', {'denoised': denoised_hsi})
psnr = PSNR(denoised_hsi, test)
ssim = SSIM(denoised_hsi, test)
sam = SAM(denoised_hsi, test)
#计算pnsr和ssim
print("=====averPSNR:{:.3f}=====averSSIM:{:.4f}=====averSAM:{:.3f}".format(psnr, ssim, sam))
def predict_lowlight_residual():
#加载模型
encam = ENCAM()
#hsid = nn.DataParallel(hsid).to(DEVICE)
#device = torch.device("cuda:1" if torch.cuda.is_available() else "cpu")
encam = encam.to(DEVICE)
encam.eval()
encam.load_state_dict(
torch.load('./checkpoints/encam_best.pth',
map_location='cuda:0')['gen'])
#加载数据
mat_src_path = '../HSID/data/test_lowlight/origin/soup_bigcorn_orange_1ms.mat'
test_label = scio.loadmat(mat_src_path)['label']
#test=test.transpose((2,0,1)) #将通道维放在最前面：191*1280*307
#test_label_tensor = torch.from_numpy(test_label)
#test_label_tensor = torch.unsqueeze(test_label_tensor, 0)
#test_label_tensor = test_label_tensor.permute(0, 3,1,2)
#test_label_tensor = F.interpolate(test_label_tensor, scale_factor=0.5, mode='bilinear')
#test_label_tensor = test_label_tensor.permute(0, 2,3,1)
#test_label = torch.squeeze(test_label_tensor).numpy()
test_data_dir = '../HSID/data/test_lowlight/origin/'
test_set = HsiLowlightTestDataset(test_data_dir)
test_dataloader = DataLoader(test_set, batch_size=1, shuffle=False)
#指定结果输出路径
test_result_output_path = './data/testresult/'
if not os.path.exists(test_result_output_path):
os.makedirs(test_result_output_path)
#逐个通道的去噪
"""
分配一个numpy数组，存储去噪后的结果
遍历所有通道，
对于每个通道，通过get_adjacent_spectral_bands获取其相邻的K个通道
调用hsid进行预测
将预测到的residual和输入的noise加起来，得到输出band
将去噪后的结果保存成mat结构
"""
for batch_idx, (noisy, label) in enumerate(test_dataloader):
noisy = noisy.type(torch.FloatTensor)
label = label.type(torch.FloatTensor)
batch_size, width, height, band_num = noisy.shape
denoised_hsi = np.zeros((width, height, band_num))
#对图像下采样
noisy_permute = noisy.permute(
0, 3, 1, 2) #交换第一维和第三维 ，shape: batch_size, band_num, height, width
label_permute = label.permute(0, 3, 1, 2)
noisy_down = F.interpolate(noisy_permute,
scale_factor=0.5,
mode='bilinear')
label_down = F.interpolate(label_permute,
scale_factor=0.5,
mode='bilinear')
#batch_size, band_num, width, height = noisy_down.shape
#denoised_hsi = np.zeros((width, height, band_num))
noisy_down = noisy_down.to(DEVICE)
label_down = label_down.to(DEVICE)
with torch.no_grad():
for i in range(band_num): #遍历每个band去处理
current_noisy_band = noisy_down[:, i, :, :]
current_noisy_band = current_noisy_band[:, None]
noisy_down = noisy_down.permute(0, 2, 3, 1)
adj_spectral_bands = get_adjacent_spectral_bands(
noisy_down, K,
i) # shape: batch_size, width, height, band_num
noisy_down = noisy_down.permute(0, 3, 1, 2)
adj_spectral_bands = adj_spectral_bands.permute(
0, 3, 1,
2) #交换第一维和第三维 ，shape: batch_size, band_num, height, width
adj_spectral_bands = torch.unsqueeze(adj_spectral_bands, 1)
adj_spectral_bands = adj_spectral_bands.to(DEVICE)
residual = encam(current_noisy_band, adj_spectral_bands)
denoised_band = current_noisy_band + residual
#对denoised_hsi进行上采样
denoised_band = F.interpolate(denoised_band,
scale_factor=2,
mode='bilinear')
denoised_band_numpy = denoised_band.cpu().numpy().astype(
np.float32)
denoised_band_numpy = np.squeeze(denoised_band_numpy)
denoised_hsi[:, :, i] = denoised_band_numpy
#mdict是python字典类型，value值需要是一个numpy数组
scio.savemat(test_result_output_path + 'result.mat',
{'denoised': denoised_hsi})
psnr = PSNR(denoised_hsi, test_label)
ssim = SSIM(denoised_hsi, test_label)
sam = SAM(denoised_hsi, test_label)
#计算pnsr和ssim
print("=====averPSNR:{:.4f}=====averSSIM:{:.4f}=====averSAM:{:.3f}".format(
psnr, ssim, sam))