DD-Net always peformed similar to UNet and DnCNN. The training time was the longest out of other 2 models. It took aprox. 24 hours to train the network.
paper: https://ieeexplore.ieee.org/document/8331861 github: https://github.com/zzc623/DD_Net
Authors write "The DD-Net was trained by the Adam algorithm [54]. The learning rate was initially set at 10−4 and slowly decreased continuously down to 10−5. The size of mini-batch was 5. DD- Net was implemented using Tensorflow [55] on a personal workstation with Intel Core i5-7400 CPU and 16GB RAM. A GPU card (Nvidia GTX Titan X) accelerated the training process. All the convolution and deconvolution filters were initialized with random Gaussian distributions with zero mean and 0.01 standard deviation."
Results
The average PSNR and SSIM values over 354 test images are displayed in the table below.
Learning rate 10^-4
40 epochs
| Low Dose Image | UNet | DnCNN | DDNet |
|---|---|---|---|
| sparseview_60 | Avg PSNR: 33.28 Avg SSIM: 0.8858 | Avg PSNR: 32.30 Avg SSIM: 0.8560 | Avg PSNR: 32.96 Avg SSIM: 0.8797 |
| sparseview_90 | Avg PSNR: 35.42 Avg SSIM: 0.9038 | Avg PSNR: 35.13 Avg SSIM: 0.8892 | Avg PSNR: 35.29 Avg SSIM: 0.9011 |
| sparseview_180 | Avg PSNR: 39.48 Avg SSIM: 0.9319 | Avg PSNR: 39.77 Avg SSIM: 0.9341 | Avg PSNR: 39.55 Avg SSIM: 0.9322 |
| ldct_7e4 | Avg PSNR: 41.78 Avg SSIM: 0.9429 | Avg PSNR: 42.00 Avg SSIM: 0.9444 | Avg PSNR: 41.84 Avg SSIM: 0.9431 |
| ldct_1e5 | Avg PSNR: 42.11 Avg SSIM: 0.9441 | Avg PSNR: 42.32 Avg SSIM: 0.9456 | Avg PSNR: 42.23 Avg SSIM: 0.9448 |
| ldct_2e5 | Avg PSNR: 42.69 Avg SSIM: 0.9466 | Avg PSNR: 42.87 Avg SSIM: 0.9477 | Avg PSNR: 42.77 Avg SSIM: 0.9471 |
def denseblock(input):
# - L1
num_filters = 16
d2_1 = BatchNormalization()(input)
d2_1 = Activation('relu')(d2_1)
d2_1 = Conv2D(num_filters*4, (1, 1), padding='same', use_bias=True, strides=(1, 1))(d2_1)
d2_1 = BatchNormalization()(d2_1)
d2_1 = Activation('relu')(d2_1)
d2_1 = Conv2D(num_filters, (5, 5), padding='same', use_bias=True, strides=(1, 1))(d2_1)
d2_1 = concatenate([input, d2_1])
# - L2
d2_2 = BatchNormalization()(d2_1)
d2_2 = Activation('relu')(d2_2)
d2_2 = Conv2D(num_filters*4, (1, 1), padding='same', use_bias=True, strides=(1, 1))(d2_2)
d2_2 = BatchNormalization()(d2_2)
d2_2 = Activation('relu')(d2_2)
d2_2 = Conv2D(num_filters, (5, 5), padding='same', use_bias=True, strides=(1, 1))(d2_2)
d2_2 = concatenate([input, d2_1, d2_2])
# - L3
d2_3 = BatchNormalization()(d2_2)
d2_3 = Activation('relu')(d2_3)
d2_3 = Conv2D(num_filters*4, (1, 1), padding='same', use_bias=True, strides=(1, 1))(d2_3)
d2_3 = BatchNormalization()(d2_3)
d2_3 = Activation('relu')(d2_3)
d2_3 = Conv2D(num_filters, (5, 5), padding='same', use_bias=True, strides=(1, 1))(d2_3)
d2_3 = concatenate([input, d2_1, d2_2, d2_3])
# - L4
d2_4 = BatchNormalization()(d2_3)
d2_4 = Activation('relu')(d2_4)
d2_4 = Conv2D(num_filters*4, (1, 1), padding='same', use_bias=True, strides=(1, 1))(d2_4)
d2_4 = BatchNormalization()(d2_4)
d2_4 = Activation('relu')(d2_4)
d2_4 = Conv2D(num_filters, (5, 5), padding='same', use_bias=True, strides=(1, 1))(d2_4)
d2_4 = concatenate([input, d2_1, d2_2, d2_3, d2_4])
return d2_4
def dd_net(ldct_img, is_training= True):
net = ldct_img
num_filter = 16
# ---A1 Layer-----------------------
h_conv1 = Conv2D(16, (7, 7), padding='same', use_bias=True, strides=(1, 1))(net)
a1 = MaxPooling2D((3, 3), strides=(2, 2), padding='same') (h_conv1)
# images 256 X 256
d1 = denseblock(a1)
a1 = BatchNormalization()(d1)
a1 = Activation('relu')(a1)
h_conv1_T = Conv2D(16, (1, 1), strides=(1, 1), use_bias=True) (a1)
# ----A2 Layer---------------------
a2 = MaxPooling2D((2, 2),strides=(2, 2), padding='same') (h_conv1_T)
# images 128 X 128 d
d2 = denseblock(a2)
a2 = BatchNormalization()(d2)
a2 = Activation('relu')(a2)
h_conv2_T = Conv2D(16, (1, 1), strides=(1, 1), use_bias=True) (a2)
# images 128 X 128
# # ----A3 Layer----------------------
a3 = MaxPooling2D((2, 2), strides=(2, 2), padding='same') (h_conv2_T)
# images 64 X 64
d3 = denseblock(a3)
a3 = BatchNormalization()(d3)
a3 = Activation('relu')(a3)
h_conv3_T = Conv2D(16, (1, 1), strides=(1, 1), use_bias=True) (a3)
# ----A4 Layer----------------------
a4 = MaxPooling2D((2, 2), strides=(2, 2), padding='same') (h_conv3_T)
# images 32 X 3
d4 = denseblock(a4)
a4 = BatchNormalization()(d4)
a4 = Activation('relu')(a4)
h_conv4_T = Conv2D(16, (1, 1), strides=(1, 1), use_bias=True) (a4)
# #----B1 Layer-----------------------
b1 = UpSampling2D((2, 2), interpolation="nearest") (h_conv4_T)
# images 64 X 64
b1 = concatenate([b1, h_conv3_T])
b1 = Conv2DTranspose(num_filter*2, (5, 5), padding='same', strides=(1, 1)) (b1)
b1 = Activation('relu')(b1)
b1 = BatchNormalization()(b1)
b1 = Conv2DTranspose(16, (1, 1), padding='same', strides=(1, 1)) (b1)
b1 = Activation('relu')(b1)
b1 = BatchNormalization()(b1)
# #----B2 Layer-----------------------
b2 = UpSampling2D((2, 2), interpolation="nearest") (b1)
# images 128 X 128
b2 = concatenate([b2, h_conv2_T])
b2 = Conv2DTranspose(num_filter*2, (5, 5), padding='same', strides=(1, 1)) (b2)
b2 = Activation('relu')(b2)
b2 = BatchNormalization()(b2)
b2 = Conv2DTranspose(16, (1, 1), padding='same', strides=(1, 1)) (b2)
b2 = Activation('relu')(b2)
b2 = BatchNormalization()(b2)
#----B3 Layer------------------------conv6
b3 = UpSampling2D((2, 2),interpolation="nearest") (b2)
# images 256 X 256
b3 = concatenate([b3, h_conv1_T])
b3 = Conv2DTranspose(num_filter*2, (5, 5), padding='same', strides=(1, 1)) (b3)
b3 = Activation('relu')(b3)
b3 = BatchNormalization()(b3)
b3 = Conv2DTranspose(16, (1, 1), padding='same', strides=(1, 1)) (b3)
b3 = Activation('relu')(b3)
b3 = BatchNormalization()(b3)
#----B4 Layer-------------------------
b4 = UpSampling2D((2, 2),interpolation="nearest") (b3)
# images 512 X 512
b4 = concatenate([b4, h_conv1])
b4 = Conv2DTranspose(num_filter*2, (5, 5),padding='same', strides=(1, 1)) (b4)
b4 = Activation('relu')(b4)
# b4 = BatchNormalization()(b4)
output_img = Conv2DTranspose(1, (1, 1), strides=(1, 1)) (b4)
# output_img = Activation('relu')(output_img) # in paper but DIDN'T CONVERGE
# ------ end B4 layer
denoised_image = Subtract()([net, output_img])
return denoised_image
Each training process took approximately 24 hours.
(base) [npovey@ka ~]$ conda create -n keras-gpu python=3.6 numpy scipy keras-gpu
(base) [npovey@ka unet4]$ conda activate keras-gpu
(keras-gpu) [npovey@ka unet4]$ pip install pandas
(keras-gpu) [npovey@ka unet4]$ pip install Pillow
(keras-gpu) [npovey@ka unet4]$ pip install matplotlib
(keras-gpu) [npovey@ka unet4]$ python main.py
In main.py (line 77) use train.
#parser.add_argument('--phase', dest='phase', default='test', help='test')
parser.add_argument('--phase', dest='phase', default='train', help ='train')Change from train to test in main.py line 77
#parser.add_argument('--phase', dest='phase', default='test', help='test')
parser.add_argument('--phase', dest='phase', default='train', help ='train')Sign in into remote linux machines
nps-MacBook-Air-2:Desktop np$ ssh npovey@ka...[npovey@ka dd_net]$ ls
main.py model.py model.py~ utils.pyGot core dumped problem as all GPUs were taken
Aborted (core dumped)
(keras-gpu) [npovey@ka dd_net]$ Check available GPUs
(keras-gpu) [npovey@ka dd_net]$ nvidia-smi -L
GPU 0: Quadro RTX 5000 (UUID: GPU-1f923d52-ea64-f463-96a4-3bece2719a8b)
GPU 1: Quadro RTX 5000 (UUID: GPU-20947179-1907-5468-1325-a3fc16f5a54e)
(keras-gpu) [npovey@ka dd_net]$ nvidia-smi
Fri Mar 6 14:25:38 2020
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 418.87.00 Driver Version: 418.87.00 CUDA Version: 10.1 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 Quadro RTX 5000 Off | 00000000:67:00.0 Off | Off |
| 43% 68C P2 212W / 230W | 15871MiB / 16095MiB | 94% Default |
+-------------------------------+----------------------+----------------------+
| 1 Quadro RTX 5000 Off | 00000000:68:00.0 Off | Off |
| 55% 78C P2 221W / 230W | 15869MiB / 16092MiB | 96% Default |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: GPU Memory |
| GPU PID Type Process name Usage |
|=============================================================================|
| 0 78799 C python 15861MiB |
| 1 26617 C python 15859MiB |
+-----------------------------------------------------------------------------+
WARNING: infoROM is corrupted at gpu 0000:67:00.0
(keras-gpu) [npovey@ka dd_net]$
Looks like both GPUs are taken
Trying a little bit later
(base) [npovey@ka ~]$ nvidia-smi
Fri Mar 6 20:23:45 2020
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 418.87.00 Driver Version: 418.87.00 CUDA Version: 10.1 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 Quadro RTX 5000 Off | 00000000:67:00.0 Off | Off |
| 43% 65C P2 74W / 230W | 15871MiB / 16095MiB | 0% Default |
+-------------------------------+----------------------+----------------------+
| 1 Quadro RTX 5000 Off | 00000000:68:00.0 Off | Off |
| 35% 35C P0 N/A / N/A | 0MiB / 16092MiB | 0% Default |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: GPU Memory |
| GPU PID Type Process name Usage |
|=============================================================================|
| 0 78799 C python 15861MiB |
+-----------------------------------------------------------------------------+
WARNING: infoROM is corrupted at gpu 0000:67:00.0
(base) [npovey@ka ~]$ We can observe that GPU1 is free to run our code.
Run the code on the remote linux machine
[npovey@ka data]$ cd dd_net/
[npovey@ka dd_net]$ source activate keras-gpu
(keras-gpu) [npovey@ka dd_net]$
(keras-gpu) [npovey@ka dd_net]$ python main.pyAs the code takes long time to run I recommend running it using screen
[npovey@ka dd_net]$ screen -S dd_net1
[npovey@ka dd_net]$ source activate keras-gpu
[npovey@ka dd_net]$ python main.py
(press ctlr+A, D to detach from screen)
[detached from 6921.dd_net1]
To copy result to a local machine and view them
nps-MacBook-Air-2:Desktop np$ scp -r npovey@ka:/home/npovey/data/dd_net/output_dd_net_60.txt DD_Net/
npovey@ka's password:
output_dd_net_60.txt 100% 32KB 321.5KB/s 00:00
nps-MacBook-Air-2:Desktop np$
[npovey@ka dd_net]$ screen -S dd_net1
[npovey@ka dd_net]$ source activate keras-gpu
(keras-gpu) [npovey@ka dncnn1]$ python main.py > output_ddnet_60.txt[npovey@ka dd_net]$ screen -S dd_net1
[npovey@ka dd_net]$ source activate keras-gpu
(keras-gpu) [npovey@ka dd_net]$ python main.py > output_ddnet_90.txt
Got good results:
[npovey@ka dd_net]$ screen -S dd_net1
[npovey@ka dd_net]$ source activate keras-gpu
(keras-gpu) [npovey@ka dd_net]$ python main.py > output_ddnet_180.txt[npovey@ka dd_net]$ screen -S dd_net1
[npovey@ka dd_net]$ source activate keras-gpu
(keras-gpu) [npovey@ka dd_net]$ python main.py > output_ddnet_ldct_7e4.txtEpoch 1/50 Avg PSNR: 40.91373649562122
Epoch 2/50 Avg PSNR: 41.007181141963734
Epoch 3/50 Avg PSNR: 32.428249836375585
Epoch 4/50 Avg PSNR: 32.42561434856395
......
Epoch 50/50 Avg PSNR: 32.427569692000205
#####run ldcd_2e5
[npovey@ka dd_net]$ screen -S dd_net1
[npovey@ka dd_net]$ source activate keras-gpu
(keras-gpu) [npovey@ka dd_net]$ python main.py > output_ddnet_ldct_2e5.txtEpoch 0: Avg PSNR: 41.967376293118136
Epoch 1: Avg PSNR: 37.859690812973035
Epoch 2: Avg PSNR: 39.812348874868206
Epoch 9: Avg PSNR: 41.398334420237376
ps aux >text.txt
[npovey@ka dd_net]$ kill -KILL 10371
[npovey@ka dd_net]$ rm -r logs/
[npovey@ka dd_net]$ rm -r checkpoints
[npovey@ka dd_net]$ rm -r ndct_train.tfrecord
[npovey@ka dd_net]$ rm -r ldct_train.tfrecord
[npovey@ka dd_net]$ rm -r ndct_test.tfrecord
[npovey@ka dd_net]$ rm -r ldct_test.tfrecord
[npovey@ka dd_net]$ rm -r __pycache__/
[npovey@ka dd_net]$ rm -r output_samples/
[npovey@ka dd_net]$ rm -r test/
[npovey@ka dd_net]$ conda activate keras-gpu
(keras-gpu) [npovey@ka dd_net]$ python main.py > output_ddnet_60_2.txt
[npovey@ka dd_net]$ screen -S dd_net1
[npovey@ka dd_net]$ conda activate keras-gpu
(keras-gpu) [npovey@ka dncnn1]$ python main.py > output_ddnet_60.txt[npovey@ka ddnet2]$ screen -S dd_net1
[npovey@ka ddnet2]$ conda activate keras-gpu
(keras-gpu) [npovey@ka ddnet2]$ python main.py > output_ddnet_ldct_7e4.txtcopy to local desktop
nps-MacBook-Air-2:Desktop np$ scp -r npovey@ka:/home/npovey/data/ddnet2/main.py .
npovey@ka's password:
main.py 100% 7690 305.0KB/s 00:00
nps-MacBook-Air-2:Desktop np$ scp -r npovey@ka:/home/npovey/data/ddnet2/model.py .
npovey@ka's password:
model.py 100% 33KB 562.2KB/s 00:00
(base) nps-MacBook-Air-2:Desktop np$
[npovey@ka ddnet2]$ screen -S dd_net1
[npovey@ka ddnet2]$ conda activate keras-gpu
(keras-gpu) [npovey@ka ddnet2]$ python main.py > output_ddnet_ldct_2e5.txt[npovey@ka ddnet2]$ screen -S dd_net1
[npovey@ka ddnet2]$ conda activate keras-gpu
(keras-gpu) [npovey@ka ddnet2]$ python main.py > output_ddnet_ldct_1e5.txt[npovey@ka ddnet2]$ screen -S dd_net1
[npovey@ka ddnet2]$ conda activate keras-gpu
(keras-gpu) [npovey@ka ddnet2]$ python main.py > output_ddnet_ldct_2e5.txt[npovey@ka ddnet2]$ screen -S dd_net1
[npovey@ka ddnet2]$ conda activate keras-gpu
(keras-gpu) [npovey@ka ddnet2]$ python main.py > output_ddnet_sparseview_60.txt
CTRL+A D exit screens mode[npovey@ka ddnet2]$ screen -S dd_net1
[npovey@ka ddnet2]$ conda activate keras-gpu
(keras-gpu) [npovey@ka ddnet2]$ python main.py > output_ddnet_sparseview_90.txt
CTRL+A D exit screens mode[npovey@ka ddnet2]$ screen -S dd_net1
[npovey@ka ddnet2]$ conda activate keras-gpu
(keras-gpu) [npovey@ka ddnet2]$ python main.py > output_ddnet_sparseview_180.txt
CTRL+A D exit screens modeChange from train to test in main.py line 77
#parser.add_argument('--phase', dest='phase', default='test', help='test')
parser.add_argument('--phase', dest='phase', default='train', help ='train')[npovey@ka ddnet2]$ conda activate keras-gpu
(keras-gpu) [npovey@ka ddnet2]$ python main.py > test_180.txt[npovey@ka ddnet2]$ conda activate keras-gpu
(keras-gpu) [npovey@ka ddnet2]$ python main.py > test_90.txt[npovey@ka ddnet2]$ conda activate keras-gpu
(keras-gpu) [npovey@ka ddnet2]$ python main.py > test_7e4.txt[npovey@ka ddnet2]$ conda activate keras-gpu
(keras-gpu) [npovey@ka ddnet2]$ python main.py > test_1e5.txt[npovey@ka ddnet2]$ conda activate keras-gpu
(keras-gpu) [npovey@ka ddnet2]$ python main.py > test_2e5.txt