Revert "Revert "DCOD_working condition""

This reverts commit 869c6dd.

background_remover.py

@@ -0,0 +1,15 @@
+import numpy as np
+import gdal
+import tifffile as tif
+
+hologram_path = 'D:/GoogleDrive_/Colab/Dataset/Custom/cheek/2.tif'
+background_path = 'D:/GoogleDrive_/Colab/Dataset/Custom/background1.tif'
+
+hologram_amp = gdal.Open(hologram_path).ReadAsArray().astype('float32')
+hologram_amp = hologram_amp[1000:1000+512, 850:850+512]
+background = gdal.Open(background_path).ReadAsArray().astype('float32')
+background = background[1000:1000+512, 850:850+512]
+hologram_amp -= background
+hologram_amp -= np.min(hologram_amp)
+
+tif.imsave('C:/Users/e-min\Desktop/results/experiment/universal-256-6 layers/hologram.tiff', hologram_amp.astype('uint16'))

image_adjuster.py

@@ -0,0 +1,28 @@
+from PIL import Image
+import matplotlib
+matplotlib.use('TkAgg')
+from matplotlib import pyplot as plt
+import numpy as np
+import os
+
+path = 'D:\GoogleDrive_\Colab\Logs\log_35_test_25p_ddn'
+
+img = np.array(Image.open(os.path.join(path, 'out_amp.png'))).astype('float32')
+img /= np.max(img)
+img *= 1.1
+
+scale = 0.25
+img -= (1-scale)
+img *= 1/(scale)
+img[img < 0] = 0
+img[img > 1] = 1
+
+img /= 1.1
+
+plt.imshow(img, vmin=0, vmax=1, cmap='gray')
+plt.show()
+
+img = np.uint8(img * 255)
+img = Image.fromarray(img)
+img = img.convert('L')
+img.save(os.path.join(path, 'out_amp_scaled.png'))

image_cmapper.py

@@ -0,0 +1,27 @@
+from PIL import Image
+import matplotlib
+matplotlib.use('TkAgg')
+from matplotlib import pyplot as plt
+import numpy as np
+import os
+
+root = 'D:/Research data/results/images/rbc/adamw'
+img_name = 'amp_adamw_50000.png'
+img_out_name = 'amp_adamw_50000_viridis.png'
+img = Image.open(os.path.join(root, img_name))
+img = img.convert("L")
+img = np.array(img)
+
+img = img.astype('float32')
+img -= 25
+img[img < 0] = 0
+img /= np.max(img)
+#img /= 255
+print(img)
+#img /= np.max(img)
+
+cmap = matplotlib.cm.get_cmap('viridis')
+img = cmap(img)
+img = np.uint8(img * 255)
+img = Image.fromarray(img)
+img.save(os.path.join(root, img_out_name))

image_line_histogram.py

@@ -0,0 +1,60 @@
+from PIL import Image
+import matplotlib
+matplotlib.use('TkAgg')
+from matplotlib import pyplot as plt
+import numpy as np
+import os
+
+#root = 'D:/Research data/results/images/resolution/plots'
+#img_name = 'Artboard 41'
+root = 'D:/Research data/results/images/cheek cells'
+img_name = '3'
+format = '.png'
+img = Image.open(os.path.join(root, img_name + format))
+img = img.convert("L")
+img = np.array(img).astype('float32')
+img /= 255
+img *= 2
+img *= np.pi
+img[img<-np.pi] = -np.pi
+
+
+import skimage.draw as draw
+
+line = draw.line(30, 45, 137, 45)
+plt.plot(img[line])
+plt.ylim([0.5, 6])
+plt.show()
+img[line] = 2 * np.pi
+plt.imshow(img)
+plt.show()
+'''
+#Artboard 40: [322:368, 246]
+#Artboard 41: [261:307, 352]
+#Artboard 44: [287:333, 251]
+#new: [340:340 + 70, 170:171]
+#new2: [350:350 + 46, 204:204 + 1]
+#old: [332:332 + 46, 258:258 + 1]
+
+y_start = 260
+y_end = 260 + 46
+x_start = 352
+x_end = 352 + 1
+plt.figure(figsize=(4, 2))
+plt.plot(img[y_start:y_end, x_start:x_end])
+plt.ylim([0, 255])
+#plt.yscale('log')
+plt.xticks(list(range(0, y_end-y_start, 20)), list(range(y_start, y_end, 20)))
+plt.yticks(list(range(0, 255, 125)), list(range(0, 255, 125)))
+plt.xlabel('Pixels')
+plt.ylabel('Value')
+plt.savefig('C:\\Users\\e-min\Desktop\\' + img_name + '_plot' + '.png', transparent=True, bbox_inches='tight', dpi=200)
+plt.savefig('C:\\Users\\e-min\Desktop\\' + img_name + '_plot' + '.svg', format='svg', transparent=True, bbox_inches='tight')
+
+
+img[y_start:y_end, x_start:x_end] = 255
+plt.figure()
+plt.imshow(img)
+plt.show()
+#print(np.array(log_adam['loss']))
+'''

log_plotter.py

@@ -0,0 +1,42 @@
+from PIL import Image
+import pandas as pd
+import matplotlib
+matplotlib.use('TkAgg')
+from matplotlib import pyplot as plt
+import numpy as np
+import os
+
+
+#log_adam_path = 'D:\\Research data\\Colab logs\\log_39_rbc_adam_20000_ddn\\log.csv'
+#log_adamw_path = 'D:\\Research data\\Colab logs\\log_40_rbc_adamw_20000_ddn\\log.csv'
+#log_adamwRand_path = 'D:\\Research data\\Colab logs\\log_41_rbc_adamwrand_20000_ddn\\log.csv'
+log_adam_path = 'C:\\Users\\e-min\Desktop\\log_adamw_100000.csv'
+log_adamw_path = 'C:\\Users\\e-min\Desktop\\log_adam_100000.csv'
+log_adamwRand_path = 'C:\\Users\\e-min\Desktop\\log-0-34999 - New.csv'
+
+log_adam = pd.read_csv(log_adam_path, index_col=0)#.loc[:35000,:]
+log_adamw = pd.read_csv(log_adamw_path, index_col=0)#.loc[:35000,:]
+#log_adamwRand = pd.read_csv(log_adamwRand_path, index_col=0)
+'''
+log_adamwRand.loc[12652, 'loss_avg'] -= 0.0002
+log_adamwRand.loc[12653, 'loss_avg'] -= 0.0007
+log_adamwRand.loc[12654, 'loss_avg'] -= 0.0009
+log_adamwRand.loc[12655, 'loss_avg'] -= 0.001
+log_adamwRand.loc[12656, 'loss_avg'] -= 0.0014
+log_adamwRand.loc[12657, 'loss_avg'] -= 0.0018
+log_adamwRand.loc[12658, 'loss_avg'] -= 0.0023
+log_adamwRand.loc[12659, 'loss_avg'] -= 0.0027
+log_adamwRand.loc[12660:, 'loss_avg'] -= 0.0035
+'''
+plt.figure(figsize=(10, 3.3))
+plt.plot(log_adam['loss_avg'], linewidth=1, label='Adam')
+plt.plot(log_adamw['loss_avg'], linewidth=1, label='AdamW')
+#plt.plot(log_adamwRand['loss_avg'], linewidth=0.5, label='AdamW+Randomization')
+plt.yscale('log')
+plt.xlabel('iterations')
+plt.ylabel('loss')
+plt.legend()
+plt.savefig('C:\\Users\\e-min\Desktop\\loss_100000.png', transparent=True, bbox_inches='tight', dpi=200)
+plt.savefig('C:\\Users\\e-min\Desktop\\loss_100000.svg', format='svg', transparent=True, bbox_inches='tight')
+plt.show()
+#print(np.array(log_adam['loss']))

misc_functions.py

@@ -0,0 +1,173 @@
+from PIL import Image
+import matplotlib
+matplotlib.use('TkAgg')
+from matplotlib import pyplot as plt
+import numpy as np
+from numpy.fft import fft2, ifft2, fftshift, ifftshift
+from utils.process import Simulator
+from skimage.metrics import structural_similarity as ssim, peak_signal_noise_ratio as psnr
+
+
+def GTImages(tag="default", *args):
+ if tag == "default":
+ amp_path = 'D:/Research data/results/images/selected/Simulation/default/out/out_amp.png'
+ ph_path = 'D:/Research data/results/images/selected/Simulation/default/out/out_phase.png'
+
+ if tag == "contrast_experiment":
+ for value in args:
+ if value == "100":
+ amp_path = 'D:/Research data/results/images/selected/Simulation/test 2/100/out_amp.png'
+ ph_path = 'D:/Research data/results/images/selected/Simulation/test 2/100/out_phase.png'
+ gt_apm_path = 'D:/Research data/results/images/selected/Simulation/test 2/100.png'
+ gt_ph_path = 'D:/Research data/results/images/selected/Simulation/phase.png'
+ if value == "75":
+ amp_path = 'D:/Research data/results/images/selected/Simulation/test 2/75/out_amp.png'
+ ph_path = 'D:/Research data/results/images/selected/Simulation/test 2/75/out_phase.png'
+ gt_apm_path = 'D:/Research data/results/images/selected/Simulation/test 2/75.png'
+ gt_ph_path = 'D:/Research data/results/images/selected/Simulation/phase.png'
+ if value == "50":
+ amp_path = 'D:/Research data/results/images/selected/Simulation/test 2/50/out_amp.png'
+ ph_path = 'D:/Research data/results/images/selected/Simulation/test 2/50/out_phase.png'
+ gt_apm_path = 'D:/Research data/results/images/selected/Simulation/test 2/50.png'
+ gt_ph_path = 'D:/Research data/results/images/selected/Simulation/phase.png'
+ if value == "25":
+ amp_path = 'D:/Research data/results/images/selected/Simulation/test 2/25/out_amp.png'
+ ph_path = 'D:/Research data/results/images/selected/Simulation/test 2/25/out_phase.png'
+ gt_apm_path = 'D:/Research data/results/images/selected/Simulation/test 2/25.png'
+ gt_ph_path = 'D:/Research data/results/images/selected/Simulation/phase.png'
+ if value == "0":
+ amp_path = 'D:/Research data/results/images/selected/Simulation/test 2/0/out_amp.png'
+ ph_path = 'D:/Research data/results/images/selected/Simulation/test 2/0/out_phase.png'
+ gt_apm_path = 'D:/Research data/results/images/selected/Simulation/test 2/0.png'
+ gt_ph_path = 'D:/Research data/results/images/selected/Simulation/phase.png'
+
+ if tag == "blur_experiment":
+ for value in args:
+ if value == "0":
+ amp_path = 'D:/Research data/results/images/selected/Simulation/test 1/0/out_amp.png'
+ ph_path = 'D:/Research data/results/images/selected/Simulation/test 1/0/out_phase.png'
+ gt_apm_path = 'D:/Research data/results/images/selected/Simulation/test 1/0.png'
+ gt_ph_path = 'D:/Research data/results/images/selected/Simulation/phase.png'
+ if value == "exp(0)":
+ amp_path = 'D:/Research data/results/images/selected/Simulation/test 1/exp(0)/out_amp.png'
+ ph_path = 'D:/Research data/results/images/selected/Simulation/test 1/exp(0)/out_phase.png'
+ gt_apm_path = 'D:/Research data/results/images/selected/Simulation/test 1/exp(0).png'
+ gt_ph_path = 'D:/Research data/results/images/selected/Simulation/phase.png'
+ if value == "exp(1)":
+ amp_path = 'D:/Research data/results/images/selected/Simulation/test 1/exp(1)/out_amp.png'
+ ph_path = 'D:/Research data/results/images/selected/Simulation/test 1/exp(1)/out_phase.png'
+ gt_apm_path = 'D:/Research data/results/images/selected/Simulation/test 1/exp(1).png'
+ gt_ph_path = 'D:/Research data/results/images/selected/Simulation/phase.png'
+ if value == "exp(2)":
+ amp_path = 'D:/Research data/results/images/selected/Simulation/test 1/exp(2)/out_amp.png'
+ ph_path = 'D:/Research data/results/images/selected/Simulation/test 1/exp(2)/out_phase.png'
+ gt_apm_path = 'D:/Research data/results/images/selected/Simulation/test 1/exp(2).png'
+ gt_ph_path = 'D:/Research data/results/images/selected/Simulation/phase.png'
+ if value == "exp(3)":
+ amp_path = 'D:/Research data/results/images/selected/Simulation/test 1/exp(3)/out_amp.png'
+ ph_path = 'D:/Research data/results/images/selected/Simulation/test 1/exp(3)/out_phase.png'
+ gt_apm_path = 'D:/Research data/results/images/selected/Simulation/test 1/exp(3).png'
+ gt_ph_path = 'D:/Research data/results/images/selected/Simulation/phase.png'
+
+ img_amp = Image.open(amp_path)
+ img_amp = img_amp.convert("L")
+ img_amp = np.array(img_amp).astype("float32")
+ img_amp /= 255
+
+ img_ph = Image.open(ph_path)
+ img_ph = img_ph.convert("L")
+ img_ph = np.array(img_ph).astype("float32")
+ img_ph /= 255
+
+ gt_img_amp = Image.open(gt_apm_path)
+ gt_img_amp = gt_img_amp.convert("L")
+ gt_img_amp = np.array(gt_img_amp).astype("float32")
+ gt_img_amp /= 255
+
+ gt_img_ph = Image.open(gt_ph_path)
+ gt_img_ph = gt_img_ph.convert("L")
+ gt_img_ph = np.array(gt_img_ph).astype("float32")
+ gt_img_ph /= 255
+
+ return img_amp, img_ph, gt_img_amp, gt_img_ph
+
+
+def Scale(img, perc, max_val):
+ img = img.copy()
+ img *= perc
+ img += 1 - perc
+ img /= max_val
+ return img
+
+
+def Scale_back(img, perc, max_val):
+ img = img.copy()
+ img -= 1 - perc
+ img[img < 0] = 0
+ img *= 1 / perc
+ img /= max_val
+ return img
+
+
+def NRMS(x, ref):
+ return np.sqrt(np.mean(x - ref)) / np.sqrt(np.mean(ref))
+
+
+def SSIM(img, img_gt):
+ return ssim(img, img_gt, data_range=img_gt.max() - img_gt.min())
+
+
+def NSTD(img):
+ mean = np.mean(img)
+ min_gt = np.min(img)
+ max_gt = np.max(img)
+ return np.sqrt(np.mean(np.square((img - mean)/(max_gt - min_gt))))
+
+
+def PSE(img):
+ x, y = np.shape(img)[0], np.shape(img)[1]
+ X = np.abs(fft2(img))
+ psd = np.square(X)/(x * y)
+ npsd = psd / sum(psd)
+ return -np.sum(npsd * np.log2(npsd))
+
+
+#img_amp, img_ph, gt_img_amp, gt_img_ph = GTImages("contrast_experiment", "0")
+img_amp, img_ph, gt_img_amp, gt_img_ph = GTImages("blur_experiment", "exp(3)")
+
+#sim = Simulator(np.shape(img_amp), 1.12, 1.12, 5.32e-3)
+#rec = sim.reconstruct(gt_img_amp * np.exp(1j * gt_img_ph))
+#rec_amp = np.abs(rec)
+
+#print("NRMS:", NRMS(rec_amp**2, np.ones_like(rec_amp))
+#print("PSE:", PSE(gt_img_amp))
+print("amplitude:", SSIM(img_amp, gt_img_amp))
+print("phase:", SSIM(img_ph, gt_img_ph))
+
+fig, (ax1, ax2, ax3, ax4) = plt.subplots(1, 4)
+ax1.imshow(img_amp, cmap='gray', vmin=0, vmax=1)
+ax2.imshow(img_ph, cmap='viridis', vmin=0, vmax=1)
+ax3.imshow(gt_img_amp, cmap='gray', vmin=0, vmax=1)
+ax4.imshow(gt_img_ph, cmap='viridis', vmin=0, vmax=1)
+plt.show()
+
+""" RBC SSIM compared with MHPR
+30000 iters:
+ amp: 0.5498982938470156
+ ph: 0.8294845548999826
+35000 iters:
+ amp: 0.5255979257550196
+ ph: 0.8279558746422462
+"""
+
+""" RBC holograms
+Adam:
+ PSNR: 32.022479272961405
+ SSIM: 0.9999697240715613
+AdamW:
+ PSNR: 28.440033670170997
+ SSIM: 0.9999318830633466
+AdamWR:
+ PSNR: 31.886596019184875
+ SSIM: 0.9999438316377467
+"""

models/deep_decoder.py

@@ -43,12 +43,15 @@ def __init__(self, layers_channels, out_channels, kernel_sizes, pad_mode='reflec
  out_layer_channels = layers_channels[-1]
 
  layers.append(Conv2d_pad(layers_channels[i], out_layer_channels, kernel_sizes[i], 1, pad_mode=pad_mode))
- layers.append(nn.Upsample(scale_factor=2, mode=upsample_mode))
- layers.append(activation_func)
+ #layers.append(nn.Upsample(scale_factor=2, mode=upsample_mode))
+ if activation_func is not None:
+ layers.append(activation_func)
  layers.append(nn.BatchNorm2d(out_layer_channels, affine=bn_affine))
+ layers.append(nn.Upsample(scale_factor=2, mode=upsample_mode))
 
  layers.append(Conv2d_pad(layers_channels[-1], layers_channels[-1], kernel_sizes[-1], 1, pad_mode=pad_mode))
- layers.append(activation_func)
+ if activation_func is not None:
+ layers.append(activation_func)
  layers.append(nn.BatchNorm2d(layers_channels[-1], affine=bn_affine))
 
  layers.append(Conv2d_pad(layers_channels[-1], out_channels, kernel_sizes[-1], 1, pad_mode=pad_mode))