Changes

main.py

@@ -2,7 +2,6 @@ import builtins
 import math
 import os
 import random
-import shutil
 import time
 import warnings
 
@@ -159,16 +158,15 @@ def worker(gpu, ngpus_per_node, args):
     eval_dir = os.path.join(args.data, 'train')
 
     # center-crop augmentation
-    eval_augmentation = aug.moco_eval()
 
-    pre_train_dataset = pcl.loader.PreImager(pre_train_dir, eval_augmentation)
+    pre_train_dataset = pcl.loader.PreImager(pre_train_dir, aug.moco_eval())
 
     train_dataset = pcl.loader.ImageFolderInstance(
         train_dir,
-        pcl.loader.TwoCropsTransform(eval_augmentation))
+        pcl.loader.TwoCropsTransform(aug.only_rotate()))
     eval_dataset = pcl.loader.ImageFolderInstance(
         eval_dir,
-        eval_augmentation)
+        aug.moco_eval())
 
     if args.distributed:
         pre_train_sampler = torch.utils.data.distributed.DistributedSampler(pre_train_dataset)

pcl/loader.py

@@ -2,7 +2,6 @@ from PIL import ImageFilter
 import random
 import torch.utils.data as tud
 import torchvision.datasets as datasets
-from torchvision.io import image
 
 
 class PreImager(tud.Dataset):

requirements.txt

@@ -0,0 +1,8 @@
+torch
+torchvision
+Pillow
+numpy
+scikit-image
+tqdm
+matplotlib
+scikit-learn

scripts/augmentation.py

@@ -1,6 +1,7 @@
 import torchvision.transforms as transforms
 import pcl.loader
 
+
 normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                  std=[0.229, 0.224, 0.225])
 
@@ -32,8 +33,16 @@ def moco_v1():
 
 def moco_eval():
     return transforms.Compose([
-        transforms.Resize([512, 512]),
+        transforms.RandomResizedCrop(224, scale=(0.2, 1.)),
         # transforms.CenterCrop(512),
         transforms.ToTensor(),
         normalize
     ])
+
+def only_rotate():
+    return transforms.Compose([
+        transforms.RandomResizedCrop(224, scale=(0.2, 1.)),
+        transforms.RandomRotation(90),
+        transforms.ToTensor(),
+        normalize
+    ])

scripts/data_process.py

@@ -1,4 +1,6 @@
 import torch
+
+
 def data_process(cluster_result, images, gpu):
     im_q = []
     im_k = []

scripts/loss.py

@@ -1,6 +1,9 @@
 import torch
 import numpy as np
 import math
+import random
+from skimage.metrics import structural_similarity as SSIM
+from skimage.metrics import peak_signal_noise_ratio as PSNR
 def proto_with_quality(output, target, output_proto, target_proto, criterion, acc_proto, images, num_cluster):
     # InfoNCE loss
     loss = criterion(output, target)
@@ -18,8 +21,30 @@ def proto_with_quality(output, target, output_proto, target_proto, criterion, ac
         loss += loss_proto
 
         # Quality loss
-        mse = np.mean((images[1]/255.0-images[2]/255.0)**2)
+        im_q = torch.split(images[0], split_size_or_sections=1, dim=0)
-        psnr = 20 * math.log10(1 / math.sqrt(mse))
+        im_q = [torch.squeeze(im, dim=0) for im in im_q]
+        im_k = torch.split(images[1], split_size_or_sections=1, dim=0)
+        im_k = [torch.squeeze(im, dim=0) for im in im_k]
+
+        l_psnr = []
+        l_ssim = []
+        for i in range(min(len(im_q), len(im_k))):
+            k = im_k[i]
+            q_index = random.randint(0,i-2)
+            if q_index >= i:
+                q_index += 1
+            q = im_q[q_index]
+            psnr_temp = PSNR(k,q)
+            if psnr_temp >= 50:
+                psnr_temp = 0
+            elif psnr_temp <= 30:
+                psnr_temp = 1
+            else:
+                psnr_temp = (50-psnr_temp)/20
+            l_psnr.append(psnr_temp)
+            l_ssim.append(1-SSIM(k,q))
+
+        loss += np.mean(l_psnr)+np.mean(l_ssim)
 
 
     return loss