trainreq

Datasets.py

@@ -0,0 +1,265 @@
+
+from torch.utils.data import Dataset
+import torch
+import os
+from PIL import Image
+import matplotlib.pyplot as plt
+import torchvision.transforms as transforms
+import pandas as pd
+import numpy as np
+import cv2
+from random import sample
+from scipy import stats
+import torch.nn.functional as F
+#%%
+def resizeAndPad(img, size, padColor=0):
+
+    h, w = img.shape[:2]
+    sh, sw = size
+
+    # interpolation method
+    if h > sh or w > sw: # shrinking image
+        interp = cv2.INTER_AREA
+    else: # stretching image
+        interp = cv2.INTER_CUBIC
+
+    # aspect ratio of image
+    aspect = w/h  # if on Python 2, you might need to cast as a float: float(w)/h
+
+    # compute scaling and pad sizing
+    if aspect > 1: # horizontal image
+        new_w = sw
+        new_h = np.round(new_w/aspect).astype(int)
+        pad_vert = (sh-new_h)/2
+        pad_top, pad_bot = np.floor(pad_vert).astype(int), np.ceil(pad_vert).astype(int)
+        pad_left, pad_right = 0, 0
+    elif aspect < 1: # vertical image
+        new_h = sh
+        new_w = np.round(new_h*aspect).astype(int)
+        pad_horz = (sw-new_w)/2
+        pad_left, pad_right = np.floor(pad_horz).astype(int), np.ceil(pad_horz).astype(int)
+        pad_top, pad_bot = 0, 0
+    else: # square image
+        new_h, new_w = sh, sw
+        pad_left, pad_right, pad_top, pad_bot = 0, 0, 0, 0
+
+    # set pad color
+    if len(img.shape) is 3 and not isinstance(padColor, (list, tuple, np.ndarray)): # color image but only one color provided
+        padColor = [padColor]*3
+
+    # scale and pad
+    scaled_img = cv2.resize(img, (new_w, new_h), interpolation=interp)
+    scaled_img = cv2.copyMakeBorder(scaled_img, pad_top, pad_bot, pad_left, pad_right, borderType=cv2.BORDER_CONSTANT, value=padColor)
+
+    return scaled_img
+def resizeAndCrop(img, size=(256,256)):
+
+    h, w = img.shape[:2]
+    sh, sw = size
+
+    # interpolation method
+    if h > sh or w > sw: # shrinking image
+        interp = cv2.INTER_AREA
+    else: # stretching image
+        interp = cv2.INTER_CUBIC
+
+    # aspect ratio of image
+    aspect = w/h  # if on Python 2, you might need to cast as a float: float(w)/h
+
+    # scaling and cropping
+    if aspect > 1: # horizontal image
+        new_h = sh
+        new_w = np.round(new_h*aspect).astype(int)
+        extra_region_w = np.floor((new_w-sw)/2).astype(int)
+        # scale and crop
+        scaled_img = cv2.resize(img, (new_w, new_h), interpolation=interp)
+        scaled_img = scaled_img[0:new_h,extra_region_w:new_w-extra_region_w]
+    elif aspect < 1: # vertical image
+        new_w = sw
+        new_h = np.round(new_w/aspect).astype(int)
+        extra_region_h = np.floor((new_h-sh)/2).astype(int)
+        # scale and crop
+        scaled_img = cv2.resize(img, (new_w, new_h), interpolation=interp)
+        scaled_img = scaled_img[extra_region_h:new_h-extra_region_h,0:new_w]
+
+    else: # square image
+        new_h, new_w = sh, sw
+
+        scaled_img = cv2.resize(img, (new_w, new_h), interpolation=interp)
+
+    scaled_img = cv2.resize(img, size, interpolation=cv2.INTER_AREA)
+    return scaled_img
+#%%
+
+def randomCropResize(img,old_size=256,new_size=224):
+    remove_reg=int((old_size-new_size)/2)
+    r = np.random.choice([1,2,3,4,5,6])
+    if r==1: #top left
+        return img[0:new_size,0:new_size]
+    elif r==2: #top right
+        return img[0:new_size,remove_reg*2:old_size]
+    elif r==3: #bottom left
+        return img[remove_reg*2:old_size,0:new_size]
+    elif r==4: #bottom right
+        return img[remove_reg*2:old_size,remove_reg*2:old_size]
+    elif r==5: # center crop
+        return img[remove_reg:new_size+remove_reg,remove_reg:new_size+remove_reg]
+    else: # resize
+        return cv2.resize(img,(new_size,new_size), interpolation = cv2.INTER_AREA)
+
+#%%
+def img_mean_stddev(dataset):
+    sum_img = [0,0,0]
+    n = len(dataset)
+    h , w = dataset[0][0][:2]
+    N = n * h * w
+    for img,label in dataset:
+        sum_img[0] += np.sum(img[:,:,0])
+        sum_img[1] += np.sum(img[:,:,1])
+        sum_img[2] += np.sum(img[:,:,2])
+    mean_val = [np.round(sum_img[0]/(N),2), np.round(sum_img[1]/(N),2), np.round(sum_img[2]/(N),2)]
+    print(mean_val)          
+    x_m_sq_sum = [0,0,0]
+    for img,label in dataset:
+        x_m_sq_sum[0] += np.sum((img[:,:,0] - mean_val[0])**2)
+        x_m_sq_sum[1] += np.sum((img[:,:,1] - mean_val[1])**2)
+        x_m_sq_sum[2] += np.sum((img[:,:,2] - mean_val[2])**2)
+
+    stddev_val = [np.round(np.sqrt(x_m_sq_sum[0]/(N)),2), 
+                  np.round(np.sqrt(x_m_sq_sum[1]/(N)),2), 
+                  np.round(np.sqrt(x_m_sq_sum[2]/(N)),2)]
+    print(stddev_val)
+    return mean_val, stddev_val
+#%%
+class ImageNet(Dataset):
+    def __init__(self, root_dir, sample_type=None, transform=None):
+        self.root_dir = root_dir
+        self.transform = transform
+        self.img_files = []
+        self.labels=[]
+        self.train_img_mean = []
+        self.train_img_std = []
+
+        if sample_type == 'train':
+            data = pd.read_csv(self.root_dir+"devkit/data/train_data.csv")
+            self.img_files = [self.root_dir+x for x in list(data['img_path'])]
+            self.labels = list(data['label'])
+        elif sample_type == 'val':
+            self.img_files = [self.root_dir+"img_val/"+x for x in os.listdir(self.root_dir+"img_val") if ".JPEG" in x]
+            self.img_files.sort()
+            self.labels = list(pd.read_csv(self.root_dir+"devkit/data/ILSVRC2012_validation_ground_truth.txt",sep=",",header=None, index_col=False,names=["label"])['label'])
+#            print(self.img_files[0])
+
+    def __len__(self):
+        return len(self.labels)
+
+    def __getitem__(self,idx):
+        img = cv2.imread(self.img_files[idx])
+#        img = np.asarray(img)
+        img = resizeAndCrop(img,size=(256,256))
+
+
+#        img = cv2.imread(self.img_files[idx])
+#        
+        if self.transform:
+            img = randomCropResize(img,old_size=256,new_size=224)#random crop or resize
+            if np.random.rand()>0.5:#random horizontal flip
+                img = cv2.flip(img,0)
+#            img = img.astype('float32')
+            img = img / 255.0
+#            print(img.shape)
+#            img = transforms.ToPILImage()(torch.tensor(img))
+            img[:,:,2] = (img[:,:,2] - 0.485)/0.229  
+            img[:,:,1] = (img[:,:,1] - 0.456)/0.224
+            img[:,:,0] = (img[:,:,0] - 0.406)/0.225
+
+            #img[:,:,2] = (img[:,:,2] - 0.485) 
+            #img[:,:,1] = (img[:,:,1] - 0.456)
+            #img[:,:,0] = (img[:,:,0] - 0.406)
+#            
+#            img[:,:,0] = (img[:,:,0]-img[:,:,0].min())/(img[:,:,0].max()-img[:,:,0].min())
+#            img[:,:,1] = (img[:,:,1]-img[:,:,1].min())/(img[:,:,1].max()-img[:,:,1].min())
+#            img[:,:,2] = (img[:,:,2]-img[:,:,2].min())/(img[:,:,2].max()-img[:,:,2].min()) 
+            img = self.transform(img)
+        label = self.labels[idx]-1
+
+        return img,label
+
+#%%
+class AVA_Aesthetics_Ranking_Dataset(Dataset):
+    def __init__(self, root_dir, sample_type=None, transform=None):
+        self.root_dir = root_dir
+        self.transform = transform
+        self.sample_type=sample_type
+        self.labels = []
+        self.files = []
+        if sample_type=='train':
+#            data=pd.read_csv(self.root_dir+"AVA_dataset/AVA_mean_rating_samples_top_10pc_bottom_10pc_train.csv")
+            data=pd.read_csv(self.root_dir+"AVA_dataset/ILGNet/AVA2/train_ilgnet.txt", sep=" ")
+
+#            self.labels = list(data.mean_rating)
+#            self.files = list(data.img_id)
+            self.labels = list(data.label)
+            self.files = list(data.img)
+
+
+        elif sample_type=='val':
+#            data=pd.read_csv(self.root_dir+"AVA_dataset/AVA_mean_rating_samples_top_10pc_bottom_10pc_test.csv")
+            data=pd.read_csv(self.root_dir+"AVA_dataset/ILGNet/AVA2/val.txt", sep=" ")
+
+
+#            self.labels = list(data.mean_rating)
+#            self.files = list(data.img_id)
+            self.labels = list(data.label)
+            self.files = list(data.img)
+
+
+    def __len__(self):
+        return len(self.labels)
+
+    def __getitem__(self,idx):
+#        try:
+#        if self.sample_type=='train':
+#            img = cv2.imread(os.path.join(self.root_dir,"top_10pc_bottom_10pc_rated_resized_images_224_224_padded_black/train/"+str(self.files[idx])+'.jpg'))
+#        elif self.sample_type=='val':
+#            img = cv2.imread(os.path.join(self.root_dir,"top_10pc_bottom_10pc_rated_resized_images_224_224_padded_black/test/"+str(self.files[idx])+'.jpg'))
+
+#        if self.sample_type=='train':
+#            img = cv2.imread(os.path.join(self.root_dir,"images/"+str(self.files[idx])+'.jpg'))
+#        elif self.sample_type=='val':
+#            img = cv2.imread(os.path.join(self.root_dir,"images/"+str(self.files[idx])+'.jpg'))
+        if self.sample_type=='train':
+            img = cv2.imread(os.path.join(self.root_dir,"images/"+str(self.files[idx])))
+        elif self.sample_type=='val':
+            img = cv2.imread(os.path.join(self.root_dir,"images/"+str(self.files[idx])))
+
+        img = resizeAndCrop(img,size=(224,224))
+#        img = np.asarray(img)
+#        if self.sample_type=='train':
+#            img = resizeAndCrop(img,size=(256,256))
+#        else:
+#            img = resizeAndCrop(img,size=(224,224))
+
+
+        if self.transform:
+#            if self.sample_type=='train':
+#                img = randomCropResize(img,old_size=256,new_size=224)#random crop or resize
+
+            if np.random.rand()>0.5:
+                img = cv2.flip(img,0)
+            img = img/255.0
+
+            img[:,:,2] = (img[:,:,2] - 0.485)/0.229  
+            img[:,:,1] = (img[:,:,1] - 0.456)/0.224
+            img[:,:,0] = (img[:,:,0] - 0.406)/0.225
+#            
+#            img[:,:,0] = (img[:,:,0]-img[:,:,0].min())/(img[:,:,0].max()-img[:,:,0].min())
+#            img[:,:,1] = (img[:,:,1]-img[:,:,1].min())/(img[:,:,1].max()-img[:,:,1].min())
+#            img[:,:,2] = (img[:,:,2]-img[:,:,2].min())/(img[:,:,2].max()-img[:,:,2].min())
+            img = self.transform(img)
+#            img = img.float()
+
+        label = self.labels[idx]
+
+        return img, label
+#%%