add kaggle model

models/extract_small_tiles.py

@@ -0,0 +1,73 @@
+import os
+import cv2
+import skimage.io
+from tqdm.notebook import tqdm
+import zipfile
+import numpy as np
+
+
+##
+## extract tiles from large image and store in zip 
+### 
+
+
+TRAIN = './kaggle/train_images/'
+MASKS = './kaggle//train_label_masks/'
+OUT_TRAIN = 'train.zip'
+OUT_MASKS = 'masks.zip'
+sz = 128
+N = 16
+
+
+def tile(img, mask):
+    result = []
+    shape = img.shape
+    pad0,pad1 = (sz - shape[0]%sz)%sz, (sz - shape[1]%sz)%sz
+    img = np.pad(img,[[pad0//2,pad0-pad0//2],[pad1//2,pad1-pad1//2],[0,0]],
+                constant_values=255)
+    mask = np.pad(mask,[[pad0//2,pad0-pad0//2],[pad1//2,pad1-pad1//2],[0,0]],
+                constant_values=0)
+    img = img.reshape(img.shape[0]//sz,sz,img.shape[1]//sz,sz,3)
+    img = img.transpose(0,2,1,3,4).reshape(-1,sz,sz,3)
+    mask = mask.reshape(mask.shape[0]//sz,sz,mask.shape[1]//sz,sz,3)
+    mask = mask.transpose(0,2,1,3,4).reshape(-1,sz,sz,3)
+    if len(img) < N:
+        mask = np.pad(mask,[[0,N-len(img)],[0,0],[0,0],[0,0]],constant_values=0)
+        img = np.pad(img,[[0,N-len(img)],[0,0],[0,0],[0,0]],constant_values=255)
+    idxs = np.argsort(img.reshape(img.shape[0],-1).sum(-1))[:N]
+    img = img[idxs]
+    mask = mask[idxs]
+    for i in range(len(img)):
+        result.append({'img':img[i], 'mask':mask[i], 'idx':i})
+    return result
+
+
+
+x_tot,x2_tot = [],[]
+names = [name[:-10] for name in os.listdir(MASKS)]
+with zipfile.ZipFile(OUT_TRAIN, 'w') as img_out,\
+ zipfile.ZipFile(OUT_MASKS, 'w') as mask_out:
+    for name in tqdm(names):
+        print (name)
+        if len(skimage.io.MultiImage(os.path.join(TRAIN,name+'.tiff'))) and len(skimage.io.MultiImage(os.path.join(MASKS,name+'_mask.tiff'))):
+            img = skimage.io.MultiImage(os.path.join(TRAIN,name+'.tiff'))[-1]
+            mask = skimage.io.MultiImage(os.path.join(MASKS,name+'_mask.tiff'))[-1]
+
+            tiles = tile(img,mask)
+            for t in tiles:
+                img,mask,idx = t['img'],t['mask'],t['idx']
+                x_tot.append((img/255.0).reshape(-1,3).mean(0))
+                x2_tot.append(((img/255.0)**2).reshape(-1,3).mean(0)) 
+                #if read with PIL RGB turns into BGR
+                img = cv2.imencode('.png',cv2.cvtColor(img, cv2.COLOR_RGB2BGR))[1]
+                img_out.writestr(f'{name}_{idx}.png', img)
+                mask = cv2.imencode('.png',mask[:,:,0])[1]
+                mask_out.writestr(f'{name}_{idx}.png', mask)
+
+
+
+
+#image stats
+img_avr =  np.array(x_tot).mean(0)
+img_std =  np.sqrt(np.array(x2_tot).mean(0) - img_avr**2)
+print('mean:',img_avr, ', std:', np.sqrt(img_std))