![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1342122%2F502c1145ac9c877556c2308650e0901d%2Fviz_img.001.jpeg?generation=1588637596178459&alt=media\")
I have been utilizing the amazing wizardry done by @iafoss throughout this competition in his notebooks and discussions around using tiles pulled from the slides as data input.
\n\nI made a notebook that displays the modules for both the tiles themselves as well as an overlay of the selected tiles on a given slide. I made them both into simple python modules with outputs as below:
\n\n
I also added some of my notes to the tile module to better understand the magic performed:\n```:::python\ndef tile_plot(base_image, N=12, sz=128, plot=True):\n \"\"\"\n Description\n ____\n Tilizer module made by @iafoss that can be found in the notebook:\n https://www.kaggle.com/iafoss/panda-concat-tile-pooling-starter-inference\n Takes a base image and returns the N tiles with the largest differnce\n from a white backgound each with a given square size of input-sz.
\n\nParameters\n__________\nbase_image: numpy array\n Image array to split into tiles and plot\nN: int\n This is the number of tiles to split the image into\nsz: int\n This is the size for each side of the square tiles\nplot: bool\n True to show plot of chosen tiles, False for silent return\n\nReturns\n__________\n- List of size N with each item being a numpy array tile.\n\"\"\"\n\n#Get the shape of the input image\nshape = base_image.shape\n\n#Find the padding such that the image divides evenly by the desired size\npad0,pad1 = (sz - shape[0]%sz)%sz, (sz - shape[1]%sz)%sz\n\n#Pad the image with blank space to reach the above found tagrgets\nbase_img = np.pad(base_image,[[pad0//2,pad0-pad0//2],\n [pad1//2,pad1-pad1//2],[0,0]],\n constant_values=255)\n\n#Reshape and Transpose to get the images into tiles\nall_tiles = base_img.reshape(base_img.shape[0]//sz,sz, base_img.shape[1]//sz,sz,3)\nall_tiles = all_tiles.transpose(0,2,1,3,4).reshape(-1,sz,sz,3)\n\n#If there are not enough tiles to meet desired N pad again\nif len(all_tiles) < N:\n all_tiles = np.pad(all_tiles,[[0,N-len(all_tiles)],[0,0],[0,0],[0,0]],\n constant_values=255) \n\n#Sort the images by those with the lowest sum (i.e the least white)\n#Return indexes to the lowest N \nidxs = np.argsort(all_tiles.reshape(all_tiles.shape[0],-1).sum(-1))[:N]\n\n#Slect by index those returned from the above funtion\ntissue_tiles = all_tiles[idxs]\n\nif plot:\n #Funciton for plotting\n line_color=[0,255,255]\n line_sz=5\n #Get the deminsions in terms of slides\n tile_cnt_size = [base_img.shape[0]//sz, base_img.shape[1]//sz]\n\n #Iterate through all images; change the border on the selected tiles\n prod_tiles = []\n for i,img in enumerate(all_tiles):\n if i in idxs:\n #If image is in the slected slides change the..\n #Left\n img[:,:line_sz,:] = [[line_color]*line_sz]*sz\n #Right\n img[:,-line_sz:,:] = [[line_color]*line_sz]*sz\n #Top\n img[:line_sz] = [[line_color]*sz]*line_sz\n #Bottom\n img[-line_sz:] = [[line_color]*sz]*line_sz\n #... boarders to the specified color\n prod_tiles.append(img)\n\n #Piece the tiles back into one image\n #Split the array of tiles into a list of rows\n rows = np.array_split(prod_tiles,tile_cnt_size[0])\n #Horizontally combine rows\n row_combine = [np.hstack(tiles) for tiles in rows]\n #Vertically stack rows back into base image\n prod_image = np.vstack(row_combine)\n #Display image\n plt.imshow(prod_image)\n plt.show()\nreturn tissue_tiles\n```
\n\nI have been utilizing the amazing wizardry done by @iafoss throughout this competition in his notebooks and discussions around using tiles pulled from the slides as data input.
\n\nI made a notebook that displays the modules for both the tiles themselves as well as an overlay of the selected tiles on a given slide. I made them both into simple python modules with outputs as below:
\n\n
I also added some of my notes to the tile module to better understand the magic performed:\n```:::python\ndef tile_plot(base_image, N=12, sz=128, plot=True):\n \"\"\"\n Description\n ____\n Tilizer module made by @iafoss that can be found in the notebook:\n https://www.kaggle.com/iafoss/panda-concat-tile-pooling-starter-inference\n Takes a base image and returns the N tiles with the largest differnce\n from a white backgound each with a given square size of input-sz.
\n\nParameters\n__________\nbase_image: numpy array\n Image array to split into tiles and plot\nN: int\n This is the number of tiles to split the image into\nsz: int\n This is the size for each side of the square tiles\nplot: bool\n True to show plot of chosen tiles, False for silent return\n\nReturns\n__________\n- List of size N with each item being a numpy array tile.\n\"\"\"\n\n#Get the shape of the input image\nshape = base_image.shape\n\n#Find the padding such that the image divides evenly by the desired size\npad0,pad1 = (sz - shape[0]%sz)%sz, (sz - shape[1]%sz)%sz\n\n#Pad the image with blank space to reach the above found tagrgets\nbase_img = np.pad(base_image,[[pad0//2,pad0-pad0//2],\n [pad1//2,pad1-pad1//2],[0,0]],\n constant_values=255)\n\n#Reshape and Transpose to get the images into tiles\nall_tiles = base_img.reshape(base_img.shape[0]//sz,sz, base_img.shape[1]//sz,sz,3)\nall_tiles = all_tiles.transpose(0,2,1,3,4).reshape(-1,sz,sz,3)\n\n#If there are not enough tiles to meet desired N pad again\nif len(all_tiles) < N:\n all_tiles = np.pad(all_tiles,[[0,N-len(all_tiles)],[0,0],[0,0],[0,0]],\n constant_values=255) \n\n#Sort the images by those with the lowest sum (i.e the least white)\n#Return indexes to the lowest N \nidxs = np.argsort(all_tiles.reshape(all_tiles.shape[0],-1).sum(-1))[:N]\n\n#Slect by index those returned from the above funtion\ntissue_tiles = all_tiles[idxs]\n\nif plot:\n #Funciton for plotting\n line_color=[0,255,255]\n line_sz=5\n #Get the deminsions in terms of slides\n tile_cnt_size = [base_img.shape[0]//sz, base_img.shape[1]//sz]\n\n #Iterate through all images; change the border on the selected tiles\n prod_tiles = []\n for i,img in enumerate(all_tiles):\n if i in idxs:\n #If image is in the slected slides change the..\n #Left\n img[:,:line_sz,:] = [[line_color]*line_sz]*sz\n #Right\n img[:,-line_sz:,:] = [[line_color]*line_sz]*sz\n #Top\n img[:line_sz] = [[line_color]*sz]*line_sz\n #Bottom\n img[-line_sz:] = [[line_color]*sz]*line_sz\n #... boarders to the specified color\n prod_tiles.append(img)\n\n #Piece the tiles back into one image\n #Split the array of tiles into a list of rows\n rows = np.array_split(prod_tiles,tile_cnt_size[0])\n #Horizontally combine rows\n row_combine = [np.hstack(tiles) for tiles in rows]\n #Vertically stack rows back into base image\n prod_image = np.vstack(row_combine)\n #Display image\n plt.imshow(prod_image)\n plt.show()\nreturn tissue_tiles\n```
\n\n