{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"pygments_lexer":"ipython3","nbconvert_exporter":"python","version":"3.6.4","file_extension":".py","codemirror_mode":{"name":"ipython","version":3},"name":"python","mimetype":"text/x-python"}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"## Tensorflow HuBMAP - Hacking the Kidney competition starter kit:\n* https://www.kaggle.com/wrrosa/hubmap-tf-with-tpu-efficientunet-512x512-tfrecs (how to create training and inference tfrecords)\n* https://www.kaggle.com/wrrosa/hubmap-tf-with-tpu-efficientunet-512x512-train (training pipeline)\n* this notebook (inference with submission)\n\n# Versions\n* V1 (V7 train notebook) 4-CV efficientunetb0 512x512 (**LB .834**)\n* V2 (V8 train notebook) loss bce (LB .835)\n* V3 (V9 train notebook) efficientunetb1 (CV .871, LB .830)\n* V4 (V10 train notebook) efficientunetb4 (CV .874, **LB .839**)\n* V5 (V12 train notebook) efficientunetb7 (CV .858, LB .835)\n* V6 (V13 train notebook) efficientunetb4 (CV .877, LB .836)\n* V7 (V14 train notebook) efficientunetb4 with overlapped train data, summing preds in inference (CV .879, **LB .843**)\n* V8 (V14 train notebook) efficientunetb4  THRESHOLD=0.4, interpolation = cv2.INTER_AREA, rle_encode_less_memory (**LB .846**)\n* V9 (V14 train notebook) efficientunetb4, MIN_OVERLAP = 300 (**LB 0.848**)\n* V10 (V14 train notebook) efficientunetb4, checksum mask before modifications (1h 11m, no need to score)\n* V11 (V14 train notebook) efficientunetb4, SUBMISSION_MODE added (generate submission from public tfrec files, almost 20m = 3.5 times faster!)\n* V12 (V14 train notebook) efficientunetb4, CHECKSUM = False (**LB 0.848**)\n* V13 (V15 train notebook) efficientunetb4, (**LB 0.849**)\n* V14 (V15 train notebook) efficientunetb4, switch public tfrec files path: from nb output to dataset (**LB 0.849**)","metadata":{}},{"cell_type":"markdown","source":"# Refferences:\n* https://www.kaggle.com/joshi98kishan/hubmap-keras-pipeline-training-inference\n* https://www.kaggle.com/bguberfain/memory-aware-rle-encoding/\n* https://www.kaggle.com/leighplt/pytorch-fcn-resnet50","metadata":{}},{"cell_type":"markdown","source":"# Parameters\nRead parameteres from notebook output, actually only **DIM** is used:","metadata":{}},{"cell_type":"code","source":"mod_path = '/kaggle/input/hubmap-tf-with-tpu-efficientunet-512x512-train/'\nimport yaml\nimport pprint\nwith open(mod_path+'params.yaml') as file:\n    P = yaml.load(file, Loader=yaml.FullLoader)\n    pprint.pprint(P)\n    \nTHRESHOLD = 0.4 # preds > THRESHOLD\nWINDOW = 1024\nMIN_OVERLAP = 300\nNEW_SIZE = P['DIM']\n\nSUBMISSION_MODE = 'PUBLIC_TFREC' \n# 'PUBLIC_TFREC' = use created tfrecords for public test set with MIN_OVERLAP = 300 tiling 1024-512, ignore other (private test) data\n# 'FULL' do not use tfrecords, just full submission \n\nCHECKSUM = False # compute mask sum for each image\n","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Metrics","metadata":{}},{"cell_type":"code","source":"import json\n\nwith open(mod_path + 'metrics.json') as json_file:\n    M = json.load(json_file)\nprint('Model run datetime: '+M['datetime'])\nprint('OOF val_dice_coe: ' + str(M['oof_dice_coe']))","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Packages","metadata":{}},{"cell_type":"code","source":"! pip install ../input/kerasapplications/keras-team-keras-applications-3b180cb -f ./ --no-index -q\n! pip install ../input/efficientnet/efficientnet-1.1.0/ -f ./ --no-index -q\nimport numpy as np\nimport pandas as pd\nimport os\nimport glob\nimport gc\n\nimport rasterio\nfrom rasterio.windows import Window\n\nimport pathlib\nfrom tqdm.notebook import tqdm\nimport cv2\n\nimport tensorflow as tf\nimport efficientnet as efn\nimport efficientnet.tfkeras","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Functions","metadata":{}},{"cell_type":"code","source":"def rle_encode_less_memory(img):\n    pixels = img.T.flatten()\n    pixels[0] = 0\n    pixels[-1] = 0\n    runs = np.where(pixels[1:] != pixels[:-1])[0] + 2\n    runs[1::2] -= runs[::2]\n    return ' '.join(str(x) for x in runs)\n\ndef make_grid(shape, window=256, min_overlap=32):\n    \"\"\"\n        Return Array of size (N,4), where N - number of tiles,\n        2nd axis represente slices: x1,x2,y1,y2 \n    \"\"\"\n    x, y = shape\n    nx = x // (window - min_overlap) + 1\n    x1 = np.linspace(0, x, num=nx, endpoint=False, dtype=np.int64)\n    x1[-1] = x - window\n    x2 = (x1 + window).clip(0, x)\n    ny = y // (window - min_overlap) + 1\n    y1 = np.linspace(0, y, num=ny, endpoint=False, dtype=np.int64)\n    y1[-1] = y - window\n    y2 = (y1 + window).clip(0, y)\n    slices = np.zeros((nx,ny, 4), dtype=np.int64)\n    \n    for i in range(nx):\n        for j in range(ny):\n            slices[i,j] = x1[i], x2[i], y1[j], y2[j]    \n    return slices.reshape(nx*ny,4)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Models","metadata":{}},{"cell_type":"code","source":"identity = rasterio.Affine(1, 0, 0, 0, 1, 0)\nfold_models = []\nfor fold_model_path in glob.glob(mod_path+'*.h5'):\n    fold_models.append(tf.keras.models.load_model(fold_model_path,compile = False))\nprint(len(fold_models))","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Tfrecords functions","metadata":{}},{"cell_type":"code","source":"AUTO = tf.data.experimental.AUTOTUNE\nimage_feature = {\n    'image': tf.io.FixedLenFeature([], tf.string),\n    'x1': tf.io.FixedLenFeature([], tf.int64),\n    'y1': tf.io.FixedLenFeature([], tf.int64)\n}\ndef _parse_image(example_proto):\n    example = tf.io.parse_single_example(example_proto, image_feature)\n    image = tf.reshape( tf.io.decode_raw(example['image'],out_type=np.dtype('uint8')), (P['DIM'],P['DIM'], 3))\n    return image, example['x1'], example['y1']\n\ndef load_dataset(filenames, ordered=True):\n    ignore_order = tf.data.Options()\n    if not ordered:\n        ignore_order.experimental_deterministic = False\n    dataset = tf.data.TFRecordDataset(filenames)\n    dataset = dataset.with_options(ignore_order)\n    dataset = dataset.map(_parse_image)\n    return dataset\n\ndef get_dataset(FILENAME):\n    dataset = load_dataset(FILENAME)\n    dataset  = dataset.batch(64)\n    dataset = dataset.prefetch(AUTO)\n    return dataset","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Results","metadata":{}},{"cell_type":"code","source":"p = pathlib.Path('../input/hubmap-kidney-segmentation')\nsubm = {}\n\nfor i, filename in tqdm(enumerate(p.glob('test/*.tiff')), \n                        total = len(list(p.glob('test/*.tiff')))):\n    \n    print(f'{i+1} Predicting {filename.stem}')\n    \n    dataset = rasterio.open(filename.as_posix(), transform = identity)\n    preds = np.zeros(dataset.shape, dtype=np.uint8)    \n    \n    if SUBMISSION_MODE == 'PUBLIC_TFREC' and MIN_OVERLAP == 300 and WINDOW == 1024 and NEW_SIZE == 512:\n        print('SUBMISSION_MODE: PUBLIC_TFREC')\n        fnames = glob.glob('/kaggle/input/hubmap-tfrecords-1024-512/test/'+filename.stem+'*.tfrec')\n        \n        if len(fnames)>0: # PUBLIC TEST SET\n            for FILENAME in fnames:\n                pred = None\n                for fold_model in fold_models:\n                    tmp = fold_model.predict(get_dataset(FILENAME))/len(fold_models)\n                    if pred is None:\n                        pred = tmp\n                    else:\n                        pred += tmp\n                    del tmp\n                    gc.collect()\n\n                pred = tf.cast((tf.image.resize(pred, (WINDOW,WINDOW)) > THRESHOLD),tf.bool).numpy().squeeze()\n\n                idx = 0\n                for img, X1, Y1 in get_dataset(FILENAME):\n                    for fi in range(X1.shape[0]):\n                        x1 = X1[fi].numpy()\n                        y1 = Y1[fi].numpy()\n                        preds[x1:(x1+WINDOW),y1:(y1+WINDOW)] += pred[idx]\n                        idx += 1\n                        \n        else: # IGNORE PRIVATE TEST SET (CREATE TFRECORDS IN FUTURE)\n            pass\n    else:\n        print('SUBMISSION_MODE: FULL')\n        slices = make_grid(dataset.shape, window=WINDOW, min_overlap=MIN_OVERLAP)\n\n\n        for (x1,x2,y1,y2) in slices:\n            image = dataset.read([1,2,3],\n                        window=Window.from_slices((x1,x2),(y1,y2)))\n            image = np.moveaxis(image, 0, -1)\n            image = cv2.resize(image, (NEW_SIZE, NEW_SIZE),interpolation = cv2.INTER_AREA)\n            image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)\n            image = np.expand_dims(image, 0)\n\n            pred = None\n\n            for fold_model in fold_models:\n                if pred is None:\n                    pred = np.squeeze(fold_model.predict(image))\n                else:\n                    pred += np.squeeze(fold_model.predict(image))\n\n            pred = pred/len(fold_models)\n\n            pred = cv2.resize(pred, (WINDOW, WINDOW))\n            preds[x1:x2,y1:y2] += (pred > THRESHOLD).astype(np.uint8)\n\n    preds = (preds > 0.5).astype(np.uint8)\n    \n    subm[i] = {'id':filename.stem, 'predicted': rle_encode_less_memory(preds)}\n    \n    if CHECKSUM:\n        print('Checksum: '+ str(np.sum(preds)))\n    \n    del preds\n    gc.collect();","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Making submission","metadata":{}},{"cell_type":"code","source":"submission = pd.DataFrame.from_dict(subm, orient='index')\nsubmission.to_csv('submission.csv', index=False)\nsubmission.head()","metadata":{"trusted":true},"execution_count":null,"outputs":[]}]}