{"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":"
\n\n
\n\n

Mayo Clinic - STRIP AI

No Tile Model


\n
CREATED BY: DARIEN SCHETTLER
\n\n
\n\n---\n\n
\n\n
\n 🛑   WARNING:

THIS IS A WORK IN PROGRESS
\n
\n\n\n
\n 👏   IF YOU FORK THIS OR FIND THIS HELPFUL   👏

PLEASE UPVOTE!

This was a lot of work for me and while it may seem silly, it makes me feel appreciated when others like my work. 😅\n
","metadata":{}},{"cell_type":"markdown","source":"
\n\n\n\n

0  IMPORTS & INSTALLS    ⤒

\n\n
\n\n**PYVIPS INSTALL CODE ONLY WORKS PINNED TO ORIGINAL ENVIRONMENT (2020)**\n- Original: https://www.kaggle.com/hirune924/fast-image-region-loading-using-pyvips\n - Thanks for @hirune924 !!\n- https://www.kaggle.com/code/yukkyo/using-pyvips-without-internet/comments\n- Install needs about 5 minutes","metadata":{}},{"cell_type":"code","source":"print(\"\\n... INSTALLING LOCAL VERSION OF PYVIPS! ...\")\n!dpkg -i --force-depends /kaggle/input/pyvips-local/libvips-apt/libvips-apt/*.deb >/dev/null 2>&1\n!pip install /kaggle/input/pyvips-local/cffi-1.14.4-cp37-cp37m-manylinux1_x86_64.whl\n!pip install /kaggle/input/pyvips-local/pycparser-2.20-py2.py3-none-any.whl\n!pip install /kaggle/input/pyvips-local/pyvips-2.1.13-py2.py3-none-any.whl\nimport pyvips\nprint(\"... INSTALL COMPLETE! ...\\n\")","metadata":{"execution":{"iopub.status.busy":"2022-07-13T22:18:43.237916Z","iopub.execute_input":"2022-07-13T22:18:43.238342Z","iopub.status.idle":"2022-07-13T22:21:03.931601Z","shell.execute_reply.started":"2022-07-13T22:18:43.238293Z","shell.execute_reply":"2022-07-13T22:21:03.930877Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"print(\"\\n... IMPORTS STARTING ...\\n\")\n\nprint(\"\\n\\tVERSION INFORMATION\")\n\n# Machine Learning and Data Science Imports\nimport tensorflow as tf; print(f\"\\t\\t– TENSORFLOW VERSION: {tf.__version__}\");\nimport pandas as pd; pd.options.mode.chained_assignment = None;\nimport numpy as np; print(f\"\\t\\t– NUMPY VERSION: {np.__version__}\");\n\n# Built In Imports\nfrom glob import glob\nimport openslide\nimport warnings\nimport requests\nimport hashlib\nimport imageio\nimport IPython\nimport sklearn\nimport urllib\nimport zipfile\nimport pickle\nimport random\nimport shutil\nimport string\nimport json\nimport math\nimport time\nimport gzip\nimport ast\nimport sys\nimport io\nimport os\nimport gc\nimport re\n\n# Visualization Imports\nfrom matplotlib.colors import ListedColormap\nfrom matplotlib.patches import Rectangle\nimport matplotlib.patches as patches\nimport plotly.graph_objects as go\nimport matplotlib.pyplot as plt\nfrom tqdm.notebook import tqdm; tqdm.pandas();\nimport plotly.express as px\nimport tifffile as tif\nimport seaborn as sns\nfrom PIL import Image, ImageEnhance; Image.MAX_IMAGE_PIXELS = 5_000_000_000;\nimport matplotlib; print(f\"\\t\\t– MATPLOTLIB VERSION: {matplotlib.__version__}\");\nfrom matplotlib import animation, rc; rc('animation', html='jshtml')\nimport plotly\nimport PIL\nimport cv2\n\nimport plotly.io as pio\nprint(pio.renderers)\n\ndef seed_it_all(seed=7):\n \"\"\" Attempt to be Reproducible \"\"\"\n os.environ['PYTHONHASHSEED'] = str(seed)\n random.seed(seed)\n np.random.seed(seed)\n tf.random.set_seed(seed)\nseed_it_all()\n \nprint(\"\\n\\n... IMPORTS COMPLETE ...\\n\")","metadata":{"execution":{"iopub.status.busy":"2022-07-13T22:21:03.936916Z","iopub.execute_input":"2022-07-13T22:21:03.937507Z","iopub.status.idle":"2022-07-13T22:21:12.037579Z","shell.execute_reply.started":"2022-07-13T22:21:03.937467Z","shell.execute_reply":"2022-07-13T22:21:12.036514Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"
\n\n\n\n

1  SETUP    ⤒

\n\n---\n","metadata":{}},{"cell_type":"code","source":"print(\"\\n... BASIC SETUP STARTING ...\\n\\n\")\n\n# Path to competition data\nDATA_DIR = \"/kaggle/input/mayo-clinic-strip-ai\"\n\n# enable XLA optmizations (10% speedup when using @tf.function calls)\ntf.config.optimizer.set_jit(True)\n\n# From dataset\nINPUT_SHAPE = (512,512,3)\nS2I_LBL_MAP = {\"CE\":0, \"LAA\":1}\nI2S_LBL_MAP = {v:k for k,v in S2I_LBL_MAP.items()}\nN_CLASSES=len(S2I_LBL_MAP)\n\n# Open the training dataframe and display the initial dataframe\nTRAIN_DIR = os.path.join(DATA_DIR, \"train\")\nTRAIN_CSV = os.path.join(DATA_DIR, \"train.csv\")\ntrain_df = pd.read_csv(TRAIN_CSV)\ntrain_df[\"image_path\"] = train_df[\"image_id\"].apply(lambda x: os.path.join(TRAIN_DIR, x+\".tif\"))\n\nprint(\"\\n... TRAINING DATAFRAME... \\n\")\ndisplay(train_df)\n\n# Open the testing dataframe and display the initial dataframe\nTEST_DIR = os.path.join(DATA_DIR, \"test\")\nTEST_CSV = os.path.join(DATA_DIR, \"test.csv\")\ntest_df = pd.read_csv(TEST_CSV)\ntest_df[\"image_path\"] = test_df[\"image_id\"].apply(lambda x: os.path.join(TEST_DIR, x+\".tif\"))\n\nprint(\"\\n... TESTING DATAFRAME... \\n\")\ndisplay(test_df)\n\n# Open the sample submission dataframe\nSS_CSV = os.path.join(DATA_DIR, \"sample_submission.csv\")\nss_df = pd.read_csv(SS_CSV)\n\nprint(\"\\n... SAMPLE SUBMISSION DATAFRAME... \\n\")\ndisplay(ss_df)\n\n# For debugging purposes when the test set hasn't been substituted we will know\nDEBUG=len(ss_df)==4\nprint(f\"\\n\\n\\n... ARE WE DEBUGGING: {DEBUG}... \\n\")\n\nAPPROX_PRIVATE_TEST_LEN = 280\nif DEBUG:\n print(\"\\n\\n... MODIFYING TEST DATAFRAME TO BE MORE SIMILAR TO PRIVATE DATASET ...\")\n test_df = train_df.sample(APPROX_PRIVATE_TEST_LEN).reset_index(drop=True).drop(columns=[\"label\"])\n print(\"\\n... UPDATED TESTING DATAFRAME... \\n\\n\")\n display(test_df)\n\n# Capture examples that span smallest, median, and large image size examples\nSMALLEST_IMAGE_ID = \"b43ebe_0\"\nSMALLEST_IMAGE_ROW = train_df[train_df.image_id==SMALLEST_IMAGE_ID]\nMEDIAN_IMAGE_ID = \"719165_0\"\nMEDIAN_IMAGE_ROW = train_df[train_df.image_id==MEDIAN_IMAGE_ID]\nLARGEST_IMAGE_ID = \"6baf51_0\"\nLARGEST_IMAGE_ROW = train_df[train_df.image_id==LARGEST_IMAGE_ID]\n\nprint(\"\\n... BASIC SETUP FINISHED ...\\n\\n\")","metadata":{"_kg_hide-output":true,"execution":{"iopub.status.busy":"2022-07-13T22:21:12.038969Z","iopub.execute_input":"2022-07-13T22:21:12.039328Z","iopub.status.idle":"2022-07-13T22:21:12.129187Z","shell.execute_reply.started":"2022-07-13T22:21:12.039288Z","shell.execute_reply":"2022-07-13T22:21:12.128331Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"
\n\n\n\n

2  HELPER FUNCTIONS    ⤒

","metadata":{}},{"cell_type":"code","source":"# vips image to numpy array\ndef vips2numpy(vi):\n # map vips formats to np dtypes\n format_to_dtype = {\n 'uchar': np.uint8,\n 'char': np.int8,\n 'ushort': np.uint16,\n 'short': np.int16,\n 'uint': np.uint32,\n 'int': np.int32,\n 'float': np.float32,\n 'double': np.float64,\n 'complex': np.complex64,\n 'dpcomplex': np.complex128,\n }\n return np.ndarray(buffer=vi.write_to_memory(),\n dtype=format_to_dtype[vi.format],\n shape=[vi.height, vi.width, vi.bands])\n\ndef pyvips_open_downsampled_slide(img_path, downsample_by=8, as_numpy=True, resize_to=(512,512)):\n \"\"\"\n \n Helper function to convert WSI into smaller downscaled version using pyvips.\n \n Timing details for MAYO CLINIC STRIP AI dataset:\n SMALLEST IMAGE BY AREA (4417, 5314)\n * Function takes ~1 seconds to run\n MEDIAN IMAGE BY AREA (17573, 38743)(~30X LARGER THAN SMALLEST IMAGE)\n * Function takes ~30 seconds to run\n LARGEST IMAGE BY AREA (48282, 101406)(~208X LARGER THAN SMALLEST IMAGE)(~7X LARGER THAN MEDIAN IMAGE)\n * Function takes ~405 seconds to run\n \n Args:\n img_path (str): Path to .tif file to be downsampled\n downsample_by (int): How many times smaller should resultant \n image be. i.e. image_size*(1/downsample_by) = new_size\n as_numpy (bool, optional): Whether to return image as numpy array (default)\n or leave as PIL.Image object for further manipulation\n \n Returns:\n Downsampled image as a numpy array of type uint8 with only 3 channels\n \"\"\"\n \n # Open the image with PIL\n tmp_img = pyvips.Image.new_from_file(img_path) \n \n print(\"\\n... APPROXIMATE TIME TO LOAD IMAGE IS AT MOST APPROXIMATELY: \" \\\n f\"{int((405/(48282*101406))*(tmp_img.width*tmp_img.height))} SECONDS ...\\n\")\n \n # if -1 than we downsample by whatever results in the image having dimensions as\n # close to 512x512 as possible so the image can be resized after\n \n if downsample_by==-1:\n _epsilon = 1e-3\n downsample_by=min(tmp_img.width, tmp_img.height)/resize_to[0]-_epsilon\n \n # Resize the image\n tmp_img = tmp_img.resize(1/downsample_by)\n tmp_img = vips2numpy(tmp_img) if as_numpy else tmp_img\n tmp_img = cv2.resize(tmp_img, resize_to) if resize_to is not None else tmp_img\n \n gc.collect(); gc.collect();\n return tmp_img\n\n# Prove that we can infer on all image sizes available in training dataset \n# --> (smallest, median and largest)\nif DEBUG:\n plt.figure(figsize=(20,15))\n \n plt.subplot(1,3,1)\n plt.imshow(pyvips_open_downsampled_slide(SMALLEST_IMAGE_ROW.image_path.values[0], downsample_by=-1, as_numpy=True, resize_to=(512,512)))\n plt.title(\"SMALLEST IMAGE\", fontweight=\"bold\")\n \n plt.subplot(1,3,2)\n plt.imshow(pyvips_open_downsampled_slide(MEDIAN_IMAGE_ROW.image_path.values[0], downsample_by=-1, as_numpy=True, resize_to=(512,512)))\n plt.title(\"MEDIAN IMAGE\", fontweight=\"bold\")\n \n plt.subplot(1,3,3)\n plt.imshow(pyvips_open_downsampled_slide(LARGEST_IMAGE_ROW.image_path.values[0], downsample_by=-1, as_numpy=True, resize_to=(512,512)))\n plt.title(\"LARGEST IMAGE\", fontweight=\"bold\")\n \n plt.tight_layout()\n plt.show()","metadata":{"execution":{"iopub.status.busy":"2022-07-13T22:21:12.130625Z","iopub.execute_input":"2022-07-13T22:21:12.131003Z","iopub.status.idle":"2022-07-13T22:21:12.143614Z","shell.execute_reply.started":"2022-07-13T22:21:12.130967Z","shell.execute_reply":"2022-07-13T22:21:12.142638Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"
\n\n\n\n\n\n

\n 3  INFER ON TEST IMAGES    ⤒\n

\n\n---\n","metadata":{}},{"cell_type":"code","source":"# def tta_augment(img_batch):\n# img_batch = tf.image.random_brightness(img_batch, 0.2)\n# img_batch = tf.image.random_contrast(img_batch, 0.5, 2.0)\n# img_batch = tf.image.random_saturation(img_batch, 0.75, 1.25)\n# img_batch = tf.image.random_hue(img_batch, 0.1)\n# return img_batch\n\nN_TEST = len(test_df)\ntest_image_paths = test_df.image_path.values\n\n# test time augmentation (not implemented yet)\nDO_TTA = False\nN_TTA = 4\n\n# Model loading\nprint(\"\\n... LOADING MODEL (~60-120 SECONDS)...\\n\")\nMODEL_PATH = \"/kaggle/input/mcsai-no-tiling-model-tpu-tf/efficientnetb6_512_notile\"\nmodel = tf.keras.models.load_model(MODEL_PATH, compile=False)\ndisplay(model.summary())\n\n# Get predictions\npreds = []\nfor img_path in tqdm(test_image_paths, total=N_TEST):\n preds.append(model(tf.expand_dims(pyvips_open_downsampled_slide(img_path, downsample_by=-1, as_numpy=True, resize_to=(512,512)), axis=0)).numpy())\n tf.keras.backend.clear_session(); gc.collect(); gc.collect();\npreds = np.concatenate(preds, axis=0)\nprint(preds.shape)","metadata":{"execution":{"iopub.status.busy":"2022-07-13T22:33:03.652719Z","iopub.execute_input":"2022-07-13T22:33:03.653084Z","iopub.status.idle":"2022-07-13T22:33:50.718634Z","shell.execute_reply.started":"2022-07-13T22:33:03.65305Z","shell.execute_reply":"2022-07-13T22:33:50.717734Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"
\n\n\n\n\n\n

\n 4  CREATE SUBMISSION.CSV    ⤒\n

\n\n---\n\nRemember we have to identify predictions for the same patient and take the average probability for each class before mapping those probabilities to the patient_id in the sample submission dataframe","metadata":{}},{"cell_type":"code","source":"for lbl, idx in S2I_LBL_MAP.items():\n test_df[lbl]=preds[:, idx]\n\nCE_test_prob_map = test_df.groupby(\"patient_id\")[\"CE\"].mean().to_dict()\nLAA_test_prob_map = test_df.groupby(\"patient_id\")[\"LAA\"].mean().to_dict()\n\nss_df[\"CE\"] = ss_df[\"patient_id\"].map(lambda x: CE_test_prob_map.get(x, 0.5))\nss_df[\"LAA\"] = ss_df[\"patient_id\"].map(lambda x: LAA_test_prob_map.get(x, 0.5))\nss_df.to_csv(\"submission.csv\", index=False)\n\ndisplay(ss_df)","metadata":{"execution":{"iopub.status.busy":"2022-07-13T22:38:10.116923Z","iopub.execute_input":"2022-07-13T22:38:10.117341Z","iopub.status.idle":"2022-07-13T22:38:10.423824Z","shell.execute_reply.started":"2022-07-13T22:38:10.117304Z","shell.execute_reply":"2022-07-13T22:38:10.422874Z"},"trusted":true},"execution_count":null,"outputs":[]}]}