Libraries
","metadata":{}},{"cell_type":"code","source":"import pandas as pd\nimport numpy as np\nimport matplotlib.pyplot as plt\nfrom matplotlib import animation\nimport seaborn as sns\nfrom tqdm.auto import tqdm\nfrom pandas.io.formats.style import Styler\nimport json\nimport os\n\nimport warnings\nwarnings.filterwarnings('ignore')\n\ntqdm.pandas()\n\nrc = {\n \"axes.facecolor\": \"#F8F8F8\",\n \"figure.facecolor\": \"#F8F8F8\",\n \"axes.edgecolor\": \"#000000\",\n \"grid.color\": \"#EBEBE7\" + \"30\",\n \"font.family\": \"serif\",\n \"axes.labelcolor\": \"#000000\",\n \"xtick.color\": \"#000000\",\n \"ytick.color\": \"#000000\",\n \"grid.alpha\": 0.4\n}\n\nsns.set(rc=rc)\npalette = ['#302c36', '#037d97', '#91013E', '#C09741',\n '#EC5B6D', '#90A6B1', '#6ca957', '#D8E3E2']\n\nfrom colorama import Style, Fore\nblk = Style.BRIGHT + Fore.BLACK\nmgt = Style.BRIGHT + Fore.MAGENTA\nred = Style.BRIGHT + Fore.RED\nblu = Style.BRIGHT + Fore.BLUE\nres = Style.RESET_ALL\n\nfrom IPython import display as disp","metadata":{"_kg_hide-input":true,"execution":{"iopub.status.busy":"2023-05-11T23:20:49.00771Z","iopub.execute_input":"2023-05-11T23:20:49.008123Z","iopub.status.idle":"2023-05-11T23:20:49.025889Z","shell.execute_reply.started":"2023-05-11T23:20:49.008078Z","shell.execute_reply":"2023-05-11T23:20:49.024589Z"},"_kg_hide-output":true,"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"\n \n
Introduction
","metadata":{}},{"cell_type":"markdown","source":"‣ The aim of this competition is to help researchers improve the accuracy of their contrail models by validating them with satellite imagery.
\n\n‣ Contrails (short for \"condensation trails\") or vapor trails are line-shaped clouds produced by aircraft engine exhaust or changes in air pressure, typically at aircraft cruising altitudes several miles above the Earth's surface. Contrails are composed primarily of water, in the form of ice crystals.
\n\n‣ Contrails can contribute to global warming by trapping heat in the atmosphere.
\n \n‣ Contrail avoidance is potentially one of the most scalable, cost-effective sustainability solutions for reducing the impact on climate change available to airlines today.
\n \n‣ Submissions are evaluated on the global Dice coefficient. The formula is given by:
\n\n$$\\frac{2 * \\left| X \\bigcap_{}^{} Y \\right|}{\\left| X \\right| + \\left| Y \\right|}$$\n\nwhere X is the entire set of predicted contrail pixels for all observations in the test data and Y is the ground truth set of all contrail pixels in the test data. The maximum score for the Dice coefficient is 1, which indicates a perfect overlap between the predicted and ground truth binary masks.
\n\n‣ The competition format requires a space delimited list of pairs. For example, '1 3 10 5' implies pixels 1,2,3,10,11,12,13,14 are to be included in the mask. The metric checks that the pairs are sorted, positive, and the decoded pixel values are not duplicated. The pixels are numbered from top to bottom, then left to right: 1 is pixel (1,1), 2 is pixel (2,1), etc..
","metadata":{}},{"cell_type":"markdown","source":"##Data
","metadata":{}},{"cell_type":"markdown","source":"‣ The dataset for this competition was obtained from the GOES-16 Advanced Baseline Imager (ABI), which is publicly available on Google Cloud Storage. Our task is to use geostationary satellite images to identify aviation contrails.
\n\n‣ The original full-disk images were reprojected using bilinear resampling to generate a local scene image. Because contrails are easier to identify with temporal context, a sequence of images at 10-minute intervals are provided. Each example (record_id) contains exactly one labeled frame.
‣ Learn more about the dataset from the preprint: OpenContrails: Benchmarking Contrail Detection on GOES-16 ABI.
\n\n‣ Labeling instructions can be found at in this supplementary material Some key labeling guidance:
\n\n‣ Ground truth was determined by (generally) 4+ different labelers annotating each image. Pixels were considered a contrail when >50% of the labelers annotated it as such. Individual annotations (human_individual_masks.npy) as well as the aggregated ground truth annotations (human_pixel_masks.npy) are included in the training data. The validation data only includes the aggregated ground truth annotations.
To get a sense of the data, here are a few illustrations from the preprint above:
\n\n\n \n
‣ The dataset provided with this challenge consists of the following folders and files:
\n\nrecord_id and contains the following data:H x W x T, where T = n_times_before + n_times_after + 1, representing the number of images in the sequence. There are n_times_before and n_times_after images before and after the labeled frame respectively. In our dataset all examples have n_times_before=4 and n_times_after=3. Each band represents an infrared channel at different wavelengths and is converted to brightness temperatures based on the calibration parameters. The number in the filename corresponds to the GOES-16 ABI band number. Details of the ABI bands can be found here.H x W x 1 x R. Each example is labeled by R individual human labelers. R is not the same for all samples. The labeled masks have value either 0 or 1 and correspond to the (n_times_before+1)-th image in band_{08-16}.npy. They are available only in the training set.H x W x 1 containing the binary ground truth. A pixel is regarded as contrail pixel in evaluation if it is labeled as contrail by more than half of the labelers. '-' in the submission. (See this notebook for details.)Preprint Notes
","metadata":{}},{"cell_type":"markdown","source":"‣ Dataset contains 20,544 examples in the train set and 1,866 examples in the validation set. The examples are randomly partitioned except for the satellites scenes that were identified as likely to have contrails by Google Street View, which are only included in the training set. 9,283 of the training examples contain at least one annotated contrail. About 1.2% of the pixels in the training set are labeled as contrails.
\n‣ The dataset contains a wide variety of times and locations, as shown in Figure 5 and 6. The examples are not uniformly distributed in space and time as the images are sampled to include more contrail examples as described above.
\n\n\n \n
‣ Contrails occur more often in cloudy scenes. We estimate the cloud cover fraction using the GOES-16 ABI L2 Cloud Top Phase product and regard all non-clear-sky pixels as clouds. We show the example counts in our dataset by the cloud coverage fraction in Figure 7. Our dataset contains more cloudy scenes than clear sky images, and the positives examples are roughly proportionally distributed at different cloud coverage fractions. It is possible that contrails are harder to be identified in cloudy scenes.
\n\n\n \n
‣ The authors used a convolutional neural network model for contrail detection. The model outputs a score (between 0 to 1) for each pixel indicating the confidence that pixel is part of a contrail. First, the authors trained a binary image segmentation model for classifying each pixel as contrail or background, then the semantic segmentation architecture DeeplabV3+ was employed experimenting with different image backbones Below is an example of network achitercture for multi-frame detection model:
\n\n\n \n
‣ The training results of the single frame models and multi-frame\n model with different backbones were summarized in Table I. All models achieve reasonable detection performance, showing that our dataset is sufficiently large for training contrail detection models. The multi-frame model slightly outperforms the single-frame based models, showing the model is able to use temporal context to improve detection
\n \n\n \n
‣ To access additional details, please, refer to the preprint OpenContrails: Benchmarking Contrail Detection on GOES-16 ABI.
","metadata":{}},{"cell_type":"code","source":"BASE_DIR = '/kaggle/input/google-research-identify-contrails-reduce-global-warming/train'\nTRAIN_META_PATH = '/kaggle/input/google-research-identify-contrails-reduce-global-warming/train_metadata.json'\nTEST_META_PATH = '/kaggle/input/google-research-identify-contrails-reduce-global-warming/validation_metadata.json'\nTRAIN_SIZES_PATH = '/kaggle/input/train-file-sizes-google-identify-contrails/train_file_sizes.csv'\n\ndef process_json_meta_data(file_path: str, verbose: bool=False):\n '''Loads the json meta file and preprocesses it.\n \n Args:\n file_path: path to the json file with metadata.\n verbose: displays count of distinct values of split attributes\n for projection_wkt feature. The difference is only appears in \"central_meridian\".\n Those are the negative integers from 2 to 3 digits.\n '''\n # Loads JSON file.\n with open(file_path, 'r') as f:\n dict_meta = json.load(f)\n \n df = pd.DataFrame.from_dict(dict_meta)\n \n if verbose:\n display(df.projection_wkt.str.split(',', expand=True).nunique())\n display(df.projection_wkt.str.split(',', expand=True)[22].value_counts())\n \n expr = r'(?<=\"central_meridian\",)([-]*\\d{2,3})(?=])'\n df['central_meridian'] = df.projection_wkt.str.extract(expr).astype(int) \n df = df.drop(columns='projection_wkt')\n df['record_id'] = df['record_id'].astype(int)\n return df\n \ntrain_meta = process_json_meta_data(TRAIN_META_PATH)\ntest_meta = process_json_meta_data(TEST_META_PATH)\ntrain_sizes = pd.read_csv(TRAIN_SIZES_PATH)","metadata":{"execution":{"iopub.status.busy":"2023-05-11T23:22:10.21567Z","iopub.execute_input":"2023-05-11T23:22:10.216068Z","iopub.status.idle":"2023-05-11T23:22:11.555457Z","shell.execute_reply.started":"2023-05-11T23:22:10.216039Z","shell.execute_reply":"2023-05-11T23:22:11.554318Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"##Metadata Overview
","metadata":{}},{"cell_type":"markdown","source":"This notebook provides a code snippet of how effectively display dataframes in a tidy format using Pandas Styler class.\n\nWe are going to leverage CSS styling language to manipulate many parameters including colors, fonts, borders, background, format and make our tables interactive.\n\n\n**Reference**:\n* [Pandas Table Visualization](https://pandas.pydata.org/pandas-docs/stable/user_guide/style.html).","metadata":{}},{"cell_type":"markdown","source":"At this point, you will require a certain level of understanding in web development.\n\nPrimarily, you will have to modify the CSS of the `td`, `tr`, and `th` tags.\n\n**You can refer** to the following materials to learn **HTML/CSS**:\n* [w3schools HTML Tutorial](https://www.w3schools.com/html/default.asp)\n* [w3schools CSS Reference](https://www.w3schools.com/cssref/index.php)","metadata":{}},{"cell_type":"code","source":" def magnify(is_test: bool = False):\n base_color = '#91013E'\n if is_test:\n highlight_target_row = []\n else:\n highlight_target_row = [dict(selector='tr:last-child',\n props=[('background-color', f'{base_color}'+'20')])]\n \n return [dict(selector=\"th\",\n props=[(\"font-size\", \"11pt\"),\n ('background-color', f'{base_color}'),\n ('color', 'white'),\n ('font-weight', 'bold'),\n ('border-bottom', '0.1px solid white'),\n ('border-left', '0.1px solid white'),\n ('text-align', 'right')]),\n\n dict(selector='th.blank.level0', \n props=[('font-weight', 'bold'),\n ('border-left', '1.7px solid white'),\n ('background-color', 'white')]),\n\n dict(selector=\"td\",\n props=[('padding', \"0.5em 1em\"),\n ('text-align', 'right')]),\n\n dict(selector=\"th:hover\",\n props=[(\"font-size\", \"14pt\")]),\n\n dict(selector=\"tr:hover td:hover\",\n props=[('max-width', '250px'),\n ('font-size', '14pt'),\n ('color', f'{base_color}'),\n ('font-weight', 'bold'),\n ('background-color', 'white'),\n ('border', f'1px dashed {base_color}')]),\n \n dict(selector=\"caption\",\n props=[(('caption-side', 'bottom'))])] + highlight_target_row\n\ndef stylize_min_max_count(pivot_table):\n \"\"\"Waps the min_max_count pivot_table into the Styler.\n\n Args:\n df: |min_train| max_train |min_test |max_test |top10_counts_train |top_10_counts_train|\n\n Returns:\n s: the dataframe wrapped into Styler.\n \"\"\"\n s = pivot_table\n # A formatting dictionary for controlling each column precision (.000 <-). \n di_frmt = {(i if i.startswith('m') else i):\n ('{:.3f}' if i.startswith('m') else '{:}') for i in s.columns}\n\n s = s.style.set_table_styles(magnify(True))\\\n .format(di_frmt)\\\n .set_caption(f\"The train and test datasets min, max, top10 values side by side (hover to magnify).\")\n return s\n \n \ndef stylize_describe(df: pd.DataFrame, dataset_name: str = 'train', is_test: bool = False) -> Styler:\n \"\"\"Applies .descibe() method to the df and wraps it into the Styler.\n \n Args:\n df: any dataframe (train/test/origin)\n dataset_name: default 'train'\n is_test: the bool parameter passed into magnify() function\n in order to control the highlighting of the last row.\n \n Returns:\n s: the dataframe wrapped into Styler.\n \"\"\"\n s = df.describe().T\n # A formatting dictionary for controlling each column precision (.000 <-). \n di_frmt = {(i if i == 'count' else i):\n ('{:.0f}' if i == 'count' else '{:.3f}') for i in s.columns}\n \n s = s.style.set_table_styles(magnify(is_test))\\\n .format(di_frmt)\\\n .set_caption(f\"The {dataset_name} dataset descriptive statistics (hover to magnify).\")\n return s\n\ndef stylize_simple(df: pd.DataFrame, caption: str) -> Styler:\n \"\"\"Waps the min_max_count pivot_table into the Styler.\n\n Args:\n df: any dataframe (train/test/origin)\n\n Returns:\n s: the dataframe wrapped into Styler.\n \"\"\"\n s = df\n s = s.style.set_table_styles(magnify(True)).set_caption(f\"{caption}\")\n return s\n\nft = ['row_min','row_size','col_min', 'col_size', 'timestamp', 'central_meridian']\ndisplay(stylize_simple(train_meta[ft].head(4), 'The train meta file 4 top rows (hover to magnify).'))\ndisplay(stylize_simple(test_meta[ft].head(4), 'The test meta file 4 top rows (hover to magnify).'))\nrename_di = {'index': 'file_size', 'file_size': 'counts'}\nagg = train_sizes.groupby(by='file_size').agg({'count', 'last'})['file_path']\nagg = agg.reset_index().sort_values(by='count', ascending=False)\n\ndisplay(stylize_simple(agg, 'The train file sizes (KB) (hover to magnify).'))\ndisplay(stylize_describe(train_sizes.drop(columns=['record_id', 'file_path'])/1e6, 'sizes (MB) of train'))","metadata":{"_kg_hide-input":true,"execution":{"iopub.status.busy":"2023-05-11T23:22:38.628467Z","iopub.execute_input":"2023-05-11T23:22:38.62944Z","iopub.status.idle":"2023-05-11T23:22:38.772943Z","shell.execute_reply.started":"2023-05-11T23:22:38.629401Z","shell.execute_reply":"2023-05-11T23:22:38.771792Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"**Observation**:\n* The train files `band_{}` are relatively bigger than `human_individual_masks` but we need to check if there any masks with the same size in the data.\n* It will be nice to sanity check a symmetric difference between **train/** `record_id` and **train_metadata.json** `record_id`.\n\nLet's check this out:","metadata":{}},{"cell_type":"code","source":"record_id_diff = set(train_sizes.record_id).symmetric_difference(train_meta.record_id)\nprint(f'{blk}[INFO] Any difference in record_id for train_metadata.json and train dir:'\n f'{blk}\\n[+] -> {red}{len(record_id_diff) > 0}')\n\nmask_sizes = train_sizes[~train_sizes.file_path.str.contains('band') & train_sizes.file_size.eq(2097280)]\n\nprint(f'\\n{blk}[INFO] There are following number of non band_ files with size 2097280 KB:'\n f'{blk}\\n[+] -> {red}{mask_sizes.shape[0]}\\n')\n\nstylize_simple(mask_sizes.head(4), 'Exampes of non- band_{} files with size 2097280 KB')","metadata":{"execution":{"iopub.status.busy":"2023-05-11T23:22:54.564622Z","iopub.execute_input":"2023-05-11T23:22:54.56501Z","iopub.status.idle":"2023-05-11T23:22:54.790981Z","shell.execute_reply.started":"2023-05-11T23:22:54.564971Z","shell.execute_reply":"2023-05-11T23:22:54.789829Z"},"_kg_hide-input":true,"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"**Note**:\n\n* There are **57** `human_individual_masks.npy` with the same size as `band_{}`. It makes sense to visualize such samples later on.","metadata":{}},{"cell_type":"code","source":"# kudos to @docxian\ns = sns.pairplot(train_meta[ft],\n plot_kws={'alpha' : 0.2, 'color' : palette[2]},\n diag_kws={'color' : palette[2]})\ns.fig.suptitle(f'Pair Plot for the train metadata file',\n fontweight='bold', fontsize=20, y=1.03);","metadata":{"_kg_hide-input":true,"execution":{"iopub.status.busy":"2023-05-11T23:23:22.095741Z","iopub.execute_input":"2023-05-11T23:23:22.096129Z","iopub.status.idle":"2023-05-11T23:23:35.515331Z","shell.execute_reply.started":"2023-05-11T23:23:22.0961Z","shell.execute_reply":"2023-05-11T23:23:35.514209Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"**Observation**:\n* `col_size` and `row_size` are not always symmetrical. It might make sense to check such examples.","metadata":{}},{"cell_type":"markdown","source":"##Train records
","metadata":{}},{"cell_type":"code","source":"# original source code https://www.kaggle.com/code/inversion/visualizing-contrails\n# I have prepared the custom functions below based on the original code.\n\nN_TIMES_BEFORE = 4\n\ndef normalize_range(data, bounds):\n \"\"\"Maps data to the range [0, 1].\"\"\"\n return (data - bounds[0]) / (bounds[1] - bounds[0])\n\ndef read_record(record_id: str):\n\n dir_files = os.listdir(os.path.join(BASE_DIR, record_id))\n # band_files = sorted([f for f in dir_files if f.startswith('band_')])[::2][:3]\n band_files = ['band_11.npy', 'band_14.npy', 'band_15.npy']\n files_to_viz = band_files + ['human_pixel_masks.npy', 'human_individual_masks.npy']\n\n loaded_files = []\n for ftz in files_to_viz:\n try:\n with open(os.path.join(BASE_DIR, record_id, ftz), 'rb') as f:\n loaded_files.append(np.load(f))\n except Exception as e:\n print(f'[ERROR] One of the files is missing in the train dir. {e}.')\n\n _T11_BOUNDS = (243, 303)\n _CLOUD_TOP_TDIFF_BOUNDS = (-4, 5)\n _TDIFF_BOUNDS = (-4, 2) \n\n r = normalize_range(loaded_files[2] - loaded_files[1], _TDIFF_BOUNDS)\n g = normalize_range(loaded_files[1] - loaded_files[0], _CLOUD_TOP_TDIFF_BOUNDS)\n b = normalize_range(loaded_files[1], _T11_BOUNDS)\n false_color = np.clip(np.stack([r, g, b], axis=2), 0, 1)\n return loaded_files, false_color\n\ndef viz_record(img: np.ndarray, hpm: np.ndarray) -> None:\n fig, ax = plt.subplots(1, 3, figsize=(18, 6))\n ax = ax.flatten()\n\n ax[0].imshow(img)\n ax[0].set_title('False color image')\n ax[0].grid(False)\n\n ax[1].imshow(hpm, interpolation='none')\n ax[1].set_title('Ground truth contrail mask')\n ax[1].grid(False)\n\n ax[2].imshow(img)\n ax[2].imshow(hpm, cmap='Reds', alpha=.4, interpolation='none')\n ax[2].set_title('Contrail mask on false color image')\n ax[2].grid(False)\n fig.suptitle(f'Record #: {record_id}')\n plt.show()\n \ndef viz_human_individual_masks(him: np.ndarray) -> None:\n \"\"\"Draws Individual human masks\n \n Args:\n him: human_individual_mask\n \n Returns:\n None\n \"\"\"\n n = human_individual_mask.shape[-1]\n fig = plt.figure(figsize=(16, 4))\n for i in range(n):\n plt.subplot(1, n, i+1)\n plt.imshow(human_individual_mask[..., i], interpolation='none')\n plt.grid(False)\n fig.suptitle(f'Record #: {record_id} \\n human_individual_masks')\n \ndef run_animation(fc: np.array): \n \"\"\"Runs matplotlib animation.\n \n Args: \n fc: false_color \n \n Return:\n animation.FuncAnimation\n \"\"\"\n fig = plt.figure(figsize=(12, 8))\n img = plt.imshow(fc[..., 0]) \n plt.axis('off')\n plt.close()\n \n def frame(i):\n img.set_array(fc[..., i])\n return [img]\n\n return animation.FuncAnimation(fig, frame, frames=false_color.shape[-1],\n interval=500, blit=True)","metadata":{"_kg_hide-input":true,"execution":{"iopub.status.busy":"2023-05-11T23:23:35.516857Z","iopub.execute_input":"2023-05-11T23:23:35.517146Z","iopub.status.idle":"2023-05-11T23:23:35.542656Z","shell.execute_reply.started":"2023-05-11T23:23:35.51712Z","shell.execute_reply":"2023-05-11T23:23:35.541085Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Combine bands into a false color image.
\n‣ In order to view contrails in GOES, we use the \"ash\" color scheme. This color scheme was originally developed for viewing volcanic ash in the atmosphere but is also useful for viewing thin cirrus, including contrails. In this color scheme, contrails appear in the image as dark blue.
\n\nNote that the Kaggle Team used a modified version of the ash color scheme here, developed by Kulik et al., which uses slightly different bands and bounds tuned for contrails.
\n\nReferences
\n\n- Original Ash RGB description
\n\n- Modified Ash Color Scheme (Kulik et al., page 22)
\n\n‣ Let's look at our first randomly chosen record 18467979031221222. We can easily load it and plot it with the hidden custom functions above.
","metadata":{}},{"cell_type":"code","source":"record_id = '18467979031221222'\n\n# Loads the data, creates img and various masks.\nloaded_files, false_color = read_record(record_id)\nimg = false_color[..., N_TIMES_BEFORE]\nhuman_pixel_masks = loaded_files[3]\nhuman_individual_mask = loaded_files[4]\nviz_record(img, human_pixel_masks)","metadata":{"execution":{"iopub.status.busy":"2023-05-11T23:23:38.635266Z","iopub.execute_input":"2023-05-11T23:23:38.635677Z","iopub.status.idle":"2023-05-11T23:23:40.104379Z","shell.execute_reply.started":"2023-05-11T23:23:38.635646Z","shell.execute_reply":"2023-05-11T23:23:40.103432Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"**Observation**:\n* This example shows really faint traces of contrails (at least to the naked eye).\n* The ground truth contrail mask is fair.\n\nLet's look at the `human_individual_masks:","metadata":{}},{"cell_type":"code","source":"viz_human_individual_masks(human_individual_mask)","metadata":{"scrolled":true,"execution":{"iopub.status.busy":"2023-05-11T23:23:42.803658Z","iopub.execute_input":"2023-05-11T23:23:42.804972Z","iopub.status.idle":"2023-05-11T23:23:43.611364Z","shell.execute_reply.started":"2023-05-11T23:23:42.804903Z","shell.execute_reply":"2023-05-11T23:23:43.610278Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"**Note**:\n* It can be clearly seen that some human labelers struggled with the task.\n\nLet's animate:","metadata":{}},{"cell_type":"code","source":"anim = run_animation(false_color)\ndisp.HTML(anim.to_jshtml())","metadata":{"execution":{"iopub.status.busy":"2023-05-11T23:23:45.058667Z","iopub.execute_input":"2023-05-11T23:23:45.059277Z","iopub.status.idle":"2023-05-11T23:23:49.006131Z","shell.execute_reply.started":"2023-05-11T23:23:45.059244Z","shell.execute_reply":"2023-05-11T23:23:49.004627Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"‣ Now Let's examine the records which draw our attention during the metadata exploration.
\n\nhuman_individual_masks files got the equals size as bank_{} files: Acknowledgement
","metadata":{}},{"cell_type":"markdown","source":"@jcaliz for .css and plotting ideas.","metadata":{}},{"cell_type":"markdown","source":"##Outro and future work
\n\nThe work is still in progress. I hope to continue working on it and add more to EDA, training and inference part. Good luck in the competition!","metadata":{}}]}