\n",
+ "
"
+ ],
+ "text/plain": [
+ "<xarray.DataArray (y: 256, x: 256)>\n", + "array([[5, 5, 5, ..., 5, 5, 5],\n", + " [5, 5, 5, ..., 5, 5, 5],\n", + " [5, 5, 5, ..., 5, 5, 5],\n", + " ...,\n", + " [7, 5, 5, ..., 5, 5, 7],\n", + " [5, 5, 5, ..., 5, 5, 5],\n", + " [5, 5, 5, ..., 7, 7, 7]], dtype=uint16)\n", + "Coordinates:\n", + " band int64 1\n", + " * x (x) float64 6.026e+05 6.026e+05 ... 6.052e+05 6.052e+05\n", + " * y (y) float64 9.239e+06 9.239e+06 ... 9.237e+06 9.237e+06\n", + " spatial_ref int64 0\n", + "Attributes:\n", + " _FillValue: 0.0\n", + " scale_factor: 1.0\n", + " add_offset: 0.0
\n",
+ "\n",
+ "This tutorial delves into building a scalable model on the LandCoverNet dataset.\n",
+ "\n",
+ "This portion of the tutorial is focused on developing a semantic segmentation model for LandCoverNet data\n",
+ "Here:\n",
+ "\n",
+ "1. We will inspect the source imagery for the labels we have\n",
+ "\n",
+ "2. We will process the source imagery in parallel using Dask\n",
+ "\n",
+ "3. We will select the labels and filtered source images from Dask to be loaded \n",
+ "\n",
+ "4. We will save the images and associated labels data as a `pickle` file ('.pkl') on our directory to be loaded for model training\n",
+ "\n",
+ "The esip-summer-2021-geospatial-ml tutorial was helpful in creating this notebook, which can be found [here](https://github.com/TomAugspurger/esip-summer-2021-geospatial-ml/blob/main/segmentation-model.ipynb). It was particularly useful for loading the STAC items and Sentinel-2 scenes using the `stackstac` library."
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "5de7e8ca-7bc6-42b7-9be8-0c00fb1a318c",
+ "metadata": {},
+ "source": [
+ "### Authentication\n",
+ "\n",
+ "As demonstrated in the [Data Exploration notebook](/1.%20Data%20Exploration.ipynb\") of this tutorial series, access to the Radiant MLHub API using the `pystac_client` library requires both an API end-point and API key. This notebook assumes that you have already followed the steps in `1. Data Exploration.ipynb` and also already have an MLHub API key that is not expired."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "4153b89b-440d-4d9d-a209-e5f37b0a5648",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "import getpass\n",
+ "\n",
+ "MLHUB_API_KEY = getpass.getpass(prompt=\"MLHub API Key: \")\n",
+ "MLHUB_ROOT_URL = \"https://api.radiant.earth/mlhub/v1\""
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "98927dce-f400-4161-8930-fb72cdec4b30",
+ "metadata": {},
+ "source": [
+ "There are a number of STAC and geospatial related libraries used in this notebook that need to be imported."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "03050794-0493-4981-bed3-7a4100120db9",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "import os\n",
+ "import pystac\n",
+ "import warnings\n",
+ "import pystac_client\n",
+ "from shapely.geometry import mapping, shape\n",
+ "import rioxarray\n",
+ "from pystac import Item\n",
+ "from typing import List, Tuple\n",
+ "\n",
+ "warnings.simplefilter(\"ignore\", UserWarning) # ignore warnings that get printed out\n",
+ "from datetime import datetime\n",
+ "\n",
+ "import stackstac\n",
+ "import rasterio as rio\n",
+ "import rasterio.plot\n",
+ "\n",
+ "warnings.simplefilter(action=\"ignore\", category=FutureWarning)\n",
+ "import numpy as np\n",
+ "\n",
+ "import pandas as pd\n",
+ "import pickle\n",
+ "import dask\n",
+ "import dask_gateway\n",
+ "from pystac.item_collection import ItemCollection\n",
+ "\n",
+ "warnings.filterwarnings(\"ignore\", \"Creating an ndarray from ragged\")"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "28f7c5db-cc3a-4237-89e9-10e0972e7de8",
+ "metadata": {},
+ "source": [
+ "### Launch a Dask gateway cluster for parallel processing"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "34b960e1-0660-452f-8a8e-2282438c9ad1",
+ "metadata": {},
+ "source": [
+ "We will use Dask to optimize our data processing of thousands of source image chips by parallelizing the workflow with a delayed computation graph. The Dask Client schedules, runs the delayed computations, and gathers the results, while the Dask Gateway provides a secure and centralized way of managing the multiple client clusters. This is especially useful for running Dask on Planetary Computer."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "6680af86-7399-460b-944f-57b6f51e62c2",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "from dask.distributed import Client\n",
+ "\n",
+ "client = Client()\n",
+ "client.run(lambda: warnings.filterwarnings(\"ignore\", \"Creating an ndarray from ragged\"))"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "1007756f-3df8-461d-a63a-d179eaaef39f",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "gateway = dask_gateway.Gateway()\n",
+ "options = gateway.cluster_options()\n",
+ "options[\"worker_cores\"] = 7"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "85490b75-fb5c-4789-a880-e05d6c5517b5",
+ "metadata": {},
+ "source": [
+ "### Instantiate an instance of the MLHub API Client"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "97ad7e45-bcf9-428d-b1a0-03efdeec7f37",
+ "metadata": {},
+ "source": [
+ "Here again we demonstrate how to instantiate an API client connected to the MLHub end-point using the `pystac_client` library."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "2703906b-fbfb-4a9c-86f0-78273b9ebb31",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "mlhub_client = pystac_client.Client.open(\n",
+ " url=MLHUB_ROOT_URL, parameters={\"key\": MLHUB_API_KEY}, ignore_conformance=True\n",
+ ")"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "531848a9-6c38-44ea-bbd6-a61aec020df5",
+ "metadata": {},
+ "source": [
+ "We set the temporary directory based on the current working directory."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "a64905eb-9dbf-4d77-8010-7d4eb5a950fa",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "tmp_dir = os.path.join(os.getcwd(), \"landcovernet\")\n",
+ "labels_dir = os.path.join(tmp_dir, \"labels\")"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "1a072c3a-69fb-4e69-bd40-f813a307d480",
+ "metadata": {},
+ "source": [
+ "We need to make sure that the labels collection has already been downloaded to the Planetary Computer instance we are running, stored in the shared directory `/home/jovyan/PlanetaryComputerExamples/landcovnet`. Please double-check the active working directory to make sure that the catalog is found when you run this notebook."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "35a62dc5-7197-4842-ae8f-bc165169a635",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "# check for data in collection file\n",
+ "catalog = pystac.read_file(\n",
+ " os.path.join(labels_dir, \"ref_landcovernet_v1_labels/collection.json\")\n",
+ ")"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "15fcd2d5-c7a7-4ea6-b3c5-30964152540e",
+ "metadata": {},
+ "source": [
+ "### Loading the source imagery\n",
+ "\n"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "5a399a24-0d45-4049-bb41-e03efa3b0527",
+ "metadata": {},
+ "source": [
+ "In order to fetch the source images from the source Item Assets, first we need to gather all of the label Items from the LandCoverNet labels collection we downloaded in the previous tutorial. This grabs all of the label Item STAC objects. "
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "b0c6e5cd-e912-4d92-a107-9bb24fa71883",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "links = catalog.get_item_links() # links from the catalog\n",
+ "label_items = [link.resolve_stac_object().target for link in links]"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "f50cd6ce-9c9e-49cd-8900-2fe583bdbbbd",
+ "metadata": {},
+ "source": [
+ "This is a helper function to calculate the percent of a raster image that is covered with clouds. It is assumed that the image input dimensions are 256 by 256 pixels. The sum of cloud cover across the image is normalized and divided by the total area of the chip. This returns an integer value of cloud cover between 0 and 100 to be passed to the STAC Item metadata. ***NOTE: This function is only called if the Item metadata does not include the `eo:cloud_cover` property.***"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "d8e3835b-5c44-489a-b4a2-da953eb031ac",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "def calculate_cloud_cover(img_arr: np.ndarray) -> int:\n",
+ "\n",
+ " \"\"\"Takes a chip cloud cover band and returns the integer score\n",
+ " by dividing the sum of normalized values by the chip area (HxW).\n",
+ "\n",
+ " Args:\n",
+ " img_arr: np.ndarray - 2d array of cloud cover mask\n",
+ "\n",
+ " Returns:\n",
+ " arr_cc: int - integer value of cloud cover score\n",
+ "\n",
+ " \"\"\"\n",
+ " CHIP_AREA = 256 * 256\n",
+ " arr_filled = np.nan_to_num(img_arr)\n",
+ " arr_norm = arr_filled / 100\n",
+ " arr_sum = arr_norm.sum()\n",
+ " arr_cc = arr_sum / CHIP_AREA * 100\n",
+ " return int(arr_cc)"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "9d3c443c-1a4b-4aa7-97aa-ab14bf168b52",
+ "metadata": {},
+ "source": [
+ "For our use-case, we decided not to train the model on the entire source LandCoverNet dataset. Instead we take chip samples that were representative of each season or quadrimester, or any custom number of bins spread over a temporal range. \n",
+ "\n",
+ "This is a helper function that returns the median date from a set of all dates in a range representing each source Item linked to a label Item."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "aa8adef1-2779-407c-8714-41f49f5391cb",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "def get_median_date(id_arr: np.ndarray) -> int:\n",
+ "\n",
+ " \"\"\"Takes a 2d array of source Item IDs for a quarter, and returns median date\n",
+ "\n",
+ " Args: id_arr: np.ndarray - 2d array of string values for source Item IDs\n",
+ "\n",
+ " Returns:\n",
+ " median_date: int - the calculated median date value for input array\n",
+ "\n",
+ " \"\"\"\n",
+ "\n",
+ " dates = [int(s[-8:]) for s in id_arr]\n",
+ " dates.sort()\n",
+ "\n",
+ " n = len(dates)\n",
+ "\n",
+ " # case in which multiple items returned\n",
+ " if n > 1:\n",
+ " if n % 2 == 0:\n",
+ " mid = int(n / 2)\n",
+ " else:\n",
+ " mid = int((n + 1) / 2)\n",
+ " median_date = dates[mid]\n",
+ " # base case there is only one source item\n",
+ " elif n == 1:\n",
+ " median_date = dates[0]\n",
+ " # base case there are no source items\n",
+ " else:\n",
+ " median_date = 0\n",
+ "\n",
+ " return median_date"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "2d6b5a0f-4b21-427a-9dc6-f1ddc7099d22",
+ "metadata": {},
+ "source": [
+ "This helper function assigns an integer value from a datetime value based on the `period_ranges` variable created by another function below called `get_date_ranges()`. The period value assigned is used later to group and rank source Items on their cloud cover value so that within each temporal period, we are only working with the images with minimal cloud cover."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "9130240e-bbeb-4c89-a322-bbd26815f6e6",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "def assign_temporal_period(dt: datetime) -> int:\n",
+ " \"\"\"Takes a datetime and returns an integer based on n_periods defined\"\"\"\n",
+ " for ix, pair in enumerate(period_ranges):\n",
+ " if dt >= pair[0] and dt <= pair[1]:\n",
+ " return ix + 1\n",
+ "\n",
+ " return None"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "48f76591-6b05-4427-8431-09ff053688a1",
+ "metadata": {},
+ "source": [
+ "This takes the DataFrame created from Item metadata in `get_season_min_cloud_cover()`, ranks the dates for each period by cloud cover value, and returns a single source chip for each datetime periods the Items are split into."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "81adfdfb-c777-42c1-a4ab-02b5f540dc2a",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "def filter_period_items(cc_df: pd.DataFrame) -> pd.DataFrame:\n",
+ "\n",
+ " \"\"\"Takes a dataframe of source Items with metadata and filters\n",
+ " on ranked cloudcover by period/season (quadrimester).\n",
+ "\n",
+ " Args:\n",
+ " cc_df: pd.DataFrame - unfiltered dataframe\n",
+ "\n",
+ " Returns:\n",
+ " filtered_df: pd.Dataframe - filtered dataframe\n",
+ "\n",
+ " \"\"\"\n",
+ " pd.options.mode.chained_assignment = None\n",
+ " # assigns quarter and rank by quarter\n",
+ " cc_df[\"date_time\"] = pd.to_datetime(cc_df[\"date_time\"])\n",
+ " cc_df[\"period\"] = cc_df[\"date_time\"].apply(assign_temporal_period)\n",
+ " cc_df[\"rank\"] = cc_df.groupby(\"period\")[\"cloud_cover\"].rank(\n",
+ " method=\"min\", ascending=True\n",
+ " )\n",
+ "\n",
+ " id_prefix = cc_df.iloc[0][\"id\"][:-8]\n",
+ " median_dates = []\n",
+ "\n",
+ " # filters DataFrame on rank\n",
+ " min_cc_df = cc_df[cc_df[\"rank\"] == 1]\n",
+ "\n",
+ " # for each quadrimester in year, get the median date of source items\n",
+ " for i in range(1, n_periods + 1):\n",
+ " quarter_df = min_cc_df[min_cc_df[\"period\"] == i]\n",
+ " quarter_median_date = get_median_date(quarter_df[\"id\"].values)\n",
+ " quarter_median_id = id_prefix + str(quarter_median_date)\n",
+ " median_dates.append(quarter_median_id)\n",
+ "\n",
+ " # filter the ranked DataFrame by median date\n",
+ " filtered_df = min_cc_df[min_cc_df[\"id\"].isin(median_dates)]\n",
+ " return filtered_df"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "ac17c5f0-91cf-45b2-b0a6-2e3fc26bd5ee",
+ "metadata": {},
+ "source": [
+ "This is a wrapper function that creates a DataFrame from a list of Items, and calls the nested filtering functions defined above."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "225d596e-d9af-4c5f-a507-5aff263f293b",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "def get_season_min_cloud_cover(item_list: List[Item]) -> ItemCollection:\n",
+ "\n",
+ " \"\"\"Takes a list of source Items and returns a single chip per season\n",
+ " ranked by the minimum cloud cover from eo:cloud_cover property\n",
+ "\n",
+ " Args:\n",
+ " item_list: List[Item] - iterable of source Items returned from search\n",
+ "\n",
+ " Returns:\n",
+ " ItemCollection - STAC Iterable containing Items filtered by cloud cover\n",
+ " \"\"\"\n",
+ "\n",
+ " # constructs a DataFrame of each source item properties\n",
+ " df_list = []\n",
+ " for ui in item_list:\n",
+ " if \"eo:cloud_cover\" in ui.properties:\n",
+ " cloud_cover = ui.properties[\"eo:cloud_cover\"]\n",
+ " else:\n",
+ " cloud_cover = calculate_cloud_cover(\n",
+ " rio.open(ui.get_assets()[\"CLD\"].href).read()\n",
+ " )\n",
+ " uid = {\n",
+ " \"item\": ui,\n",
+ " \"id\": ui.id,\n",
+ " \"cloud_cover\": cloud_cover,\n",
+ " \"date_time\": ui.datetime,\n",
+ " }\n",
+ " df_list.append(uid)\n",
+ "\n",
+ " cc_df = pd.DataFrame(df_list)\n",
+ "\n",
+ " # filters source items by cloud cover rank and returns ItemCollection\n",
+ " if not cc_df.empty:\n",
+ " filtered_df = filter_period_items(cc_df)\n",
+ "\n",
+ " return ItemCollection(filtered_df[\"item\"].tolist())\n",
+ "\n",
+ " return None"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "3f7e5351-f592-45d7-91cd-d3ccda073626",
+ "metadata": {},
+ "source": [
+ "This will take the temporal and spatial extent of an Item to query MLHub API client for matching source Items."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "0b3ad2c5-0ea6-47a0-afe0-6e305a68720f",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "def get_label_item_collection(label_item: Item) -> ItemCollection:\n",
+ "\n",
+ " \"\"\"Takes a label Item from the LandCoverNet Collection and searches\n",
+ " for source imagery for chips that match spatial and temporal criteria\n",
+ "\n",
+ " Args:\n",
+ " label_item: Item - item of current iteration in the get_item() Dask parallelization\n",
+ "\n",
+ " Returns:\n",
+ " ItemCollection - STAC Iterable containing Items that match search criteria\n",
+ " \"\"\"\n",
+ "\n",
+ " n = 0\n",
+ " cc_thresh = 10\n",
+ " year_collection = ItemCollection([])\n",
+ "\n",
+ " # iterate over each start and end date per quarter\n",
+ " for start, end in period_ranges:\n",
+ "\n",
+ " while n == 0:\n",
+ "\n",
+ " # performs a temporal and spatial search for each label item\n",
+ " search = mlhub_client.search(\n",
+ " collections=[\"ref_landcovernet_v1_source\"],\n",
+ " intersects=mapping(shape(label_item.geometry)),\n",
+ " datetime=[start, end],\n",
+ " query={\"eo:cloud_cover\": {\"lt\": cc_thresh}},\n",
+ " )\n",
+ "\n",
+ " # converts search results to ItemCollection\n",
+ " item_results = search.get_all_items()\n",
+ "\n",
+ " if not item_results:\n",
+ " cc_thresh += 5\n",
+ " else:\n",
+ " n = len(item_results)\n",
+ "\n",
+ " year_collection += item_results # concatenate ItemCollections for each quarter\n",
+ " n = 0 # reset the length criteria for search results\n",
+ "\n",
+ " filtered_items = get_season_min_cloud_cover(year_collection.items)\n",
+ "\n",
+ " return filtered_items"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "56a27886-df90-4143-b966-834b2295c702",
+ "metadata": {},
+ "source": [
+ "This is the primary function that drives all the processing required to filter and load source imagery and label data into a stack of Xarray DataArrays for further processing, e.g. splitting the dataset into training and validation sets prior to training a machine learning model."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "102fbbe5-4ed8-4c43-b7b9-2a29bac7c386",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "def get_item(label_item: Item, assets: Tuple[str]) -> (np.ndarray, np.ndarray):\n",
+ "\n",
+ " \"\"\"Takes label Item and asset bands to construct n-darrays for model training\n",
+ "\n",
+ " Args:\n",
+ " label_item: Item - item of current iteration in the get_item() Dask parallelization\n",
+ " assets: Tuple[str] - a set of strings corresponding to the Asset band names\n",
+ "\n",
+ " Returns:\n",
+ " data: np.ndarray, labels: np.ndarray - X and y n-darrays for model training\n",
+ " \"\"\"\n",
+ " warnings.simplefilter(action=\"ignore\", category=FutureWarning)\n",
+ " assets = list(assets)\n",
+ " labels = rioxarray.open_rasterio(\n",
+ " tmp_dir + \"/labels/ref_landcovernet_v1_labels/\" + label_item.id + \"/labels.tif\",\n",
+ " ).squeeze()\n",
+ "\n",
+ " source_item_collection = get_label_item_collection(label_item)\n",
+ "\n",
+ " if len(source_item_collection) > 0:\n",
+ "\n",
+ " bounds = tuple(round(x, 0) for x in labels.rio.bounds())\n",
+ "\n",
+ " data = stackstac.stack(\n",
+ " items=source_item_collection,\n",
+ " assets=assets,\n",
+ " dtype=\"float32\",\n",
+ " resolution=10,\n",
+ " bounds=bounds,\n",
+ " epsg=labels.rio.crs.to_epsg(),\n",
+ " )\n",
+ "\n",
+ " data = data.assign_coords(x=labels.x.data, y=labels.y.data)\n",
+ " data /= 4000\n",
+ " data = np.clip(data, 0, 1)\n",
+ "\n",
+ " return data, labels.astype(\"int64\")"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "f4a759a7-14f7-4331-a9ea-a7768a986edf",
+ "metadata": {},
+ "source": [
+ "This takes in the temporal range of the Collection as well as a global variable `n_periods` defined below to return a list of datetime ranges split up into equal sized buckets based on the designated number of periods. E.g. `n_periods=3` will return quadrimesters."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "1173b773-3ce5-4de1-966b-b23243aa6d4c",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "def get_date_ranges(\n",
+ " start: datetime, end: datetime, periods: int\n",
+ ") -> List[List[datetime]]:\n",
+ "\n",
+ " \"\"\"Builds a list of start and end date ranges for every four in the year\n",
+ "\n",
+ " Args: None\n",
+ " Returns:\n",
+ " quarter_ranges: List[List[datetime]] - a list of pairs of strings representing\n",
+ " the start and end dates.\n",
+ "\n",
+ " \"\"\"\n",
+ "\n",
+ " period_ranges = []\n",
+ " all_dates = pd.DataFrame(\n",
+ " pd.date_range(start=temporal_start, end=temporal_end, freq=\"1D\"),\n",
+ " columns=[\"Date\"],\n",
+ " )\n",
+ " date_bins = pd.cut(all_dates.Date, bins=periods).drop_duplicates()\n",
+ "\n",
+ " for interval in date_bins:\n",
+ " period_ranges.append(\n",
+ " [\n",
+ " interval.left.tz_localize(\"UTC\").to_pydatetime(),\n",
+ " interval.right.tz_localize(\"UTC\").to_pydatetime(),\n",
+ " ]\n",
+ " )\n",
+ "\n",
+ " return period_ranges"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "4a4da8ac-371e-4a11-a5a7-845ca96dee11",
+ "metadata": {},
+ "source": [
+ "Here we specify the temporal extent of the Catalog, n periods to divide the temporal range into, and bands to fetch for each source Item."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "fdda7a62-cd3a-41e4-ab52-d0ba985b4d8e",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "temporal_start = catalog.extent.temporal.intervals[0][0].strftime(\n",
+ " \"%Y-%m-%d\"\n",
+ ") # global starting datetime for label Collection\n",
+ "temporal_end = catalog.extent.temporal.intervals[0][1].strftime(\n",
+ " \"%Y-%m-%d\"\n",
+ ") # global ending datetime for label Collection\n",
+ "n_periods = 5\n",
+ "\n",
+ "period_ranges = get_date_ranges(temporal_start, temporal_end, n_periods)\n",
+ "assets = (\"B04\", \"B03\", \"B02\") # we will make use of the RGB bands"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "12727fc1-67ed-4ee3-a9ff-87ce643576c3",
+ "metadata": {},
+ "source": [
+ "### Load the source imagery"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "ec271c06-48be-4dd4-b7e0-98a17cbcddf1",
+ "metadata": {},
+ "source": [
+ "Now we will bring everything together. We setup the `get_item()` function defined above as a Dask delayed function, and append the lazy results from fetching N source images (`n_periods` variable above) for each label Item in the Catalog. Then the actual computation occurs in parallel, and the results are appended to a list of DataArrays containing the aligned images and labels (X and y features)."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "9c255ac3-223e-4fc5-a480-1a836865fb72",
+ "metadata": {
+ "tags": []
+ },
+ "outputs": [],
+ "source": [
+ "%%time\n",
+ "\n",
+ "Xys_list = []\n",
+ "import logging\n",
+ "\n",
+ "logger = logging.getLogger(\"distributed.utils_perf\")\n",
+ "logger.setLevel(logging.ERROR)\n",
+ "\n",
+ "chunk_size = 20\n",
+ "for i in range(0, len(label_items), chunk_size):\n",
+ " label_chunk = label_items[i : i + chunk_size]\n",
+ "\n",
+ " Xys = []\n",
+ " get_item_ = dask.delayed(get_item, nout=5)\n",
+ "\n",
+ " Xys.append([get_item_(label, assets) for label in label_chunk])\n",
+ " Xys = dask.persist(*Xys)\n",
+ " Xys = dask.compute(*Xys)\n",
+ " Xys_list.append(Xys[0])"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "144d6a14-fb78-4f4e-a98e-ee3077a26bb2",
+ "metadata": {},
+ "source": [
+ "The Dask client can be shutdown with the following command"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "5cb07ea9-8e36-43fd-92d4-0d072aab1747",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "client.shutdown()"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "cdbe1876-dc80-4cdc-85a1-cf06ba2ca576",
+ "metadata": {},
+ "source": [
+ "We stacked the results of our parallelized function into chunks of 20 Items at a time, so this will flatten the list."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "ca0d6014-40e8-4441-9b80-9905d3dd5f62",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "flat_list = [item for sublist in Xys_list for item in sublist]"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "e88b2c15-3536-4d7f-9c1f-59809dc1504e",
+ "metadata": {},
+ "source": [
+ "This confirms that every item was extracted, e.g. the flattened list has the length of label Items fetched."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "24f19675-5f2e-4bbd-89ae-0c565ba92a75",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "len(flat_list) # confirm every item was extracted"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "28482cc9-c091-4286-8aa0-53b0907a1ed1",
+ "metadata": {},
+ "source": [
+ "This confirms that the shape of data for each label item has n source items."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "542ce4a4-80ee-4262-8e0f-2fd1afe42047",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "flat_list[0][0].shape # confirm that we have the desired shape for a chip"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "a1bccfdf-8553-4a34-a6c0-252df4dff865",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "%rm -rf labels #clear up labels to clear PC space. may choose to leave the labels"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "6d433db3-a018-48b6-8a4e-022b7a47f979",
+ "metadata": {},
+ "source": [
+ "The last step before training a neural network is to dump the image dataset we just created into a pickle file stored locally on the Planetary Computer instance running. This is an efficient way to store and load the dataset in the next notebook and to conserve memory resources."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "06e349c3-3ece-4fd6-b800-5ac360dd874f",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "pickle.dump((flat_list), open(f\"{tmp_dir}/items\" + \".pkl\", \"ab\"))"
+ ]
+ }
+ ],
+ "metadata": {
+ "kernelspec": {
+ "display_name": "Python 3 (ipykernel)",
+ "language": "python",
+ "name": "python3"
+ },
+ "language_info": {
+ "codemirror_mode": {
+ "name": "ipython",
+ "version": 3
+ },
+ "file_extension": ".py",
+ "mimetype": "text/x-python",
+ "name": "python",
+ "nbconvert_exporter": "python",
+ "pygments_lexer": "ipython3",
+ "version": "3.8.12"
+ }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}
diff --git a/tutorials/model_training/03-model-training.ipynb b/tutorials/model_training/03-model-training.ipynb
new file mode 100644
index 00000000..5c14f30f
--- /dev/null
+++ b/tutorials/model_training/03-model-training.ipynb
@@ -0,0 +1,1456 @@
+{
+ "cells": [
+ {
+ "cell_type": "markdown",
+ "id": "815cb424-89e3-45ed-b4d6-5389391efa99",
+ "metadata": {},
+ "source": [
+ "## LandCoverNet Model Training\n",
+ "\n",
+ "\n",
+ "\n",
+ "This portion of the tutorial is focused on developing a semantic segmentation model for LandCoverNet data.\n",
+ "\n", + "Here:\n", + "\n", + "1. We will load the prepared data as training, test and validation data\n", + "\n", + "2. We will build a segmentation model, train it on the training data, inspect the results on the validation data and make predictions on the test data" + ] + }, + { + "cell_type": "markdown", + "id": "da3ffed5-7a26-4f7f-9a19-06af0f3ed349", + "metadata": {}, + "source": [ + "### Importing the needed libraries\n", + "\n", + "We will use TensorFlow to build the model. Hence, we will import a number of TensorFlow modules to be used, in addition to our other data related libraries.\n", + "\n", + "For faster model training, it is advisable to use Planetary Computer's TensorFlow GPU server to run the notebook." + ] + }, + { + "cell_type": "code", + "execution_count": 1, + "id": "36d9a2cc-d448-4383-b6c1-80f06cc9f00e", + "metadata": {}, + "outputs": [], + "source": [ + "import os\n", + "import numpy as np\n", + "\n", + "import warnings\n", + "\n", + "from tensorflow.keras.layers import (\n", + " Conv2D,\n", + " Conv2DTranspose,\n", + " Dropout,\n", + " MaxPooling2D,\n", + " concatenate,\n", + ")\n", + "from tensorflow.keras.callbacks import (\n", + " ModelCheckpoint,\n", + " EarlyStopping,\n", + ")\n", + "from sklearn.model_selection import train_test_split\n", + "import tensorflow as tf\n", + "import matplotlib.pyplot as plt\n", + "import pickle\n", + "import random\n", + "\n", + "warnings.filterwarnings(\"ignore\", \"Creating an ndarray from ragged\")" + ] + }, + { + "cell_type": "markdown", + "id": "845be0db-9d65-4b40-b5ec-e1dcde2d171a", + "metadata": {}, + "source": [ + "### Data Preprocessing\n", + "\n", + "The next few steps will address:\n", + "\n", + "1. Loading and preprocessing the prepared data consisting of the source images and labels\n", + "\n", + "2. Splitting the preprocessed images and labels into the training, validation and test data sets\n", + "\n", + "3. A sanity check to visualise our preprocessed data sets. This ensures everything from the previous two steps adds up." + ] + }, + { + "cell_type": "markdown", + "id": "e0c7bb1f-d48f-4a6c-8579-67864abafcdd", + "metadata": {}, + "source": [ + "#### Loading and Preprocessing the data\n", + "\n", + "We will load the prepared data defined by the least cloudy RGB source images across all labels taken for every quadrimester throughout the year (every three months).\n", + "\n", + "We set up the temporary directory `tmp_dir` to point to the prepared data and loaded it onto `Xys`" + ] + }, + { + "cell_type": "code", + "execution_count": 2, + "id": "f42d4d29-507b-43ca-a7bf-08aaaae975da", + "metadata": {}, + "outputs": [], + "source": [ + "tmp_dir = os.path.join(os.getcwd(), \"landcovernet\")\n", + "labels_dir = os.path.join(tmp_dir, \"labels\")" + ] + }, + { + "cell_type": "code", + "execution_count": 4, + "id": "df58b4c9-0642-487d-bffc-636a037c5fad", + "metadata": {}, + "outputs": [], + "source": [ + "with open(os.path.join(tmp_dir, \"items.pkl\"), \"rb\") as fp:\n", + " Xys = pickle.load(fp) # load data" + ] + }, + { + "cell_type": "markdown", + "id": "46a54833-14d2-4f01-b21c-d4c845902446", + "metadata": {}, + "source": [ + "Next, we will preprocess and extract the images and labels respectively.\n", + "\n", + "Since we obtained three RGB images from each season in the data preparation notebook, we will combine/concatenate those images into a single numpy array consisting of (3x3)= 9 bands\n", + "\n", + "We define the function `process_xy(Xys)` to take care of this.\n", + "The axis for the images were also swapped, as seen in `temp.swapaxes(i,j)` and the three images concatenated.\n", + "\n", + "This produces the 'X' data for images in the shape `(length of Xys, image height, image width, number of channels)`, which in our case will be `(1980, 256, 256, 9)`.\n", + "\n", + "Likewise, we have the 'Y' data in the shape `(length of Xys, image height, image width)`, which in our case will be `(1980, 256, 256)`" + ] + }, + { + "cell_type": "code", + "execution_count": 5, + "id": "17431331-874a-46e7-a681-ebd0e3a76b1c", + "metadata": {}, + "outputs": [], + "source": [ + "def process_xy(Xys): # process x and y\n", + " \"\"\"\n", + " Ths function will process Xys into the desired X and Y\n", + " variables which we need for model development. It returns\n", + " them as numpy arrays\n", + " \"\"\"\n", + " total = 0\n", + " temp = []\n", + " total_list = []\n", + " for i in range(0, len(Xys)):\n", + " temp.append(Xys[i][0].data)\n", + " for i in range(0, len(Xys)):\n", + " total += len(Xys[i][0])\n", + " # saves the index of the label's first image, then the next label's\n", + " # first image, etc.\n", + " total_list.append(total)\n", + " # print(temp.shape)\n", + " for i in range(0, len(total_list) - 1):\n", + " if (\n", + " total_list[i + 1] != total_list[i] + 3\n", + " ): # the label does not contain three source images as desired\n", + " outlier = total_list[i]\n", + " temp = [it for sub in temp for it in sub]\n", + " temp = np.asarray(temp)\n", + " temp = temp.swapaxes(1, 2) # transform axis to desired shape\n", + " temp = temp.swapaxes(2, 3)\n", + " X_mod = []\n", + " for item in range(0, len(temp)):\n", + " if item == outlier:\n", + " pass\n", + " else:\n", + " X_mod.append(list(temp[item]))\n", + " del temp\n", + " del total_list\n", + " X_mod = np.asarray(X_mod)\n", + " flat_list = []\n", + " for i in range(0, len(X_mod), 3):\n", + " flat_list.append(\n", + " list(np.concatenate(X_mod[i : i + 3], axis=2))\n", + " ) # concatenate the data\n", + " flat_list = np.asarray(flat_list)\n", + " del X_mod\n", + " Y = np.zeros((len(Xys), 256, 256), dtype=np.float32)\n", + " for j in range(\n", + " 0, len(Xys)\n", + " ): # labels for the source images belonging to a label item\n", + " Y[j] = Xys[j][1][0].data\n", + " return flat_list, Y" + ] + }, + { + "cell_type": "code", + "execution_count": 6, + "id": "b9bc9bd3-e74d-46f1-840f-c369e7aaa1d7", + "metadata": {}, + "outputs": [], + "source": [ + "X, Y = process_xy(Xys) # extract x and y" + ] + }, + { + "cell_type": "markdown", + "id": "b145887a-e5ba-436c-83a2-697bc7a7a2ea", + "metadata": {}, + "source": [ + "#### Splitting the data\n", + "\n", + "We will split the data into three non-overlapping data sets; the training, validation and test data sets.\n", + "\n", + "Much of the training will be on the training set, containing 70% of the data, while the validation and test sets will contain 30% of the data (15% each).\n", + "\n", + "We will use Scikit-learn's `train_test_split` for this. We also delete the X and Y variables thereafter to free some memory as they have served their purposes." + ] + }, + { + "cell_type": "code", + "execution_count": 7, + "id": "683e6248-18fb-4190-bca2-5b633ddf1971", + "metadata": {}, + "outputs": [], + "source": [ + "x_train, x_test, y_train, y_test = train_test_split(\n", + " X, Y, test_size=0.3\n", + ") # 70% split for training data\n", + "x_test, x_val, y_test, y_val = train_test_split(\n", + " x_test, y_test, test_size=0.5\n", + ") # 15% each for test and validation data\n", + "del X # clear memory\n", + "del Y" + ] + }, + { + "cell_type": "code", + "execution_count": 8, + "id": "325359b0-4adf-439f-8af7-e5f54272d739", + "metadata": {}, + "outputs": [ + { + "data": { + "text/plain": [ + "0" + ] + }, + "execution_count": 8, + "metadata": {}, + "output_type": "execute_result" + } + ], + "source": [ + "import gc\n", + "\n", + "gc.collect()" + ] + }, + { + "cell_type": "code", + "execution_count": 8, + "id": "d7a5eb69-e9ba-4603-afd1-6e040b14ff4a", + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "(1386, 256, 256, 9)\n", + "(1386, 256, 256)\n", + "(297, 256, 256, 9)\n", + "(297, 256, 256)\n", + "(297, 256, 256, 9)\n", + "(297, 256, 256)\n" + ] + } + ], + "source": [ + "print(x_train.shape)\n", + "print(y_train.shape)\n", + "print(x_val.shape)\n", + "print(y_val.shape)\n", + "print(x_test.shape)\n", + "print(y_test.shape)" + ] + }, + { + "cell_type": "markdown", + "id": "42d1a6ff-6aae-48dd-9fa7-4b1b162e8c5a", + "metadata": {}, + "source": [ + "#### Sanity Check\n", + "\n", + "We will just make a sanity check by visualising a random training image and its label to make sure they fit.\n", + "\n", + "We will choose the second image within the concantenated image to visualise" + ] + }, + { + "cell_type": "code", + "execution_count": 28, + "id": "2286f2c1-549d-45e0-8b69-5e10d6d342de", + "metadata": {}, + "outputs": [ + { + "name": "stderr", + "output_type": "stream", + "text": [ + "Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).\n" + ] + }, + { + "data": { + "text/plain": [ + "
\n", + "This represents the encoder portion of the model.\n", + "
\n", + "Hence, the `Conv2D` layer is used for this. We will choose the default number of filters, `n_filters` to be 32, using the 'relu' activation function.\n", + "\n", + "We will also define the upsampling block for the expanding path. The expanding path combines optimally, the high-level feature information and the information from the skip-connections, which copies a cropped version of the feature maps in the contracting path.\n", + "
\n", + "To implement this, the `Conv2DTranspose` layer was used, with the parameters defined.\n", + "\n", + "For both paths, we chose the `kernel_size`, as 3, which is the size of a convolutional filter." + ] + }, + { + "cell_type": "code", + "execution_count": 10, + "id": "9680b0fc-26cb-4c55-87f7-7ce7c5189d84", + "metadata": {}, + "outputs": [], + "source": [ + "def conv_block(inputs=None, n_filters=32, dropout_prob=0, max_pooling=True):\n", + " # code reference: https://www.kaggle.com/yesa911/aerial-semantic-segmentation-96-acc\n", + " conv = Conv2D(\n", + " n_filters,\n", + " kernel_size=3,\n", + " activation=\"relu\",\n", + " padding=\"same\",\n", + " kernel_initializer=tf.keras.initializers.HeNormal(),\n", + " )(inputs)\n", + " conv = Conv2D(\n", + " n_filters,\n", + " kernel_size=3,\n", + " activation=\"relu\",\n", + " padding=\"same\",\n", + " kernel_initializer=tf.keras.initializers.HeNormal(),\n", + " )(conv)\n", + "\n", + " if dropout_prob > 0:\n", + " conv = Dropout(dropout_prob)(conv)\n", + " if max_pooling:\n", + " next_layer = MaxPooling2D(pool_size=(2, 2))(conv)\n", + "\n", + " else:\n", + " next_layer = conv\n", + "\n", + " skip_connection = conv\n", + "\n", + " return next_layer, skip_connection\n", + "\n", + "\n", + "def upsampling_block(expansive_input, contractive_input, n_filters=32):\n", + "\n", + " up = Conv2DTranspose(n_filters, kernel_size=3, strides=(2, 2), padding=\"same\")(\n", + " expansive_input\n", + " )\n", + "\n", + " merge = concatenate([up, contractive_input], axis=3)\n", + " conv = Conv2D(\n", + " n_filters,\n", + " kernel_size=3,\n", + " activation=\"relu\",\n", + " padding=\"same\",\n", + " kernel_initializer=tf.keras.initializers.HeNormal(),\n", + " )(merge)\n", + "\n", + " conv = Conv2D(\n", + " n_filters,\n", + " kernel_size=3,\n", + " activation=\"relu\",\n", + " padding=\"same\",\n", + " kernel_initializer=tf.keras.initializers.HeNormal(),\n", + " )(conv)\n", + "\n", + " return conv" + ] + }, + { + "cell_type": "code", + "execution_count": 11, + "id": "dd2a2434-de3c-46a1-99fb-bb4a3d0faeb3", + "metadata": {}, + "outputs": [], + "source": [ + "def unet_model(input_size=(256, 256, 3), n_filters=32, n_classes=8):\n", + " # code reference: https://www.kaggle.com/yesa911/aerial-semantic-segmentation-96-acc\n", + " inputs = tf.keras.Input(input_size)\n", + " # inputs = BatchNormalization()(inputs) # we can learn the normalization step\n", + " # inputs = Dropout(0.5)(inputs)\n", + " # contracting path\n", + " cblock1 = conv_block(inputs, n_filters)\n", + "\n", + " cblock2 = conv_block(cblock1[0], 2 * n_filters)\n", + "\n", + " cblock3 = conv_block(cblock2[0], 4 * n_filters)\n", + "\n", + " cblock4 = conv_block(cblock3[0], 8 * n_filters, dropout_prob=0.3)\n", + "\n", + " cblock5 = conv_block(cblock4[0], 16 * n_filters, dropout_prob=0.3, max_pooling=None)\n", + "\n", + " # expanding path\n", + " ublock6 = upsampling_block(cblock5[0], cblock4[1], 8 * n_filters)\n", + " ublock7 = upsampling_block(ublock6, cblock3[1], n_filters * 4)\n", + " ublock8 = upsampling_block(ublock7, cblock2[1], n_filters * 2)\n", + " ublock9 = upsampling_block(ublock8, cblock1[1], n_filters)\n", + "\n", + " conv9 = Conv2D(\n", + " n_filters, 3, activation=\"relu\", padding=\"same\", kernel_initializer=\"he_normal\"\n", + " )(ublock9)\n", + " # conv9 = BatchNormalization()(conv9)\n", + " conv10 = Conv2D(n_classes, kernel_size=1, padding=\"same\")(conv9)\n", + " model = tf.keras.Model(inputs=inputs, outputs=conv10)\n", + "\n", + " return model" + ] + }, + { + "cell_type": "markdown", + "id": "6911c090-6422-420e-83b8-58452de9a8e8", + "metadata": {}, + "source": [ + "We will then build the model based on those defined paths and display the built model.\n", + "\n", + "The image height, image width and number of channels (`img_height`, `img_width`, `num_channels`) as 256, 256, 9 respectively, as defined by the shape of the input image described." + ] + }, + { + "cell_type": "code", + "execution_count": 13, + "id": "0f1c87d9-0446-4f7b-9bba-67721351d3e7", + "metadata": {}, + "outputs": [], + "source": [ + "img_height = 256\n", + "img_width = 256\n", + "num_channels = 9 # modify since we have 9 channels instead\n", + "\n", + "unet = unet_model((img_height, img_width, num_channels))" + ] + }, + { + "cell_type": "code", + "execution_count": 14, + "id": "610477ac-ba64-4760-b257-7b1e78475ba8", + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "Model: \"model_1\"\n", + "__________________________________________________________________________________________________\n", + "Layer (type) Output Shape Param # Connected to \n", + "==================================================================================================\n", + "input_2 (InputLayer) [(None, 256, 256, 9) 0 \n", + "__________________________________________________________________________________________________\n", + "conv2d_20 (Conv2D) (None, 256, 256, 32) 2624 input_2[0][0] \n", + "__________________________________________________________________________________________________\n", + "conv2d_21 (Conv2D) (None, 256, 256, 32) 9248 conv2d_20[0][0] \n", + "__________________________________________________________________________________________________\n", + "max_pooling2d_4 (MaxPooling2D) (None, 128, 128, 32) 0 conv2d_21[0][0] \n", + "__________________________________________________________________________________________________\n", + "conv2d_22 (Conv2D) (None, 128, 128, 64) 18496 max_pooling2d_4[0][0] \n", + "__________________________________________________________________________________________________\n", + "conv2d_23 (Conv2D) (None, 128, 128, 64) 36928 conv2d_22[0][0] \n", + "__________________________________________________________________________________________________\n", + "max_pooling2d_5 (MaxPooling2D) (None, 64, 64, 64) 0 conv2d_23[0][0] \n", + "__________________________________________________________________________________________________\n", + "conv2d_24 (Conv2D) (None, 64, 64, 128) 73856 max_pooling2d_5[0][0] \n", + "__________________________________________________________________________________________________\n", + "conv2d_25 (Conv2D) (None, 64, 64, 128) 147584 conv2d_24[0][0] \n", + "__________________________________________________________________________________________________\n", + "max_pooling2d_6 (MaxPooling2D) (None, 32, 32, 128) 0 conv2d_25[0][0] \n", + "__________________________________________________________________________________________________\n", + "conv2d_26 (Conv2D) (None, 32, 32, 256) 295168 max_pooling2d_6[0][0] \n", + "__________________________________________________________________________________________________\n", + "conv2d_27 (Conv2D) (None, 32, 32, 256) 590080 conv2d_26[0][0] \n", + "__________________________________________________________________________________________________\n", + "dropout_2 (Dropout) (None, 32, 32, 256) 0 conv2d_27[0][0] \n", + "__________________________________________________________________________________________________\n", + "max_pooling2d_7 (MaxPooling2D) (None, 16, 16, 256) 0 dropout_2[0][0] \n", + "__________________________________________________________________________________________________\n", + "conv2d_28 (Conv2D) (None, 16, 16, 512) 1180160 max_pooling2d_7[0][0] \n", + "__________________________________________________________________________________________________\n", + "conv2d_29 (Conv2D) (None, 16, 16, 512) 2359808 conv2d_28[0][0] \n", + "__________________________________________________________________________________________________\n", + "dropout_3 (Dropout) (None, 16, 16, 512) 0 conv2d_29[0][0] \n", + "__________________________________________________________________________________________________\n", + "conv2d_transpose_4 (Conv2DTrans (None, 32, 32, 256) 1179904 dropout_3[0][0] \n", + "__________________________________________________________________________________________________\n", + "concatenate_4 (Concatenate) (None, 32, 32, 512) 0 conv2d_transpose_4[0][0] \n", + " dropout_2[0][0] \n", + "__________________________________________________________________________________________________\n", + "conv2d_30 (Conv2D) (None, 32, 32, 256) 1179904 concatenate_4[0][0] \n", + "__________________________________________________________________________________________________\n", + "conv2d_31 (Conv2D) (None, 32, 32, 256) 590080 conv2d_30[0][0] \n", + "__________________________________________________________________________________________________\n", + "conv2d_transpose_5 (Conv2DTrans (None, 64, 64, 128) 295040 conv2d_31[0][0] \n", + "__________________________________________________________________________________________________\n", + "concatenate_5 (Concatenate) (None, 64, 64, 256) 0 conv2d_transpose_5[0][0] \n", + " conv2d_25[0][0] \n", + "__________________________________________________________________________________________________\n", + "conv2d_32 (Conv2D) (None, 64, 64, 128) 295040 concatenate_5[0][0] \n", + "__________________________________________________________________________________________________\n", + "conv2d_33 (Conv2D) (None, 64, 64, 128) 147584 conv2d_32[0][0] \n", + "__________________________________________________________________________________________________\n", + "conv2d_transpose_6 (Conv2DTrans (None, 128, 128, 64) 73792 conv2d_33[0][0] \n", + "__________________________________________________________________________________________________\n", + "concatenate_6 (Concatenate) (None, 128, 128, 128 0 conv2d_transpose_6[0][0] \n", + " conv2d_23[0][0] \n", + "__________________________________________________________________________________________________\n", + "conv2d_34 (Conv2D) (None, 128, 128, 64) 73792 concatenate_6[0][0] \n", + "__________________________________________________________________________________________________\n", + "conv2d_35 (Conv2D) (None, 128, 128, 64) 36928 conv2d_34[0][0] \n", + "__________________________________________________________________________________________________\n", + "conv2d_transpose_7 (Conv2DTrans (None, 256, 256, 32) 18464 conv2d_35[0][0] \n", + "__________________________________________________________________________________________________\n", + "concatenate_7 (Concatenate) (None, 256, 256, 64) 0 conv2d_transpose_7[0][0] \n", + " conv2d_21[0][0] \n", + "__________________________________________________________________________________________________\n", + "conv2d_36 (Conv2D) (None, 256, 256, 32) 18464 concatenate_7[0][0] \n", + "__________________________________________________________________________________________________\n", + "conv2d_37 (Conv2D) (None, 256, 256, 32) 9248 conv2d_36[0][0] \n", + "__________________________________________________________________________________________________\n", + "conv2d_38 (Conv2D) (None, 256, 256, 32) 9248 conv2d_37[0][0] \n", + "__________________________________________________________________________________________________\n", + "conv2d_39 (Conv2D) (None, 256, 256, 8) 264 conv2d_38[0][0] \n", + "==================================================================================================\n", + "Total params: 8,641,704\n", + "Trainable params: 8,641,704\n", + "Non-trainable params: 0\n", + "__________________________________________________________________________________________________\n" + ] + } + ], + "source": [ + "unet.summary()" + ] + }, + { + "cell_type": "markdown", + "id": "9c2ec848-f37b-40c8-affb-6b09dd0aca4e", + "metadata": {}, + "source": [ + "We will define a `model_checkpoint` which will be used to save the model as the validation loss declines.\n", + "\n", + "`model_earlyStopping` also terminates the training process if the validation loss doesn't increase after a certain number of epochs represented by the `patience`. In this case, we're using 50 epochs" + ] + }, + { + "cell_type": "code", + "execution_count": 19, + "id": "9802e03d-6b2a-4d23-bc93-7406d8346e74", + "metadata": {}, + "outputs": [], + "source": [ + "# save the model based on best validation loss\n", + "checkpoint_path = os.path.join(tmp_dir, \"unet_model.hdf5\")\n", + "model_checkpoint = ModelCheckpoint(\n", + " checkpoint_path, monitor=\"val_loss\", verbose=1, save_best_only=True\n", + ")\n", + "# stop training if no improvements after 50 epochs\n", + "model_earlyStopping = EarlyStopping(min_delta=0.001, patience=50)" + ] + }, + { + "cell_type": "markdown", + "id": "33b0d70f-58cf-410a-9d8d-21b4a9cc056e", + "metadata": {}, + "source": [ + "Next, we compile the model. We will use the Adam optimizer, with a learning rate of 3e-4.\n", + "The loss defined is the sparse categorical cross-entropy loss (since we have the label dimension of 256x256) and we chose the performance to be judged by the model's accuracy." + ] + }, + { + "cell_type": "code", + "execution_count": 22, + "id": "006987da-64d3-4f49-9b75-2af9fe2e5315", + "metadata": {}, + "outputs": [], + "source": [ + "# setting up and compiling the model\n", + "optimizer = tf.keras.optimizers.Adam(learning_rate=0.0003, name=\"Adam\")\n", + "loss = loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)\n", + "unet.compile(optimizer=optimizer, loss=loss, metrics=[\"accuracy\"])" + ] + }, + { + "cell_type": "markdown", + "id": "708d34e5-3bab-4998-b3b5-3aa746e5b548", + "metadata": {}, + "source": [ + "### Training the Model\n", + "\n", + "We will train the model on the data sets. We will choose a batch size of 16. The steps per epoch was calculated as the `(length of the training data)//batch size`.\n", + "\n", + "The validation steps as `(length of the validation data)//batch size`.\n", + "\n", + "The `model_checkpoint` and `model_earlyStopping` were also defined into the `callbacks`." + ] + }, + { + "cell_type": "code", + "execution_count": 23, + "id": "2e81f5fb-9723-4905-9d47-250c9f93e8e9", + "metadata": {}, + "outputs": [ + { + "name": "stderr", + "output_type": "stream", + "text": [ + "2022-02-28 23:48:55.446867: W tensorflow/core/framework/cpu_allocator_impl.cc:80] Allocation of 3269984256 exceeds 10% of free system memory.\n", + "2022-02-28 23:48:55.943699: I tensorflow/compiler/mlir/mlir_graph_optimization_pass.cc:116] None of the MLIR optimization passes are enabled (registered 2)\n", + "2022-02-28 23:48:55.965384: I tensorflow/core/platform/profile_utils/cpu_utils.cc:112] CPU Frequency: 2445410000 Hz\n" + ] + }, + { + "name": "stdout", + "output_type": "stream", + "text": [ + "Epoch 1/100\n" + ] + }, + { + "name": "stderr", + "output_type": "stream", + "text": [ + "2022-02-28 23:48:56.933316: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcudnn.so.7\n", + "2022-02-28 23:48:59.621233: W tensorflow/stream_executor/gpu/asm_compiler.cc:63] Running ptxas --version returned 256\n", + "2022-02-28 23:48:59.772307: W tensorflow/stream_executor/gpu/redzone_allocator.cc:314] Internal: ptxas exited with non-zero error code 256, output: \n", + "Relying on driver to perform ptx compilation. \n", + "Modify $PATH to customize ptxas location.\n", + "This message will be only logged once.\n", + "2022-02-28 23:49:01.455656: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcublas.so.10\n" + ] + }, + { + "name": "stdout", + "output_type": "stream", + "text": [ + "86/86 [==============================] - 123s 337ms/step - loss: 1.7958 - accuracy: 0.3338 - val_loss: 1.5575 - val_accuracy: 0.3477\n", + "\n", + "Epoch 00001: val_loss improved from inf to 1.55745, saving model to /home/jovyan/PlanetaryComputerExamples/landcovnet/unet_model.hdf5\n", + "Epoch 2/100\n", + "86/86 [==============================] - 118s 328ms/step - loss: 1.5182 - accuracy: 0.3648 - val_loss: 1.5356 - val_accuracy: 0.3547\n", + "\n", + "Epoch 00002: val_loss improved from 1.55745 to 1.53557, saving model to /home/jovyan/PlanetaryComputerExamples/landcovnet/unet_model.hdf5\n", + "Epoch 3/100\n", + "86/86 [==============================] - 28s 329ms/step - loss: 1.4885 - accuracy: 0.3681 - val_loss: 1.5318 - val_accuracy: 0.3553\n", + "\n", + "Epoch 00003: val_loss improved from 1.53557 to 1.53175, saving model to /home/jovyan/PlanetaryComputerExamples/landcovnet/unet_model.hdf5\n", + "Epoch 4/100\n", + "86/86 [==============================] - 28s 332ms/step - loss: 1.4986 - accuracy: 0.3635 - val_loss: 1.5306 - val_accuracy: 0.3552\n", + "\n", + "Epoch 00004: val_loss improved from 1.53175 to 1.53057, saving model to /home/jovyan/PlanetaryComputerExamples/landcovnet/unet_model.hdf5\n", + "Epoch 5/100\n", + "86/86 [==============================] - 29s 334ms/step - loss: 1.5225 - accuracy: 0.3477 - val_loss: 1.5499 - val_accuracy: 0.3554\n", + "\n", + "Epoch 00005: val_loss did not improve from 1.53057\n", + "Epoch 6/100\n", + "86/86 [==============================] - 29s 335ms/step - loss: 1.5080 - accuracy: 0.3558 - val_loss: 1.5315 - val_accuracy: 0.3556\n", + "\n", + "Epoch 00006: val_loss did not improve from 1.53057\n", + "Epoch 7/100\n", + "86/86 [==============================] - 29s 337ms/step - loss: 1.4908 - accuracy: 0.3514 - val_loss: 1.5334 - val_accuracy: 0.3552\n", + "\n", + "Epoch 00007: val_loss did not improve from 1.53057\n", + "Epoch 8/100\n", + "86/86 [==============================] - 29s 338ms/step - loss: 1.5140 - accuracy: 0.3699 - val_loss: 1.5305 - val_accuracy: 0.3554\n", + "\n", + "Epoch 00008: val_loss improved from 1.53057 to 1.53046, saving model to /home/jovyan/PlanetaryComputerExamples/landcovnet/unet_model.hdf5\n", + "Epoch 9/100\n", + "86/86 [==============================] - 29s 338ms/step - loss: 1.5134 - accuracy: 0.3577 - val_loss: 1.5278 - val_accuracy: 0.3557\n", + "\n", + "Epoch 00009: val_loss improved from 1.53046 to 1.52785, saving model to /home/jovyan/PlanetaryComputerExamples/landcovnet/unet_model.hdf5\n", + "Epoch 10/100\n", + "86/86 [==============================] - 29s 339ms/step - loss: 1.4930 - accuracy: 0.3672 - val_loss: 1.5282 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00010: val_loss did not improve from 1.52785\n", + "Epoch 11/100\n", + "86/86 [==============================] - 29s 340ms/step - loss: 1.4827 - accuracy: 0.3503 - val_loss: 1.5396 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00011: val_loss did not improve from 1.52785\n", + "Epoch 12/100\n", + "86/86 [==============================] - 29s 341ms/step - loss: 1.4833 - accuracy: 0.3644 - val_loss: 1.5364 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00012: val_loss did not improve from 1.52785\n", + "Epoch 13/100\n", + "86/86 [==============================] - 29s 341ms/step - loss: 1.5142 - accuracy: 0.3597 - val_loss: 1.5288 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00013: val_loss did not improve from 1.52785\n", + "Epoch 14/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.5037 - accuracy: 0.3479 - val_loss: 1.5338 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00014: val_loss did not improve from 1.52785\n", + "Epoch 15/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.5177 - accuracy: 0.3591 - val_loss: 1.5279 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00015: val_loss did not improve from 1.52785\n", + "Epoch 16/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.5020 - accuracy: 0.3713 - val_loss: 1.5321 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00016: val_loss did not improve from 1.52785\n", + "Epoch 17/100\n", + "86/86 [==============================] - 29s 343ms/step - loss: 1.4694 - accuracy: 0.3572 - val_loss: 1.5266 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00017: val_loss improved from 1.52785 to 1.52662, saving model to /home/jovyan/PlanetaryComputerExamples/landcovnet/unet_model.hdf5\n", + "Epoch 18/100\n", + "86/86 [==============================] - 30s 343ms/step - loss: 1.5468 - accuracy: 0.3270 - val_loss: 1.5268 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00018: val_loss did not improve from 1.52662\n", + "Epoch 19/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.5024 - accuracy: 0.3405 - val_loss: 1.5391 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00019: val_loss did not improve from 1.52662\n", + "Epoch 20/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.4870 - accuracy: 0.3576 - val_loss: 1.5312 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00020: val_loss did not improve from 1.52662\n", + "Epoch 21/100\n", + "86/86 [==============================] - 29s 343ms/step - loss: 1.5126 - accuracy: 0.3619 - val_loss: 1.5260 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00021: val_loss improved from 1.52662 to 1.52603, saving model to /home/jovyan/PlanetaryComputerExamples/landcovnet/unet_model.hdf5\n", + "Epoch 22/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.4966 - accuracy: 0.3416 - val_loss: 1.5349 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00022: val_loss did not improve from 1.52603\n", + "Epoch 23/100\n", + "86/86 [==============================] - 29s 343ms/step - loss: 1.4935 - accuracy: 0.3623 - val_loss: 1.5266 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00023: val_loss did not improve from 1.52603\n", + "Epoch 24/100\n", + "86/86 [==============================] - 29s 343ms/step - loss: 1.4942 - accuracy: 0.3442 - val_loss: 1.5466 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00024: val_loss did not improve from 1.52603\n", + "Epoch 25/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.4941 - accuracy: 0.3644 - val_loss: 1.5322 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00025: val_loss did not improve from 1.52603\n", + "Epoch 26/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.5024 - accuracy: 0.3500 - val_loss: 1.5286 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00026: val_loss did not improve from 1.52603\n", + "Epoch 27/100\n", + "86/86 [==============================] - 29s 343ms/step - loss: 1.4976 - accuracy: 0.3629 - val_loss: 1.5290 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00027: val_loss did not improve from 1.52603\n", + "Epoch 28/100\n", + "86/86 [==============================] - 29s 343ms/step - loss: 1.4727 - accuracy: 0.3910 - val_loss: 1.5271 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00028: val_loss did not improve from 1.52603\n", + "Epoch 29/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.5095 - accuracy: 0.3594 - val_loss: 1.5282 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00029: val_loss did not improve from 1.52603\n", + "Epoch 30/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.4996 - accuracy: 0.3420 - val_loss: 1.5279 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00030: val_loss did not improve from 1.52603\n", + "Epoch 31/100\n", + "86/86 [==============================] - 29s 343ms/step - loss: 1.4839 - accuracy: 0.3629 - val_loss: 1.5262 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00031: val_loss did not improve from 1.52603\n", + "Epoch 32/100\n", + "86/86 [==============================] - 29s 343ms/step - loss: 1.4861 - accuracy: 0.3486 - val_loss: 1.5268 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00032: val_loss did not improve from 1.52603\n", + "Epoch 33/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.4885 - accuracy: 0.3490 - val_loss: 1.5284 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00033: val_loss did not improve from 1.52603\n", + "Epoch 34/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.5166 - accuracy: 0.3577 - val_loss: 1.5262 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00034: val_loss did not improve from 1.52603\n", + "Epoch 35/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.4855 - accuracy: 0.3493 - val_loss: 1.5260 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00035: val_loss improved from 1.52603 to 1.52601, saving model to /home/jovyan/PlanetaryComputerExamples/landcovnet/unet_model.hdf5\n", + "Epoch 36/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.4740 - accuracy: 0.3662 - val_loss: 1.5286 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00036: val_loss did not improve from 1.52601\n", + "Epoch 37/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.4804 - accuracy: 0.3652 - val_loss: 1.5302 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00037: val_loss did not improve from 1.52601\n", + "Epoch 38/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.4947 - accuracy: 0.3388 - val_loss: 1.5296 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00038: val_loss did not improve from 1.52601\n", + "Epoch 39/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.4901 - accuracy: 0.3592 - val_loss: 1.5259 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00039: val_loss improved from 1.52601 to 1.52589, saving model to /home/jovyan/PlanetaryComputerExamples/landcovnet/unet_model.hdf5\n", + "Epoch 40/100\n", + "86/86 [==============================] - 29s 341ms/step - loss: 1.4699 - accuracy: 0.3735 - val_loss: 1.5296 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00040: val_loss did not improve from 1.52589\n", + "Epoch 41/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.5091 - accuracy: 0.3508 - val_loss: 1.5290 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00041: val_loss did not improve from 1.52589\n", + "Epoch 42/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.4662 - accuracy: 0.3759 - val_loss: 1.5271 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00042: val_loss did not improve from 1.52589\n", + "Epoch 43/100\n", + "86/86 [==============================] - 29s 341ms/step - loss: 1.5121 - accuracy: 0.3477 - val_loss: 1.5361 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00043: val_loss did not improve from 1.52589\n", + "Epoch 44/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.5049 - accuracy: 0.3656 - val_loss: 1.5318 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00044: val_loss did not improve from 1.52589\n", + "Epoch 45/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.4975 - accuracy: 0.3524 - val_loss: 1.5300 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00045: val_loss did not improve from 1.52589\n", + "Epoch 46/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.4920 - accuracy: 0.3686 - val_loss: 1.5275 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00046: val_loss did not improve from 1.52589\n", + "Epoch 47/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.4966 - accuracy: 0.3530 - val_loss: 1.5259 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00047: val_loss improved from 1.52589 to 1.52586, saving model to /home/jovyan/PlanetaryComputerExamples/landcovnet/unet_model.hdf5\n", + "Epoch 48/100\n", + "86/86 [==============================] - 29s 341ms/step - loss: 1.5025 - accuracy: 0.3617 - val_loss: 1.5254 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00048: val_loss improved from 1.52586 to 1.52541, saving model to /home/jovyan/PlanetaryComputerExamples/landcovnet/unet_model.hdf5\n", + "Epoch 49/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.4785 - accuracy: 0.3665 - val_loss: 1.5277 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00049: val_loss did not improve from 1.52541\n", + "Epoch 50/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.4744 - accuracy: 0.3572 - val_loss: 1.5270 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00050: val_loss did not improve from 1.52541\n", + "Epoch 51/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.4842 - accuracy: 0.3572 - val_loss: 1.5276 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00051: val_loss did not improve from 1.52541\n", + "Epoch 52/100\n", + "86/86 [==============================] - 29s 341ms/step - loss: 1.4725 - accuracy: 0.3760 - val_loss: 1.5256 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00052: val_loss did not improve from 1.52541\n", + "Epoch 53/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.4905 - accuracy: 0.3561 - val_loss: 1.5275 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00053: val_loss did not improve from 1.52541\n", + "Epoch 54/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.5079 - accuracy: 0.3531 - val_loss: 1.5285 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00054: val_loss did not improve from 1.52541\n", + "Epoch 55/100\n", + "86/86 [==============================] - 29s 341ms/step - loss: 1.4643 - accuracy: 0.3860 - val_loss: 1.5254 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00055: val_loss improved from 1.52541 to 1.52536, saving model to /home/jovyan/PlanetaryComputerExamples/landcovnet/unet_model.hdf5\n", + "Epoch 56/100\n", + "86/86 [==============================] - 29s 341ms/step - loss: 1.4870 - accuracy: 0.3449 - val_loss: 1.5267 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00056: val_loss did not improve from 1.52536\n", + "Epoch 57/100\n", + "86/86 [==============================] - 29s 341ms/step - loss: 1.4967 - accuracy: 0.3423 - val_loss: 1.5264 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00057: val_loss did not improve from 1.52536\n", + "Epoch 58/100\n", + "86/86 [==============================] - 29s 341ms/step - loss: 1.4949 - accuracy: 0.3458 - val_loss: 1.5270 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00058: val_loss did not improve from 1.52536\n", + "Epoch 59/100\n", + "86/86 [==============================] - 29s 341ms/step - loss: 1.4858 - accuracy: 0.3632 - val_loss: 1.5271 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00059: val_loss did not improve from 1.52536\n", + "Epoch 60/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.5012 - accuracy: 0.3445 - val_loss: 1.5270 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00060: val_loss did not improve from 1.52536\n", + "Epoch 61/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.4769 - accuracy: 0.3585 - val_loss: 1.5277 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00061: val_loss did not improve from 1.52536\n", + "Epoch 62/100\n", + "86/86 [==============================] - 29s 341ms/step - loss: 1.5104 - accuracy: 0.3449 - val_loss: 1.5285 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00062: val_loss did not improve from 1.52536\n", + "Epoch 63/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.4938 - accuracy: 0.3607 - val_loss: 1.5292 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00063: val_loss did not improve from 1.52536\n", + "Epoch 64/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.4853 - accuracy: 0.3552 - val_loss: 1.5281 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00064: val_loss did not improve from 1.52536\n", + "Epoch 65/100\n", + "86/86 [==============================] - 29s 342ms/step - loss: 1.4846 - accuracy: 0.3536 - val_loss: 1.5295 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00065: val_loss did not improve from 1.52536\n", + "Epoch 66/100\n", + "86/86 [==============================] - 29s 341ms/step - loss: 1.4920 - accuracy: 0.3530 - val_loss: 1.5273 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00066: val_loss did not improve from 1.52536\n", + "Epoch 67/100\n", + "86/86 [==============================] - 29s 341ms/step - loss: 1.4785 - accuracy: 0.3473 - val_loss: 1.5271 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00067: val_loss did not improve from 1.52536\n", + "Epoch 68/100\n", + "86/86 [==============================] - 29s 341ms/step - loss: 1.4956 - accuracy: 0.3626 - val_loss: 1.5255 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00068: val_loss did not improve from 1.52536\n", + "Epoch 69/100\n", + "86/86 [==============================] - 29s 341ms/step - loss: 1.4699 - accuracy: 0.3719 - val_loss: 1.5296 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00069: val_loss did not improve from 1.52536\n", + "Epoch 70/100\n", + "86/86 [==============================] - 29s 341ms/step - loss: 1.4865 - accuracy: 0.3686 - val_loss: 1.5266 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00070: val_loss did not improve from 1.52536\n", + "Epoch 71/100\n", + "86/86 [==============================] - 29s 341ms/step - loss: 1.4808 - accuracy: 0.3780 - val_loss: 1.5256 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00071: val_loss did not improve from 1.52536\n", + "Epoch 72/100\n", + "86/86 [==============================] - 29s 341ms/step - loss: 1.4826 - accuracy: 0.3619 - val_loss: 1.5256 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00072: val_loss did not improve from 1.52536\n", + "Epoch 73/100\n", + "86/86 [==============================] - 29s 341ms/step - loss: 1.5173 - accuracy: 0.3407 - val_loss: 1.5277 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00073: val_loss did not improve from 1.52536\n", + "Epoch 74/100\n", + "86/86 [==============================] - 29s 341ms/step - loss: 1.5002 - accuracy: 0.3504 - val_loss: 1.5322 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00074: val_loss did not improve from 1.52536\n", + "Epoch 75/100\n", + "86/86 [==============================] - 29s 341ms/step - loss: 1.4901 - accuracy: 0.3482 - val_loss: 1.5261 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00075: val_loss did not improve from 1.52536\n", + "Epoch 76/100\n", + "86/86 [==============================] - 29s 341ms/step - loss: 1.5001 - accuracy: 0.3589 - val_loss: 1.5305 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00076: val_loss did not improve from 1.52536\n", + "Epoch 77/100\n", + "86/86 [==============================] - 29s 340ms/step - loss: 1.4884 - accuracy: 0.3553 - val_loss: 1.5272 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00077: val_loss did not improve from 1.52536\n", + "Epoch 78/100\n", + "86/86 [==============================] - 29s 340ms/step - loss: 1.4937 - accuracy: 0.3631 - val_loss: 1.5319 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00078: val_loss did not improve from 1.52536\n", + "Epoch 79/100\n", + "86/86 [==============================] - 29s 339ms/step - loss: 1.4878 - accuracy: 0.3509 - val_loss: 1.5274 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00079: val_loss did not improve from 1.52536\n", + "Epoch 80/100\n", + "86/86 [==============================] - 29s 339ms/step - loss: 1.5072 - accuracy: 0.3551 - val_loss: 1.5257 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00080: val_loss did not improve from 1.52536\n", + "Epoch 81/100\n", + "86/86 [==============================] - 29s 339ms/step - loss: 1.4803 - accuracy: 0.3519 - val_loss: 1.5311 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00081: val_loss did not improve from 1.52536\n", + "Epoch 82/100\n", + "86/86 [==============================] - 29s 338ms/step - loss: 1.4972 - accuracy: 0.3528 - val_loss: 1.5261 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00082: val_loss did not improve from 1.52536\n", + "Epoch 83/100\n", + "86/86 [==============================] - 29s 338ms/step - loss: 1.4818 - accuracy: 0.3620 - val_loss: 1.5272 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00083: val_loss did not improve from 1.52536\n", + "Epoch 84/100\n", + "86/86 [==============================] - 29s 339ms/step - loss: 1.5038 - accuracy: 0.3569 - val_loss: 1.5276 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00084: val_loss did not improve from 1.52536\n", + "Epoch 85/100\n", + "86/86 [==============================] - 29s 339ms/step - loss: 1.4974 - accuracy: 0.3505 - val_loss: 1.5338 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00085: val_loss did not improve from 1.52536\n", + "Epoch 86/100\n", + "86/86 [==============================] - 29s 339ms/step - loss: 1.4645 - accuracy: 0.3675 - val_loss: 1.5257 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00086: val_loss did not improve from 1.52536\n", + "Epoch 87/100\n", + "86/86 [==============================] - 29s 339ms/step - loss: 1.4907 - accuracy: 0.3640 - val_loss: 1.5271 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00087: val_loss did not improve from 1.52536\n", + "Epoch 88/100\n", + "86/86 [==============================] - 29s 340ms/step - loss: 1.4890 - accuracy: 0.3578 - val_loss: 1.5277 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00088: val_loss did not improve from 1.52536\n", + "Epoch 89/100\n", + "86/86 [==============================] - 29s 340ms/step - loss: 1.4965 - accuracy: 0.3382 - val_loss: 1.5308 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00089: val_loss did not improve from 1.52536\n", + "Epoch 90/100\n", + "86/86 [==============================] - 29s 339ms/step - loss: 1.5022 - accuracy: 0.3543 - val_loss: 1.5269 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00090: val_loss did not improve from 1.52536\n", + "Epoch 91/100\n", + "86/86 [==============================] - 29s 339ms/step - loss: 1.4952 - accuracy: 0.3704 - val_loss: 1.5301 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00091: val_loss did not improve from 1.52536\n", + "Epoch 92/100\n", + "86/86 [==============================] - 29s 339ms/step - loss: 1.4951 - accuracy: 0.3597 - val_loss: 1.5264 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00092: val_loss did not improve from 1.52536\n", + "Epoch 93/100\n", + "86/86 [==============================] - 29s 339ms/step - loss: 1.4864 - accuracy: 0.3565 - val_loss: 1.5257 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00093: val_loss did not improve from 1.52536\n", + "Epoch 94/100\n", + "86/86 [==============================] - 29s 339ms/step - loss: 1.5263 - accuracy: 0.3303 - val_loss: 1.5283 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00094: val_loss did not improve from 1.52536\n", + "Epoch 95/100\n", + "86/86 [==============================] - 29s 339ms/step - loss: 1.4925 - accuracy: 0.3593 - val_loss: 1.5255 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00095: val_loss did not improve from 1.52536\n", + "Epoch 96/100\n", + "86/86 [==============================] - 29s 339ms/step - loss: 1.4633 - accuracy: 0.3698 - val_loss: 1.5259 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00096: val_loss did not improve from 1.52536\n", + "Epoch 97/100\n", + "86/86 [==============================] - 29s 338ms/step - loss: 1.4966 - accuracy: 0.3655 - val_loss: 1.5260 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00097: val_loss did not improve from 1.52536\n", + "Epoch 98/100\n", + "86/86 [==============================] - 29s 339ms/step - loss: 1.4984 - accuracy: 0.3582 - val_loss: 1.5271 - val_accuracy: 0.3559\n", + "\n", + "Epoch 00098: val_loss did not improve from 1.52536\n" + ] + } + ], + "source": [ + "batch_size = 16\n", + "history = unet.fit(\n", + " x=x_train,\n", + " y=y_train,\n", + " validation_data=(x_val, y_val),\n", + " steps_per_epoch=len(x_train) // batch_size,\n", + " validation_steps=len(x_val) // batch_size,\n", + " batch_size=batch_size,\n", + " epochs=100,\n", + " callbacks=[model_checkpoint, model_earlyStopping],\n", + ")" + ] + }, + { + "cell_type": "markdown", + "id": "3892a194-bd3d-403a-a996-d5df5ec5f2e2", + "metadata": {}, + "source": [ + "### Visualising the Model Results\n", + "\n", + "First, we can inspect the loss and accuracy curves for the number of epochs defined in the training process.\n", + "\n", + "This was done for both the training and test losses and accuracies, shown in the figures below" + ] + }, + { + "cell_type": "code", + "execution_count": 24, + "id": "d51e6e57-4e59-4c9a-91d2-e7958175be4c", + "metadata": {}, + "outputs": [ + { + "data": { + "image/png": "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\n", + "text/plain": [ + "
+
+This tutorial delves into building a scalable model on the LandCoverNet dataset.
+
+The tutorial is split into three notebooks:
+
+### 1. Data Exploration
+In this notebook, a portion of the tutorial:
+
+1. We will authenticate to MLHub via the API key to make requests on the LandCoverNet label collection
+
+2. We will download the label collection into PC
+
+3. We will perform some exploratory data analysis on these labels to get a better understanding of the LandCoverNet data
+
+### 2. Data Preparation
+
+This portion of the tutorial is focused on preparing the data for training a semantic segmentation model for LandCoverNet data.
+Here:
+
+1. We will inspect the source imagery for the labels we have
+
+2. We will process the source imagery in parallel using Dask
+
+3. We will select the labels and filtered source images from Dask to be loaded
+
+4. We will save the images and associated labels data as a pickle file ('.pkl') on our directory to be loaded for model training
+
+### 3. Model Training
+
+This portion of the tutorial is focused on developing a semantic segmentation model for LandCoverNet data.
+Here:
+
+1. We will load the prepared data as training, test and validation data
+
+2. We will build a segmentation model, train it on the training data, inspect the results on the validation data and make predictions on the test data
+
+## How To Get Started
+
+1. Login to Planetary Computer's JupyterLab hub.
+2. CD to the directory you're working on
+3. Run the command: `pip install -r requirements.txt`
+4. Use the `1. Data Exploration` notebook and follow the instructions
+5. Use the `2. Data Preparation` notebook and follow the instructions
+6. Use the `3. Model Training` notebook and follow the instructions (preferably use the Tensorflow GPU server for faster training results)
diff --git a/tutorials/model_training/requirements.txt b/tutorials/model_training/requirements.txt
new file mode 100644
index 00000000..15f4970c
--- /dev/null
+++ b/tutorials/model_training/requirements.txt
@@ -0,0 +1,2 @@
+tensorflow
+keras==2.3.1