Experiments_monthly.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "1d0fb88f",
   "metadata": {},
   "outputs": [],
   "source": [
    "from datasets import *\n",
    "from train import *\n",
    "from predict import *\n",
    "from evaluate import *\n",
    "from util import *\n",
    "import numpy as np"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "09965741",
   "metadata": {},
   "source": [
    "# Experiments on monthly time scale:\n",
    "\n",
    "To conduct the experiments on the monthly time scale we first create data sets from the climate model data.\n",
    "\n",
    "We test 3 configurations:\n",
    "- Training a separate UNet for each month (thus doesn't have to learn yearly cycle)\n",
    "- Training one UNet to predict on all months (thus possible benefits from better generalization)\n",
    "- Training one UNet to predict on all months, while using not only the current but also the previous month as input to the emulation"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "0a0fa8a2",
   "metadata": {},
   "source": [
    "## Create datasets:\n",
    "\n",
    "Creating monthly data sets is time intensive because they are significantly bigger than the yearly data sets."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "a368fa41",
   "metadata": {},
   "source": [
    "All months:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "a501fd68",
   "metadata": {},
   "outputs": [],
   "source": [
    "base_folder = \"Datasets\"\n",
    "output_folder = \"Output/Reproduce_monthly\"\n",
    "\n",
    "description = {}\n",
    "\n",
    "description[\"DATASETS_USED\"] = [\"isotopes\", \n",
    "                                \"tsurf\", \n",
    "                                \"prec\"]\n",
    "\n",
    "description[\"PREDICTOR_VARIABLES\"] = {\"tsurf\": [\"tsurf\"],\n",
    "                                      \"prec\": [\"prec\"]}\n",
    "\n",
    "description[\"TARGET_VARIABLES\"] = {\"isotopes\": [\"d18O\"]}\n",
    "\n",
    "description[\"DATASETS_NO_GAPS\"] = [\"isotopes\", \n",
    "                                   \"tsurf\", \n",
    "                                   \"prec\", \n",
    "                                   \"slp\"]\n",
    "\n",
    "description[\"CLIMATE_MODEL\"] = \"iHadCM3\"\n",
    "description[\"GRID_TYPE\"] = \"Flat\"\n",
    "\n",
    "description[\"START_YEAR\"] = 850\n",
    "description[\"END_YEAR\"] = 1850\n",
    "description[\"LATITUDES_SLICE\"] = [1,-1]\n",
    "\n",
    "description[\"SPLIT_YEAR\"] = 1750\n",
    "description[\"DO_SHUFFLE\"] = False\n",
    "\n",
    "description[\"PRECIP_WEIGHTING\"] = False\n",
    "\n",
    "description[\"TIMESCALE\"] = \"MONTHLY\"\n",
    "description[\"MONTHS_USED\"] = np.sort([0,1,2,3,4,5,6,7,8,9,10,11]).tolist() # months of the year to use in this data set - Jan: 0 to Dec: 11\n",
    "description[\"MONTHS_USED_IN_PREDICTION\"] = np.sort([0]).tolist() # What months are used in prediction, if [0], only the current time step is used in prediction, if [0,-1] also the previous month is used."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "7eb1a606",
   "metadata": {},
   "outputs": [],
   "source": [
    "create_monthly_dataset(description, base_folder, output_folder)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "21b8361d",
   "metadata": {},
   "source": [
    "All months, use lagged months in predictions:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "abe3e2ad",
   "metadata": {},
   "outputs": [],
   "source": [
    "base_folder = \"Datasets\"\n",
    "output_folder = \"Output/Reproduce_monthly\"\n",
    "\n",
    "description = {}\n",
    "\n",
    "description[\"DATASETS_USED\"] = [\"isotopes\", \n",
    "                                \"tsurf\", \n",
    "                                \"prec\"]\n",
    "\n",
    "description[\"PREDICTOR_VARIABLES\"] = {\"tsurf\": [\"tsurf\"],\n",
    "                                      \"prec\": [\"prec\"]}\n",
    "\n",
    "description[\"TARGET_VARIABLES\"] = {\"isotopes\": [\"d18O\"]}\n",
    "\n",
    "description[\"DATASETS_NO_GAPS\"] = [\"isotopes\", \n",
    "                                   \"tsurf\", \n",
    "                                   \"prec\", \n",
    "                                   \"slp\"]\n",
    "\n",
    "description[\"CLIMATE_MODEL\"] = \"iHadCM3\"\n",
    "description[\"GRID_TYPE\"] = \"Flat\"\n",
    "\n",
    "description[\"START_YEAR\"] = 850\n",
    "description[\"END_YEAR\"] = 1850\n",
    "description[\"LATITUDES_SLICE\"] = [1,-1]\n",
    "\n",
    "description[\"SPLIT_YEAR\"] = 1750\n",
    "description[\"DO_SHUFFLE\"] = False\n",
    "\n",
    "description[\"PRECIP_WEIGHTING\"] = False\n",
    "\n",
    "description[\"TIMESCALE\"] = \"MONTHLY\"\n",
    "description[\"MONTHS_USED\"] = np.sort([0,1,2,3,4,5,6,7,8,9,10,11]).tolist()\n",
    "description[\"MONTHS_USED_IN_PREDICTION\"] = np.sort([0,-1]).tolist() # use current and previous timestep in prediction"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "8ba53bc6",
   "metadata": {},
   "outputs": [],
   "source": [
    "create_monthly_dataset(description, base_folder, output_folder)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "312ad231",
   "metadata": {},
   "source": [
    "Individual dataset for each month:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "99fcd58f",
   "metadata": {},
   "outputs": [],
   "source": [
    "base_folder = \"Datasets\"\n",
    "output_folder = \"Output/Reproduce_monthly\"\n",
    "\n",
    "description = {}\n",
    "\n",
    "description[\"DATASETS_USED\"] = [\"isotopes\", \n",
    "                                \"tsurf\", \n",
    "                                \"prec\"]\n",
    "\n",
    "description[\"PREDICTOR_VARIABLES\"] = {\"tsurf\": [\"tsurf\"],\n",
    "                                      \"prec\": [\"prec\"]}\n",
    "\n",
    "description[\"TARGET_VARIABLES\"] = {\"isotopes\": [\"d18O\"]}\n",
    "\n",
    "description[\"DATASETS_NO_GAPS\"] = [\"isotopes\", \n",
    "                                   \"tsurf\", \n",
    "                                   \"prec\", \n",
    "                                   \"slp\"]\n",
    "\n",
    "description[\"CLIMATE_MODEL\"] = \"iHadCM3\"\n",
    "description[\"GRID_TYPE\"] = \"Flat\"\n",
    "\n",
    "description[\"START_YEAR\"] = 850\n",
    "description[\"END_YEAR\"] = 1850\n",
    "description[\"LATITUDES_SLICE\"] = [1,-1]\n",
    "\n",
    "description[\"SPLIT_YEAR\"] = 1750\n",
    "description[\"DO_SHUFFLE\"] = False\n",
    "\n",
    "description[\"PRECIP_WEIGHTING\"] = False\n",
    "\n",
    "description[\"TIMESCALE\"] = \"MONTHLY\"\n",
    "description[\"MONTHS_USED_IN_PREDICTION\"] = np.sort([0]).tolist()\n",
    "\n",
    "for i in range(12):\n",
    "    description[\"MONTHS_USED\"] = np.sort([i]).tolist() # create a separate data set for each month\n",
    "    create_monthly_dataset(description, base_folder, output_folder)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d24ab8a5",
   "metadata": {},
   "source": [
    "# Run experiments on these datasets:"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "734434f9",
   "metadata": {},
   "source": [
    "All months."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "ea7519c8",
   "metadata": {},
   "outputs": [],
   "source": [
    "base_folder = \"Datasets\"\n",
    "output_folder = \"Output/Reproduce_monthly\"\n",
    "\n",
    "description = {}\n",
    "\n",
    "description[\"DATASETS_USED\"] = [\"isotopes\", \n",
    "                                \"tsurf\", \n",
    "                                \"prec\"]\n",
    "\n",
    "description[\"PREDICTOR_VARIABLES\"] = {\"tsurf\": [\"tsurf\"],\n",
    "                                      \"prec\": [\"prec\"]}\n",
    "\n",
    "description[\"TARGET_VARIABLES\"] = {\"isotopes\": [\"d18O\"]}\n",
    "\n",
    "description[\"DATASETS_NO_GAPS\"] = [\"isotopes\", \n",
    "                                   \"tsurf\", \n",
    "                                   \"prec\", \n",
    "                                   \"slp\"]\n",
    "\n",
    "description[\"CLIMATE_MODEL\"] = \"iHadCM3\"\n",
    "description[\"GRID_TYPE\"] = \"Flat\"\n",
    "\n",
    "description[\"START_YEAR\"] = 850\n",
    "description[\"END_YEAR\"] = 1850\n",
    "description[\"LATITUDES_SLICE\"] = [1,-1]\n",
    "\n",
    "description[\"SPLIT_YEAR\"] = 1750\n",
    "description[\"DO_SHUFFLE\"] = False\n",
    "\n",
    "description[\"PRECIP_WEIGHTING\"] = False\n",
    "\n",
    "description[\"TIMESCALE\"] = \"MONTHLY\"\n",
    "description[\"MONTHS_USED\"] = np.sort([0,1,2,3,4,5,6,7,8,9,10,11]).tolist()\n",
    "description[\"MONTHS_USED_IN_PREDICTION\"] = np.sort([0]).tolist()\n",
    "\n",
    "### MODEL_TRAINING ###############################################\n",
    "\n",
    "model_training_description = {}\n",
    "model_training_description[\"S_MODE_PREDICTORS\"] = [\"Pixelwise\",\"Pixelwise\"] # how to standardize the given variables\n",
    "model_training_description[\"S_MODE_TARGETS\"] = [\"Pixelwise\"]\n",
    "\n",
    "model_training_description[\"DATASET_FOLDER\"] = output_folder\n",
    "\n",
    "model_training_description[\"MODEL_TYPE\"] = \"UNet_Flat\"\n",
    "model_training_description[\"CREATE_VALIDATIONSET\"] = True\n",
    "model_training_description[\"SHUFFLE_VALIDATIONSET\"] = True\n",
    "\n",
    "\n",
    "# training parameters\n",
    "\n",
    "model_training_description[\"DEPTH\"] = 3\n",
    "model_training_description[\"NUM_EPOCHS\"] = \"early_stopping\"  # 20\n",
    "model_training_description[\"PATIENCE\"] = 5\n",
    "model_training_description[\"BATCH_SIZE\"] = 8\n",
    "model_training_description[\"LEARNING_RATE\"] = 1e-3  # 0.002637 # 5e-3  # use either this or default ADAM learning rate\n",
    "\n",
    "# model parameters\n",
    "model_training_description[\"DEPTH\"] = 3\n",
    "model_training_description[\"IN_CHANNELS\"] = len(util.flatten(description[\"PREDICTOR_VARIABLES\"].values()))\n",
    "model_training_description[\"CHANNELS_FIRST_CONV\"] = 32\n",
    "model_training_description[\"OUT_CHANNELS\"] = len(util.flatten(description[\"TARGET_VARIABLES\"].values()))\n",
    "model_training_description[\"FMAPS\"] = (32,32,64,64)\n",
    "\n",
    "\n",
    "model_training_description[\"ACTIVATION\"] = torch.nn.ReLU\n",
    "model_training_description[\"NORMALIZATION\"] = torch.nn.BatchNorm2d  # IcoBatchNorm2d \n",
    "\n",
    "\n",
    "model_training_description[\"OPTIMIZER\"] = \"Adam\"\n",
    "\n",
    "model_training_description[\"DEVICE\"] = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n",
    "\n",
    "model_training_description[\"USE_CYLINDRICAL_PADDING\"] = True\n",
    "model_training_description[\"USE_COORD_CONV\"] = True\n",
    "model_training_description[\"LOSS\"] = \"Masked_AreaWeightedMSELoss\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "8eca9806",
   "metadata": {},
   "outputs": [],
   "source": [
    "n_runs = 5  # run 5 experiments for each configuration\n",
    "for i in range(n_runs):\n",
    "    model_training_description[\"RUN_NR\"] = i\n",
    "    unet = train_unet(description, model_training_description, output_folder)\n",
    "    predict_save_unet(description, model_training_description, output_folder, unet, output_folder)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "908a542a",
   "metadata": {},
   "source": [
    "All months, use previous timestep in prediction"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "26ffd7ff",
   "metadata": {},
   "outputs": [],
   "source": [
    "base_folder = \"Datasets\"\n",
    "output_folder = \"Output/Reproduce_monthly\"\n",
    "\n",
    "description = {}\n",
    "\n",
    "description[\"DATASETS_USED\"] = [\"isotopes\", \n",
    "                                \"tsurf\",\n",
    "                                \"prec\"]\n",
    "\n",
    "description[\"PREDICTOR_VARIABLES\"] = {\"tsurf\": [\"tsurf\"],\n",
    "                                      \"prec\": [\"prec\"]}\n",
    "\n",
    "description[\"TARGET_VARIABLES\"] = {\"isotopes\": [\"d18O\"]}\n",
    "\n",
    "description[\"DATASETS_NO_GAPS\"] = [\"isotopes\", \n",
    "                                   \"tsurf\", \n",
    "                                   \"prec\", \n",
    "                                   \"slp\"]\n",
    "\n",
    "description[\"CLIMATE_MODEL\"] = \"iHadCM3\"\n",
    "description[\"GRID_TYPE\"] = \"Flat\"\n",
    "\n",
    "description[\"START_YEAR\"] = 850\n",
    "description[\"END_YEAR\"] = 1850\n",
    "description[\"LATITUDES_SLICE\"] = [1,-1]\n",
    "\n",
    "description[\"SPLIT_YEAR\"] = 1750\n",
    "description[\"DO_SHUFFLE\"] = False\n",
    "description[\"PRECIP_WEIGHTING\"] = False\n",
    "\n",
    "\n",
    "description[\"TIMESCALE\"] = \"MONTHLY\"\n",
    "description[\"MONTHS_USED\"] = np.sort([0,1,2,3,4,5,6,7,8,9,10,11]).tolist()\n",
    "description[\"MONTHS_USED_IN_PREDICTION\"] = np.sort([0,-1]).tolist()\n",
    "\n",
    "### MODEL_TRAINING ###############################################\n",
    "\n",
    "model_training_description = {}\n",
    "model_training_description[\"S_MODE_PREDICTORS\"] = [\"Pixelwise\",\"Pixelwise\",\"Pixelwise\",\"Pixelwise\"] # how to standardize the given variables\n",
    "model_training_description[\"S_MODE_TARGETS\"] = [\"Pixelwise\"]\n",
    "\n",
    "model_training_description[\"DATASET_FOLDER\"] = output_folder\n",
    "\n",
    "model_training_description[\"MODEL_TYPE\"] = \"UNet_Flat\"\n",
    "model_training_description[\"CREATE_VALIDATIONSET\"] = True\n",
    "model_training_description[\"SHUFFLE_VALIDATIONSET\"] = True\n",
    "\n",
    "\n",
    "# training parameters\n",
    "\n",
    "model_training_description[\"DEPTH\"] = 3\n",
    "model_training_description[\"NUM_EPOCHS\"] = \"early_stopping\"  # 20\n",
    "model_training_description[\"PATIENCE\"] = 5\n",
    "model_training_description[\"BATCH_SIZE\"] = 8\n",
    "model_training_description[\"LEARNING_RATE\"] = 1e-3  # 0.002637 # 5e-3  # use either this or default ADAM learning rate\n",
    "\n",
    "# model parameters\n",
    "model_training_description[\"DEPTH\"] = 3\n",
    "model_training_description[\"IN_CHANNELS\"] = len(util.flatten(description[\"PREDICTOR_VARIABLES\"].values()))*2\n",
    "model_training_description[\"CHANNELS_FIRST_CONV\"] = 32\n",
    "model_training_description[\"OUT_CHANNELS\"] = len(util.flatten(description[\"TARGET_VARIABLES\"].values()))\n",
    "model_training_description[\"FMAPS\"] = (32,32,64,64)\n",
    "\n",
    "\n",
    "model_training_description[\"ACTIVATION\"] = torch.nn.ReLU\n",
    "model_training_description[\"NORMALIZATION\"] = torch.nn.BatchNorm2d  # IcoBatchNorm2d \n",
    "\n",
    "\n",
    "model_training_description[\"OPTIMIZER\"] = \"Adam\"\n",
    "\n",
    "model_training_description[\"DEVICE\"] = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n",
    "\n",
    "model_training_description[\"USE_CYLINDRICAL_PADDING\"] = True\n",
    "model_training_description[\"USE_COORD_CONV\"] = True\n",
    "model_training_description[\"LOSS\"] = \"Masked_AreaWeightedMSELoss\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "711a1525",
   "metadata": {},
   "outputs": [],
   "source": [
    "n_runs = 5\n",
    "for i in range(n_runs):\n",
    "    model_training_description[\"RUN_NR\"] = i\n",
    "    unet = train_unet(description, model_training_description, output_folder)\n",
    "    predict_save_unet(description, model_training_description, output_folder, unet, output_folder)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "224a831a",
   "metadata": {},
   "source": [
    "Individual models for different cases:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "e6d73afb",
   "metadata": {},
   "outputs": [],
   "source": [
    "base_folder = \"Datasets\"\n",
    "output_folder = \"Output/Reproduce_monthly\"\n",
    "\n",
    "description = {}\n",
    "\n",
    "description[\"DATASETS_USED\"] = [\"isotopes\", \n",
    "                                \"tsurf\",\n",
    "                                \"prec\"]\n",
    "\n",
    "description[\"PREDICTOR_VARIABLES\"] = {\"tsurf\": [\"tsurf\"],\n",
    "                                      \"prec\": [\"prec\"]}\n",
    "\n",
    "description[\"TARGET_VARIABLES\"] = {\"isotopes\": [\"d18O\"]}\n",
    "\n",
    "description[\"DATASETS_NO_GAPS\"] = [\"isotopes\", \n",
    "                                   \"tsurf\", \n",
    "                                   \"prec\", \n",
    "                                   \"slp\"]\n",
    "\n",
    "description[\"CLIMATE_MODEL\"] = \"iHadCM3\"\n",
    "description[\"GRID_TYPE\"] = \"Flat\"\n",
    "\n",
    "description[\"START_YEAR\"] = 850\n",
    "description[\"END_YEAR\"] = 1850\n",
    "description[\"LATITUDES_SLICE\"] = [1,-1]\n",
    "\n",
    "description[\"SPLIT_YEAR\"] = 1750\n",
    "description[\"DO_SHUFFLE\"] = False\n",
    "description[\"PRECIP_WEIGHTING\"] = False\n",
    "\n",
    "\n",
    "description[\"TIMESCALE\"] = \"MONTHLY\"\n",
    "description[\"MONTHS_USED_IN_PREDICTION\"] = np.sort([0]).tolist()\n",
    "\n",
    "### MODEL_TRAINING ###############################################\n",
    "\n",
    "model_training_description = {}\n",
    "model_training_description[\"S_MODE_PREDICTORS\"] = [\"Pixelwise\",\"Pixelwise\"] # how to standardize the given variables\n",
    "model_training_description[\"S_MODE_TARGETS\"] = [\"Pixelwise\"]\n",
    "\n",
    "model_training_description[\"DATASET_FOLDER\"] = output_folder\n",
    "\n",
    "model_training_description[\"MODEL_TYPE\"] = \"UNet_Flat\"\n",
    "model_training_description[\"CREATE_VALIDATIONSET\"] = True\n",
    "model_training_description[\"SHUFFLE_VALIDATIONSET\"] = True\n",
    "\n",
    "\n",
    "# training parameters\n",
    "\n",
    "model_training_description[\"DEPTH\"] = 3\n",
    "model_training_description[\"NUM_EPOCHS\"] = \"early_stopping\"  # 20\n",
    "model_training_description[\"PATIENCE\"] = 5\n",
    "model_training_description[\"BATCH_SIZE\"] = 8\n",
    "model_training_description[\"LEARNING_RATE\"] = 1e-3  # 0.002637 # 5e-3  # use either this or default ADAM learning rate\n",
    "\n",
    "# model parameters\n",
    "model_training_description[\"DEPTH\"] = 3\n",
    "model_training_description[\"IN_CHANNELS\"] = len(util.flatten(description[\"PREDICTOR_VARIABLES\"].values()))\n",
    "model_training_description[\"CHANNELS_FIRST_CONV\"] = 32\n",
    "model_training_description[\"OUT_CHANNELS\"] = len(util.flatten(description[\"TARGET_VARIABLES\"].values()))\n",
    "model_training_description[\"FMAPS\"] = (32,32,64,64)\n",
    "\n",
    "\n",
    "model_training_description[\"ACTIVATION\"] = torch.nn.ReLU\n",
    "model_training_description[\"NORMALIZATION\"] = torch.nn.BatchNorm2d  # IcoBatchNorm2d \n",
    "\n",
    "\n",
    "model_training_description[\"OPTIMIZER\"] = \"Adam\"\n",
    "\n",
    "model_training_description[\"DEVICE\"] = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n",
    "\n",
    "model_training_description[\"USE_CYLINDRICAL_PADDING\"] = True\n",
    "model_training_description[\"USE_COORD_CONV\"] = True\n",
    "model_training_description[\"LOSS\"] = \"Masked_AreaWeightedMSELoss\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "45c8a3c9",
   "metadata": {},
   "outputs": [],
   "source": [
    "n_runs = 5\n",
    "for j in range(n_runs):\n",
    "    model_training_description[\"RUN_NR\"] = j\n",
    "    for i in range(12):\n",
    "        description[\"MONTHS_USED\"] = np.sort([i]).tolist()\n",
    "        unet = train_unet(description, model_training_description, output_folder)\n",
    "        predict_save_unet(description, model_training_description, output_folder, unet, output_folder)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "d8acd915",
   "metadata": {},
   "source": [
    "# Evaluate the results:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "c79cde5f",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "5 matching runs found\n",
      "5 matching runs found\n"
     ]
    }
   ],
   "source": [
    "base_folder = \"Datasets\"\n",
    "output_folder = \"Output/Reproduce_monthly\"\n",
    "\n",
    "c_all = {\n",
    "    \"DATASET_DESCRIPTION\": {'MONTHS_USED': [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11],\n",
    "                            'MONTHS_USED_IN_PREDICTION': [0]},\n",
    "    \"MODEL_TRAINING_DESCRIPTION\": {}\n",
    "}\n",
    "\n",
    "descriptions, predictions, gt, masks = load_data_for_comparison(output_folder, c_all, do_split=True)\n",
    "\n",
    "c_all_use_p = {\n",
    "    \"DATASET_DESCRIPTION\": {'MONTHS_USED': [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11],\n",
    "                            'MONTHS_USED_IN_PREDICTION': [-1, 0]},\n",
    "    \"MODEL_TRAINING_DESCRIPTION\": {}\n",
    "}\n",
    "\n",
    "descriptions_p, predictions_p, gt_p, masks_p = load_data_for_comparison(output_folder, c_all_use_p, do_split=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "8775e992",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "5 matching runs found\n",
      "5 matching runs found\n",
      "5 matching runs found\n",
      "5 matching runs found\n",
      "5 matching runs found\n",
      "5 matching runs found\n",
      "5 matching runs found\n",
      "5 matching runs found\n",
      "5 matching runs found\n",
      "5 matching runs found\n",
      "5 matching runs found\n",
      "5 matching runs found\n"
     ]
    }
   ],
   "source": [
    "r2_single = []\n",
    "\n",
    "for i in range(12):\n",
    "    r2_single.append([])\n",
    "    c_single = {\n",
    "        \"DATASET_DESCRIPTION\": {'MONTHS_USED': [i],\n",
    "                                'MONTHS_USED_IN_PREDICTION': [0]},\n",
    "        \"MODEL_TRAINING_DESCRIPTION\": {}\n",
    "    }\n",
    "    descriptions_s, predictions_s, gt_s, masks_s = load_data_for_comparison(output_folder, c_single, do_split=True)\n",
    "    for j in range(3):\n",
    "        r2 = get_r2(predictions_s[j][i], gt_s[j][i], masks_s[j][i])\n",
    "        r2_single[-1].append(get_weighted_average(r2, descriptions[j][\"DATASET_DESCRIPTION\"])[0])\n",
    "        \n",
    "r2_single = np.mean(np.array(r2_single),axis=1)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "a5606c10",
   "metadata": {},
   "source": [
    "On the monthly time scale there is an annual cycle in the isotopes. This cycle can be trivially explained by just using the climatological mean. In order to not include this part into the computation of the $R^2$ score, we compute the $R^2$ score separately for each month."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "7723e752",
   "metadata": {},
   "outputs": [],
   "source": [
    "r2_all = []\n",
    "r2_all_use_p = []\n",
    "\n",
    "for j in range(3):\n",
    "    r2_all.append([])\n",
    "    r2_all_use_p.append([])\n",
    "    for i in range(12):\n",
    "        r2 = get_r2(predictions[j][i], gt[j][i], masks[j][i])\n",
    "        r2_p = get_r2(predictions_p[j][i], gt_p[j][i], masks_p[j][i])\n",
    "        r2_all[-1].append(get_weighted_average(r2, descriptions[j][\"DATASET_DESCRIPTION\"])[0])\n",
    "        r2_all_use_p[-1].append(get_weighted_average(r2_p, descriptions_p[j][\"DATASET_DESCRIPTION\"])[0])\n",
    "        \n",
    "r2_all = np.mean(np.array(r2_all),axis=0)\n",
    "r2_all_use_p = np.mean(np.array(r2_all_use_p),axis=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "77fbfe3a",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.legend.Legend at 0x20ac963a340>"
      ]
     },
     "execution_count": 7,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "\n",
    "plt.plot(r2_all, label=\"UNet, all\")\n",
    "plt.xlabel(\"Month in year\")\n",
    "plt.ylabel(\"$R^2$ score\")\n",
    "plt.plot(r2_all_use_p, label=\"UNet, all, use previous month\")\n",
    "plt.plot(r2_single, label=\"UNet, single\")\n",
    "plt.legend()"
   ]
  }
 ],
 "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": "3.8.0"
  },
  "vscode": {
   "interpreter": {
    "hash": "90b409070de10f2c3c5a8c504ecede612c695480a2027d21529fbe1a33a534d8"
   }
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}