Upload 3 files

load_ssl4eo_s.ipynb

@@ -0,0 +1,426 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### Load from raw data (geotif)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "import kornia\n",
+    "import numpy as np\n",
+    "from kornia.augmentation import AugmentationSequential\n",
+    "from torch.utils.data import Dataset, DataLoader\n",
+    "from ssl4eo_s_dataset import SSL4EO_S\n",
+    "import time\n",
+    "import os\n",
+    "from tqdm import tqdm"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "fnames_path = '../data/example_100_grids/fnames_sampled_union.json.gz'\n",
+    "root_dir = '../data/example_100_grids/'"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Grid ID: ('0913063_-154.25_68.50',)\n",
+      "dict_keys(['s1_grd', 's2_toa', 's3_olci', 's5p_co', 's5p_no2', 's5p_o3', 's5p_so2', 'dem'])\n",
+      "dict_keys(['s1_grd', 's2_toa', 's3_olci', 's5p_co', 's5p_no2', 's5p_o3', 's5p_so2', 'dem'])\n",
+      "### S1 GRD ###\n",
+      "Number of s1 local patches: 1    Number of time stamps for first local patch: 1\n",
+      "Example for one image: torch.Size([1, 2, 224, 224]) torch.float32 ('0913063_-154.25_68.50/1092252_-154.25_68.50/20210530',)\n",
+      "### S2 TOA ###\n",
+      "Number of s2 local patches: 1    Number of time stamps for first local patch: 4\n",
+      "Example for one image: torch.Size([1, 13, 224, 224]) torch.int16 ('0913063_-154.25_68.50/1092252_-154.25_68.50/20191002',)\n",
+      "### S3 OLCI ###\n",
+      "Number of s3 time stamps: 7\n",
+      "Example for one image: torch.Size([1, 21, 96, 96]) torch.float32 ('0913063_-154.25_68.50/20210411',)\n",
+      "### S5P ###\n",
+      "Number of s5p time stamps for CO/NO2/O3/SO2: 7 5 11 4\n",
+      "Example for one CO image: torch.Size([1, 1, 28, 28]) torch.float32 ('0913063_-154.25_68.50/20210401',)\n",
+      "Example for one NO2 image: torch.Size([1, 1, 28, 28]) torch.float32 ('0913063_-154.25_68.50/20210401',)\n",
+      "Example for one O3 image: torch.Size([1, 1, 28, 28]) torch.float32 ('0913063_-154.25_68.50/20210101',)\n",
+      "Example for one SO2 image: torch.Size([1, 1, 28, 28]) torch.float32 ('0913063_-154.25_68.50/20210501',)\n",
+      "### DEM ###\n",
+      "One DEM image for the grid: torch.Size([1, 1, 960, 960]) torch.float32 ('0913063_-154.25_68.50',)\n",
+      "Time: 94.70435643196106\n"
+     ]
+    }
+   ],
+   "source": [
+    "transform_s1 = AugmentationSequential(\n",
+    "    #kornia.augmentation.SmallestMaxSize(264),\n",
+    "    kornia.augmentation.CenterCrop(224),\n",
+    ")\n",
+    "transform_s2 = AugmentationSequential(\n",
+    "    #kornia.augmentation.SmallestMaxSize(264),\n",
+    "    kornia.augmentation.CenterCrop(224),\n",
+    ")\n",
+    "transform_s3 = AugmentationSequential(\n",
+    "    kornia.augmentation.SmallestMaxSize(96),\n",
+    "    kornia.augmentation.CenterCrop(96),\n",
+    ")\n",
+    "transform_s5p = AugmentationSequential(\n",
+    "    kornia.augmentation.SmallestMaxSize(28),\n",
+    "    kornia.augmentation.CenterCrop(28),\n",
+    ")\n",
+    "transform_dem = AugmentationSequential(\n",
+    "    kornia.augmentation.SmallestMaxSize(960),\n",
+    "    kornia.augmentation.CenterCrop(960),\n",
+    ")\n",
+    "\n",
+    "ssl4eo_s = SSL4EO_S(fnames_path, root_dir, transform_s1=transform_s1, transform_s2=transform_s2, transform_s3=transform_s3, transform_s5p=transform_s5p, transform_dem=transform_dem)\n",
+    "dataloader = DataLoader(ssl4eo_s, batch_size=1, shuffle=True, num_workers=4) # batch size can only be 1 because of varying number of images per grid\n",
+    "\n",
+    "start_time = time.time()\n",
+    "\n",
+    "for i, (sample, meta_data) in enumerate(dataloader):\n",
+    "    if i == 0:\n",
+    "        print('Grid ID:', meta_data['dem'][0])\n",
+    "        print(sample.keys())\n",
+    "        print(meta_data.keys())\n",
+    "\n",
+    "        print('### S1 GRD ###')\n",
+    "        print('Number of s1 local patches:', len(meta_data['s1_grd']), '  ', 'Number of time stamps for first local patch:', len(meta_data['s1_grd'][0]))\n",
+    "        print('Example for one image:', sample['s1_grd'][0][0].shape, sample['s1_grd'][0][0].dtype, meta_data['s1_grd'][0][0])\n",
+    "        print('### S2 TOA ###')\n",
+    "        print('Number of s2 local patches:', len(meta_data['s2_toa']), '  ', 'Number of time stamps for first local patch:', len(meta_data['s2_toa'][0]))\n",
+    "        print('Example for one image:', sample['s2_toa'][0][0].shape, sample['s2_toa'][0][0].dtype, meta_data['s2_toa'][0][0])\n",
+    "        print('### S3 OLCI ###')\n",
+    "        print('Number of s3 time stamps:', len(meta_data['s3_olci']))\n",
+    "        print('Example for one image:', sample['s3_olci'][0].shape, sample['s3_olci'][0].dtype, meta_data['s3_olci'][0])\n",
+    "        print('### S5P ###')\n",
+    "        print('Number of s5p time stamps for CO/NO2/O3/SO2:', len(meta_data['s5p_co']), len(meta_data['s5p_no2']), len(meta_data['s5p_o3']), len(meta_data['s5p_so2']))\n",
+    "        print('Example for one CO image:', sample['s5p_co'][0].shape, sample['s5p_co'][0].dtype, meta_data['s5p_co'][0])\n",
+    "        print('Example for one NO2 image:', sample['s5p_no2'][0].shape, sample['s5p_no2'][0].dtype, meta_data['s5p_no2'][0])\n",
+    "        print('Example for one O3 image:', sample['s5p_o3'][0].shape, sample['s5p_o3'][0].dtype, meta_data['s5p_o3'][0])\n",
+    "        print('Example for one SO2 image:', sample['s5p_so2'][0].shape, sample['s5p_so2'][0].dtype, meta_data['s5p_so2'][0])\n",
+    "        print('### DEM ###')\n",
+    "        print('One DEM image for the grid:', sample['dem'].shape, sample['dem'].dtype, meta_data['dem'][0])\n",
+    "    else:\n",
+    "        pass\n",
+    "\n",
+    "print('Time:', time.time()-start_time)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### Load from webdataset (npy)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import webdataset as wds\n",
+    "\n",
+    "## 85 grids version\n",
+    "webdataset_npy_dir = '../data/example_100_grids_cleaned/webdataset_obj_grid_pad_npy/'\n",
+    "# shards path\n",
+    "shards_path = os.path.join(webdataset_npy_dir, 'example-{000000..000008}.tar')\n",
+    "batch_size = 1\n",
+    "#shuffle = 8\n",
+    "train_transform = None\n",
+    "\n",
+    "# pytorch dataset\n",
+    "dataset = (\n",
+    "        wds.WebDataset(shards_path,shardshuffle=True)\n",
+    "        #.shuffle(shuffle)\n",
+    "        .decode()\n",
+    "        #.to_tuple(\"json\")\n",
+    "        .to_tuple(\"s1_grd.npy\", \n",
+    "                \"s2_toa.npy\", \n",
+    "                \"s3_olci.npy\", \n",
+    "                \"s5p.npy\", \n",
+    "                \"dem.npy\",\n",
+    "                \"json\") # also possible to only extract part of the data\n",
+    "        #.map(sample_one_grid)\n",
+    "        #.map_tuple(train_transform, identity)\n",
+    "    ).batched(batch_size, partial=False)\n",
+    "\n",
+    "# pytorch dataloader\n",
+    "dataloader = torch.utils.data.DataLoader(dataset, num_workers=4, batch_size=None)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 7,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "85it [00:36,  2.30it/s]"
+     ]
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Time: 36.96s\n"
+     ]
+    },
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "\n"
+     ]
+    }
+   ],
+   "source": [
+    "start_time = time.time()\n",
+    "for i, data in tqdm(enumerate(dataloader)):\n",
+    "    #key,img = data\n",
+    "    pass\n",
+    "print(f\"Time: {time.time()-start_time:.2f}s\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### Load from webdataset (pth)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "metadata": {},
+   "outputs": [
+    {
+     "ename": "NameError",
+     "evalue": "name 'wds' is not defined",
+     "output_type": "error",
+     "traceback": [
+      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
+      "\u001b[0;31mNameError\u001b[0m                                 Traceback (most recent call last)",
+      "Cell \u001b[0;32mIn[8], line 44\u001b[0m\n\u001b[1;32m     39\u001b[0m     \u001b[38;5;28;01mreturn\u001b[39;00m sample\n\u001b[1;32m     42\u001b[0m \u001b[38;5;66;03m# pytorch dataset\u001b[39;00m\n\u001b[1;32m     43\u001b[0m dataset \u001b[38;5;241m=\u001b[39m (\n\u001b[0;32m---> 44\u001b[0m         \u001b[43mwds\u001b[49m\u001b[38;5;241m.\u001b[39mWebDataset(shards_path,shardshuffle\u001b[38;5;241m=\u001b[39m\u001b[38;5;28;01mTrue\u001b[39;00m)\n\u001b[1;32m     45\u001b[0m         \u001b[38;5;66;03m#.shuffle(shuffle)\u001b[39;00m\n\u001b[1;32m     46\u001b[0m         \u001b[38;5;241m.\u001b[39mdecode()\n\u001b[1;32m     47\u001b[0m         \u001b[38;5;241m.\u001b[39mselect(has_all_modalities) \u001b[38;5;66;03m# only keep samples with all modalities\u001b[39;00m\n\u001b[1;32m     48\u001b[0m         \u001b[38;5;241m.\u001b[39mmap(sample_one_local_patch) \u001b[38;5;66;03m# sample one local patch for S1 and S2\u001b[39;00m\n\u001b[1;32m     49\u001b[0m         \u001b[38;5;66;03m#.to_tuple(\"json\")\u001b[39;00m\n\u001b[1;32m     50\u001b[0m         \u001b[38;5;241m.\u001b[39mto_tuple(\u001b[38;5;124m\"\u001b[39m\u001b[38;5;124ms1_grd.pth\u001b[39m\u001b[38;5;124m\"\u001b[39m, \n\u001b[1;32m     51\u001b[0m                 \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124ms2_toa.pth\u001b[39m\u001b[38;5;124m\"\u001b[39m, \n\u001b[1;32m     52\u001b[0m                 \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124ms3_olci.pth\u001b[39m\u001b[38;5;124m\"\u001b[39m, \n\u001b[1;32m     53\u001b[0m                 \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124ms5p_co.pth\u001b[39m\u001b[38;5;124m\"\u001b[39m, \n\u001b[1;32m     54\u001b[0m                 \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124ms5p_no2.pth\u001b[39m\u001b[38;5;124m\"\u001b[39m,\n\u001b[1;32m     55\u001b[0m                 \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124ms5p_o3.pth\u001b[39m\u001b[38;5;124m\"\u001b[39m,\n\u001b[1;32m     56\u001b[0m                 \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124ms5p_so2.pth\u001b[39m\u001b[38;5;124m\"\u001b[39m,\n\u001b[1;32m     57\u001b[0m                 \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mdem.pth\u001b[39m\u001b[38;5;124m\"\u001b[39m,\n\u001b[1;32m     58\u001b[0m                 \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mjson\u001b[39m\u001b[38;5;124m\"\u001b[39m) \u001b[38;5;66;03m# also possible to only extract part of the data\u001b[39;00m\n\u001b[1;32m     59\u001b[0m         \u001b[38;5;66;03m#.map(sample_one_grid)\u001b[39;00m\n\u001b[1;32m     60\u001b[0m         \u001b[38;5;66;03m#.map_tuple(train_transform, identity)\u001b[39;00m\n\u001b[1;32m     61\u001b[0m     )\u001b[38;5;241m.\u001b[39mbatched(batch_size, partial\u001b[38;5;241m=\u001b[39m\u001b[38;5;28;01mFalse\u001b[39;00m)\n\u001b[1;32m     63\u001b[0m \u001b[38;5;66;03m# pytorch dataloader\u001b[39;00m\n\u001b[1;32m     64\u001b[0m dataloader \u001b[38;5;241m=\u001b[39m torch\u001b[38;5;241m.\u001b[39mutils\u001b[38;5;241m.\u001b[39mdata\u001b[38;5;241m.\u001b[39mDataLoader(dataset, num_workers\u001b[38;5;241m=\u001b[39m\u001b[38;5;241m4\u001b[39m, batch_size\u001b[38;5;241m=\u001b[39m\u001b[38;5;28;01mNone\u001b[39;00m)\n",
+      "\u001b[0;31mNameError\u001b[0m: name 'wds' is not defined"
+     ]
+    }
+   ],
+   "source": [
+    "import random\n",
+    "\n",
+    "webdataset_pth_dir = '../data/example_100_grids/reformat_webdataset_pth/'\n",
+    "# shards path\n",
+    "shards_path = os.path.join(webdataset_pth_dir, 'example-{000000..000009}.tar')\n",
+    "batch_size = 1\n",
+    "#shuffle = 8\n",
+    "#train_transform = None\n",
+    "\n",
+    "def has_all_modalities(sample):\n",
+    "    required_keys = [\n",
+    "        \"s1_grd.pth\", \n",
+    "        \"s2_toa.pth\", \n",
+    "        \"s3_olci.pth\", \n",
+    "        \"s5p_co.pth\", \n",
+    "        \"s5p_no2.pth\",\n",
+    "        \"s5p_o3.pth\",\n",
+    "        \"s5p_so2.pth\",\n",
+    "        \"dem.pth\",\n",
+    "        \"json\"\n",
+    "    ]\n",
+    "    return all(key in sample for key in required_keys)\n",
+    "\n",
+    "def sample_one_local_patch(sample):\n",
+    "    s1 = sample[\"s1_grd.pth\"]\n",
+    "    s2 = sample[\"s2_toa.pth\"]\n",
+    "    meta_s1 = sample[\"json\"][\"s1_grd\"]\n",
+    "    meta_s2 = sample[\"json\"][\"s2_toa\"]\n",
+    "    #idx = torch.randint(0, s1.shape[0], (1,))\n",
+    "    idx = random.randint(0, s1.shape[0]-1)\n",
+    "    s1_new = s1[idx]\n",
+    "    s2_new = s2[idx]\n",
+    "    meta_s1_new = meta_s1[idx]\n",
+    "    meta_s2_new = meta_s2[idx]\n",
+    "    sample[\"s1_grd.pth\"] = s1_new\n",
+    "    sample[\"s2_toa.pth\"] = s2_new\n",
+    "    sample[\"json\"][\"s1_grd\"] = meta_s1_new\n",
+    "    sample[\"json\"][\"s2_toa\"] = meta_s2_new\n",
+    "    return sample\n",
+    "\n",
+    "\n",
+    "# pytorch dataset\n",
+    "dataset = (\n",
+    "        wds.WebDataset(shards_path,shardshuffle=True)\n",
+    "        #.shuffle(shuffle)\n",
+    "        .decode()\n",
+    "        .select(has_all_modalities) # only keep samples with all modalities\n",
+    "        .map(sample_one_local_patch) # sample one local patch for S1 and S2\n",
+    "        #.to_tuple(\"json\")\n",
+    "        .to_tuple(\"s1_grd.pth\", \n",
+    "                \"s2_toa.pth\", \n",
+    "                \"s3_olci.pth\", \n",
+    "                \"s5p_co.pth\", \n",
+    "                \"s5p_no2.pth\",\n",
+    "                \"s5p_o3.pth\",\n",
+    "                \"s5p_so2.pth\",\n",
+    "                \"dem.pth\",\n",
+    "                \"json\") # also possible to only extract part of the data\n",
+    "        #.map(sample_one_grid)\n",
+    "        #.map_tuple(train_transform, identity)\n",
+    "    ).batched(batch_size, partial=False)\n",
+    "\n",
+    "# pytorch dataloader\n",
+    "dataloader = torch.utils.data.DataLoader(dataset, num_workers=4, batch_size=None)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 24,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "0it [00:00, ?it/s]"
+     ]
+    },
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "100it [00:51,  1.94it/s]"
+     ]
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Time: 51.68s\n"
+     ]
+    },
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "\n"
+     ]
+    }
+   ],
+   "source": [
+    "start_time = time.time()\n",
+    "for i, data in tqdm(enumerate(dataloader)):\n",
+    "    #key,img = data\n",
+    "    pass\n",
+    "print(f\"Time: {time.time()-start_time:.2f}s\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### Load from lmdb"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from ssl4eo_s_lmdb_dataset import SSL4EO_S_lmdb\n",
+    "\n",
+    "root_dir = '../data/example_100_grids/'\n",
+    "lmdb_path = root_dir + 'reformat_lmdb/' + 'ssl4eo_s_data.lmdb'\n",
+    "key_path = root_dir + 'reformat_lmdb/' + 'ssl4eo_s_key.csv'\n",
+    "mode = ['s1_grd', 's2_toa', 's3_olci', 's5p_co', 's5p_no2', 's5p_so2', 's5p_o3', 'dem']\n",
+    "\n",
+    "dataset = SSL4EO_S_lmdb(lmdb_path=lmdb_path, key_path=key_path, mode=mode)\n",
+    "dataloader = DataLoader(dataset, batch_size=1, shuffle=True, num_workers=4)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 7,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "/home/wangyi111/miniconda3/envs/pytorch/lib/python3.10/site-packages/torch/utils/data/_utils/collate.py:171: UserWarning: The given NumPy array is not writable, and PyTorch does not support non-writable tensors. This means writing to this tensor will result in undefined behavior. You may want to copy the array to protect its data or make it writable before converting it to a tensor. This type of warning will be suppressed for the rest of this program. (Triggered internally at ../torch/csrc/utils/tensor_numpy.cpp:206.)\n",
+      "  return collate([torch.as_tensor(b) for b in batch], collate_fn_map=collate_fn_map)\n",
+      "0it [00:00, ?it/s]/home/wangyi111/miniconda3/envs/pytorch/lib/python3.10/site-packages/torch/utils/data/_utils/collate.py:171: UserWarning: The given NumPy array is not writable, and PyTorch does not support non-writable tensors. This means writing to this tensor will result in undefined behavior. You may want to copy the array to protect its data or make it writable before converting it to a tensor. This type of warning will be suppressed for the rest of this program. (Triggered internally at ../torch/csrc/utils/tensor_numpy.cpp:206.)\n",
+      "  return collate([torch.as_tensor(b) for b in batch], collate_fn_map=collate_fn_map)\n",
+      "/home/wangyi111/miniconda3/envs/pytorch/lib/python3.10/site-packages/torch/utils/data/_utils/collate.py:171: UserWarning: The given NumPy array is not writable, and PyTorch does not support non-writable tensors. This means writing to this tensor will result in undefined behavior. You may want to copy the array to protect its data or make it writable before converting it to a tensor. This type of warning will be suppressed for the rest of this program. (Triggered internally at ../torch/csrc/utils/tensor_numpy.cpp:206.)\n",
+      "  return collate([torch.as_tensor(b) for b in batch], collate_fn_map=collate_fn_map)\n",
+      "/home/wangyi111/miniconda3/envs/pytorch/lib/python3.10/site-packages/torch/utils/data/_utils/collate.py:171: UserWarning: The given NumPy array is not writable, and PyTorch does not support non-writable tensors. This means writing to this tensor will result in undefined behavior. You may want to copy the array to protect its data or make it writable before converting it to a tensor. This type of warning will be suppressed for the rest of this program. (Triggered internally at ../torch/csrc/utils/tensor_numpy.cpp:206.)\n",
+      "  return collate([torch.as_tensor(b) for b in batch], collate_fn_map=collate_fn_map)\n",
+      "100it [00:04, 23.99it/s]"
+     ]
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Time: 4.21s\n"
+     ]
+    },
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "\n"
+     ]
+    }
+   ],
+   "source": [
+    "start_time = time.time()\n",
+    "for i, data in tqdm(enumerate(dataloader)):\n",
+    "    #key,img = data\n",
+    "    pass\n",
+    "print(f\"Time: {time.time()-start_time:.2f}s\")"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "pytorch",
+   "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": "3.10.13"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}

ssl4eo_s_dataset.py

@@ -0,0 +1,234 @@
+import os
+import gzip
+import json
+import numpy as np
+import rasterio
+import re
+from torch.utils.data import Dataset, DataLoader
+import torch
+#from cvtorchvision import cvtransforms
+from kornia.augmentation import AugmentationSequential
+import kornia
+import argparse
+
+
+class SSL4EO_S(Dataset):
+    def __init__(self, fnames_path, root_dir, modality=['s1_grd', 's2_toa', 's3_olci', 's5p_co', 's5p_no2', 's5p_so2', 's5p_o3', 'dem'], transform_s1=None, transform_s2=None, transform_s3=None, transform_s5p=None, transform_dem=None):
+        with gzip.open(fnames_path, 'rt', encoding='utf-8') as gz_file:
+            self.fnames_json = json.load(gz_file)
+        self.grid_ids = list(self.fnames_json.keys())
+        self.root_dir = root_dir
+        self.transform_s1 = transform_s1
+        self.transform_s2 = transform_s2
+        self.transform_s3 = transform_s3
+        self.transform_s5p = transform_s5p
+        self.transform_dem = transform_dem
+        self.modality = modality
+
+    def __len__(self):
+        return len(self.grid_ids)
+
+    def get_s1_s2(self,grid_id,modality):
+        arrays = []
+        meta_data = []
+        local_grids = list(self.fnames_json[grid_id][modality].keys())
+        grid_id_coord = self.fnames_json[grid_id]['grid_id_coord']
+        for local_grid in local_grids:
+            local_fpaths = self.fnames_json[grid_id][modality][local_grid]
+            imgs = []
+            meta = []
+            for local_fpath in local_fpaths:
+                with rasterio.open(os.path.join(self.root_dir, local_fpath)) as src:
+                    img = src.read()
+                if modality=='s1_grd' and self.transform_s1:
+                    #img = self.transform_s1(np.transpose(img, (1, 2, 0)))
+                    img = torch.from_numpy(img).unsqueeze(0)
+                    img = self.transform_s1(img).squeeze(0)
+                elif modality=='s2_toa' and self.transform_s2:
+                    #img = self.transform_s2(np.transpose(img.astype(np.int16), (1, 2, 0)))
+                    img = torch.from_numpy(img.astype(np.int16)).unsqueeze(0)
+                    img = self.transform_s2(img.to(torch.float16)).squeeze(0)
+                imgs.append(img)
+                fname = local_fpath.split('/')[-1]
+                date = re.search(r'(\d{8})T', fname).group(1)
+                meta_info = f"{grid_id_coord}/{local_grid}/{date}"
+                meta.append(meta_info)
+            arrays.append(imgs)
+            meta_data.append(meta)
+        return arrays, meta_data
+
+    def get_s3(self,grid_id):
+        arrays = []
+        meta_data = []
+        fpaths = self.fnames_json[grid_id]['s3_olci']
+        grid_id_coord = self.fnames_json[grid_id]['grid_id_coord']
+        for fpath in fpaths:
+            with rasterio.open(os.path.join(self.root_dir, fpath)) as src:
+                img = src.read()
+            if self.transform_s3:
+                #img = self.transform_s3(np.transpose(img, (1, 2, 0)))
+                img = torch.from_numpy(img).unsqueeze(0)
+                img = self.transform_s3(img).squeeze(0)
+            arrays.append(img)
+            fname = fpath.split('/')[-1]
+            date = re.search(r'(\d{8})T', fname).group(1)
+            meta_info = f"{grid_id_coord}/{date}"
+            meta_data.append(meta_info)
+        return arrays, meta_data
+
+    def get_s5p(self,grid_id,modality):
+        arrays = []
+        meta_data = []
+        fpaths = self.fnames_json[grid_id][modality]
+        grid_id_coord = self.fnames_json[grid_id]['grid_id_coord']
+        for fpath in fpaths:
+            with rasterio.open(os.path.join(self.root_dir, fpath)) as src:
+                img = src.read()
+            if self.transform_s5p:
+                #img = self.transform_s5p(np.transpose(img, (1, 2, 0)))
+                img = torch.from_numpy(img).unsqueeze(0)
+                img = self.transform_s5p(img).squeeze(0)
+            arrays.append(img)
+            fname = fpath.split('/')[-1]
+            match = re.search(r'(\d{4})-(\d{2})-(\d{2})', fname)
+            date = f"{match.group(1)}{match.group(2)}{match.group(3)}"
+            meta_info = f"{grid_id_coord}/{date}"
+            meta_data.append(meta_info)
+        return arrays, meta_data
+
+    def get_dem(self,grid_id):
+        fpath = self.fnames_json[grid_id]['dem'][0]
+        with rasterio.open(os.path.join(self.root_dir, fpath)) as src:
+            img = src.read()
+        if self.transform_dem:
+            #img = self.transform_dem(np.transpose(img, (1, 2, 0)))
+            img = torch.from_numpy(img).unsqueeze(0)
+            img = self.transform_dem(img).squeeze(0)
+        return img
+
+    def __getitem__(self, idx):
+        grid_id = self.grid_ids[idx]
+        grid_id_coord = self.fnames_json[grid_id]['grid_id_coord']
+        sample = {}
+        meta_data = {}
+        # s1
+        if 's1_grd' in self.modality:
+            arr_s1, meta_s1 = self.get_s1_s2(grid_id,'s1_grd')
+            sample['s1_grd'] = arr_s1
+            meta_data['s1_grd'] = meta_s1
+        # s2
+        if 's2_toa' in self.modality:
+            arr_s2, meta_s2 = self.get_s1_s2(grid_id,'s2_toa')
+            sample['s2_toa'] = arr_s2
+            meta_data['s2_toa'] = meta_s2
+        # s3
+        if 's3_olci' in self.modality:
+            arr_s3, meta_s3 = self.get_s3(grid_id)
+            sample['s3_olci'] = arr_s3
+            meta_data['s3_olci'] = meta_s3
+        # s5p_co
+        if 's5p_co' in self.modality:
+            arr_s5p_co, meta_s5p_co = self.get_s5p(grid_id,'s5p_co')
+            sample['s5p_co'] = arr_s5p_co
+            meta_data['s5p_co'] = meta_s5p_co
+        # s5p_no2
+        if 's5p_no2' in self.modality:
+            arr_s5p_no2, meta_s5p_no2 = self.get_s5p(grid_id,'s5p_no2')
+            sample['s5p_no2'] = arr_s5p_no2
+            meta_data['s5p_no2'] = meta_s5p_no2
+        # s5p_o3
+        if 's5p_o3' in self.modality:
+            arr_s5p_o3, meta_s5p_o3 = self.get_s5p(grid_id,'s5p_o3')
+            sample['s5p_o3'] = arr_s5p_o3
+            meta_data['s5p_o3'] = meta_s5p_o3
+        # s5p_so2
+        if 's5p_so2' in self.modality:
+            arr_s5p_so2, meta_s5p_so2 = self.get_s5p(grid_id,'s5p_so2')
+            sample['s5p_so2'] = arr_s5p_so2
+            meta_data['s5p_so2'] = meta_s5p_so2
+        # dem
+        if 'dem' in self.modality:
+            arr_dem = self.get_dem(grid_id)
+            sample['dem'] = arr_dem
+            meta_data['dem'] = grid_id_coord
+
+        return sample, meta_data
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--fnames_path', type=str, default='data_loading/fnames.json.gz')
+    parser.add_argument('--root_dir', type=str, default='data_loading/data')
+    args = parser.parse_args()
+
+    # transform_s1 = cvtransforms.Compose([
+    #     cvtransforms.CenterCrop(224),
+    #     cvtransforms.ToTensor()
+    # ])
+    # transform_s2 = cvtransforms.Compose([
+    #     cvtransforms.CenterCrop(224),
+    #     cvtransforms.ToTensor()
+    # ])
+    # transform_s3 = cvtransforms.Compose([
+    #     cvtransforms.CenterCrop(96),
+    #     cvtransforms.ToTensor()
+    # ])
+    # transform_s5p = cvtransforms.Compose([
+    #     cvtransforms.CenterCrop(28),
+    #     cvtransforms.ToTensor()
+    # ])
+    # transform_dem = cvtransforms.Compose([
+    #     cvtransforms.CenterCrop(960),
+    #     cvtransforms.ToTensor()
+    # ])
+    transform_s1 = AugmentationSequential(
+        #kornia.augmentation.SmallestMaxSize(264),
+        kornia.augmentation.CenterCrop(224),
+    )
+    transform_s2 = AugmentationSequential(
+        #kornia.augmentation.SmallestMaxSize(264),
+        kornia.augmentation.CenterCrop(224),
+    )
+    transform_s3 = AugmentationSequential(
+        kornia.augmentation.SmallestMaxSize(96),
+        kornia.augmentation.CenterCrop(96),
+    )
+    transform_s5p = AugmentationSequential(
+        kornia.augmentation.SmallestMaxSize(28),
+        kornia.augmentation.CenterCrop(28),
+    )
+    transform_dem = AugmentationSequential(
+        kornia.augmentation.SmallestMaxSize(960),
+        kornia.augmentation.CenterCrop(960),
+    )
+
+
+    ssl4eo_s = SSL4EO_S(args.fnames_path, args.root_dir, transform_s1=transform_s1, transform_s2=transform_s2, transform_s3=transform_s3, transform_s5p=transform_s5p, transform_dem=transform_dem)
+    dataloader = DataLoader(ssl4eo_s, batch_size=1, shuffle=True, num_workers=0) # batch size can only be 1 because of varying number of images per grid
+
+    for i, (sample, meta_data) in enumerate(dataloader):
+        #print(i)
+        print('Grid ID:', meta_data['dem'][0])
+        print(sample.keys())
+        print(meta_data.keys())
+
+        
+        print('### S1 GRD ###')
+        print('Number of s1 local patches:', len(meta_data['s1_grd']), '  ', 'Number of time stamps for first local patch:', len(meta_data['s1_grd'][0]))
+        print('Example for one image:', sample['s1_grd'][0][0].shape, meta_data['s1_grd'][0][0])
+        print('### S2 TOA ###')
+        print('Number of s2 local patches:', len(meta_data['s2_toa']), '  ', 'Number of time stamps for first local patch:', len(meta_data['s2_toa'][0]))
+        print('Example for one image:', sample['s2_toa'][0][0].shape, meta_data['s2_toa'][0][0])
+        print('### S3 OLCI ###')
+        print('Number of s3 time stamps:', len(meta_data['s3_olci']))
+        print('Example for one image:', sample['s3_olci'][0].shape, meta_data['s3_olci'][0])
+        print('### S5P ###')
+        print('Number of s5p time stamps for CO/NO2/O3/SO2:', len(meta_data['s5p_co']), len(meta_data['s5p_no2']), len(meta_data['s5p_o3']), len(meta_data['s5p_so2']))
+        print('Example for one CO image:', sample['s5p_co'][0].shape, meta_data['s5p_co'][0])
+        print('Example for one NO2 image:', sample['s5p_no2'][0].shape, meta_data['s5p_no2'][0])
+        print('Example for one O3 image:', sample['s5p_o3'][0].shape, meta_data['s5p_o3'][0])
+        print('Example for one SO2 image:', sample['s5p_so2'][0].shape, meta_data['s5p_so2'][0])
+        print('### DEM ###')
+        print('One DEM image for the grid:', sample['dem'].shape, meta_data['dem'][0])
+
+        break

ssl4eo_s_lmdb_dataset.py

@@ -0,0 +1,200 @@
+import lmdb
+import os
+import torch
+import numpy as np
+from torch.utils.data import Dataset, DataLoader
+import csv
+
+
+class SSL4EO_S_lmdb(Dataset):
+    def __init__(self, lmdb_path, key_path, slurm_job=False, mode=['s1_grd','s2_toa','s3_olci','s5p_co','s5p_no2','s5p_so2','s5p_o3','dem'], s1_transform=None, s2_transform=None, s3_transform=None, s5p_transform=None, dem_transform=None):
+        self.lmdb_path = lmdb_path
+        self.key_path = key_path
+        self.slurm_job = slurm_job
+        self.mode = mode
+        self.s1_transform = s1_transform
+        self.s2_transform = s2_transform
+        self.s3_transform = s3_transform
+        self.s5p_transform = s5p_transform
+        self.dem_transform = dem_transform
+
+        if not self.slurm_job:
+            self.env = lmdb.open(lmdb_path, readonly=True, lock=False, readahead=False, meminit=False)
+        #self.txn = self.env.begin(write=False) # Q: when to close the txn? # 
+        self.keys = {}
+        with open(key_path, 'r') as f:
+            reader = csv.reader(f)
+            for row in reader:
+                modality, meta_info = row[0], row[1]
+                if modality=='s1_grd' or modality=='s2_toa':
+                    _, grid_id, local_grid_id, date = meta_info.split('/')
+                    if grid_id not in self.keys:
+                        self.keys[grid_id] = {}
+                    if modality not in self.keys[grid_id]:
+                        self.keys[grid_id][modality] = {}
+                    if local_grid_id not in self.keys[grid_id][modality]:
+                        self.keys[grid_id][modality][local_grid_id] = []
+                    self.keys[grid_id][modality][local_grid_id].append(meta_info)
+                elif modality=='s3_olci' or modality=='s5p_co' or modality=='s5p_no2' or modality=='s5p_so2' or modality=='s5p_o3':
+                    _, grid_id, date = meta_info.split('/')
+                    if grid_id not in self.keys:
+                        self.keys[grid_id] = {}
+                    if modality not in self.keys[grid_id]:
+                        self.keys[grid_id][modality] = []
+                    self.keys[grid_id][modality].append(meta_info)
+                elif modality=='dem':
+                    _, grid_id = meta_info.split('/')
+                    if grid_id not in self.keys:
+                        self.keys[grid_id] = {}
+                    if modality not in self.keys[grid_id]:
+                        self.keys[grid_id][modality] = []
+                    self.keys[grid_id][modality].append(meta_info)
+        self.indices = list(self.keys.keys())
+
+    def __len__(self):
+        return len(self.indices)
+
+    def _init_db(self):
+        self.env = lmdb.open(self.lmdb_path, max_readers=1, readonly=True, lock=False, readahead=False, meminit=False)
+
+    def __getitem__(self, idx): 
+        if self.slurm_job:
+            # Delay loading LMDB data until after initialization
+            if self.env is None:
+                self._init_db()
+        # get all images in a random local grid in one era5 grid (for batch loading)
+        grid_id = self.indices[idx]
+        grid_keys = self.keys[grid_id]
+        sample = {}
+        meta_info = {}
+
+        with self.env.begin(write=False) as txn:
+            # s1
+            if 's1_grd' in self.mode:
+                sample['s1_grd'] = []
+                meta_info['s1_grd'] = []
+                if 's1_grd' in grid_keys:
+                    local_grids = list(grid_keys['s1_grd'].keys()) # list of local grid ids
+                    for local_grid_id in local_grids:
+                        local_keys = grid_keys['s1_grd'][local_grid_id] # list of 4 keys
+                        local_meta_info = []
+                        local_imgs = []
+                        for key in local_keys:
+                            #print(key)
+                            img_bytes = txn.get(key.encode('utf-8'))
+                            img = np.frombuffer(img_bytes, dtype=np.float32).reshape(264, 264, 2)
+                            if self.s1_transform:
+                                img = self.s1_transform(img)
+                            local_meta_info.append(key)
+                            local_imgs.append(img)
+                        ## pad time stamps to 4
+                        #if len(s1_meta_info) < 4:
+                        #    s1_meta_info += [s1_meta_info[-1]] * (4 - len(s1_meta_info))
+                        #    s1_imgs += [s1_imgs[-1]] * (4 - len(s1_imgs))
+                        sample['s1_grd'].append(local_imgs)
+                        meta_info['s1_grd'].append(local_meta_info)
+
+            # s2
+            if 's2_toa' in self.mode:
+                sample['s2_toa'] = []
+                meta_info['s2_toa'] = []
+                if 's2_toa' in grid_keys:
+                    local_grids = list(grid_keys['s2_toa'].keys())
+                    for local_grid_id in local_grids:
+                        local_keys = grid_keys['s2_toa'][local_grid_id]
+                        local_meta_info = []
+                        local_imgs = []
+                        for key in local_keys:
+                            img_bytes = txn.get(key.encode('utf-8'))
+                            img = np.frombuffer(img_bytes, dtype=np.int16).reshape(264, 264, 13)
+                            if self.s2_transform:
+                                img = self.s2_transform(img)
+                            local_meta_info.append(key)
+                            local_imgs.append(img)
+                        sample['s2_toa'].append(local_imgs)
+                        meta_info['s2_toa'].append(local_meta_info)
+
+            # s3
+            if 's3_olci' in self.mode:
+                sample['s3_olci'] = []
+                meta_info['s3_olci'] = []
+                if 's3_olci' in grid_keys:
+                    local_keys = grid_keys['s3_olci']
+                    for key in local_keys:
+                        img_bytes = txn.get(key.encode('utf-8'))
+                        img = np.frombuffer(img_bytes, dtype=np.float32).reshape(96, 96, 21)
+                        if self.s3_transform:
+                            img = self.s3_transform(img)
+                        meta_info['s3_olci'].append(key)
+                        sample['s3_olci'].append(img)
+
+            # s5p
+            if 's5p_co' in self.mode:
+                sample['s5p_co'] = []
+                meta_info['s5p_co'] = []
+                if 's5p_co' in grid_keys:
+                    local_keys = grid_keys['s5p_co']
+                    for key in local_keys:
+                        img_bytes = txn.get(key.encode('utf-8'))
+                        img = np.frombuffer(img_bytes, dtype=np.float32).reshape(28, 28, 1)
+                        if self.s5p_transform:
+                            img = self.s5p_transform(img)
+                        meta_info['s5p_co'].append(key)
+                        sample['s5p_co'].append(img)
+
+            if 's5p_no2' in self.mode:
+                sample['s5p_no2'] = []
+                meta_info['s5p_no2'] = []
+                if 's5p_no2' in grid_keys:
+                    local_keys = grid_keys['s5p_no2']
+                    for key in local_keys:
+                        img_bytes = txn.get(key.encode('utf-8'))
+                        img = np.frombuffer(img_bytes, dtype=np.float32).reshape(28, 28, 1)
+                        if self.s5p_transform:
+                            img = self.s5p_transform(img)
+                        meta_info['s5p_no2'].append(key)
+                        sample['s5p_no2'].append(img)
+
+            if 's5p_so2' in self.mode:
+                sample['s5p_so2'] = []
+                meta_info['s5p_so2'] = []
+                if 's5p_so2' in grid_keys:
+                    local_keys = grid_keys['s5p_so2']
+                    for key in local_keys:
+                        img_bytes = txn.get(key.encode('utf-8'))
+                        img = np.frombuffer(img_bytes, dtype=np.float32).reshape(28, 28, 1)
+                        if self.s5p_transform:
+                            img = self.s5p_transform(img)
+                        meta_info['s5p_so2'].append(key)
+                        sample['s5p_so2'].append(img)
+
+            if 's5p_o3' in self.mode:
+                sample['s5p_o3'] = []
+                meta_info['s5p_o3'] = []
+                if 's5p_o3' in grid_keys:
+                    local_keys = grid_keys['s5p_o3']
+                    for key in local_keys:
+                        img_bytes = txn.get(key.encode('utf-8'))
+                        img = np.frombuffer(img_bytes, dtype=np.float32).reshape(28, 28, 1)
+                        if self.s5p_transform:
+                            img = self.s5p_transform(img)
+                        meta_info['s5p_o3'].append(key)
+                        sample['s5p_o3'].append(img)
+            
+            # dem
+            if 'dem' in self.mode:
+                sample['dem'] = []
+                meta_info['dem'] = []
+                if 'dem' in grid_keys:
+                    local_keys = grid_keys['dem']
+                    for key in local_keys:
+                        img_bytes = txn.get(key.encode('utf-8'))
+                        img = np.frombuffer(img_bytes, dtype=np.float32).reshape(960,960,1)
+                        if self.dem_transform:
+                            img = self.dem_transform(img)
+                        meta_info['dem'].append(key)
+                        sample['dem'].append(img)
+
+
+
+        return sample, meta_info