def main(args):
config = load_config(args.config)
config["dataset"][
"path"] = args.dataset if args.dataset else config["dataset"]["path"]
config["model"]["lr"] = args.lr if args.lr else config["model"]["lr"]
config["model"][
"epochs"] = args.epochs if args.epochs else config["model"]["epochs"]
config["model"][
"batch_size"] = args.batch_size if args.batch_size else config[
"model"]["batch_size"]
log = Logs(os.path.join(args.out, "log"))
if torch.cuda.is_available():
device = torch.device("cuda")
torch.backends.cudnn.benchmark = True
log.log("RoboSat - training on {} GPUs, with {} workers".format(
torch.cuda.device_count(), args.workers))
else:
device = torch.device("cpu")
log.log("RoboSat - training on CPU, with {} workers".format(
args.workers))
num_classes = len(config["classes"])
num_channels = 0
for channel in config["channels"]:
num_channels += len(channel["bands"])
pretrained = config["model"]["pretrained"]
encoder = config["model"]["encoder"]
models = [
name for _, name, _ in pkgutil.iter_modules(
[os.path.dirname(robosat_pink.models.__file__)])
]
if config["model"]["name"] not in [model for model in models]:
sys.exit("Unknown model, thoses available are {}".format(
[model for model in models]))
model_module = import_module("robosat_pink.models.{}".format(
config["model"]["name"]))
net = getattr(model_module, "{}".format(config["model"]["name"].title()))(
num_classes=num_classes,
num_channels=num_channels,
encoder=encoder,
pretrained=pretrained).to(device)
net = torch.nn.DataParallel(net)
optimizer = Adam(net.parameters(),
lr=config["model"]["lr"],
weight_decay=config["model"]["decay"])
resume = 0
if args.checkpoint:
def map_location(storage, _):
return storage.cuda() if torch.cuda.is_available(
) else storage.cpu()
# https://github.com/pytorch/pytorch/issues/7178
chkpt = torch.load(args.checkpoint, map_location=map_location)
net.load_state_dict(chkpt["state_dict"])
log.log("Using checkpoint: {}".format(args.checkpoint))
if args.resume:
optimizer.load_state_dict(chkpt["optimizer"])
resume = chkpt["epoch"]
losses = [
name for _, name, _ in pkgutil.iter_modules(
[os.path.dirname(robosat_pink.losses.__file__)])
]
if config["model"]["loss"] not in [loss for loss in losses]:
sys.exit("Unknown loss, thoses available are {}".format(
[loss for loss in losses]))
loss_module = import_module("robosat_pink.losses.{}".format(
config["model"]["loss"]))
criterion = getattr(loss_module, "{}".format(
config["model"]["loss"].title()))().to(device)
train_loader, val_loader = get_dataset_loaders(config["dataset"]["path"],
config, args.workers)
if resume >= config["model"]["epochs"]:
sys.exit(
"Error: Epoch {} set in {} already reached by the checkpoint provided"
.format(config["model"]["epochs"], args.config))
log.log("")
log.log("--- Input tensor from Dataset: {} ---".format(
config["dataset"]["path"]))
num_channel = 1
for channel in config["channels"]:
for band in channel["bands"]:
log.log("Channel {}:\t\t {}[band: {}]".format(
num_channel, channel["sub"], band))
num_channel += 1
log.log("")
log.log("--- Hyper Parameters ---")
log.log("Model:\t\t\t {}".format(config["model"]["name"]))
log.log("Encoder model:\t\t {}".format(config["model"]["encoder"]))
log.log("Loss function:\t\t {}".format(config["model"]["loss"]))
log.log("ResNet pre-trained:\t {}".format(config["model"]["pretrained"]))
log.log("Batch Size:\t\t {}".format(config["model"]["batch_size"]))
log.log("Tile Size:\t\t {}".format(config["model"]["tile_size"]))
log.log("Data Augmentation:\t {}".format(
config["model"]["data_augmentation"]))
log.log("Learning Rate:\t\t {}".format(config["model"]["lr"]))
log.log("Weight Decay:\t\t {}".format(config["model"]["decay"]))
log.log("")
for epoch in range(resume, config["model"]["epochs"]):
log.log("---")
log.log("Epoch: {}/{}".format(epoch + 1, config["model"]["epochs"]))
train_hist = train(train_loader, num_classes, device, net, optimizer,
criterion)
log.log(
"Train loss: {:.4f}, mIoU: {:.3f}, {} IoU: {:.3f}, MCC: {:.3f}".
format(
train_hist["loss"],
train_hist["miou"],
config["classes"][1]["title"],
train_hist["fg_iou"],
train_hist["mcc"],
))
val_hist = validate(val_loader, num_classes, device, net, criterion)
log.log(
"Validate loss: {:.4f}, mIoU: {:.3f}, {} IoU: {:.3f}, MCC: {:.3f}".
format(val_hist["loss"], val_hist["miou"],
config["classes"][1]["title"], val_hist["fg_iou"],
val_hist["mcc"]))
states = {
"epoch": epoch + 1,
"state_dict": net.state_dict(),
"optimizer": optimizer.state_dict()
}
checkpoint_path = os.path.join(
args.out, "checkpoint-{:05d}-of-{:05d}.pth".format(
epoch + 1, config["model"]["epochs"]))
torch.save(states, checkpoint_path)
def main(args):
tiles = list(tiles_from_csv(args.tiles))
already_dl = 0
dl = 0
with requests.Session() as session:
num_workers = args.rate
os.makedirs(os.path.join(args.out), exist_ok=True)
log = Logs(os.path.join(args.out, "log"), out=sys.stderr)
log.log("Begin download from {}".format(args.url))
progress = tqdm(total=len(tiles), ascii=True, unit="image")
with futures.ThreadPoolExecutor(num_workers) as executor:
def worker(tile):
tick = time.monotonic()
x, y, z = map(str, [tile.x, tile.y, tile.z])
os.makedirs(os.path.join(args.out, z, x), exist_ok=True)
path = os.path.join(args.out, z, x,
"{}.{}".format(y, args.ext))
if os.path.isfile(path):
progress.update()
return tile, None, True
if args.type == "XYZ":
url = args.url.format(x=tile.x, y=tile.y, z=tile.z)
elif args.type == "TMS":
tile.y = (2**tile.z) - tile.y - 1
url = args.url.format(x=tile.x, y=tile.y, z=tile.z)
elif args.type == "WMS":
xmin, ymin, xmax, ymax = xy_bounds(tile)
url = args.url.format(xmin=xmin,
ymin=ymin,
xmax=xmax,
ymax=ymax)
res = tile_image_from_url(session, url, args.timeout)
if not res:
return tile, url, False
try:
cv2.imwrite(
path,
cv2.imdecode(np.fromstring(res.read(), np.uint8),
cv2.IMREAD_COLOR))
progress.update()
except OSError:
return tile, url, False
tock = time.monotonic()
time_for_req = tock - tick
time_per_worker = num_workers / args.rate
if time_for_req < time_per_worker:
time.sleep(time_per_worker - time_for_req)
return tile, url, True
for tile, url, ok in executor.map(worker, tiles):
if url and ok:
dl += 1
elif not url and ok:
already_dl += 1
else:
log.log("Warning:\n {} failed, skipping.\n {}\n".format(
tile, url))
if already_dl:
log.log(
"Notice:\n {} tiles were already downloaded previously, and so skipped now."
.format(already_dl))
if already_dl + dl == len(tiles):
log.log(" Coverage is fully downloaded.")
if args.web_ui:
template = "leaflet.html" if not args.web_ui_template else args.web_ui_template
base_url = args.web_ui_base_url if args.web_ui_base_url else "./"
web_ui(args.out, base_url, tiles, tiles, args.ext, template)
def main(args):
config = load_config(args.config)
print(config)
log = Logs(os.path.join(args.out, "log"))
if torch.cuda.is_available():
device = torch.device("cuda")
torch.backends.cudnn.benchmark = True
log.log("RoboSat - training on {} GPUs, with {} workers".format(
torch.cuda.device_count(), args.workers))
else:
device = torch.device("cpu")
log.log("RoboSat - training on CPU, with {} workers".format(
args.workers))
num_classes = len(config["classes"])
num_channels = 0
for channel in config["channels"]:
num_channels += len(channel["bands"])
pretrained = config["model"]["pretrained"]
encoder = config["model"]["encoder"]
models = [
name for _, name, _ in pkgutil.iter_modules(
[os.path.dirname(robosat_pink.models.__file__)])
]
if config["model"]["name"] not in [model for model in models]:
sys.exit("Unknown model, thoses available are {}".format(
[model for model in models]))
model_module = import_module("robosat_pink.models.{}".format(
config["model"]["name"]))
net = getattr(model_module, "{}".format(config["model"]["name"].title()))(
num_classes=num_classes,
num_channels=num_channels,
encoder=encoder,
pretrained=pretrained).to(device)
net = torch.nn.DataParallel(net)
optimizer = Adam(net.parameters(),
lr=config["model"]["lr"],
weight_decay=config["model"]["decay"])
resume = 0
# check checkpoint situation + load if ncessary
checkpoint = None # no checkpoint
if args.checkpoint: # command line checkpoint
checkpoint = args.checkpoint
try: # config file checkpoint
checkpoint = config["checkpoint"]['path']
except:
# no checkpoint in config file
pass
S3_CHECKPOINT = False
if checkpoint:
if checkpoint.startswith("s3://"):
S3_CHECKPOINT = True
# load from s3
checkpoint = checkpoint[5:]
sess = boto3.Session(profile_name=config['dataset']['aws_profile'])
fs = s3fs.S3FileSystem(session=sess)
s3ckpt = s3fs.S3File(fs, checkpoint, 'rb')
def map_location(storage, _):
return storage.cuda() if torch.cuda.is_available(
) else storage.cpu()
if checkpoint is not None:
def map_location(storage, _):
return storage.cuda() if torch.cuda.is_available(
) else storage.cpu()
try:
if S3_CHECKPOINT:
with s3fs.S3File(fs, checkpoint, 'rb') as C:
state = torch.load(io.BytesIO(C.read()),
map_location=map_location)
else:
state = torch.load(checkpoint)
optimizer.load_state_dict(state['optimizer'])
net.load_state_dict(state['state_dict'])
net.to(device)
except FileNotFoundError as f:
print("{} checkpoint not found.".format(CHECKPOINT))
log.log("Using checkpoint: {}".format(checkpoint))
losses = [
name for _, name, _ in pkgutil.iter_modules(
[os.path.dirname(robosat_pink.losses.__file__)])
]
if config["model"]["loss"] not in [loss for loss in losses]:
sys.exit("Unknown loss, thoses available are {}".format(
[loss for loss in losses]))
loss_module = import_module("robosat_pink.losses.{}".format(
config["model"]["loss"]))
criterion = getattr(loss_module, "{}".format(
config["model"]["loss"].title()))().to(device)
train_loader, val_loader = get_dataset_loaders(config,
args.workers,
idDir=args.out)
if resume >= config["model"]["epochs"]:
sys.exit(
"Error: Epoch {} set in {} already reached by the checkpoint provided"
.format(config["model"]["epochs"], args.config))
log.log("")
log.log("--- Input tensor from Dataset: {} ---".format(
config["dataset"]["image_bucket"] + '/' +
config['dataset']['imagery_directory_regex']))
log.log("")
log.log("--- Hyper Parameters ---")
log.log("Model:\t\t\t {}".format(config["model"]["name"]))
log.log("Encoder model:\t\t {}".format(config["model"]["encoder"]))
log.log("Loss function:\t\t {}".format(config["model"]["loss"]))
log.log("ResNet pre-trained:\t {}".format(config["model"]["pretrained"]))
log.log("Batch Size:\t\t {}".format(config["model"]["batch_size"]))
log.log("Tile Size:\t\t {}".format(config["model"]["tile_size"]))
log.log("Data Augmentation:\t {}".format(
config["model"]["data_augmentation"]))
log.log("Learning Rate:\t\t {}".format(config["model"]["lr"]))
log.log("Weight Decay:\t\t {}".format(config["model"]["decay"]))
log.log("")
for epoch in range(resume, config["model"]["epochs"]):
log.log("---")
log.log("Epoch: {}/{}".format(epoch + 1, config["model"]["epochs"]))
train_hist = train(train_loader, num_classes, device, net, optimizer,
criterion)
log.log(
"Train loss: {:.4f}, mIoU: {:.3f}, IoU: {:.3f}, precision: {:.3f}, recall: {:.3f}"
.format(
train_hist["loss"],
train_hist["miou"],
train_hist["fg_iou"],
train_hist["precision"],
train_hist["recall"],
))
val_hist = validate(val_loader, num_classes, device, net, criterion)
log.log(
"Validate loss: {:.4f}, mIoU: {:.3f}, IoU: {:.3f}, precision: {:.3f}, recall: {:.3f}"
.format(
train_hist["loss"],
train_hist["miou"],
train_hist["fg_iou"],
train_hist["precision"],
train_hist["recall"],
))
states = {
"epoch": epoch + 1,
"state_dict": net.state_dict(),
"optimizer": optimizer.state_dict()
}
checkpoint_path = os.path.join(
args.out, "checkpoint-{:05d}-of-{:05d}.pth".format(
epoch + 1, config["model"]["epochs"]))
torch.save(states, checkpoint_path)
def main(args):
args.out = os.path.expanduser(args.out)
if not args.workers:
args.workers = max(1, math.floor(os.cpu_count() * 0.5))
print("RoboSat.pink - compare {} on CPU, with {} workers".format(
args.mode, args.workers))
if not args.masks or not args.labels:
if args.mode == "list":
sys.exit(
"ERROR: Parameters masks and labels are mandatories in list mode."
)
if args.minimum_fg > 0 or args.maximum_fg < 100 or args.minimum_qod > 0 or args.maximum_qod < 100:
sys.exit(
"ERROR: Parameters masks and labels are mandatories in QoD filtering."
)
try:
if args.images:
tiles = [tile for tile, _ in tiles_from_slippy_map(args.images[0])]
for image in args.images[1:]:
assert sorted(tiles) == sorted(
[tile for tile, _ in tiles_from_slippy_map(image)])
if args.labels and args.masks:
tiles_masks = [
tile for tile, _ in tiles_from_slippy_map(args.masks)
]
tiles_labels = [
tile for tile, _ in tiles_from_slippy_map(args.labels)
]
if args.images:
assert sorted(tiles) == sorted(tiles_masks) == sorted(
tiles_labels)
else:
assert sorted(tiles_masks) == sorted(tiles_labels)
tiles = tiles_masks
except:
sys.exit("ERROR: inconsistent input coverage")
tiles_list = []
tiles_compare = []
progress = tqdm(total=len(tiles), ascii=True, unit="tile")
log = False if args.geojson else Logs(os.path.join(args.out, "log"))
with futures.ThreadPoolExecutor(args.workers) as executor:
def worker(tile):
x, y, z = list(map(str, tile))
if args.masks and args.labels:
try:
dist, fg_ratio, qod = compare(args.masks, args.labels,
tile)
except:
progress.update()
return False, tile
if not args.minimum_fg <= fg_ratio <= args.maximum_fg or not args.minimum_qod <= qod <= args.maximum_qod:
progress.update()
return True, tile
tiles_compare.append(tile)
if args.mode == "side":
for i, root in enumerate(args.images):
img = tile_image_from_file(
tile_from_slippy_map(root, x, y, z)[1])
if i == 0:
side = np.zeros(
(img.shape[0], img.shape[1] * len(args.images), 3))
side = np.swapaxes(side, 0,
1) if args.vertical else side
image_shape = img.shape
else:
assert image_shape[0:2] == img.shape[
0:2], "Unconsistent image size to compare"
if args.vertical:
side[i * image_shape[0]:(i + 1) *
image_shape[0], :, :] = img
else:
side[:, i * image_shape[0]:(i + 1) *
image_shape[0], :] = img
tile_image_to_file(args.out, tile, np.uint8(side))
elif args.mode == "stack":
for i, root in enumerate(args.images):
tile_image = tile_image_from_file(
tile_from_slippy_map(root, x, y, z)[1])
if i == 0:
image_shape = tile_image.shape[0:2]
stack = tile_image / len(args.images)
else:
assert image_shape == tile_image.shape[
0:2], "Unconsistent image size to compare"
stack = stack + (tile_image / len(args.images))
tile_image_to_file(args.out, tile, np.uint8(stack))
elif args.mode == "list":
tiles_list.append([tile, fg_ratio, qod])
progress.update()
return True, tile
for tile, ok in executor.map(worker, tiles):
if not ok and log:
log.log("Warning: skipping. {}".format(str(tile)))
if args.mode == "list":
with open(args.out, mode="w") as out:
if args.geojson:
out.write('{"type":"FeatureCollection","features":[')
first = True
for tile_list in tiles_list:
tile, fg_ratio, qod = tile_list
x, y, z = list(map(str, tile))
if args.geojson:
prop = '"properties":{{"x":{},"y":{},"z":{},"fg":{:.1f},"qod":{:.1f}}}'.format(
x, y, z, fg_ratio, qod)
geom = '"geometry":{}'.format(
json.dumps(feature(tile, precision=6)["geometry"]))
out.write('{}{{"type":"Feature",{},{}}}'.format(
"," if not first else "", geom, prop))
first = False
else:
out.write("{},{},{}\t{:.1f}\t{:.1f}{}".format(
x, y, z, fg_ratio, qod, os.linesep))
if args.geojson:
out.write("]}")
out.close()
base_url = args.web_ui_base_url if args.web_ui_base_url else "./"
if args.mode == "side" and not args.no_web_ui:
template = "compare.html" if not args.web_ui_template else args.web_ui_template
web_ui(args.out, base_url, None, tiles_compare, args.format, template)
if args.mode == "stack" and not args.no_web_ui:
template = "leaflet.html" if not args.web_ui_template else args.web_ui_template
tiles = [tile for tile, _ in tiles_from_slippy_map(args.images[0])]
web_ui(args.out, base_url, tiles, tiles_compare, args.format, template)
def main(args):
if (args.geojson and args.postgis) or (not args.geojson and not args.postgis):
sys.exit("Input features to rasterize must be either GeoJSON or PostGIS")
config = load_config(args.config)
tile_size = args.tile_size if args.tile_size else config["model"]["tile_size"]
colors = [classe["color"] for classe in config["classes"]]
burn_value = 1
os.makedirs(args.out, exist_ok=True)
log = Logs(os.path.join(args.out, "log"), out=sys.stderr)
def geojson_parse_polygon(zoom, feature_map, polygon, i):
try:
for i, ring in enumerate(polygon["coordinates"]): # GeoJSON coordinates could be N dimensionals
polygon["coordinates"][i] = [[x, y] for point in ring for x, y in zip([point[0]], [point[1]])]
for tile in burntiles.burn([{"type": "feature", "geometry": polygon}], zoom=zoom):
feature_map[mercantile.Tile(*tile)].append({"type": "feature", "geometry": polygon})
except ValueError:
log.log("Warning: invalid feature {}, skipping".format(i))
return feature_map
def geojson_parse_geometry(zoom, feature_map, geometry, i):
if geometry["type"] == "Polygon":
feature_map = geojson_parse_polygon(zoom, feature_map, geometry, i)
elif geometry["type"] == "MultiPolygon":
for polygon in geometry["coordinates"]:
feature_map = geojson_parse_polygon(zoom, feature_map, {"type": "Polygon", "coordinates": polygon}, i)
else:
log.log("Notice: {} is a non surfacic geometry type, skipping feature {}".format(geometry["type"], i))
return feature_map
if args.geojson:
tiles = [tile for tile in tiles_from_csv(args.cover)]
zoom = tiles[0].z
if [tile for tile in tiles if tile.z != zoom]:
sys.exit("With GeoJson input, all tiles z values have to be the same, in the csv cover file.")
feature_map = collections.defaultdict(list)
# Compute a spatial index like
for geojson_file in args.geojson:
with open(geojson_file) as geojson:
feature_collection = json.load(geojson)
for i, feature in enumerate(tqdm(feature_collection["features"], ascii=True, unit="feature")):
if feature["geometry"]["type"] == "GeometryCollection":
for geometry in feature["geometry"]["geometries"]:
feature_map = geojson_parse_geometry(zoom, feature_map, geometry, i)
else:
feature_map = geojson_parse_geometry(zoom, feature_map, feature["geometry"], i)
# Rasterize tiles
for tile in tqdm(list(tiles_from_csv(args.cover)), ascii=True, unit="tile"):
if tile in feature_map:
out = geojson_tile_burn(tile, feature_map[tile], tile_size, burn_value)
else:
out = np.zeros(shape=(tile_size, tile_size), dtype=np.uint8)
write_tile(args.out, tile, colors, out)
if args.postgis:
try:
pg_conn = psycopg2.connect(config["dataset"]["pg_dsn"])
pg = pg_conn.cursor()
except Exception:
sys.exit("Unable to connect PostgreSQL: {}".format(config["dataset"]["pg_dsn"]))
try:
pg.execute("SELECT ST_Srid(geom) AS srid FROM ({} LIMIT 1) AS sub".format(args.postgis))
srid = pg.fetchone()[0]
except Exception:
sys.exit("Unable to retrieve geometry SRID.")
for tile in tqdm(list(tiles_from_csv(args.cover)), ascii=True, unit="tile"):
s, w, e, n = mercantile.bounds(tile)
raster = np.zeros((tile_size, tile_size))
query = """
WITH
bbox AS (SELECT ST_Transform(ST_MakeEnvelope({},{},{},{}, 4326), {} ) AS bbox),
bbox_merc AS (SELECT ST_Transform(ST_MakeEnvelope({},{},{},{}, 4326), 3857) AS bbox),
rast_a AS (SELECT ST_AddBand(
ST_SetSRID(
ST_MakeEmptyRaster({}, {}, ST_Xmin(bbox), ST_Ymax(bbox), (ST_YMax(bbox) - ST_YMin(bbox)) / {}),
3857),
'8BUI'::text, 0) AS rast
FROM bbox_merc),
features AS (SELECT ST_Union(ST_Transform(ST_Force2D(geom), 3857)) AS geom
FROM ({}) AS sub, bbox
WHERE ST_Intersects(geom, bbox)),
rast_b AS (SELECT ST_AsRaster(geom, rast, '8BUI', {}) AS rast
FROM features, rast_a
WHERE NOT ST_IsEmpty(geom))
SELECT ST_AsBinary(ST_MapAlgebra(rast_a.rast, rast_b.rast, '{}', NULL, 'FIRST')) AS wkb FROM rast_a, rast_b
""".format(
s, w, e, n, srid, s, w, e, n, tile_size, tile_size, tile_size, args.postgis, burn_value, burn_value
)
try:
pg.execute(query)
row = pg.fetchone()
if row:
raster = np.squeeze(wkb_to_numpy(io.BytesIO(row[0])), axis=2)
except Exception:
log.log("Warning: Invalid geometries, skipping {}".format(tile))
pg_conn = psycopg2.connect(config["dataset"]["pg_dsn"])
pg = pg_conn.cursor()
write_tile(args.out, tile, colors, raster)
if args.web_ui:
template = "leaflet.html" if not args.web_ui_template else args.web_ui_template
base_url = args.web_ui_base_url if args.web_ui_base_url else "./"
tiles = [tile for tile in tiles_from_csv(args.cover)]
web_ui(args.out, base_url, tiles, tiles, "png", template)
def main(args):
config = load_config(args.config)
args.out = os.path.expanduser(args.out)
args.workers = torch.cuda.device_count() * 2 if torch.device(
"cuda") and not args.workers else args.workers
config["model"][
"loader"] = args.loader if args.loader else config["model"]["loader"]
config["model"]["bs"] = args.bs if args.bs else config["model"]["bs"]
config["model"]["lr"] = args.lr if args.lr else config["model"]["lr"]
config["model"]["ts"] = args.ts if args.ts else config["model"]["ts"]
config["model"]["nn"] = args.nn if args.nn else config["model"]["nn"]
config["model"][
"loss"] = args.loss if args.loss else config["model"]["loss"]
config["model"]["da"] = args.da if args.da else config["model"]["da"]
config["model"]["dap"] = args.dap if args.dap else config["model"]["dap"]
check_classes(config)
check_channels(config)
check_model(config)
if not os.path.isdir(os.path.expanduser(args.dataset)):
sys.exit("ERROR: dataset {} is not a directory".format(args.dataset))
log = Logs(os.path.join(args.out, "log"))
if torch.cuda.is_available():
log.log("RoboSat.pink - training on {} GPUs, with {} workers".format(
torch.cuda.device_count(), args.workers))
log.log("(Torch:{} Cuda:{} CudNN:{})".format(
torch.__version__, torch.version.cuda,
torch.backends.cudnn.version()))
device = torch.device("cuda")
torch.backends.cudnn.benchmark = True
else:
log.log("RoboSat.pink - training on CPU, with {} workers - (Torch:{})".
format(args.workers, torch.__version__))
log.log(
"WARNING: Are you really sure sure about not training on GPU ?")
device = torch.device("cpu")
try:
loader = import_module("robosat_pink.loaders.{}".format(
config["model"]["loader"].lower()))
loader_train = getattr(loader, config["model"]["loader"])(
config, config["model"]["ts"],
os.path.join(args.dataset, "training"), "train")
loader_val = getattr(loader, config["model"]["loader"])(
config, config["model"]["ts"],
os.path.join(args.dataset, "validation"), "train")
except:
sys.exit("ERROR: Unable to load data loaders")
try:
model_module = import_module("robosat_pink.models.{}".format(
config["model"]["nn"].lower()))
except:
sys.exit("ERROR: Unable to load {} model".format(
config["model"]["nn"]))
nn = getattr(model_module, config["model"]["nn"])(
loader_train.shape_in, loader_train.shape_out,
config["model"]["pretrained"]).to(device)
nn = torch.nn.DataParallel(nn)
optimizer = Adam(nn.parameters(), lr=config["model"]["lr"])
resume = 0
if args.checkpoint:
try:
chkpt = torch.load(os.path.expanduser(args.checkpoint),
map_location=device)
nn.load_state_dict(chkpt["state_dict"])
log.log("Using checkpoint: {}".format(args.checkpoint))
except:
sys.exit("ERROR: Unable to load {} checkpoint".format(
args.checkpoint))
if args.resume:
optimizer.load_state_dict(chkpt["optimizer"])
resume = chkpt["epoch"]
if resume >= args.epochs:
sys.exit(
"ERROR: Epoch {} already reached by the given checkpoint".
format(config["model"]["epochs"]))
try:
loss_module = import_module("robosat_pink.losses.{}".format(
config["model"]["loss"].lower()))
criterion = getattr(loss_module, config["model"]["loss"])().to(device)
except:
sys.exit("ERROR: Unable to load {} loss".format(
config["model"]["loss"]))
bs = config["model"]["bs"]
train_loader = DataLoader(loader_train,
batch_size=bs,
shuffle=True,
drop_last=True,
num_workers=args.workers)
val_loader = DataLoader(loader_val,
batch_size=bs,
shuffle=False,
drop_last=True,
num_workers=args.workers)
log.log("--- Input tensor from Dataset: {} ---".format(args.dataset))
num_channel = 1 # 1-based numerotation
for channel in config["channels"]:
for band in channel["bands"]:
log.log("Channel {}:\t\t {}[band: {}]".format(
num_channel, channel["name"], band))
num_channel += 1
log.log("--- Hyper Parameters ---")
for hp in config["model"]:
log.log("{}{}".format(hp.ljust(25, " "), config["model"][hp]))
for epoch in range(resume, args.epochs):
UUID = uuid.uuid1()
log.log("---{}Epoch: {}/{} -- UUID: {}".format(os.linesep, epoch + 1,
args.epochs, UUID))
train(train_loader, config, log, device, nn, optimizer, criterion)
validate(val_loader, config, log, device, nn, criterion)
try: # https://github.com/pytorch/pytorch/issues/9176
nn_doc = nn.module.doc
nn_version = nn.module.version
except AttributeError:
nn_version = nn.version
nn_doc == nn.doc
states = {
"uuid": UUID,
"model_version": nn_version,
"producer_name": "RoboSat.pink",
"producer_version": "0.4.0",
"model_licence": "MIT",
"domain": "pink.RoboSat", # reverse-DNS
"doc_string": nn_doc,
"shape_in": loader_train.shape_in,
"shape_out": loader_train.shape_out,
"state_dict": nn.state_dict(),
"epoch": epoch + 1,
"nn": config["model"]["nn"],
"optimizer": optimizer.state_dict(),
"loader": config["model"]["loader"],
}
checkpoint_path = os.path.join(
args.out, "checkpoint-{:05d}.pth".format(epoch + 1))
try:
torch.save(states, checkpoint_path)
except:
sys.exit(
"ERROR: Unable to save checkpoint {}".format(checkpoint_path))
def main(args):
try:
tiles = list(tiles_from_csv(args.cover))
except:
sys.exit("ERROR: unable to read cover file: {}".format(args.cover))
if not args.workers:
args.workers = max(1, math.floor(os.cpu_count() * 0.5))
try:
os.makedirs(os.path.expanduser(args.out), exist_ok=True)
except:
sys.exit("ERROR: unable to create output dir: {}".format(args.out))
log = Logs(os.path.join(args.out, "log"), out=sys.stderr)
log.log(
"RoboSat.pink - download with {} workers, at max {} req/s, from: {}".
format(args.workers, args.rate, args.url))
already_dl = 0
dl = 0
with requests.Session() as session:
progress = tqdm(total=len(tiles), ascii=True, unit="image")
with futures.ThreadPoolExecutor(args.workers) as executor:
def worker(tile):
tick = time.monotonic()
progress.update()
x, y, z = map(str, [tile.x, tile.y, tile.z])
try:
os.makedirs(os.path.join(args.out, z, x), exist_ok=True)
except:
return tile, None, False
path = os.path.join(args.out, z, x,
"{}.{}".format(y, args.format))
if os.path.isfile(path): # already downloaded
return tile, None, True
if args.type == "XYZ":
url = args.url.format(x=tile.x, y=tile.y, z=tile.z)
elif args.type == "TMS":
tile.y = (2**tile.z) - tile.y - 1
url = args.url.format(x=tile.x, y=tile.y, z=tile.z)
elif args.type == "WMS":
xmin, ymin, xmax, ymax = xy_bounds(tile)
url = args.url.format(xmin=xmin,
ymin=ymin,
xmax=xmax,
ymax=ymax)
res = tile_image_from_url(session, url, args.timeout)
if res is None: # let's retry once
res = tile_image_from_url(session, url, args.timeout)
if res is None:
return tile, url, False
try:
tile_image_to_file(args.out, tile, res)
# cv2.imwrite(path, cv2.imdecode(np.fromstring(res.read(), np.uint8), cv2.IMREAD_COLOR))
except OSError:
return tile, url, False
tock = time.monotonic()
time_for_req = tock - tick
time_per_worker = args.workers / args.rate
if time_for_req < time_per_worker:
time.sleep(time_per_worker - time_for_req)
return tile, url, True
for tile, url, ok in executor.map(worker, tiles):
if url and ok:
dl += 1
elif not url and ok:
already_dl += 1
else:
log.log("Warning:\n {} failed, skipping.\n {}\n".format(
tile, url))
if already_dl:
log.log(
"Notice: {} tiles were already downloaded previously, and so skipped now."
.format(already_dl))
if already_dl + dl == len(tiles):
log.log("Notice: Coverage is fully downloaded.")
if not args.no_web_ui:
template = "leaflet.html" if not args.web_ui_template else args.web_ui_template
base_url = args.web_ui_base_url if args.web_ui_base_url else "./"
web_ui(args.out, base_url, tiles, tiles, args.format, template)
def main(args):
if (args.geojson and args.postgis) or (not args.geojson
and not args.postgis):
sys.exit(
"ERROR: Input features to rasterize must be either GeoJSON or PostGIS"
)
if args.postgis and not args.pg_dsn:
sys.exit(
"ERROR: With PostGIS input features, --pg_dsn must be provided")
config = load_config(args.config)
check_classes(config)
palette = make_palette(*[classe["color"] for classe in config["classes"]],
complementary=True)
burn_value = 1
args.out = os.path.expanduser(args.out)
os.makedirs(args.out, exist_ok=True)
log = Logs(os.path.join(args.out, "log"), out=sys.stderr)
def geojson_parse_polygon(zoom, srid, feature_map, polygon, i):
try:
if srid != 4326:
polygon = [
xy for xy in geojson_reproject(
{
"type": "feature",
"geometry": polygon
}, srid, 4326)
][0]
for i, ring in enumerate(
polygon["coordinates"]
): # GeoJSON coordinates could be N dimensionals
polygon["coordinates"][i] = [[
x, y
] for point in ring for x, y in zip([point[0]], [point[1]])]
if polygon["coordinates"]:
for tile in burntiles.burn([{
"type": "feature",
"geometry": polygon
}],
zoom=zoom):
feature_map[mercantile.Tile(*tile)].append({
"type":
"feature",
"geometry":
polygon
})
except ValueError:
log.log("Warning: invalid feature {}, skipping".format(i))
return feature_map
def geojson_parse_geometry(zoom, srid, feature_map, geometry, i):
if geometry["type"] == "Polygon":
feature_map = geojson_parse_polygon(zoom, srid, feature_map,
geometry, i)
elif geometry["type"] == "MultiPolygon":
for polygon in geometry["coordinates"]:
feature_map = geojson_parse_polygon(zoom, srid, feature_map, {
"type": "Polygon",
"coordinates": polygon
}, i)
else:
log.log(
"Notice: {} is a non surfacic geometry type, skipping feature {}"
.format(geometry["type"], i))
return feature_map
if args.geojson:
try:
tiles = [
tile for tile in tiles_from_csv(os.path.expanduser(args.cover))
]
zoom = tiles[0].z
assert not [tile for tile in tiles if tile.z != zoom]
except:
sys.exit("ERROR: Inconsistent cover {}".format(args.cover))
feature_map = collections.defaultdict(list)
log.log("RoboSat.pink - rasterize - Compute spatial index")
for geojson_file in args.geojson:
with open(os.path.expanduser(geojson_file)) as geojson:
try:
feature_collection = json.load(geojson)
except:
sys.exit("ERROR: {} is not a valid JSON file.".format(
geojson_file))
try:
crs_mapping = {"CRS84": "4326", "900913": "3857"}
srid = feature_collection["crs"]["properties"][
"name"].split(":")[-1]
srid = int(srid) if srid not in crs_mapping else int(
crs_mapping[srid])
except:
srid = int(4326)
for i, feature in enumerate(
tqdm(feature_collection["features"],
ascii=True,
unit="feature")):
try:
if feature["geometry"]["type"] == "GeometryCollection":
for geometry in feature["geometry"]["geometries"]:
feature_map = geojson_parse_geometry(
zoom, srid, feature_map, geometry, i)
else:
feature_map = geojson_parse_geometry(
zoom, srid, feature_map, feature["geometry"],
i)
except:
sys.exit(
"ERROR: Unable to parse {} file. Seems not a valid GEOJSON file."
.format(geojson_file))
log.log(
"RoboSat.pink - rasterize - rasterizing tiles from {} on cover {}".
format(args.geojson, args.cover))
with open(os.path.join(os.path.expanduser(args.out),
"instances.cover"),
mode="w") as cover:
for tile in tqdm(list(
tiles_from_csv(os.path.expanduser(args.cover))),
ascii=True,
unit="tile"):
try:
if tile in feature_map:
cover.write("{},{},{} {}{}".format(
tile.x, tile.y, tile.z, len(feature_map[tile]),
os.linesep))
out = geojson_tile_burn(tile, feature_map[tile], 4326,
args.ts, burn_value)
else:
cover.write("{},{},{} {}{}".format(
tile.x, tile.y, tile.z, 0, os.linesep))
out = np.zeros(shape=(args.ts, args.ts),
dtype=np.uint8)
tile_label_to_file(args.out, tile, palette, out)
except:
log.log("Warning: Unable to rasterize tile. Skipping {}".
format(str(tile)))
if args.postgis:
try:
pg_conn = psycopg2.connect(args.pg_dsn)
pg = pg_conn.cursor()
except Exception:
sys.exit("Unable to connect PostgreSQL: {}".format(args.pg_dsn))
log.log(
"RoboSat.pink - rasterize - rasterizing tiles from PostGIS on cover {}"
.format(args.cover))
log.log(" SQL {}".format(args.postgis))
try:
pg.execute(
"SELECT ST_Srid(geom) AS srid FROM ({} LIMIT 1) AS sub".format(
args.postgis))
srid = pg.fetchone()[0]
except Exception:
sys.exit("Unable to retrieve geometry SRID.")
for tile in tqdm(list(tiles_from_csv(args.cover)),
ascii=True,
unit="tile"):
s, w, e, n = mercantile.bounds(tile)
raster = np.zeros((args.ts, args.ts))
query = """
WITH
bbox AS (SELECT ST_Transform(ST_MakeEnvelope({},{},{},{}, 4326), {} ) AS bbox),
bbox_merc AS (SELECT ST_Transform(ST_MakeEnvelope({},{},{},{}, 4326), 3857) AS bbox),
rast_a AS (SELECT ST_AddBand(
ST_SetSRID(
ST_MakeEmptyRaster({}, {}, ST_Xmin(bbox), ST_Ymax(bbox), (ST_YMax(bbox) - ST_YMin(bbox)) / {}),
3857),
'8BUI'::text, 0) AS rast
FROM bbox_merc),
features AS (SELECT ST_Union(ST_Transform(ST_Force2D(geom), 3857)) AS geom
FROM ({}) AS sub, bbox
WHERE ST_Intersects(geom, bbox)),
rast_b AS (SELECT ST_AsRaster(geom, rast, '8BUI', {}) AS rast
FROM features, rast_a
WHERE NOT ST_IsEmpty(geom))
SELECT ST_AsBinary(ST_MapAlgebra(rast_a.rast, rast_b.rast, '{}', NULL, 'FIRST')) AS wkb FROM rast_a, rast_b
""".format(s, w, e, n, srid, s, w, e, n, args.ts, args.ts, args.ts,
args.postgis, burn_value, burn_value)
try:
pg.execute(query)
row = pg.fetchone()
if row:
raster = np.squeeze(wkb_to_numpy(io.BytesIO(row[0])),
axis=2)
except Exception:
log.log(
"Warning: Invalid geometries, skipping {}".format(tile))
pg_conn = psycopg2.connect(args.pg_dsn)
pg = pg_conn.cursor()
try:
tile_label_to_file(args.out, tile, palette, raster)
except:
log.log(
"Warning: Unable to rasterize tile. Skipping {}".format(
str(tile)))
if not args.no_web_ui:
template = "leaflet.html" if not args.web_ui_template else args.web_ui_template
base_url = args.web_ui_base_url if args.web_ui_base_url else "./"
tiles = [tile for tile in tiles_from_csv(args.cover)]
web_ui(args.out, base_url, tiles, tiles, "png", template)
def main(args):
config = load_config(args.config)
tile_size = args.tile_size if args.tile_size else config["model"]["tile_size"]
colors = [classe["color"] for classe in config["classes"]]
os.makedirs(args.out, exist_ok=True)
# We can only rasterize all tiles at a single zoom.
assert all(tile.z == args.zoom for tile in tiles_from_csv(args.cover))
# Find all tiles the features cover and make a map object for quick lookup.
feature_map = collections.defaultdict(list)
log = Logs(os.path.join(args.out, "log"), out=sys.stderr)
def parse_polygon(feature_map, polygon, i):
try:
for i, ring in enumerate(polygon["coordinates"]): # GeoJSON coordinates could be N dimensionals
polygon["coordinates"][i] = [[x, y] for point in ring for x, y in zip([point[0]], [point[1]])]
for tile in burntiles.burn([{"type": "feature", "geometry": polygon}], zoom=args.zoom):
feature_map[mercantile.Tile(*tile)].append({"type": "feature", "geometry": polygon})
except ValueError:
log.log("Warning: invalid feature {}, skipping".format(i))
return feature_map
def parse_geometry(feature_map, geometry, i):
if geometry["type"] == "Polygon":
feature_map = parse_polygon(feature_map, geometry, i)
elif geometry["type"] == "MultiPolygon":
for polygon in geometry["coordinates"]:
feature_map = parse_polygon(feature_map, {"type": "Polygon", "coordinates": polygon}, i)
else:
log.log("Notice: {} is a non surfacic geometry type, skipping feature {}".format(geometry["type"], i))
return feature_map
for feature in args.features:
with open(feature) as f:
fc = json.load(f)
for i, feature in enumerate(tqdm(fc["features"], ascii=True, unit="feature")):
if feature["geometry"]["type"] == "GeometryCollection":
for geometry in feature["geometry"]["geometries"]:
feature_map = parse_geometry(feature_map, geometry, i)
else:
feature_map = parse_geometry(feature_map, feature["geometry"], i)
# Burn features to tiles and write to a slippy map directory.
for tile in tqdm(list(tiles_from_csv(args.cover)), ascii=True, unit="tile"):
if tile in feature_map:
out = burn(tile, feature_map[tile], tile_size)
else:
out = np.zeros(shape=(tile_size, tile_size), dtype=np.uint8)
out_dir = os.path.join(args.out, str(tile.z), str(tile.x))
os.makedirs(out_dir, exist_ok=True)
out_path = os.path.join(out_dir, "{}.png".format(tile.y))
if os.path.exists(out_path):
prev = np.array(Image.open(out_path))
out = np.maximum(out, prev)
out = Image.fromarray(out, mode="P")
out_path = os.path.join(args.out, str(tile.z), str(tile.x))
os.makedirs(out_path, exist_ok=True)
out.putpalette(complementary_palette(make_palette(colors[0], colors[1])))
out.save(os.path.join(out_path, "{}.png".format(tile.y)), optimize=True)
if args.web_ui:
template = "leaflet.html" if not args.web_ui_template else args.web_ui_template
base_url = args.web_ui_base_url if args.web_ui_base_url else "./"
tiles = [tile for tile in tiles_from_csv(args.cover)]
web_ui(args.out, base_url, tiles, tiles, "png", template)
def main(args):
config = load_config(args.config)
check_channels(config)
check_classes(config)
args.workers = torch.cuda.device_count() * 2 if torch.device(
"cuda") and not args.workers else args.workers
log = Logs(os.path.join(args.out, "log"))
if torch.cuda.is_available():
log.log("RoboSat.pink - predict on {} GPUs, with {} workers".format(
torch.cuda.device_count(), args.workers))
log.log("(Torch:{} Cuda:{} CudNN:{})".format(
torch.__version__, torch.version.cuda,
torch.backends.cudnn.version()))
device = torch.device("cuda")
torch.backends.cudnn.benchmark = True
else:
log.log("RoboSat.pink - predict on CPU, with {} workers".format(
args.workers))
device = torch.device("cpu")
try:
chkpt = torch.load(args.checkpoint, map_location=device)
assert chkpt["producer_name"] == "RoboSat.pink"
model_module = import_module("robosat_pink.models.{}".format(
chkpt["nn"].lower()))
nn = getattr(model_module, chkpt["nn"])(chkpt["shape_in"],
chkpt["shape_out"]).to(device)
nn = torch.nn.DataParallel(nn)
nn.load_state_dict(chkpt["state_dict"])
nn.eval()
except:
sys.exit("ERROR: Unable to load {} checkpoint.".format(
args.checkpoint))
log.log("Model {} - UUID: {}".format(chkpt["nn"], chkpt["uuid"]))
try:
loader_module = import_module("robosat_pink.loaders.{}".format(
chkpt["loader"].lower()))
loader_predict = getattr(loader_module,
chkpt["loader"])(config,
chkpt["shape_in"][1:3],
args.tiles,
mode="predict")
except:
sys.exit("ERROR: Unable to load {} data loader.".format(
chkpt["loader"]))
loader = DataLoader(loader_predict,
batch_size=args.bs,
num_workers=args.workers)
palette = make_palette(config["classes"][0]["color"],
config["classes"][1]["color"])
with torch.no_grad(
): # don't track tensors with autograd during prediction
for images, tiles in tqdm(loader,
desc="Eval",
unit="batch",
ascii=True):
images = images.to(device)
try:
outputs = nn(images)
probs = torch.nn.functional.softmax(outputs,
dim=1).data.cpu().numpy()
except:
log.log("WARNING: Skipping batch:")
for tile, prob in zip(tiles, probs):
log.log(" - {}".format(str(tile)))
continue
for tile, prob in zip(tiles, probs):
try:
x, y, z = list(map(int, tile))
mask = np.around(prob[1:, :, :]).astype(np.uint8).squeeze()
tile_label_to_file(args.out, mercantile.Tile(x, y, z),
palette, mask)
except:
log.log("WARNING: Skipping tile {}".format(str(tile)))
if not args.no_web_ui:
template = "leaflet.html" if not args.web_ui_template else args.web_ui_template
base_url = args.web_ui_base_url if args.web_ui_base_url else "./"
tiles = [tile for tile, _ in tiles_from_slippy_map(args.out)]
web_ui(args.out, base_url, tiles, tiles, "png", template)