dlr: add cross-entropy + dice loss fn option

This commit is contained in:
Starbeamrainbowlabs 2023-01-13 17:58:00 +00:00
parent b2a5acaf4e
commit be77f035c8
Signed by: sbrl
GPG key ID: 1BE5172E637709C2
3 changed files with 59 additions and 6 deletions

View file

@ -30,10 +30,11 @@ show_help() {
echo -e " DIR_RAINFALLWATER The path to the directory the .tfrecord files containing the rainfall radar / water depth data." >&2;
echo -e " PATH_HEIGHTMAP The path to the heightmap jsonl file to read in." >&2;
echo -e " PATH_COLOURMAP The path to the colourmap for predictive purposes." >&2;
echo -e " PATH_CHECKPOINT The path to a checkcpoint to load. If specified, a model will be loaded instead of being trained." >&2;
echo -e " STEPS_PER_EPOCH The number of steps to consider an epoch. Defaults to None, which means use the entire dataset." >&2;
echo -e " NO_REMOVE_ISOLATED_PIXELS Set to any value to avoid the engine from removing isolated pixels - that is, water pixels with no other surrounding pixels, either side to side to diagonally." >&2;
echo -e " EPOCHS The number of epochs to train for." >&2;
echo -e " LOSS The loss function to use. Default: cross-entropy (possible values: cross-entropy, cross-entropy-dice)." >&2;
echo -e " PATH_CHECKPOINT The path to a checkcpoint to load. If specified, a model will be loaded instead of being trained." >&2;
echo -e " PREDICT_COUNT The number of items from the (SCRAMBLED) dataset to make a prediction for." >&2;
echo -e " POSTFIX Postfix to append to the output dir (auto calculated)." >&2;
echo -e " ARGS Optional. Any additional arguments to pass to the python program." >&2;
@ -60,7 +61,7 @@ DIR_OUTPUT="output/$(date -u --rfc-3339=date)_${CODE}";
echo -e ">>> Additional args: ${ARGS}";
export PATH=$HOME/software/bin:$PATH;
export IMAGE_SIZE BATCH_SIZE DIR_RAINFALLWATER PATH_HEIGHTMAP PATH_COLOURMAP STEPS_PER_EPOCH DIR_OUTPUT PATH_CHECKPOINT EPOCHS PREDICT_COUNT NO_REMOVE_ISOLATED_PIXELS;
export IMAGE_SIZE BATCH_SIZE DIR_RAINFALLWATER PATH_HEIGHTMAP PATH_COLOURMAP STEPS_PER_EPOCH DIR_OUTPUT PATH_CHECKPOINT EPOCHS PREDICT_COUNT NO_REMOVE_ISOLATED_PIXELS LOSS;
echo ">>> Installing requirements";
conda run -n py38 pip install -q -r requirements.txt;

View file

@ -17,6 +17,7 @@ import matplotlib.pyplot as plt
import tensorflow as tf
from lib.dataset.dataset_mono import dataset_mono
from lib.ai.components.LossCrossEntropyDice import LossCrossEntropyDice
IMAGE_SIZE = int(os.environ["IMAGE_SIZE"]) if "IMAGE_SIZE" in os.environ else 128 # was 512; 128 is the highest power of 2 that fits the data
BATCH_SIZE = int(os.environ["BATCH_SIZE"]) if "BATCH_SIZE" in os.environ else 64
@ -27,12 +28,12 @@ PATH_COLOURMAP = os.environ["PATH_COLOURMAP"]
STEPS_PER_EPOCH = int(os.environ["STEPS_PER_EPOCH"]) if "STEPS_PER_EPOCH" in os.environ else None
REMOVE_ISOLATED_PIXELS = FALSE if "NO_REMOVE_ISOLATED_PIXELS" in os.environ else True
EPOCHS = int(os.environ["EPOCHS"]) if "EPOCHS" in os.environ else 25
PREDICT_COUNT = int(os.environ["PREDICT_COUNT"]) if "PREDICT_COUNT" in os.environ else 4
LOSS = os.environ["LOSS"] if "LOSS" in os.environ else "cross-entropy"
DIR_OUTPUT=os.environ["DIR_OUTPUT"] if "DIR_OUTPUT" in os.environ else f"output/{datetime.utcnow().date().isoformat()}_deeplabv3plus_rainfall_TEST"
PATH_CHECKPOINT = os.environ["PATH_CHECKPOINT"] if "PATH_CHECKPOINT" in os.environ else None
PREDICT_COUNT = int(os.environ["PREDICT_COUNT"]) if "PREDICT_COUNT" in os.environ else 4
if not os.path.exists(DIR_OUTPUT):
os.makedirs(os.path.join(DIR_OUTPUT, "checkpoints"))
@ -45,7 +46,10 @@ logger.info(f"> PATH_COLOURMAP {PATH_COLOURMAP}")
logger.info(f"> STEPS_PER_EPOCH {STEPS_PER_EPOCH}")
logger.info(f"> REMOVE_ISOLATED_PIXELS {REMOVE_ISOLATED_PIXELS} [NO_REMOVE_ISOLATED_PIXELS]")
logger.info(f"> EPOCHS {EPOCHS}")
logger.info(f"> LOSS {LOSS}")
logger.info(f"> DIR_OUTPUT {DIR_OUTPUT}")
logger.info(f"> PATH_CHECKPOINT {PATH_CHECKPOINT}")
logger.info(f"> PREDICT_COUNT {PREDICT_COUNT}")
@ -152,10 +156,17 @@ else:
# ██ ██ ██ ██ ██ ██ ██ ████ ██ ██ ████ ██████
if PATH_CHECKPOINT is None:
loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
loss_fn = None
if LOSS == "cross-entropy-dice":
loss_fn = LossCrossEntropyDice()
elif LOSS == "cross-entropy":
tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
else:
raise Exception(f"Error: Unknown loss function '{LOSS}' (possible values: cross-entropy, cross-entropy-dice).")
model.compile(
optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),
loss=loss,
loss=loss_fn,
metrics=["accuracy"],
)
logger.info(">>> Beginning training")

View file

@ -0,0 +1,41 @@
import math
import tensorflow as tf
def dice_loss(y_true, y_pred):
"""Compute Dice loss.
@source https://lars76.github.io/2018/09/27/loss-functions-for-segmentation.html#9
Args:
y_true (tf.Tensor): The ground truth label.
y_pred (tf.Tensor): The output predicted by the model.
Returns:
tf.Tensor: The computed Dice loss.
"""
y_pred = tf.math.sigmoid(y_pred)
numerator = 2 * tf.reduce_sum(y_true * y_pred)
denominator = tf.reduce_sum(y_true + y_pred)
return 1 - numerator / denominator
class LossCrossEntropyDice(tf.keras.losses.Loss):
"""Cross-entropy loss and dice loss combined together into one nice neat package.
Combines the two with mean.
@source https://lars76.github.io/2018/09/27/loss-functions-for-segmentation.html#9
"""
def __init__(self, **kwargs):
super(LossDice, self).__init__(**kwargs)
def call(self, y_true, y_pred):
y_true = tf.cast(y_true, tf.float32)
o = tf.nn.sigmoid_cross_entropy_with_logits(y_true, y_pred) + dice_loss(y_true, y_pred)
return tf.reduce_mean(o)
def get_config(self):
config = super(LossDice, self).get_config()
config.update({
})
return config