def test_reduce_to_binary_crossentropy():
"""Focal loss with gamma=0 should be the same as cross-entropy."""
# From probabilities
for y_true, y_pred in product(Y_TRUE, Y_PRED_PROB):
# tf.keras.losses.binary_crossentropy averages its output along the last
# axis, so we do the same here
focal_loss = binary_focal_loss(y_true=y_true, y_pred=y_pred, gamma=0)
focal_loss = tf.math.reduce_mean(focal_loss, axis=-1)
ce = tf.keras.losses.binary_crossentropy(y_true=y_true, y_pred=y_pred)
tf.debugging.assert_near(focal_loss, ce)
# From logits
for y_true, y_pred in product(Y_TRUE, Y_PRED_LOGITS):
focal_loss = binary_focal_loss(y_true=y_true, y_pred=y_pred, gamma=0,
from_logits=True)
ce = tf.nn.sigmoid_cross_entropy_with_logits(
labels=tf.dtypes.cast(y_true, dtype=tf.float32),
logits=tf.dtypes.cast(y_pred, dtype=tf.float32),
)
tf.debugging.assert_near(focal_loss, ce)
# From logits, with positive class weighting
for y_true, y_pred, pos_weight in product(Y_TRUE, Y_PRED_LOGITS, (1, 2)):
focal_loss = binary_focal_loss(y_true=y_true, y_pred=y_pred, gamma=0,
from_logits=True, pos_weight=pos_weight)
ce = tf.nn.weighted_cross_entropy_with_logits(
labels=tf.dtypes.cast(y_true, dtype=tf.float32),
logits=tf.dtypes.cast(y_pred, dtype=tf.float32),
pos_weight=pos_weight,
)
tf.debugging.assert_near(focal_loss, ce)
def test_computation_sanity_checks():
"""Make sure the focal loss computation behaves as expected."""
for y_true, y_pred_logits in product(Y_TRUE, Y_PRED_LOGITS):
for y_pred_prob, pos_weight in product((Y_PRED_PROB), (None, 1, 2)):
for gamma, label_smoothing in product((0, 1, 2), (None, 0.1, 0.5)):
focal_loss_prob = binary_focal_loss(
y_true=y_true,
y_pred=y_pred_prob,
gamma=gamma,
from_logits=False,
pos_weight=pos_weight,
label_smoothing=label_smoothing,
)
focal_loss_logits = binary_focal_loss(
y_true=y_true,
y_pred=y_pred_logits,
gamma=gamma,
from_logits=True,
pos_weight=pos_weight,
label_smoothing=label_smoothing,
)
losses = [focal_loss_prob, focal_loss_logits]
if not (isinstance(y_true, tf.Tensor)
or isinstance(y_pred_logits, tf.Tensor)):
numpy_focal_loss_logits = numpy_binary_focal_loss(
y_true=y_true,
y_pred=y_pred_logits,
gamma=gamma,
from_logits=True,
pos_weight=pos_weight,
label_smoothing=label_smoothing,
)
losses.append(numpy_focal_loss_logits)
if not (isinstance(y_true, tf.Tensor)
or isinstance(y_pred_prob, tf.Tensor)):
numpy_focal_loss_prob = numpy_binary_focal_loss(
y_true=y_true,
y_pred=y_pred_prob,
gamma=gamma,
from_logits=False,
pos_weight=pos_weight,
label_smoothing=label_smoothing,
)
losses.append(numpy_focal_loss_prob)
for i, loss_1 in enumerate(losses):
for loss_2 in losses[(i + 1):]:
tf.debugging.assert_near(loss_1, loss_2, atol=1e-5,
rtol=1e-5)
def rpn_class_anchorless_loss_graph(rpn_match, rpn_class_logits):
# Squeeze last dim to simplify
rpn_match = tf.squeeze(rpn_match, -1)
indices = tf.where(K.not_equal(rpn_match, 0))
target_class = tf.gather_nd(rpn_match, indices)
rpn_class_logits = tf.gather_nd(rpn_class_logits, indices)
target_class = K.cast(K.equal(target_class, 1), tf.int32)
pos_counts = K.sum(K.cast(K.equal(target_class, 1), tf.float32))
loss = binary_focal_loss(y_true=target_class,
y_pred=rpn_class_logits[..., -1],
gamma=2,
pos_weight=0.75,
from_logits=True)
loss = K.sum(loss) / pos_counts
"""
loss = BinaryFocalLoss(gamma=2, pos_weight=0.75, from_logits=True)(
y_true = target_class,
y_pred = rpn_class_logits[..., -1])
loss = focal_loss(gamma=2.0, alpha=0.75, from_logits=True)(
y_true=target_class,
y_pred=rpn_class_logits[..., -1])
loss = K.sum(loss)/pos_counts
"""
return loss
def test_computation_sanity_checks(self, y_true, y_pred_logits,
y_pred_prob, pos_weight, gamma,
label_smoothing):
"""Make sure the focal loss computation behaves as expected."""
focal_loss_prob = binary_focal_loss(
y_true=y_true,
y_pred=y_pred_prob,
gamma=gamma,
from_logits=False,
pos_weight=pos_weight,
label_smoothing=label_smoothing,
)
focal_loss_logits = binary_focal_loss(
y_true=y_true,
y_pred=y_pred_logits,
gamma=gamma,
from_logits=True,
pos_weight=pos_weight,
label_smoothing=label_smoothing,
)
losses = [focal_loss_prob, focal_loss_logits]
if not (isinstance(y_true, tf.Tensor)
or isinstance(y_pred_logits, tf.Tensor)):
numpy_focal_loss_logits = numpy_binary_focal_loss(
y_true=y_true,
y_pred=y_pred_logits,
gamma=gamma,
from_logits=True,
pos_weight=pos_weight,
label_smoothing=label_smoothing,
)
losses.append(numpy_focal_loss_logits)
if not (isinstance(y_true, tf.Tensor)
or isinstance(y_pred_prob, tf.Tensor)):
numpy_focal_loss_prob = numpy_binary_focal_loss(
y_true=y_true,
y_pred=y_pred_prob,
gamma=gamma,
from_logits=False,
pos_weight=pos_weight,
label_smoothing=label_smoothing,
)
losses.append(numpy_focal_loss_prob)
for i, loss_1 in enumerate(losses):
for loss_2 in losses[(i + 1):]:
self.assertAllClose(loss_1, loss_2, atol=1e-5, rtol=1e-5)
def train():
sample_input = np.empty(
[FRAMES_PER_VIDEO, FRAME_HEIGHT, FRAME_WIDTH, FRAME_CHANNEL], dtype=np.uint8)
# Read Dataset
d_train = pd.read_csv(os.path.join('train.csv'))
d_valid = pd.read_csv(os.path.join('test.csv'))
# Split data into random training and validation sets
nb_classes = len(set(d_train['class']))
video_train_generator = video_gen(
d_train, FRAMES_PER_VIDEO, FRAME_HEIGHT, FRAME_WIDTH, FRAME_CHANNEL, nb_classes, batch_size=BATCH_SIZE, augmentations=True)
video_val_generator = video_gen(
d_valid, FRAMES_PER_VIDEO, FRAME_HEIGHT, FRAME_WIDTH, FRAME_CHANNEL, nb_classes, batch_size=BATCH_SIZE, augmentations=False)
# Get Model
# model = densenet121_3D_DropOut(sample_input.shape, nb_classes)
model = T3D169_Dropout(sample_input.shape, nb_classes, d_rate=0.2)
checkpoint = ModelCheckpoint('T3D_weights.hdf5', monitor='val_loss',
verbose=1, save_best_only=True, mode='min', save_weights_only=True)
# reduceLROnPlat = ReduceLROnPlateau(monitor='val_loss', factor=0.1,
# patience=20,
# verbose=1, mode='min', min_delta=0.0001, cooldown=2, min_lr=1e-6)
csvLogger = CSVLogger('T3D_history.csv', append=True)
tensorboard = TensorBoard(log_dir=f'./logs/T3D-{int(time.time())}')
callbacks_list = [checkpoint, csvLogger, tensorboard]
# compile model
#optim = Adam(lr=1e-4, decay=1e-6)
#optim = SGD(lr = 0.1, momentum=0.9, decay=1e-4, nesterov=True)
optim = AdaBound()
model.compile(optimizer=optim, loss=[binary_focal_loss(alpha=.25, gamma=2)], metrics=['accuracy'])
if os.path.exists('./T3D_weights.hdf5'):
print('Pre-existing model weights found, loading weights.......')
model.load_weights('./T3D_weights.hdf5')
print('Weights loaded')
# train model
print('Training started....')
train_steps = len(d_train)//BATCH_SIZE
val_steps = len(d_valid)//BATCH_SIZE
history = model.fit_generator(
video_train_generator,
steps_per_epoch=train_steps,
epochs=EPOCHS,
validation_data=video_val_generator,
validation_steps=val_steps,
verbose=1,
callbacks=callbacks_list,
workers=1,
use_multiprocessing=True
)
model.save(MODEL_FILE_NAME)
def test_train_dummy_binary_classifier():
"""Train a simple model to make sure that BinaryFocalLoss works."""
# Data/model parameters
n_examples = 100
n_features = 16
epochs = 3
random_state = np.random.RandomState(0)
# Generate some fake data
x = random_state.binomial(n=1, p=0.5, size=(n_examples, n_features))
x = 2.0 * x.astype(np.float32) - 1.0
weights = 100 * np.ones(shape=(n_features, 1)).astype(np.float32)
y = (x.dot(weights) > 0).astype(np.int8)
# Number of positive and negative examples
n_pos = y.sum()
n_neg = n_examples - n_pos
for pos_weight in (None, (n_neg / n_pos)):
for gamma, label_smoothing in product((0, 2), (None, 0.1)):
for from_logits in (True, False):
if from_logits:
activation = None
else:
activation = 'sigmoid'
# Just a linear classifier (without bias term)
model = tf.keras.Sequential(layers=[
tf.keras.layers.Input(shape=n_features),
tf.keras.layers.Dense(units=1, use_bias=False,
activation=activation),
])
model.compile(
optimizer='sgd',
loss=BinaryFocalLoss(gamma=gamma, pos_weight=pos_weight,
from_logits=from_logits,
label_smoothing=label_smoothing),
metrics=['accuracy'],
)
stop_on_nan = tf.keras.callbacks.TerminateOnNaN()
history = model.fit(x, y, batch_size=n_examples, epochs=epochs,
callbacks=[stop_on_nan])
history = history.history
# Check that we didn't stop early: if we did then we
# encountered NaNs during training, and that shouldn't happen
assert len(history['loss']) == epochs
# Check that BinaryFocalLoss and binary_focal_loss agree (at
# least when averaged)
model_loss, *_ = model.evaluate(x, y)
y_pred = model.predict(x)
loss = binary_focal_loss(y_true=y, y_pred=y_pred,
gamma=gamma, pos_weight=pos_weight,
from_logits=from_logits,
label_smoothing=label_smoothing)
loss = tf.math.reduce_mean(loss)
tf.debugging.assert_near(loss, model_loss)
def test_reduce_to_binary_crossentropy_from_probabilities(
self, y_true, y_pred):
"""Focal loss with gamma=0 should be the same as cross-entropy."""
# tf.keras.losses.binary_crossentropy averages its output along the last
# axis, so we do the same here
focal_loss = binary_focal_loss(y_true=y_true, y_pred=y_pred, gamma=0)
focal_loss = tf.math.reduce_mean(focal_loss, axis=-1)
ce = tf.keras.losses.binary_crossentropy(y_true=y_true, y_pred=y_pred)
self.assertAllClose(focal_loss, ce)
def test_reduce_to_binary_crossentropy_from_logits(self, y_true, y_pred):
"""Focal loss with gamma=0 should be the same as cross-entropy."""
focal_loss = binary_focal_loss(y_true=y_true,
y_pred=y_pred,
gamma=0,
from_logits=True)
ce = tf.nn.sigmoid_cross_entropy_with_logits(
labels=tf.dtypes.cast(y_true, dtype=tf.dtypes.float32),
logits=tf.dtypes.cast(y_pred, dtype=tf.dtypes.float32),
)
self.assertAllClose(focal_loss, ce)
def test_train_dummy_binary_classifier(self, n_examples, n_features,
epochs, pos_weight, gamma,
label_smoothing, from_logits,
random_state):
# Generate some fake data
x = random_state.binomial(n=1, p=0.5, size=(n_examples, n_features))
x = 2.0 * x.astype(np.float32) - 1.0
weights = 100 * np.ones(shape=(n_features, 1)).astype(np.float32)
y = (x.dot(weights) > 0).astype(np.int8)
model = get_dummy_binary_classifier(n_features=n_features,
gamma=gamma,
pos_weight=pos_weight,
label_smoothing=label_smoothing,
from_logits=from_logits)
history = model.fit(x,
y,
batch_size=n_examples,
epochs=epochs,
callbacks=[tf.keras.callbacks.TerminateOnNaN()])
history = history.history
# Check that we didn't stop early: if we did then we
# encountered NaNs during training, and that shouldn't happen
self.assertEqual(len(history['loss']), epochs)
# Check that BinaryFocalLoss and binary_focal_loss agree (at
# least when averaged)
model_loss, *_ = model.evaluate(x, y)
y_pred = model.predict(x)
loss = binary_focal_loss(y_true=y,
y_pred=y_pred,
gamma=gamma,
pos_weight=pos_weight,
from_logits=from_logits,
label_smoothing=label_smoothing)
loss = tf.math.reduce_mean(loss)
self.assertAllClose(loss, model_loss)