def __init__(self, args, data_loader, valid_loader=None):
self.data_loader = data_loader
self.valid_loader = valid_loader
self.mode = args.mode
self.batch_size = args.batch_size
self.mixed_training = args.mixed_training
self.n_epochs = args.n_epochs
self.save_dir = args.save_dir
if args.mixed_training:
# 計算を早くする為にfloat16で計算する.
# kerasのmixed_precicion.policyで設定可能．
policy = mixed_precicion.Policy('mixed_float16')
mixed_precicion.set_policy(policy)
self.n_classes = len(np.unique(data_loader.y_train))
"""
cifar10からMNISTに変更する為，以下の用にmodelのgetの仕方を変更した．
"""
#self.model = get_model((None, None, 3), self.n_classes)
w = data_loader.x_test.shape[1]
h = data_loader.x_test.shape[2]
self.model = get_model((w, h, 1), 10)
print("model input : " + str(self.model.input))
print("model output : " + str(self.model.output))
self.model.compile(loss=[
losses.SparseCategoricalCrossentropy(),
losses.SparseCategoricalCrossentropy()
],
optimizer=optimizers.Adam(lr=args.lr),
metrics=['acc'])
def newModel(self) -> tf.keras.Model:
if self.modelstr == "MobileNetV2":
base_model = MobileNetV2(weights='imagenet', include_top=False,
input_shape=(224, 224, 3))
base_model.trainable = True
inputs = tf.keras.Input(shape=(224, 224, 3), name="image_input")
x = base_model(inputs, training=True, )
x = GlobalAveragePooling2D()(x)
x = BatchNormalization()(x)
x = Dropout(0.3)(x)
x = Dense(1024, activation='relu')(x)
x = Dense(1024, activation='relu')(x)
x = Dense(512, activation='relu')(x)
x = Dense(128, activation='relu')(x)
preds = Dense(3, activation='softmax')(x)
self.model = tf.keras.Model(inputs=inputs, outputs=[preds], name="mobileNetV2")
self.model.compile(loss=losses.SparseCategoricalCrossentropy(from_logits=False),
optimizer=optimizers.Adam(learning_rate=1e-3),
metrics=['accuracy'])
else: # EfficientNet
def unfreeze_model(model):
# We unfreeze the top 20 layers while leaving BatchNorm layers frozen
for layer in model.layers[-20:]:
if not isinstance(layer, layers.BatchNormalization):
layer.trainable = True
inputs = tf.keras.Input(shape=(224, 224, 3), name="image_input")
conv_base = EfficientNetB0(input_shape=(224, 224, 3),
input_tensor=inputs, drop_connect_rate=0.4,
include_top=False)
conv_base.trainable = False
unfreeze_model(conv_base)
x = GlobalAveragePooling2D()(conv_base.output)
x = BatchNormalization()(x)
x = Dropout(0.3)(x)
x = Dense(512, activation='relu')(x)
x = Dense(256, activation='relu')(x)
x = Dense(128, activation='relu')(x)
preds = Dense(3, activation='softmax')(x)
self.model = tf.keras.Model(inputs=inputs, outputs=[preds], name="efficientNetB0")
self.model.compile(loss=losses.SparseCategoricalCrossentropy(from_logits=False),
optimizer=optimizers.Adam(learning_rate=1e-3),
metrics=['accuracy'])
return self.model
def actor_loss(self, acts_and_advs, logits):
actions, advantages = tf.split(acts_and_advs, 2, axis=-1)
weighted_sparse_ce = kls.SparseCategoricalCrossentropy(from_logits=True)
actions = tf.cast(actions, tf.int32)
policy_loss = weighted_sparse_ce(actions, logits, sample_weight=advantages)
return policy_loss
def train():
num_words = 20000
sequence_length = 100
depth = 6
filters = 64
channels = 128
block_filters = [filters] * depth
num_classes = 2
inputs = layers.Input(shape=(sequence_length, ), name="inputs")
x = layers.Embedding(num_words, channels)(inputs)
x = tcn.TCN(block_filters, kernel_size=8)(x)
outputs = layers.Dense(num_classes, activation="softmax", name="output")(x)
model = Model(inputs, outputs)
model.compile(optimizer="Adam",
metrics=[metrics.SparseCategoricalAccuracy()],
loss=losses.SparseCategoricalCrossentropy())
print(model.summary())
train_dataset, test_dataset = load_dataset(num_words, sequence_length)
model.fit(train_dataset.batch(32),
validation_data=test_dataset.batch(32),
callbacks=[
TensorBoard(
str(
Path("logs") /
datetime.now().strftime("%Y-%m-%dT%H-%M_%S")))
],
epochs=5)
def create_model(lr=0.0004, beta1=0.75, beta2=0.95, dropout=0.4):
model = models.Sequential()
model.add(
layers.Conv2D(32, (3, 3),
activation='relu',
input_shape=(28, 28, 1),
padding="same"))
model.add(layers.Conv2D(64, (3, 3), activation='relu', padding="same"))
model.add(layers.MaxPooling2D((2, 2), 2))
model.add(layers.Dropout(dropout))
model.add(layers.Conv2D(64, (3, 3), activation='relu', padding="same"))
model.add(layers.Conv2D(128, (3, 3), activation='relu', padding="same"))
model.add(layers.MaxPooling2D((2, 2), 2))
model.add(layers.Dropout(dropout))
model.add(layers.Flatten())
model.add(layers.Dense(256, activation='relu'))
model.add(layers.Dense(10))
adam = optimizers.Adam(lr, beta1, beta2)
model.compile(optimizer=adam,
loss=losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])
return model
def actor_loss(self, states, actions, values, rewards, next_values, dones):
policy = self.actor_model(
tf.convert_to_tensor(np.vstack(states), dtype=tf.float32))
advantages = []
for i in range(len(states)):
reward = np.array(rewards[i])
value = np.array(values[i])
next_value = np.array(next_values[i])
if dones[i]:
advantages.append(reward - value)
else:
advantages.append(reward + self.df * next_value - value)
advantages = tf.reshape(advantages, [len(states)])
tf.convert_to_tensor(advantages, dtype=tf.float32)
# SparseCategoricalCrossentropy
entropy = losses.categorical_crossentropy(policy,
policy,
from_logits=True)
ce_loss = losses.SparseCategoricalCrossentropy(from_logits=True)
#policy_loss = ce_loss(actions, policy, sample_weight=np.array(advantages)) # same way
log_pi = ce_loss(actions, policy)
policy_loss = log_pi * np.array(advantages)
policy_loss = tf.reduce_mean(policy_loss)
log_pi = tf.reduce_mean(log_pi)
return policy_loss - self.en * entropy, log_pi
def setUp(self):
"""Setup shared by all tests"""
self.scce = losses.SparseCategoricalCrossentropy()
self.bscce_equal = BalancedSCCE([1, 1, 1])
self.class_weights = [0.2, 0.3, 0.5]
self.bscce_unequal = BalancedSCCE(self.class_weights)
def train():
depth = 6
filters = 25
block_filters = [filters] * depth
sequence_length = 601
train_dataset, test_dataset = load_dataset(30000, sequence_length)
model = tcn.build_model(sequence_length=sequence_length,
channels=1,
num_classes=10,
filters=block_filters,
kernel_size=8,
return_sequence=True)
model.compile(optimizer=optimizers.RMSprop(lr=5e-4, clipnorm=1.),
metrics=[metrics.SparseCategoricalAccuracy()],
loss=losses.SparseCategoricalCrossentropy())
print(model.summary())
model.fit(train_dataset.batch(32),
validation_data=test_dataset.batch(32),
callbacks=[
TensorBoard(
str(
Path("logs") /
datetime.now().strftime("%Y-%m-%dT%H-%M_%S")))
],
epochs=10)
def __init__(self,
policymodel,
valuemodel,
data_sz=256,
batch_sz=80,
lr=0.000085,
entropy_const=1e-6,
epochs=20):
#self.model = model
self.policymodel = policymodel
self.valuemodel = valuemodel
self.policymodel.compile(
optimizer=tf.keras.optimizers.RMSprop(learning_rate=lr),
loss=[self._logits_loss])
self.valuemodel.compile(
optimizer=tf.keras.optimizers.RMSprop(learning_rate=lr),
loss=[self._value_loss])
self.gamma = 1 #discounting
self.data_sz = data_sz
self.batch_sz = batch_sz #batch size
self.epochs = epochs
self.entropy_const = entropy_const #constant for entropy maximization term in logit loss function
self.logit2logprob = kls.SparseCategoricalCrossentropy(
from_logits=True
) #tensorflow built in converts logits to log probability
def action_loss(self, actions, advantages, policy_prediction):
actions = tf.cast(actions, tf.int32)
policy_loss = kls.SparseCategoricalCrossentropy(from_logits=True)(
actions, policy_prediction, sample_weight=advantages)
policy_2 = tf.nn.softmax(policy_prediction)
entropy_loss = kls.categorical_crossentropy(policy_2, policy_2)
return policy_loss - self.params['entropy'] * entropy_loss
def test_logits_2d(self):
"""Testing logits for 2D data"""
y_true = [[1, 2], [0, 2]]
y_pred = [[[0.05, 0.95, 0], [0.1, 0.8, 0.1]],
[[0.1, 0.2, 0.7], [0.3, 0.5, 0.2]]]
self.scce = losses.SparseCategoricalCrossentropy(
from_logits=True, reduction=losses.Reduction.NONE)
scce = self.scce(y_true, y_pred).numpy()
self.sfl1 = SparseFocalLoss(gamma=self.gamma1, from_logits=True)
self.sfl2 = SparseFocalLoss(gamma=self.gamma2, from_logits=True)
scce1 = scce * \
(1 - np.where(get_one_hot(y_true), softmax(y_pred), 0).sum(axis=-1))\
**self.gamma1
scce1 = scce1.mean()
sfl1 = self.sfl1(y_true, y_pred).numpy()
np.testing.assert_allclose(scce1, sfl1, rtol=1e-7)
scce2 = scce * \
(1 - np.where(get_one_hot(y_true), softmax(y_pred), 0).sum(axis=-1))\
**self.gamma2
scce2 = scce2.mean()
sfl2 = self.sfl2(y_true, y_pred).numpy()
np.testing.assert_allclose(scce2, sfl2, rtol=1e-6)
def train():
depth = 6
filters = 25
block_filters = [filters] * depth
model = tcn.build_model(sequence_length=28 * 28,
channels=1,
num_classes=10,
filters=block_filters,
kernel_size=8)
model.compile(optimizer="Adam",
metrics=[metrics.SparseCategoricalAccuracy()],
loss=losses.SparseCategoricalCrossentropy())
print(model.summary())
train_dataset, test_dataset = load_dataset()
model.fit(train_dataset.batch(32),
validation_data=test_dataset.batch(32),
callbacks=[
TensorBoard(
str(
Path("logs") /
datetime.now().strftime("%Y-%m-%dT%H-%M_%S")))
],
epochs=10)
def train_model():
model = build_model()
print(model.summary())
optimizer = optimizers.Adam(learning_rate=0.001)
loss = losses.SparseCategoricalCrossentropy(from_logits=True)
model.compile(optimizer=optimizer, loss=loss, metrics=['acc'])
(x_train, y_train), (x_test, y_test) = load_preprocess_data()
epochs = 10
n_train = 60000
n_test = 10000
batch_size = 32
steps_per_epoch = n_train // batch_size
validation_steps = n_test // batch_size
train_data_set = convert_to_data_set(x_train,
y_train,
repeat_times=epochs,
shuffle_buffer_size=n_train,
batch_size=batch_size)
val_data_set = convert_to_data_set(x_test,
y_test,
repeat_times=epochs,
shuffle_buffer_size=n_test,
batch_size=batch_size)
my_callbacks = []
early_stopping_cb = callbacks.EarlyStopping(monitor='val_loss',
patience=5,
restore_best_weights=True)
my_callbacks.append(early_stopping_cb)
tensorboard_cb = callbacks.TensorBoard(log_dir='logs')
my_callbacks.append(tensorboard_cb)
checkpoint_path = 'models/base_cnn/ckpt'
checkpoint_cb = callbacks.ModelCheckpoint(filepath=checkpoint_path,
save_weights_only=True,
save_best_only=True)
my_callbacks.append(checkpoint_cb)
history = model.fit(train_data_set,
epochs=epochs,
steps_per_epoch=steps_per_epoch,
validation_data=val_data_set,
validation_steps=validation_steps,
callbacks=my_callbacks)
print('\n\n')
train_result = model.evaluate(x_train, y_train)
format_result(train_result, name='train')
val_result = model.evaluate(x_test, y_test)
format_result(val_result, name='val')
return history
def CNN(self, model):
model.add(
layers.Conv2D(filters=32,
kernel_size=(3, 3),
activation="relu",
input_shape=(28, 28, 1)))
model.add(layers.MaxPooling2D(pool_size=(2, 2)))
model.add(
layers.Conv2D(filters=64,
kernel_size=(3, 3),
activation="relu"))
model.add(layers.MaxPooling2D(pool_size=(2, 2)))
model.add(layers.Flatten())
model.add(layers.Dropout(0.5))
model.add(layers.Dense(128, activation="relu"))
model.add(layers.Dropout(0.5))
model.add(layers.Dense(64, activation="relu"))
model.add(layers.Dropout(0.5))
model.add(layers.Dense(10))
model.summary()
model.compile(
optimizer="adam",
loss=losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['acc'])
print("Optimizer: ", model.optimizer)
def _actor_loss(self, acts_and_advs, actor_logits):
"""The custom loss function for the actor
For explanation of how tf/keras calls it, see critic loss above and reference.
y_true = targets are actions and advantages. The y_pred = policy = output (logits) of
the actor network (not normalized)
"""
actions, advantages = tf.split(acts_and_advs, 2, axis=-1)
# sparse categorical CE loss obj that supports sample_weight arg on call()
# from_logits argument ensures transformation into normalized probabilities
weighted_sparse_ce = kls.SparseCategoricalCrossentropy(
from_logits=True)
# policy loss is defined by policy gradients, weighted by advantages
# note: we only calculate the loss on the actions we've actually taken
# that and because A2C is on policy method is the reason why policy gardients
# ususally require many episodeds to converge
actions = tf.cast(actions, tf.int32)
policy_loss = weighted_sparse_ce(actions,
actor_logits,
sample_weight=advantages)
# entropy loss can be calculated via CE over itself
entropy_loss = kls.categorical_crossentropy(actor_logits,
actor_logits,
from_logits=True)
#signs are flipped because optimizer minimizes
return policy_loss - self.ENTROPY_FACTOR * entropy_loss
def __init__(self,args,config_path,database,network_cls):
self.config = parse_config(config_path)
config = self.config
self.args = args
self.database = database
self.network_cls = network_cls
# Initialize an optimizer and an loss func
if self.config["clip_norm"]==.0:
opt_dict = {"learning_rate":self.config["lr"]}
elif self.config["clip_norm"]>.0:
opt_dict = {"learning_rate":self.config["lr"],"clipnorm":self.config["clip_norm"]}
else:
raise ValueError("clip_norm should be 0(No clipping) or greater float")
self.optimizer = getattr(optimizers, args.optimizer)(**opt_dict)
if self.args.binary:
self.loss_func = losses.BinaryCrossentropy(from_logits=True)
else:
self.loss_func = losses.SparseCategoricalCrossentropy(from_logits=True)
self.ckpt_cb = tf.keras.callbacks.ModelCheckpoint(
os.path.join(config["model_path"],config["log_path"],config["ckpt_path"],f"{args.encoder_name}.ckpt"))
self.csv_log_cb = tf.keras.callbacks.CSVLogger(
os.path.join(config["model_path"],config["log_path"],config["csv_path"],"log.csv"))
self.tb_cb = tf.keras.callbacks.TensorBoard(
os.path.join(config["model_path"],config["log_path"],config["tb_path"]))
def train():
unet_model = unet.build_model(*oxford_iiit_pet.IMAGE_SIZE,
channels=oxford_iiit_pet.channels,
num_classes=oxford_iiit_pet.classes,
layer_depth=4,
filters_root=64,
padding="same")
unet.finalize_model(unet_model,
loss=losses.SparseCategoricalCrossentropy(),
metrics=[metrics.SparseCategoricalAccuracy()],
auc=False,
learning_rate=LEARNING_RATE)
trainer = unet.Trainer(name="oxford_iiit_pet")
train_dataset, validation_dataset = oxford_iiit_pet.load_data()
trainer.fit(unet_model,
train_dataset,
validation_dataset,
epochs=25,
batch_size=1)
return unet_model
def _logits_loss(self, acts_and_advs, logits):
# a trick to input actions and advantages through same API
actions, advantages = tf.split(acts_and_advs, 2, axis=-1)
# sparse categorical CE loss obj that supports sample_weight arg on call()
# from_logits argument ensures transformation into normalized probabilities
weighted_sparse_ce = kls.SparseCategoricalCrossentropy(
from_logits=True)
# policy loss is defined by policy gradients, weighted by advantages
# note: we only calculate the loss on the actions we've actually taken
actions = tf.cast(actions, tf.int32)
policy_loss = weighted_sparse_ce(actions,
logits,
sample_weight=advantages)
# entropy loss can be calculated via CE over itself
entropy_loss = kls.categorical_crossentropy(logits,
logits,
from_logits=True)
# here signs are flipped because optimizer minimizes
return policy_loss - self.params['entropy'] * entropy_loss
def _logits_loss(self, actions_and_advantages, logits):
# A trick to input actions and advantages through the same API.
actions, advantages = tf.split(actions_and_advantages, 2, axis=-1)
# Sparse categorical CE loss obj that supports sample_weight arg on `call()`.
# `from_logits` argument ensures transformation into normalized probabilities.
weighted_sparse_ce = kls.SparseCategoricalCrossentropy(
from_logits=True)
# Policy loss is defined by policy gradients, weighted by advantages.
# Note: we only calculate the loss on the actions we've actually taken.
actions = tf.cast(actions, tf.int32)
### NOT SURE, but to ignore the case where advantages equal negative, which causing neg loss
# adv2 = tf.where( advantages > 0 , advantages, 0)
policy_loss = weighted_sparse_ce(actions,
logits,
sample_weight=advantages)
# policy_loss = weighted_sparse_ce(actions, logits, sample_weight=adv2)
# Entropy loss can be calculated as cross-entropy over itself.
probs = tf.nn.softmax(logits)
entropy_loss = kls.categorical_crossentropy(probs, probs)
# We want to minimize policy and maximize entropy losses.
# Here signs are flipped because the optimizer minimizes.
policy_loss2 = tf.where(policy_loss > 0, policy_loss,
tf.math.maximum(policy_loss, -10))
# return policy_loss - self.entropy_c * entropy_loss
return policy_loss2 - self.entropy_c * entropy_loss
def __init__(self, model):
self.model = model
self.gamma = .99 #discounting learning_rate = 3e-8
self.model.compile(optimizer=tf.keras.optimizers.RMSprop(lr=7e-3),
loss=[self._logits_loss, self._value_loss])
#I set learning rate small because rewards are pretty big, can try changing
self.logitloss = kls.SparseCategoricalCrossentropy(from_logits=True)
def train_step(model, optim, X, Y):
with tf.GradientTape() as tape:
Y_cap = model(X, training=True)
loss = losses.SparseCategoricalCrossentropy()(Y, Y_cap)
variables = model.trainable_variables
gradeints = tape.gradient(loss, variables)
optim.apply_gradients(zip(gradeints, variables))
return loss, Y_cap
def setUp(self):
"""Setup shared by all tests"""
self.scce = losses.SparseCategoricalCrossentropy(
reduction=losses.Reduction.NONE)
self.gamma1 = 2.0
self.gamma2 = 4.0
self.sfl1 = SparseFocalLoss(gamma=self.gamma1)
self.sfl2 = SparseFocalLoss(gamma=self.gamma2)
def __init__(self, max_len):
self.model_name = 'distilbert-base-uncased'
self.max_len = max_len
self.tkzr = DistilBertTokenizer.from_pretrained(self.model_name)
self.model = TFDistilBertForSequenceClassification.from_pretrained(
self.model_name)
self.optimizer = optimizers.Adam(learning_rate=3e-5)
self.loss = losses.SparseCategoricalCrossentropy(from_logits=True)
def compile_model(model):
model.compile(optimizer=optimizers.Nadam(),
loss=losses.SparseCategoricalCrossentropy(),
metrics=[
metrics.SparseCategoricalAccuracy(),
metrics.SparseTopKCategoricalAccuracy(5)
])
return model
def preprocessing(self, seq_length, embedding_dim, units, dropout_rate,
batch_size):
"""
Parameters
----------
seq_length : int
Input length in the network.
embedding_dim : int
Embedding Dimension.
units : int
Dimensionality of the LSTM output space.
dropout_rate : float
Dropout rate between 0 and 1.
batch_size : int
Batch size.
"""
self.seq_length = seq_length
self.embedding_dim = embedding_dim
self.units = units
self.dropout_rate = dropout_rate
self.batch_size = batch_size
# Extract all sentences from json files
list_json = glob.glob(self.input_folder + '\*.json')
dic_senders = preprocess(list_json)
all_sentences = ''
for key in dic_senders: # select all senders
sentences_with_sender = [
key + ' : ' + s + '\n' for s in dic_senders[key]
]
for s in sentences_with_sender:
all_sentences += s # and add their message together
vocab = sorted(set(all_sentences))
# Creating a mapping from unique characters to indices
self.__char2idx = {u: i for i, u in enumerate(vocab)}
self.__idx2char = np.array(vocab)
self.nb_char = len(vocab)
print('{} unique characters'.format(self.nb_char))
# Create dataset for training
text_as_int = np.array([self.__char2idx[c] for c in all_sentences])
dataset = tf.data.Dataset.from_tensor_slices(text_as_int)
sequences = dataset.batch(self.seq_length + 1, drop_remainder=True)
dataset = sequences.map(self.__split_input_target)
self.dataset = dataset.shuffle(self.BUFFER_SIZE).batch(
self.batch_size, drop_remainder=True)
# Create model
self.model = self.__create_model(embedding_dim, units, dropout_rate,
self.batch_size)
self.model.summary()
self.model.compile(
optimizer='adam',
loss=losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])
def compile(self, learning_rate):
'''
Compiles this model.
'''
optimizer = optimizers.Adam(learning_rate=learning_rate)
loss = losses.SparseCategoricalCrossentropy(from_logits=True)
super().compile(optimizer=optimizer, loss=loss, metrics=['accuracy'])
def model_fn():
# We _must_ create a new model here, and _not_ capture it from an external
# scope. TFF will call this within different graph contexts.
keras_model = create_keras_model()
return tff.learning.from_keras_model(
keras_model,
input_spec=preprocessed_sample_dataset.element_spec,
loss=losses.SparseCategoricalCrossentropy(),
metrics=[metrics.SparseCategoricalAccuracy()])
def test_equal_weights_logits_1d(self):
"""Testing equal weights logits for 1D data"""
y_true = [1, 2]
y_pred = [[-0.05, 0.3, 0.19], [0.2, -0.4, 0.12]]
self.scce = losses.SparseCategoricalCrossentropy(from_logits=True)
self.bscce_equal = BalancedSCCE([1, 1, 1], from_logits=True)
scce = self.scce(y_true, y_pred).numpy()
bscce = self.bscce_equal(y_true, y_pred).numpy()
np.testing.assert_array_equal(scce, bscce)
def setUp(self):
"""Setup shared by all tests"""
self.scce = losses.SparseCategoricalCrossentropy(
reduction=losses.Reduction.NONE)
self.gamma = 4.0
self.sfl_equal = BalancedSparseFocalLoss([1, 1, 1], gamma=self.gamma)
self.class_weights = [0.2, 0.3, 0.5]
self.sfl_unequal = BalancedSparseFocalLoss(self.class_weights,
gamma=self.gamma)
def test_equal_weights_reduction_1d(self):
"""Testing equal weights reductions for 1D data"""
y_true = [1, 2]
y_pred = [[0.05, 0.95, 0], [0.1, 0.8, 0.1]]
self.scce = losses.SparseCategoricalCrossentropy(
reduction=losses.Reduction.SUM)
self.bscce_equal = BalancedSCCE([1, 1, 1],
reduction=losses.Reduction.SUM)
scce = self.scce(y_true, y_pred).numpy()
bscce = self.bscce_equal(y_true, y_pred).numpy()
np.testing.assert_array_equal(scce, bscce)
self.scce = losses.SparseCategoricalCrossentropy(
reduction=losses.Reduction.NONE)
self.bscce_equal = BalancedSCCE([1, 1, 1],
reduction=losses.Reduction.NONE)
scce = self.scce(y_true, y_pred).numpy()
bscce = self.bscce_equal(y_true, y_pred).numpy()
np.testing.assert_array_equal(scce, bscce)