model.build((None, HEIGHT, WIDTH, 3))
# load_checkpoint("./drive/MyDrive/models/ImageNet-86/ckpt", model=backbone)
criterion = DetectionLoss(box_loss_fn=iou_loss(mode='ciou', offset=True),
cls_loss_fn=focal_loss(alpha=0.25, gamma=2.0),
centerness=True)
base_lr = 0.0025
epochs = 60
lr_schedule = CosineLR(base_lr * mul,
steps_per_epoch,
epochs,
min_lr=0,
warmup_min_lr=base_lr,
warmup_epoch=5)
optimizer = SGD(lr_schedule, momentum=0.9, nesterov=True, weight_decay=1e-4)
train_metrics = {
'loss': Mean(),
}
eval_metrics = {
'loss': MeanMetricWrapper(criterion),
}
def output_transform(output):
bbox_preds, cls_scores, centerness = get(
['bbox_pred', 'cls_score', 'centerness'], output)
return postprocess(bbox_preds,
cls_scores,
bbox_coder,
batch_size, eval_batch_size, transform, zip_transform)
setup_runtime(fp16=True)
ds_train, ds_test = distribute_datasets(ds_train, ds_test)
model = ResNet(depth=16, k=8, num_classes=100)
model.build((None, 32, 32, 3))
model.summary()
criterion = CrossEntropy(label_smoothing=params['label_smoothing'])
base_lr = params['base_lr']
epochs = 50
lr_schedule = CosineLR(base_lr, steps_per_epoch, epochs=epochs, min_lr=0)
optimizer = SGD(lr_schedule,
momentum=0.9,
weight_decay=params['weight_decay'],
nesterov=True)
train_metrics = {
'loss': Mean(),
'acc': CategoricalAccuracy(),
}
eval_metrics = {
'loss': CategoricalCrossentropy(from_logits=True),
'acc': CategoricalAccuracy(),
}
learner = SuperLearner(model,
criterion,
optimizer,
train_metrics=train_metrics,
eval_metrics=eval_metrics,
def objective(trial: optuna.Trial):
base_lr = trial.suggest_float("base_lr", 0.001, 0.05, step=0.001)
weight_decay = trial.suggest_loguniform("weight_decay", 1e-5, 1e-3)
ema = trial.suggest_categorical("ema", ["true", "false"])
ema_decay = trial.suggest_loguniform("ema_decay", 0.99,
0.9999) if ema == 'true' else None
@curry
def transform(image, label, training):
image = pad(image, 2)
image, label = to_tensor(image, label)
image = normalize(image, [0.1307], [0.3081])
label = tf.one_hot(label, 10)
return image, label
batch_size = 128
eval_batch_size = 256
ds_train, ds_test, steps_per_epoch, test_steps = make_mnist_dataset(
batch_size, eval_batch_size, transform, sub_ratio=0.01)
model = LeNet5()
model.build((None, 32, 32, 1))
criterion = CrossEntropy()
epochs = 20
lr_shcedule = CosineLR(base_lr, steps_per_epoch, epochs=epochs, min_lr=0)
optimizer = SGD(lr_shcedule,
momentum=0.9,
nesterov=True,
weight_decay=weight_decay)
train_metrics = {
'loss': Mean(),
'acc': CategoricalAccuracy(),
}
eval_metrics = {
'loss': CategoricalCrossentropy(from_logits=True),
'acc': CategoricalAccuracy(),
}
learner = SuperLearner(model,
criterion,
optimizer,
train_metrics=train_metrics,
eval_metrics=eval_metrics,
work_dir=f"./MNIST",
multiple_steps=True)
callbacks = [OptunaReportIntermediateResult('acc', trial)]
# if ema == 'true':
# callbacks.append(EMA(ema_decay))
learner.fit(ds_train,
epochs,
ds_test,
val_freq=2,
steps_per_epoch=steps_per_epoch,
val_steps=test_steps,
callbacks=callbacks)
return learner.metric_history.get_metric('acc', "eval")[-1]
'affine': False,
'track_running_stats': False,
},
})
set_primitives('tiny')
model = Network(4, 5)
model.build((None, 32, 32, 3))
criterion = CrossEntropy()
base_lr = 0.025
epochs = 240
lr_schedule = CosineLR(base_lr, steps_per_epoch, epochs=epochs, min_lr=1e-3)
optimizer_model = SGD(lr_schedule, momentum=0.9, weight_decay=3e-4)
optimizer_arch = AdamW(learning_rate=3e-4, beta_1=0.5, weight_decay=1e-3)
train_metrics = {
'loss': Mean(),
'acc': CategoricalAccuracy(),
}
eval_metrics = {
'loss': CategoricalCrossentropy(from_logits=True),
'acc': CategoricalAccuracy(),
}
learner = DARTSLearner(model,
criterion,
optimizer_arch,
optimizer_model,
x = self.normal2(x, hardwts, index)
x = self.reduce2(x)
x = self.normal3(x, hardwts, index)
x = self.avg_pool(x)
x = self.fc(x)
return x
model = ConvNet()
model.build((None, 32, 32, 1))
epochs = 200
lr_schedule = CosineLR(0.05, steps_per_epoch, epochs, min_lr=0)
optimizer_model = SGD(lr_schedule,
momentum=0.9,
weight_decay=1e-4,
nesterov=True,
exclude_from_weight_decay=['alpha_normal'])
optimizer_arch = Adam(1e-3)
train_loss = tf.keras.metrics.Mean()
train_acc = tf.keras.metrics.SparseCategoricalAccuracy()
@tf.function(jit_compile=True)
def train_step(batch):
x, y = batch
with tf.GradientTape() as tape:
p = model(x, training=True)
per_example_loss = tf.keras.losses.sparse_categorical_crossentropy(
y, p, from_logits=True)
def objective(trial: optuna.Trial):
cutout_prob = trial.suggest_float("cutout_prob", 0, 1.0, step=0.1)
mixup_alpha = trial.suggest_float("mixup_alpha", 0, 0.5, step=0.1)
label_smoothing = trial.suggest_uniform("label_smoothing", 0, 0.2)
base_lr = trial.suggest_float("base_lr", 0.01, 0.2, step=0.01)
weight_decay = trial.suggest_loguniform("weight_decay", 1e-5, 1e-3)
ema = trial.suggest_categorical("ema", ["true", "false"])
ema_decay = trial.suggest_loguniform("ema_decay", 0.995,
0.9999) if ema == 'true' else None
@curry
def transform(image, label, training):
if training:
image = random_crop(image, (32, 32), (4, 4))
image = tf.image.random_flip_left_right(image)
image = autoaugment(image, "CIFAR10")
image, label = to_tensor(image, label)
image = normalize(image, [0.491, 0.482, 0.447], [0.247, 0.243, 0.262])
if training:
image = random_apply(cutout(length=16), cutout_prob, image)
label = tf.one_hot(label, 100)
return image, label
def zip_transform(data1, data2):
return mixup(data1, data2, alpha=mixup_alpha)
batch_size = 128
eval_batch_size = 2048
ds_train, ds_test, steps_per_epoch, test_steps = make_cifar100_dataset(
batch_size, eval_batch_size, transform, zip_transform)
setup_runtime(fp16=True)
ds_train, ds_test = distribute_datasets(ds_train, ds_test)
model = ResNet(depth=16, k=8, num_classes=100)
model.build((None, 32, 32, 3))
model.summary()
criterion = CrossEntropy(label_smoothing=label_smoothing)
epochs = 50
lr_schedule = CosineLR(base_lr, steps_per_epoch, epochs=epochs, min_lr=0)
optimizer = SGD(lr_schedule,
momentum=0.9,
weight_decay=weight_decay,
nesterov=True)
train_metrics = {
'loss': Mean(),
'acc': CategoricalAccuracy(),
}
eval_metrics = {
'loss': CategoricalCrossentropy(from_logits=True),
'acc': CategoricalAccuracy(),
}
learner = SuperLearner(model,
criterion,
optimizer,
train_metrics=train_metrics,
eval_metrics=eval_metrics,
work_dir=f"./CIFAR100-NNI",
multiple_steps=True)
callbacks = [OptunaReportIntermediateResult('acc', trial)]
if ema == 'true':
callbacks.append(EMA(ema_decay))
learner.fit(ds_train,
epochs,
ds_test,
val_freq=1,
steps_per_epoch=steps_per_epoch,
val_steps=test_steps,
callbacks=callbacks)
return learner.metric_history.get_metric('acc', "eval")[-1]
elif "Depth" in str(type(m)):
w = np.transpose(m.depthwise_kernel.numpy(), [2, 3, 0, 1])
else:
w = np.transpose(m.kernel.numpy(), [3, 2, 0, 1])
wt = mt.weight.detach().numpy()
dw = w - wt
print(dw.mean(), dw.std())
weight_decay = 0
num_layers = 8
channels = 16
net1 = NASNet(channels, num_layers, False, 0, 10, FTSO)
net1.build((None, 32, 32, 3))
net2 = DARTS(channels, num_layers, False, 0, 10, FTSO)
optimizer1 = SGD(
0.025, momentum=0.9, weight_decay=weight_decay)
optimizer2 = torch.optim.SGD(
net2.parameters(), 0.025, momentum=0.9, weight_decay=weight_decay)
criterion1 = CrossEntropy()
criterion2 = nn.CrossEntropyLoss()
# net1.summary()
# summary(net2, (3, 32,32))
copy_conv(net1.stem.layers[0], net2.stem[0])
copy_bn(net1.stem.layers[1], net2.stem[1])
for i in range(num_layers):
cell1 = net1.cells[i]
cell2 = net2.cells[i]
copy_preprocess(cell1.preprocess0, cell2.preprocess0)
batch_size = 128
eval_batch_size = 256
ds_train, ds_test, steps_per_epoch, test_steps = make_mnist_dataset(
batch_size, eval_batch_size, transform, sub_ratio=0.1)
model = LeNet5()
model.build((None, 32, 32, 1))
criterion = CrossEntropy()
epochs = 20
base_lr = params["learning_rate"]
lr_shcedule = CosineLR(base_lr, steps_per_epoch, epochs=epochs, min_lr=0)
optimizer = SGD(lr_shcedule,
momentum=0.9,
nesterov=True,
weight_decay=params["weight_decay"])
train_metrics = {
'loss': Mean(),
'acc': CategoricalAccuracy(),
}
eval_metrics = {
'loss': CategoricalCrossentropy(from_logits=True),
'acc': CategoricalAccuracy(),
}
learner = SuperLearner(model,
criterion,
optimizer,
train_metrics=train_metrics,