def objective(trial):
# Data Augmentation
apply_tfms = trial.suggest_categorical("apply_tfms", [True, False])
if apply_tfms:
# MNIST is a hand-written digit dataset. Thus horizontal and vertical flipping are
# disabled. However, the two flipping will be important when the dataset is CIFAR or
# ImageNet.
tfms = aug_transforms(
do_flip=False,
flip_vert=False,
max_rotate=trial.suggest_int("max_rotate", 0, 45),
max_zoom=trial.suggest_float("max_zoom", 1, 2),
p_affine=trial.suggest_float("p_affine", 0.1, 1.0, step=0.1),
)
data = ImageDataLoaders.from_folder(
path, bs=BATCHSIZE, batch_tfms=tfms if apply_tfms else None
)
n_layers = trial.suggest_int("n_layers", 2, 5)
n_channels = [3]
for i in range(n_layers):
out_channels = trial.suggest_int("n_channels_{}".format(i), 3, 32)
n_channels.append(out_channels)
n_channels.append(2)
model = SimpleCNN(n_channels)
learn = Learner(
data,
model,
metrics=[accuracy],
# You could as FastAIPruningCallback in the fit function
cbs=FastAIPruningCallback(trial),
)
# See https://forums.fast.ai/t/how-to-diable-progress-bar-completely/65249/3
# to disable progress bar and logging info
with learn.no_bar():
with learn.no_logging():
learn.fit(EPOCHS)
return learn.validate()[-1]
data = ImageDataLoaders.from_dsets(train_ds,
valid_ds,
bs=cfg.batch_size,
num_workers=4 * 4,
pin_memory=True)
if torch.cuda.is_available():
data.cuda(), model.cuda()
cbs = [SaveModelCallback(monitor='dice_th', comp=np.greater)]
learn = Learner(data,
model,
metrics=cfg.metrics,
wd=cfg.weight_decay,
loss_func=cfg.loss_func,
opt_func=ranger,
cbs=cbs,
model_dir='models' + '_' + name + cfg.cv_method +
str(fold))
if cfg.mixed_precision_training:
learn.to_fp16()
# make learner to use all GPUs
learn.model = torch.nn.DataParallel(learn.model)
if fold != args.fold:
continue
# Fit
learn.fit_one_cycle(cfg.epochs, lr_max=cfg.max_learning_rate)
from src.callbacks import *
# import dataloader
from src.build_dataloader import dataloader_s224_bs16, dataloader_s512_bs4, dataloader_s312_bs12
# import misc
import argparse
# original train
# cbs = [ChooseTask, SaveModelCallback1(monitor='accuracy'), ReduceLROnPlateau(factor=3)]
# iterative prune
# cbs = [ChooseTask, LoadCkpt(ckpt_name='tst_bs16_origin', every_batch=True), Prune]
# test
cbs = [ChooseTask, LoadCkpt(ckpt_name='tst_bs16_origin'), LoadMask(mask_name='mask_tst.pth'), CountParams]
# train
if torch.cuda.is_available():
learner1 = Learner(dataloader_s512_bs4.cuda(), xse_resnet50.cuda(), metrics=accuracy, cbs=cbs)
learner2 = Learner(dataloader_s512_bs4.cuda(), efficient_b6.cuda(), metrics=accuracy, cbs=cbs)
learner3 = Learner(dataloader_s224_bs16.cuda(), xse_resnet50.cuda(), metrics=accuracy, cbs=cbs)
learner4 = Learner(dataloader_s224_bs16.cuda(), efficient_b6.cuda(), metrics=accuracy, cbs=cbs)
learner5 = Learner(dataloader_s312_bs12.cuda(), xse_resnet50.cuda(), metrics=accuracy, cbs=cbs)
learner6 = Learner(dataloader_s312_bs12.cuda(), efficient_b6.cuda(), metrics=accuracy, cbs=cbs)
learner7 = Learner(dataloader_s224_bs16.cuda(), tst_model.cuda(), metrics=accuracy, cbs=cbs)
else:
learner1 = Learner(dataloader_s512_bs4, xse_resnet50, metrics=accuracy, cbs=cbs)
learner2 = Learner(dataloader_s512_bs4, efficient_b6, metrics=accuracy, cbs=cbs)
learner3 = Learner(dataloader_s224_bs16, xse_resnet50, metrics=accuracy, cbs=cbs)
learner4 = Learner(dataloader_s224_bs16, efficient_b6, metrics=accuracy, cbs=cbs)
learner5 = Learner(dataloader_s312_bs12, xse_resnet50, metrics=accuracy, cbs=cbs)
learner6 = Learner(dataloader_s312_bs12, efficient_b6, metrics=accuracy, cbs=cbs)
learner7 = Learner(dataloader_s224_bs16, tst_model, metrics=accuracy, cbs=cbs)
def get_embs(self, x):
return self.convs(x)
def forward(self, x):
x = self.get_embs(x)
x = self.classifier(x)
return x
dls = ImageDataLoaders.from_folder(untar_data(URLs.MNIST),
train="training",
valid="testing",
num_workers=0)
learn = Learner(dls,
SimpleConv(ArcFaceClassifier(3, 10)),
metrics=accuracy,
loss_func=arcface_loss)
# %%
learn.fit_one_cycle(5, 5e-3)
# %%
def get_embs(model, dl):
embs = []
ys = []
for bx, by in tqdm(dl):
with torch.no_grad():
embs.append(model.get_embs(bx))
ys.append(by)