In [ ]:
%load_ext autoreload
%autoreload 2
In [ ]:
from nb_005 import *
STL-10¶
Basic data aug¶
In [ ]:
PATH = Path('data/stl10')
data_mean, data_std = map(tensor, ([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]))
data_norm,data_denorm = normalize_funcs(data_mean,data_std)
train_ds = ImageDataset.from_folder(PATH/'train')
valid_ds = ImageDataset.from_folder(PATH/'valid')
In [ ]:
x=valid_ds[0][0]
x.show()
x.shape
Out[ ]:
torch.Size([3, 96, 96])
In [ ]:
size=96
tfms = get_transforms(do_flip=True, max_rotate=10, max_lighting=0.2, max_warp=0.15, max_zoom=1.2)
# tfms = get_transforms(do_flip=True, max_rotate=10, max_lighting=0.2)
tds = transform_datasets(train_ds, valid_ds, tfms, size=size)#, padding_mode='zeros')
data = DataBunch.create(*tds, bs=32, num_workers=8, tfms=data_norm)
In [ ]:
(x,y) = next(iter(data.valid_dl))
_,axs = plt.subplots(4,4,figsize=(12,12))
for i,ax in enumerate(axs.flatten()): show_image(data_denorm(x[i].cpu()), ax)
In [ ]:
(x,y) = next(iter(data.train_dl))
_,axs = plt.subplots(4,4,figsize=(12,12))
for i,ax in enumerate(axs.flatten()): show_image(data_denorm(x[i].cpu()), ax)
In [ ]:
_,axs = plt.subplots(4,4,figsize=(12,12))
for i,ax in enumerate(axs.flat): tds[0][1][0].show(ax)
Train¶
In [ ]:
from torchvision.models import resnet18, resnet34, resnet50
arch = resnet50
In [ ]:
lr = 2e-3
In [ ]:
learn = ConvLearner(data, arch, 2 , wd=1e-1) #, train_bn=False #, callback_fns=[BnFreeze]
# , opt_fn=partial(optim.SGD, momentum=0.9))
learn.metrics = [accuracy]
learn.split(lambda m: (m[0][6], m[1]))
In [ ]:
lr_find(learn)
learn.recorder.plot()
VBox(children=(HBox(children=(IntProgress(value=0, max=1), HTML(value=''))), HTML(value='epoch train loss va…
In [ ]:
learn.fit_one_cycle(6, slice(lr))
VBox(children=(HBox(children=(IntProgress(value=0, max=6), HTML(value=''))), HTML(value='epoch train loss va…
Total time: 01:30 epoch train loss valid loss accuracy 0 0.550218 0.369369 0.872692 (00:14) 1 0.389043 0.287710 0.903462 (00:14) 2 0.354424 0.225025 0.926154 (00:15) 3 0.257456 0.202688 0.930000 (00:15) 4 0.205147 0.180714 0.942308 (00:15) 5 0.176819 0.176531 0.941538 (00:15)
In [ ]:
learn.save('0')
Gradual unfreezing¶
In [ ]:
learn.load('0')
In [ ]:
learn.unfreeze()
In [ ]:
lr = 1e-3
In [ ]:
learn.fit_one_cycle(6, slice(lr/25,lr))
VBox(children=(HBox(children=(IntProgress(value=0, max=6), HTML(value=''))), HTML(value='epoch train loss va…
Total time: 02:01 epoch train loss valid loss accuracy 0 0.230329 0.228656 0.924615 (00:20) 1 0.305863 0.254649 0.916154 (00:19) 2 0.324937 0.338827 0.886154 (00:20) 3 0.276075 0.260066 0.916538 (00:20) 4 0.136763 0.194284 0.940385 (00:20) 5 0.074670 0.176807 0.945769 (00:20)
In [ ]:
learn.save('1')
In [ ]:
learn.recorder.plot_losses()
Fin¶
In [ ]:
import pandas as pd
csv = pd.read_csv(PATH/'default.csv')
is_valid = csv['2']=='valid'
valid_df,train_df = csv[is_valid],csv[~is_valid]
len(valid_df),len(train_df)
In [ ]:
len(valid_ds)
Out[ ]:
2600
In [ ]:
train_fns,train_lbls,valid_fns,valid_lbls = map(np.array,
(train_df['0'],train_df['1'],valid_df['0'],valid_df['1']))
train_fns = [PATH/o for o in train_fns]
valid_fns = [PATH/o for o in valid_fns]
train_ds = ImageDataset(train_fns,train_lbls)
valid_ds = ImageDataset(valid_fns,valid_lbls, classes=train_ds.classes)
In [ ]: