values)):
if fold == 0:
break
imgs = bp.unpack_ndarray_from_file(
'../features/train_images_size128_pad8_max_noclean.bloscpack')
lbls = pd.read_csv('../input/train.csv').iloc[:, 1:4].values
trn_imgs = imgs[trn_ndx]
trn_lbls = lbls[trn_ndx]
vld_imgs = imgs[vld_ndx]
vld_lbls = lbls[vld_ndx]
training_set = Bengaliai_DS(trn_imgs,
trn_lbls,
transform=augs,
split_label=True,
RGB=True)
validation_set = Bengaliai_DS(vld_imgs, vld_lbls, split_label=True, RGB=True)
batch_size = 64
training_loader = DataLoader(training_set,
batch_size=batch_size,
num_workers=4,
shuffle=True)
validation_loader = DataLoader(validation_set,
batch_size=batch_size,
num_workers=4,
shuffle=False)
(1, 3),
[
iaa.Affine(scale={"x": (0.8, 1.), "y": (0.8, 1.)}, rotate=(-15, 15), shear=(-15, 15)),
iaa.PiecewiseAffine(scale=(0.02, 0.04)),
iaa.DirectedEdgeDetect(alpha=(.01, .99), direction=(0.0, 1.0)),
],
random_order=True
)
# In[5]:
batch_size = 64 # 64 is important as the fit_one_cycle arguments are probably tuned for this batch size
training_set = Bengaliai_DS(trn_imgs, trn_lbls, transform=augs, RGB=False)
validation_set = Bengaliai_DS(vld_imgs, vld_lbls, RGB=False)
training_loader = DataLoader(training_set, batch_size=batch_size, num_workers=6, shuffle=True) # , sampler=sampler , shuffle=True
validation_loader = DataLoader(validation_set, batch_size=batch_size, num_workers=6, shuffle=False)
data_bunch = DataBunch(train_dl=training_loader, valid_dl=validation_loader)
# ---
# ### model
# In[6]:
device = 'cuda:0'
trn_pdf.reset_index(inplace=True, drop=True)
imgs = bp.unpack_ndarray_from_file('../features/train_images.bloscpack')
lbls = pd.read_csv('../input/train.csv').iloc[:, 1:4].values
trn_imgs = imgs[trn_ndx]
trn_lbls = lbls[trn_ndx, 0:1]
vld_imgs = imgs[vld_ndx]
vld_lbls = lbls[vld_ndx, 0:1]
# In[5]:
#sampler = Balanced_Sampler(trn_pdf, count_column='image_id', primary_group='grapheme_root', secondary_group=['vowel_diacritic', 'consonant_diacritic'], size=trn_imgs.shape[0])
training_set = Bengaliai_DS(trn_imgs, trn_lbls, transform=None)
validation_set = Bengaliai_DS(vld_imgs, vld_lbls)
training_loader = DataLoader(training_set, batch_size=64, num_workers=6, shuffle=True) # , sampler=sampler , shuffle=True
validation_loader = DataLoader(validation_set, batch_size=64, num_workers=6, shuffle=False)
data_bunch = DataBunch(train_dl=training_loader, valid_dl=validation_loader)
# ---
# ### model
# In[6]:
device = 'cuda:0'