def predict_from_model(image):
image = cv2.cvtColor(image, cv2.COLOR_GRAY2RGB)
trn_tfms, val_tfms = tfms_from_model(resnext29_8_64(),
IMAGE_SIZE) # get transformations
# im = val_tfms(open_image('./data/test/8/758.jpg'))
im = val_tfms(image)
model.precompute = False # We'll pass in a raw image, not activations
preds = model.predict_array(im[None])
prediction = np.argmax(preds) # preds are log probabilities of CLASSES
return prediction
def init_model():
global model
os.makedirs(PATH, exist_ok=True)
data = get_data(
BATCH_SIZE,
IMAGE_SIZE) # data generator for batch size=32, image size=32x32
mod = resnext29_8_64()
basemodel = BasicModel(mod.cuda(), name='cifar10_rn29_8_64')
model = ConvLearner(data, basemodel)
model.load("32x32_8")
files = os.listdir(f'{PATH}valid/cats')[:5]
img = plt.imread(f'{PATH}valid/cats/{files[0]}')
''' Model and learning rate schedule
'''
''' Data augmentation
'''
sz = 112
tfms = tfms_from_model(resnet34, sz, aug_tfms=transforms_side_on, max_zoom=1.1)
#ims = np.stack([get_augs() for i in range(6)])
#plots(ims, rows=2)
data = ImageClassifierData.from_paths(PATH, tfms=tfms)
# By default when we create a learner, it sets all but the last layer to *frozen*. That means that it's still only updating the weights in the last layer when we call `fit`.
from fastai.models.cifar10.resnext import resnext29_8_64
m = resnext29_8_64()
bm = BasicModel(m.cuda(), name='cifar10_rn29_8_64')
#arch=resnet34
#learn = ConvLearner.pretrained(arch, data, precompute=True)
learn = ConvLearner(data, bm)
learn.unfreeze()
lrf = learn.lr_find()
#learn.sched.plot()
wd = 5e-4
learn.fit(lr, 1)
learn.fit(lr, 2, cycle_len=1)
learn.fit(lr, 3, cycle_len=1, cycle_mult=2, wds=wd)
#learn.sched.plot_lr()
''' Now that final layer is trained, fine-tuning the other layers to unfreeze the remaining layers