In [ ]:

    

# automatic reloading and inline plotting
%reload_ext autoreload
%autoreload 2
%matplotlib inline


    

In [ ]:

    

# imports
# This file contains all the main external libs we'll use
from fastai.imports import *

from fastai.transforms import *
from fastai.conv_learner import *
from fastai.model import *
from fastai.dataset import *
from fastai.sgdr import *
from fastai.plots import *


    

In [ ]:

    

PATH = "data/dog_breed_identification/"
sz = 128
bs = 128


    

In [ ]:

    

# check images
files = !ls {PATH}train/ | head
img = plt.imread(f'{PATH}train/{files[0]}')
plt.imshow(img)


    

In [ ]:

    

# check format that labels are given in, choose validation labels
labels = pd.read_csv(PATH+'labels.csv')

labels.pivot_table(columns='breed', aggfunc=len).T.sort_values(by='id')


    

In [ ]:

    

# make val indices
val_idxs = get_cv_idxs(len(labels))


    

In [ ]:

    

def get_data(sz, bs, arch):
    # specify transforms I want
    # can play with max zoom after basic train
    tfms = tfms_from_model(arch, sz, aug_tfms=transforms_side_on, max_zoom=1.1)
    data = ImageClassifierData.from_csv(path=PATH, folder='train', csv_fname=PATH +
                                        'labels.csv', bs=bs, tfms=tfms, val_idxs=val_idxs, suffix='.jpg', test_name='test')
    return data if sz > 300 else data.resize(340, 'tmp')


    

In [ ]:

    

tfms = tfms_from_model(arch, sz, aug_tfms=transforms_side_on, max_zoom=1.1)
data = ImageClassifierData.from_csv(path=PATH, folder='train', csv_fname=PATH +'labels.csv', bs=bs, tfms=tfms, val_idxs=val_idxs, suffix='.jpg', test_name='test')


    

In [ ]:

    

arch = resnext101_64
data = get_data(sz, bs, arch)
learn = ConvLearner.pretrained(arch, data, precompute=True)


    

In [ ]:

    

# this block was for checking original size
# size_d = {k: Image.open(PATH+k).size for k in data.trn_ds.fnames}
# widths, heights = list(zip(*size_d.values()))
# widths = np.array(widths)
# heights = np.array(heights)
# plt.hist(widths[widths < 1000])
# plt.hist(heights[heights < 1000])


    

In [ ]:

    

learn.lr_find()


    

In [ ]:

    

learn.sched.plot_lr()


    

In [ ]:

    

learn.sched.plot()


    

In [ ]:

    

lr=0.12


    

In [ ]:

    

learn.fit(lr, n_cycle=3)


    

In [ ]:

    

# now disable precompute because data augmentation
learn.precompute=False


    

In [ ]:

    

learn.fit(lr, n_cycle=3, cycle_len=1)


    

In [ ]:

    

# unfreeze layers
learn.unfreeze()
lr_rates = np.array([lr/100, lr/10, lr])


    

In [ ]:

    

learn.fit(lr_rates, n_cycle=3, cycle_len=1)


    

In [ ]: