Getting the Data

from fastai.vision import *

path = untar_data('https://s3.amazonaws.com/fast-ai-imageclas/imagewoof2') 

We will decode the dataset labels to be more meaningful to a human.

lbl_dict = dict(
  n02093754='Australian terrier', 
  n02089973='Border terrier', 
  n02099601='Samoyed', 
  n02087394='Beagle', 
  n02105641='Shih-Tzu', 
  n02096294='English foxhound', 
  n02088364='Rhodesian ridgeback', 
  n02115641='Dingo', 
  n02111889='Golden retriever', 
  n02086240='Old English sheepdog'
)

def lbl_trfm(o): 
  return lbl_dict[str(o).split('/')[-2]] # Use folder name of image object `o`

Data augmentation

Instead of feeding the model with the same pictures every time, we do small random transformations (a bit of rotation, zoom, translation, etc...) that don't change what's inside the image (to the human eye) but do change its pixel values. Models trained with data augmentation will then generalize better.

tfms = get_transforms()
tfms

([RandTransform(tfm=TfmCrop (crop_pad), kwargs={'row_pct': (0, 1), 'col_pct': (0, 1), 'padding_mode': 'reflection'}, p=1.0, resolved={}, do_run=True, is_random=True, use_on_y=True),
  RandTransform(tfm=TfmPixel (flip_lr), kwargs={}, p=0.5, resolved={}, do_run=True, is_random=True, use_on_y=True),
  RandTransform(tfm=TfmCoord (symmetric_warp), kwargs={'magnitude': (-0.2, 0.2)}, p=0.75, resolved={}, do_run=True, is_random=True, use_on_y=True),
  RandTransform(tfm=TfmAffine (rotate), kwargs={'degrees': (-10.0, 10.0)}, p=0.75, resolved={}, do_run=True, is_random=True, use_on_y=True),
  RandTransform(tfm=TfmAffine (zoom), kwargs={'scale': (1.0, 1.1), 'row_pct': (0, 1), 'col_pct': (0, 1)}, p=0.75, resolved={}, do_run=True, is_random=True, use_on_y=True),
  RandTransform(tfm=TfmLighting (brightness), kwargs={'change': (0.4, 0.6)}, p=0.75, resolved={}, do_run=True, is_random=True, use_on_y=True),
  RandTransform(tfm=TfmLighting (contrast), kwargs={'scale': (0.8, 1.25)}, p=0.75, resolved={}, do_run=True, is_random=True, use_on_y=True)],
 [RandTransform(tfm=TfmCrop (crop_pad), kwargs={}, p=1.0, resolved={}, do_run=True, is_random=True, use_on_y=True)])

Let's pull together the data into a singe ImageDataBunch object.

data = (ImageList.from_folder(path)
        .split_by_folder(valid='val')
        .label_from_func(lbl_trfm)
        .transform(tfms, size=160)
        .databunch(bs=32)
        .normalize(imagenet_stats))

data

ImageDataBunch;

Train: LabelList (9025 items)
x: ImageList
Image (3, 160, 160),Image (3, 160, 160),Image (3, 160, 160),Image (3, 160, 160),Image (3, 160, 160)
y: CategoryList
Australian terrier,Australian terrier,Australian terrier,Australian terrier,Australian terrier
Path: /root/.fastai/data/imagewoof2;

Valid: LabelList (3929 items)
x: ImageList
Image (3, 160, 160),Image (3, 160, 160),Image (3, 160, 160),Image (3, 160, 160),Image (3, 160, 160)
y: CategoryList
Australian terrier,Australian terrier,Australian terrier,Australian terrier,Australian terrier
Path: /root/.fastai/data/imagewoof2;

Test: None

data.show_batch(5, figsize=(16,16))

Randomly Initialized CONV Model

results = pd.DataFrame(index=range(5))

int(torch.cuda.max_memory_allocated(0)/(1024*1024))

4627

from fastai.callbacks import *

learn = cnn_learner(data, models.resnet50, metrics=[accuracy], pretrained=False, callback_fns=[CSVLogger,ShowGraph])

learn.lr_find()
learn.recorder.plot()

LR Finder is complete, type {learner_name}.recorder.plot() to see the graph.

learn.fit_one_cycle(5,max_lr=5e-3)

epoch	train_loss	valid_loss	accuracy	time
0	2.748060	3.147827	0.166455	03:03
1	2.625360	3.451415	0.173836	03:02
2	2.470364	2.222389	0.196997	03:02
3	2.258959	2.040264	0.260372	03:00
4	2.075738	1.975030	0.312548	03:02

results['resnet50'] = pd.read_csv(path/'history.csv',usecols=['accuracy'])
results.head()

results.plot()

<matplotlib.axes._subplots.AxesSubplot at 0x7fe84cc26518>

Pretrained CONV net Model

learn = cnn_learner(data, models.resnet50, metrics=[accuracy], pretrained=True, callback_fns=[CSVLogger,ShowGraph])

learn.lr_find()
learn.recorder.plot()

LR Finder is complete, type {learner_name}.recorder.plot() to see the graph.

learn.fit_one_cycle(5,max_lr=slice(2e-4,2e-3,2e-2))

epoch	train_loss	valid_loss	accuracy	time
0	0.567119	0.337504	0.903029	03:01
1	0.414245	0.303489	0.909137	03:02
2	0.296347	0.235700	0.929244	03:01
3	0.226097	0.204954	0.937389	03:00
4	0.189701	0.203582	0.937898	02:59

results['resnet50 pretrained diff LR 2'] = pd.read_csv(path/'history.csv',usecols=['accuracy'])
results.plot()

<matplotlib.axes._subplots.AxesSubplot at 0x7fe87408b630>

Results

interp = ClassificationInterpretation.from_learner(learn)
losses,idxs = interp.top_losses()

interp.plot_confusion_matrix(figsize=(8,8))

interp.plot_top_losses(9, figsize=(16,16))