In [ ]:

!pip install fastai --upgrade

In [1]:

import fastai.vision.all as vision
import torch

import numpy as np
from PIL import Image
from pathlib import Path
import os
import matplotlib.pyplot as plt

In [ ]:

path = vision.untar_data(vision.URLs.IMAGENETTE)
path.ls()

In [ ]:

path.ls()[0].ls()

In [3]:

def splitter(path): return str(path.parent.parent.name) == 'val'

db = vision.DataBlock(
    blocks=(vision.ImageBlock, vision.CategoryBlock),
    get_items=vision.get_image_files,
    get_y=vision.parent_label, 
    item_tfms=vision.Resize(460),
    batch_tfms=vision.aug_transforms(size=224, min_scale=0.75),
    splitter=vision.FuncSplitter(splitter)
)

In [7]:

ds = db.datasets(path)

In [ ]:

ds.train, ds.valid, ds.vocab

In [124]:

dls = db.dataloaders(path, bs=32)

In [125]:

dls.show_batch()

In [129]:

learn = vision.Learner(dls, vision.xresnet50(), loss_func=torch.nn.CrossEntropyLoss(), metrics=vision.accuracy, pretrained=False)

In [130]:

learn.fit_one_cycle(5, 3e-3)

epoch	train_loss	valid_loss	accuracy	time
0	1.658774	2.103384	0.401529	02:44
1	1.247024	2.213533	0.491210	02:40
2	0.914718	0.945459	0.707771	02:40
3	0.737189	0.707571	0.771720	02:40
4	0.621317	0.567915	0.819108	02:40

Normalizing¶

In [138]:

db1 = vision.DataBlock(
    blocks=(vision.ImageBlock, vision.CategoryBlock),
    get_items=vision.get_image_files,
    get_y=vision.parent_label, 
    item_tfms=vision.Resize(460),
    batch_tfms=[*vision.aug_transforms(size=224, min_scale=0.75),
                vision.Normalize.from_stats(*vision.imagenet_stats)],
    splitter=vision.FuncSplitter(splitter)
)

In [ ]:

db1.summary(path)

In [140]:

dls1 = db1.dataloaders(path, bs=32)
dls1.show_batch()

In [141]:

learn1 = vision.Learner(dls1, vision.xresnet50(), loss_func=vision.CrossEntropyLossFlat(), metrics=vision.accuracy)

In [142]:

learn1.fit_one_cycle(5, 3e-3)

epoch	train_loss	valid_loss	accuracy	time
0	0.829856	0.837526	0.725605	02:40
1	0.824502	1.359232	0.628025	02:41
2	0.652435	0.677327	0.786497	02:40
3	0.519262	0.505769	0.838981	02:40
4	0.443815	0.428318	0.863185	02:40

Presizing¶

Progressive resizing

In [143]:

def get_dls(bs, size):
  db = vision.DataBlock(
      blocks=(vision.ImageBlock, vision.CategoryBlock),
      get_items=vision.get_image_files,
      get_y=vision.parent_label,
      splitter=vision.FuncSplitter(splitter),
      item_tfms=vision.Resize(460),
      batch_tfms=[*vision.aug_transforms(size=size, min_scale=0.75),
                  vision.Normalize.from_stats(*vision.imagenet_stats)]
  )

  return db.dataloaders(path, bs=bs)

In [145]:

dls = get_dls(32, 128)
learn = vision.Learner(dls, vision.xresnet50(), loss_func=vision.CrossEntropyLossFlat(), metrics=vision.accuracy)

In [146]:

learn.fit_one_cycle(4, 3e-3)

epoch	train_loss	valid_loss	accuracy	time
0	1.722279	2.013101	0.432866	02:09
1	1.256942	1.393829	0.578599	02:07
2	0.900626	0.809706	0.746242	02:07
3	0.733233	0.689337	0.780382	02:07

In [148]:

learn.dls = get_dls(16, 224)
learn.fine_tune(3, 1e-3)

epoch	train_loss	valid_loss	accuracy	time
0	0.739479	0.866746	0.746242	02:48

epoch	train_loss	valid_loss	accuracy	time
0	0.678550	0.720628	0.776815	02:49
1	0.581601	0.513642	0.833121	02:50
2	0.449859	0.436976	0.862675	02:50

In [ ]:

pred, target = learn.tta()

In [152]:

vision.accuracy(pred, target).item()

Out[152]:

0.8657324910163879

Mix Up¶

In [25]:

img1 = Image.open(path.ls()[0].ls()[0].ls()[0]).resize((256, 256))
img2 = Image.open(path.ls()[0].ls()[1].ls()[1]).resize((256, 256))

golf = np.array(img1)
truck = np.array(img2)
adv_im = torch.from_numpy(0.4*golf)/255.0 + torch.from_numpy(0.6*truck)/255.0

In [34]:

_, axs = plt.subplots(1, 3, figsize=(12, 4))

vision.show_image(golf, ax=axs[0])
vision.show_image(truck, ax=axs[1])
vision.show_image(adv_im, ax=axs[2])

Out[34]:

<matplotlib.axes._subplots.AxesSubplot at 0x7f57f3b2b3c8>

In [4]:

db = vision.DataBlock(
    blocks=(vision.ImageBlock, vision.CategoryBlock),
    get_items=vision.get_image_files,
    get_y=vision.parent_label, 
    item_tfms=vision.Resize(460),
    batch_tfms=vision.aug_transforms(size=224, min_scale=0.75),
    splitter=vision.FuncSplitter(splitter)
)

In [10]:

dls = db.dataloaders(path, bs=32)

In [19]:

learn = vision.Learner(dls, vision.xresnet50(), loss_func=vision.CrossEntropyLossFlat(), metrics=vision.accuracy)

In [20]:

learn.fit_one_cycle(1, 3e-3, cbs=vision.MixUp())

epoch	train_loss	valid_loss	accuracy	time
0	1.718072	1.311549	0.601783	02:35

Label smoothing¶

In [21]:

learn = vision.Learner(dls, vision.xresnet50(), loss_func=vision.LabelSmoothingCrossEntropy(), metrics=vision.accuracy)
learn.fit_one_cycle(1, 3e-3)

epoch	train_loss	valid_loss	accuracy	time
0	2.231875	2.110342	0.623185	02:34