Tutorial from fastai doc

!git clone --recurse-submodules https://github.com/fastai/fastai
!pip install -e "fastai[dev]"

kernel should be re-started after installing fastai-v2

check fastai-v2 version

import fastai
print(fastai.__version__)

2.0.2

vision - beginner (dog or cat)

from fastai.vision.all import *

grasp dataset & inspect a bit

path = untar_data(URLs.PETS)
print(path.ls())

(#2) [Path('/home/user/.fastai/data/oxford-iiit-pet/images'),Path('/home/user/.fastai/data/oxford-iiit-pet/annotations')]

# get_image_files
# - def get_image_files(path, recurse=True, folders=None):
#   - Get image files in `path` recursively, only in `folders`, if specified.
files = get_image_files(path/"images")
print(len(files))

7390

# just see filename examples for Cat and dog.
# - Cat - first capital letter 
# - dog - all lowercase
files[0],files[100]

(Path('/home/user/.fastai/data/oxford-iiit-pet/images/german_shorthaired_49.jpg'),
 Path('/home/user/.fastai/data/oxford-iiit-pet/images/Egyptian_Mau_156.jpg'))

define a label function

all about how to label an item

# takes filename
# return how to label it
# - if filename is in uppercase => it's a Cat => it is labeled as True
# - if filename is in lowercase => it's a dog => it is labeled as False
def label_func(filename): 
    return filename[0].isupper()

ImageDataLoaders with from_name_func

from forum about 'Why does ImageDataBunch.from_name_re() require a path argument?'

After reading the source code, my current understanding is that path is a required property of the DataBunch parent class, since methods like DataBunch.save() will save directly to path. Additionally, the Learner class usually copies Learner.path from its data.path. This is used for things like Learner.save(), Learner.load(), and Learner.export(), which write to self.path/self.model_dir or just self.path.

# ImageDatLoader"s"
# - why plurals? => multiple DataLoaders (from PyTorch) are in it
#   - Training/Validation/Testing DataLoader

# from_name_func
# - path? => root path of dataset
# - fnames? => list of filenames (image filenames)
# - label_func? => how to label each data?
# - item_tfms? => item transforms
dls = ImageDataLoaders.from_name_func(path=path, 
                                      fnames=files, 
                                      label_func=label_func,
                                      item_tfms=Resize(224))

dls.show_batch(max_n=12)

cnn_learner with fine_tune

learn = cnn_learner(dls, resnet34, metrics=error_rate)

Downloading: "https://download.pytorch.org/models/resnet34-333f7ec4.pth" to /home/user/.cache/torch/hub/checkpoints/resnet34-333f7ec4.pth

# fine_tune is not a simple method
# - but all the details are hidden, but also you can adjust the details
""" 
  - Learner.fine_tune(epochs, 
                      base_lr=0.002, 
                      freeze_epochs=1, 
                      lr_mult=100, 
                      pct_start=0.3, 
                      div=5.0, 
                      lr_max=None, 
                      div_final=100000.0, 
                      wd=None, 
                      moms=None, 
                      cbs=None, 
                      reset_opt=False)
"""

learn.fine_tune(epochs=1)

epoch	train_loss	valid_loss	error_rate	time
0	0.138832	0.016947	0.006766	00:16

epoch	train_loss	valid_loss	error_rate	time
0	0.072697	0.023958	0.008796	00:15

predict

# results
# - first item? => "predicted label"
# - second item? => index or predicted result tensor
# - third item? => actual output tensor

learn.predict(files[0])

('False', tensor(0), tensor([1.0000e+00, 8.6281e-07]))

# just show some randomly chosen data from validation set
# - but we can actually choose the target dataset
#   - with ds_idx argument
#     - ds_idx=1 is default (validation), 0 is for training set
learn.show_results(max_n=12)

# on training dataset
learn.show_results(ds_idx=0, max_n=12)

fastai Interpreter

interp = Interpretation.from_learner(learn)

interp.top_losses(k=20)

torch.return_types.topk(
values=tensor([5.8659, 4.7590, 4.5521, 1.8459, 1.6657, 1.6071, 1.6005, 1.4597, 1.4470,
        1.4424, 1.1674, 1.0468, 0.9538, 0.5710, 0.4560, 0.4243, 0.4099, 0.3973,
        0.3943, 0.3182]),
indices=tensor([ 873, 1236,  961,  947, 1287,  912,  298, 1154,  662,  307,  361, 1156,
        1041,  227,  878, 1308,   94, 1433,  859,  187]))

interp.plot_top_losses(k=20)

classification_interp = ClassificationInterpretation.from_learner(learn)

classification_interp.print_classification_report()

              precision    recall  f1-score   support

       False       0.99      1.00      0.99      1020
        True       0.99      0.98      0.99       458

    accuracy                           0.99      1478
   macro avg       0.99      0.99      0.99      1478
weighted avg       0.99      0.99      0.99      1478

classification_interp.most_confused()

[(True, False, 8), (False, True, 5)]

classification_interp.plot_confusion_matrix()

Using W&B (Weights & Bias)

# install wandb
!pip install wandb

# login
import wandb
wandb.login()

# initialize wandb project
# - it will create the project with endpoint
wandb.init(project='my_project')

# import wandb callback in fastai
from fastai.callback.wandb import *

# leave logs for everything
learn = cnn_learner(dls, resnet34, metrics=error_rate, cbs=WandbCallback())
learn.fine_tune(epochs=4)
# leave logs for training process only
# - learn.fine_tune(epochs=4, cbs=WandbCallback())

vision - beginner (dog breed)

from fastai.vision.all import *

path = untar_data(URLs.PETS)
files = get_image_files(path/"images")

pat = r'^(.*)_\d+.jpg'

dls = ImageDataLoaders.from_name_re(path=path,
                                    fnames=files,
                                    pat=pat,
                                    item_tfms=Resize(224))
dls.show_batch()

dls = ImageDataLoaders.from_name_re(path=path, 
                                    fnames=files, 
                                    pat=pat, 
                                    item_tfms=Resize(460),
                                    batch_tfms=aug_transforms(size=224))

dls.show_batch()

learn = cnn_learner(dls, resnet34, metrics=error_rate)

learn.lr_find()

SuggestedLRs(lr_min=0.010000000149011612, lr_steep=0.005248074419796467)

learn.fine_tune(4, 5e-3)

epoch	train_loss	valid_loss	error_rate	time
0	1.116782	0.310870	0.099459	00:21

epoch	train_loss	valid_loss	error_rate	time
0	0.520518	0.547198	0.158999	00:21
1	0.541868	0.341918	0.100812	00:21
2	0.318754	0.291615	0.081867	00:21
3	0.162863	0.236776	0.070365	00:21

learn.show_results()

interp = Interpretation.from_learner(learn)
interp.plot_top_losses(9, figsize=(15,10))

classification_interp = ClassificationInterpretation.from_learner(learn)
classification_interp.print_classification_report()

                            precision    recall  f1-score   support

                Abyssinian       0.93      0.91      0.92        45
                    Bengal       0.85      0.85      0.85        39
                    Birman       0.90      0.97      0.94        38
                    Bombay       0.97      1.00      0.98        32
         British_Shorthair       0.91      0.98      0.94        41
              Egyptian_Mau       0.88      0.90      0.89        42
                Maine_Coon       0.92      0.92      0.92        52
                   Persian       1.00      0.89      0.94        36
                   Ragdoll       0.86      0.88      0.87        43
              Russian_Blue       0.97      0.90      0.94        40
                   Siamese       0.93      0.93      0.93        40
                    Sphynx       0.98      0.93      0.96        46
          american_bulldog       0.84      0.79      0.81        33
 american_pit_bull_terrier       0.73      0.75      0.74        40
              basset_hound       1.00      0.97      0.99        35
                    beagle       0.86      0.95      0.90        40
                     boxer       0.85      0.90      0.88        39
                 chihuahua       0.80      0.93      0.86        30
    english_cocker_spaniel       0.91      1.00      0.95        39
            english_setter       0.97      0.91      0.94        34
        german_shorthaired       0.94      1.00      0.97        34
            great_pyrenees       0.98      0.98      0.98        52
                  havanese       0.89      0.98      0.93        42
             japanese_chin       1.00      1.00      1.00        47
                  keeshond       1.00      0.97      0.99        36
                leonberger       0.97      1.00      0.99        36
        miniature_pinscher       0.89      0.89      0.89        45
              newfoundland       1.00      1.00      1.00        40
                pomeranian       1.00      0.97      0.99        40
                       pug       1.00      0.95      0.97        37
             saint_bernard       0.90      0.96      0.93        28
                   samoyed       1.00      0.97      0.99        40
          scottish_terrier       0.98      0.98      0.98        46
                 shiba_inu       1.00      0.94      0.97        33
staffordshire_bull_terrier       0.86      0.79      0.82        61
           wheaten_terrier       1.00      0.95      0.97        39
         yorkshire_terrier       1.00      0.87      0.93        38

                  accuracy                           0.93      1478
                 macro avg       0.93      0.93      0.93      1478
              weighted avg       0.93      0.93      0.93      1478

classification_interp.most_confused(min_val=3)

[('staffordshire_bull_terrier', 'american_pit_bull_terrier', 5),
 ('Bengal', 'Egyptian_Mau', 4),
 ('Egyptian_Mau', 'Bengal', 4),
 ('Maine_Coon', 'Ragdoll', 4),
 ('american_pit_bull_terrier', 'staffordshire_bull_terrier', 4),
 ('Russian_Blue', 'British_Shorthair', 3),
 ('Sphynx', 'chihuahua', 3),
 ('american_bulldog', 'american_pit_bull_terrier', 3),
 ('american_pit_bull_terrier', 'miniature_pinscher', 3),
 ('staffordshire_bull_terrier', 'american_bulldog', 3),
 ('yorkshire_terrier', 'havanese', 3)]

classification_interp.plot_confusion_matrix(figsize=(20,15))