Notes on fastai Book Ch. 6

ai
fastai
notes
pytorch
Chapter 6 covers multi-label classification and image regression.
Author

Christian Mills

Published

March 14, 2022

This post is part of the following series:

Multi-Label Classification

  • the problem of identifying the categories of objects in images that may not contain exactly one type of object
    • there may be more than one kind of object or none at all that belong to the target classes
  • single-label classifiers cannot properly handle input that either does not contain an object of a target class or contains multiple objects of a different target classes
    • a single-label classifier trained to recognize cats and dogs could not handle an image that contains both cats and dogs
  • models deployed in production are more likely to encounter input with zero matches or more than one match

The Data

from fastai.vision.all import *

The PASCAL Visual Object Classes Challenge 2007 Dataset


path = untar_data(URLs.PASCAL_2007)
path
Path('/home/innom-dt/.fastai/data/pascal_2007')

path.ls()
(#8) [Path('/home/innom-dt/.fastai/data/pascal_2007/segmentation'),Path('/home/innom-dt/.fastai/data/pascal_2007/test'),Path('/home/innom-dt/.fastai/data/pascal_2007/train.csv'),Path('/home/innom-dt/.fastai/data/pascal_2007/valid.json'),Path('/home/innom-dt/.fastai/data/pascal_2007/train'),Path('/home/innom-dt/.fastai/data/pascal_2007/train.json'),Path('/home/innom-dt/.fastai/data/pascal_2007/test.csv'),Path('/home/innom-dt/.fastai/data/pascal_2007/test.json')]

df = pd.read_csv(path/'train.csv')
df.head()
fname labels is_valid
0 000005.jpg chair True
1 000007.jpg car True
2 000009.jpg horse person True
3 000012.jpg car False
4 000016.jpg bicycle True

Class lables are stored in a space-delimited string

Pandas and DataFrames


# Access rows and columns using the `iloc` property 
df.iloc[:,0]
0       000005.jpg
1       000007.jpg
2       000009.jpg
3       000012.jpg
4       000016.jpg
           ...    
5006    009954.jpg
5007    009955.jpg
5008    009958.jpg
5009    009959.jpg
5010    009961.jpg
Name: fname, Length: 5011, dtype: object

df.iloc[0,:]
# Trailing :s are always optional (in numpy, pytorch, pandas, etc.),
#   so this is equivalent:
df.iloc[0]
fname       000005.jpg
labels           chair
is_valid          True
Name: 0, dtype: object

# Get a column by name
df['fname']
0       000005.jpg
1       000007.jpg
2       000009.jpg
3       000012.jpg
4       000016.jpg
           ...    
5006    009954.jpg
5007    009955.jpg
5008    009958.jpg
5009    009959.jpg
5010    009961.jpg
Name: fname, Length: 5011, dtype: object

# Initialize a new data frame using a dictionary
tmp_df = pd.DataFrame({'a':[1,2], 'b':[3,4]})
tmp_df
a b
0 1 3
1 2 4

# Perform calculations using columns
tmp_df['c'] = tmp_df['a']+tmp_df['b']
tmp_df
a b c
0 1 3 4
1 2 4 6

Constructing a DataBlock

  • Dataset: a collection that returns a tuple of your independent and dependent variable for a single item
  • DataLoader: An iterator that provides a stream of mini-batches, where each mini-batch is a tuple of a batch of independent variables and a batch of dependent variables
  • Datasets: an iterator that contains a training Dataset and a validation Dataset
  • DataLoaders: an object that contains a training DataLoader and a validation DataLoader
  • By default, a DataBlock assumes we have an input and a target

Python Lambda Functions

One-hot encoding: using a vector of 0s, with a 1 in each location that is represented in the data


# Start with a data block created with no parameters
dblock = DataBlock()

# Add a Datasets object using the DataFrame
dsets = dblock.datasets(df)

len(dsets.train),len(dsets.valid)
(4009, 1002)

# Grabs the same thing twice
# Need to specify an input and a target
x,y = dsets.train[0]
x,y
(fname       008663.jpg
 labels      car person
 is_valid         False
 Name: 4346, dtype: object,
 fname       008663.jpg
 labels      car person
 is_valid         False
 Name: 4346, dtype: object)

x['fname']
'008663.jpg'

# Tell the DataBlock how to extract the input and target from the DataFrame
# Using lamda functions
dblock = DataBlock(get_x = lambda r: r['fname'], get_y = lambda r: r['labels'])
dsets = dblock.datasets(df)
dsets.train[0]
('005620.jpg', 'aeroplane')

Note: Do not use lambda functions if you need to export the Learner


# Tell the DataBlock how to extract the input and target from the DataFrame
# Using standard functions
def get_x(r): return r['fname']
def get_y(r): return r['labels']
dblock = DataBlock(get_x = get_x, get_y = get_y)
dsets = dblock.datasets(df)
dsets.train[0]
('002549.jpg', 'tvmonitor')

Note: Need the full file path for the dependent variable and need to split the dependent variables on the space character


def get_x(r): return path/'train'/r['fname']
def get_y(r): return r['labels'].split(' ')
dblock = DataBlock(get_x = get_x, get_y = get_y)
dsets = dblock.datasets(df)
dsets.train[0]
(Path('/home/innom-dt/.fastai/data/pascal_2007/train/002844.jpg'), ['train'])

ImageBlock

MultiCategoryBlock


ImageBlock
<function fastai.vision.data.ImageBlock(cls=<class 'fastai.vision.core.PILImage'>)>

MultiCategoryBlock
<function fastai.data.block.MultiCategoryBlock(encoded=False, vocab=None, add_na=False)>

dblock = DataBlock(blocks=(ImageBlock, MultiCategoryBlock),
                   get_x = get_x, get_y = get_y)
dsets = dblock.datasets(df)
dsets.train[0]
(PILImage mode=RGB size=500x375,
 TensorMultiCategory([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0.]))

# Check which object class is represented by the above one-hot encoding
idxs = torch.where(dsets.train[0][1]==1.)[0]
dsets.train.vocab[idxs]
(#1) ['dog']

# Define a function to split the dataset based on the is_valid column
def splitter(df):
    train = df.index[~df['is_valid']].tolist()
    valid = df.index[df['is_valid']].tolist()
    return train,valid

dblock = DataBlock(blocks=(ImageBlock, MultiCategoryBlock),
                   splitter=splitter,
                   get_x=get_x, 
                   get_y=get_y)

dsets = dblock.datasets(df)
dsets.train[0]
(PILImage mode=RGB size=500x333,
 TensorMultiCategory([0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]))

dblock = DataBlock(blocks=(ImageBlock, MultiCategoryBlock),
                   splitter=splitter,
                   get_x=get_x, 
                   get_y=get_y,
                   item_tfms = RandomResizedCrop(128, min_scale=0.35))
dls = dblock.dataloaders(df)

dls.show_batch(nrows=1, ncols=3)


dblock.summary(df)
Setting-up type transforms pipelines
Collecting items from            fname          labels  is_valid
0     000005.jpg           chair      True
1     000007.jpg             car      True
2     000009.jpg    horse person      True
3     000012.jpg             car     False
4     000016.jpg         bicycle      True
...          ...             ...       ...
5006  009954.jpg    horse person      True
5007  009955.jpg            boat      True
5008  009958.jpg  person bicycle      True
5009  009959.jpg             car     False
5010  009961.jpg             dog     False

[5011 rows x 3 columns]
Found 5011 items
2 datasets of sizes 2501,2510
Setting up Pipeline: get_x -> PILBase.create
Setting up Pipeline: get_y -> MultiCategorize -- {'vocab': None, 'sort': True, 'add_na': False} -> OneHotEncode -- {'c': None}

Building one sample
  Pipeline: get_x -> PILBase.create
    starting from
      fname       000012.jpg
labels             car
is_valid         False
Name: 3, dtype: object
    applying get_x gives
      /home/innom-dt/.fastai/data/pascal_2007/train/000012.jpg
    applying PILBase.create gives
      PILImage mode=RGB size=500x333
  Pipeline: get_y -> MultiCategorize -- {'vocab': None, 'sort': True, 'add_na': False} -> OneHotEncode -- {'c': None}
    starting from
      fname       000012.jpg
labels             car
is_valid         False
Name: 3, dtype: object
    applying get_y gives
      [car]
    applying MultiCategorize -- {'vocab': None, 'sort': True, 'add_na': False} gives
      TensorMultiCategory([6])
    applying OneHotEncode -- {'c': None} gives
      TensorMultiCategory([0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.])

Final sample: (PILImage mode=RGB size=500x333, TensorMultiCategory([0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]))

Collecting items from            fname          labels  is_valid
0     000005.jpg           chair      True
1     000007.jpg             car      True
2     000009.jpg    horse person      True
3     000012.jpg             car     False
4     000016.jpg         bicycle      True
...          ...             ...       ...
5006  009954.jpg    horse person      True
5007  009955.jpg            boat      True
5008  009958.jpg  person bicycle      True
5009  009959.jpg             car     False
5010  009961.jpg             dog     False

[5011 rows x 3 columns]
Found 5011 items
2 datasets of sizes 2501,2510
Setting up Pipeline: get_x -> PILBase.create
Setting up Pipeline: get_y -> MultiCategorize -- {'vocab': None, 'sort': True, 'add_na': False} -> OneHotEncode -- {'c': None}
Setting up after_item: Pipeline: RandomResizedCrop -- {'size': (128, 128), 'min_scale': 0.35, 'ratio': (0.75, 1.3333333333333333), 'resamples': (2, 0), 'val_xtra': 0.14, 'max_scale': 1.0, 'p': 1.0} -> ToTensor
Setting up before_batch: Pipeline: 
Setting up after_batch: Pipeline: IntToFloatTensor -- {'div': 255.0, 'div_mask': 1}

Building one batch
Applying item_tfms to the first sample:
  Pipeline: RandomResizedCrop -- {'size': (128, 128), 'min_scale': 0.35, 'ratio': (0.75, 1.3333333333333333), 'resamples': (2, 0), 'val_xtra': 0.14, 'max_scale': 1.0, 'p': 1.0} -> ToTensor
    starting from
      (PILImage mode=RGB size=500x333, TensorMultiCategory([0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]))
    applying RandomResizedCrop -- {'size': (128, 128), 'min_scale': 0.35, 'ratio': (0.75, 1.3333333333333333), 'resamples': (2, 0), 'val_xtra': 0.14, 'max_scale': 1.0, 'p': 1.0} gives
      (PILImage mode=RGB size=128x128, TensorMultiCategory([0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]))
    applying ToTensor gives
      (TensorImage of size 3x128x128, TensorMultiCategory([0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]))

Adding the next 3 samples

No before_batch transform to apply

Collating items in a batch

Applying batch_tfms to the batch built
  Pipeline: IntToFloatTensor -- {'div': 255.0, 'div_mask': 1}
    starting from
      (TensorImage of size 4x3x128x128, TensorMultiCategory of size 4x20)
    applying IntToFloatTensor -- {'div': 255.0, 'div_mask': 1} gives
      (TensorImage of size 4x3x128x128, TensorMultiCategory of size 4x20)

Binary Cross-Entropy

  • Getting Model Activations
    • it is important to know how to manually get a mini-batch, pass it into a model, and look at the activations
  • Can’t directly use nll_loss or softmax for a one-hot-encoded dependent variable
    • softmax requires all predictions sum to 1 and tends to push one activation to be much larger than all the other
      • not desirable when there may be multiple objects or none at all in a single image
    • nll_loss returns the value of just one activation
  • binary cross-entropy combines mnist_loss with log

learn = cnn_learner(dls, resnet18)

to_cpu(b)


to_cpu
<function fastai.torch_core.to_cpu(b)>

x,y = to_cpu(dls.train.one_batch())
activs = learn.model(x)
activs.shape
torch.Size([64, 20])

activs[0]
TensorBase([ 0.5674, -1.2013,  4.5409, -1.5284, -0.6600,  0.0999, -2.4757, -0.8773, -0.2934, -1.4746, -0.1738,  2.1763, -3.4473, -1.1407,  0.1783, -1.6922, -2.3396,  0.7602, -1.4213, -0.4334],
       grad_fn=<AliasBackward0>)

Note: The raw model activations are not scaled between [0,1]


def binary_cross_entropy(inputs, targets):
    inputs = inputs.sigmoid()
    return -torch.where(targets==1, inputs, 1-inputs).log().mean()

binary_cross_entropy(activs, y)
TensorMultiCategory(1.0367, grad_fn=<AliasBackward0>)

nn.BCELoss

nn.BCEWithLogitsLoss


nn.BCEWithLogitsLoss
torch.nn.modules.loss.BCEWithLogitsLoss

loss_func = nn.BCEWithLogitsLoss()
loss = loss_func(activs, y)
loss
TensorMultiCategory(1.0367, grad_fn=<AliasBackward0>)

Python Partial Functions


partial
functools.partial

def say_hello(name, say_what="Hello"): return f"{say_what} {name}."
say_hello('Jeremy'),say_hello('Jeremy', 'Ahoy!')
('Hello Jeremy.', 'Ahoy! Jeremy.')

f = partial(say_hello, say_what="Bonjour")
f("Jeremy"),f("Sylvain")
('Bonjour Jeremy.', 'Bonjour Sylvain.')

accuracy_multi


accuracy_multi
<function fastai.metrics.accuracy_multi(inp, targ, thresh=0.5, sigmoid=True)>

learn = cnn_learner(dls, resnet50, metrics=partial(accuracy_multi, thresh=0.2))
learn.fine_tune(3, base_lr=3e-3, freeze_epochs=4)
epoch train_loss valid_loss accuracy_multi time
0 0.942860 0.704590 0.234223 00:06
1 0.821557 0.550972 0.293825 00:06
2 0.604402 0.202164 0.813645 00:06
3 0.359336 0.122809 0.943466 00:06
epoch train_loss valid_loss accuracy_multi time
0 0.135016 0.122502 0.944601 00:07
1 0.118378 0.107208 0.950478 00:07
2 0.098511 0.103568 0.951613 00:07

learn.metrics = partial(accuracy_multi, thresh=0.1)
learn.validate()
(#2) [0.10356765240430832,0.9294222593307495]

learn.metrics = partial(accuracy_multi, thresh=0.99)
learn.validate()
(#2) [0.10356765240430832,0.9427291750907898]

preds,targs = learn.get_preds()

accuracy_multi(preds, targs, thresh=0.9, sigmoid=False)
TensorBase(0.9566)

# Try a few different threshold values to see which works best
xs = torch.linspace(0.05,0.95,29)
accs = [accuracy_multi(preds, targs, thresh=i, sigmoid=False) for i in xs]
plt.plot(xs,accs);

Regression

  • a model is defined by it independent and dependent variables, along with its loss function
  • image regression: the independent variable is an image and the dependent variable is one or more floating point numbers
  • key point model:
    • a key point refers to a specific location represented in an image

Assemble the Data

BIWI Kinect Head Pose Database

  • https://icu.ee.ethz.ch/research/datsets.html
  • over 15k images of 20 people recorded with a Kinect while turning their heads around freely
  • Depth and rgb images are provided for each frame
  • ground in the form of the 3D location of the head and its rotation angles
  • contains 24 directories numbered from 01 to 24 which correspond to the different people photographed
    • each directory has a corresponding .obj file
    • each directory contains .cal files containing the calibration data for the depth and color cameras
    • each image has a corresponding _pose.txt file containing the location of center of the head in 3D and the head rotation encoded as 3D rotation matrix

path = untar_data(URLs.BIWI_HEAD_POSE)
path
Path('/home/innom-dt/.fastai/data/biwi_head_pose')

path.ls().sorted()
(#50) [Path('/home/innom-dt/.fastai/data/biwi_head_pose/01'),Path('/home/innom-dt/.fastai/data/biwi_head_pose/01.obj'),Path('/home/innom-dt/.fastai/data/biwi_head_pose/02'),Path('/home/innom-dt/.fastai/data/biwi_head_pose/02.obj'),Path('/home/innom-dt/.fastai/data/biwi_head_pose/03'),Path('/home/innom-dt/.fastai/data/biwi_head_pose/03.obj'),Path('/home/innom-dt/.fastai/data/biwi_head_pose/04'),Path('/home/innom-dt/.fastai/data/biwi_head_pose/04.obj'),Path('/home/innom-dt/.fastai/data/biwi_head_pose/05'),Path('/home/innom-dt/.fastai/data/biwi_head_pose/05.obj')...]

(path/'01').ls().sorted()
(#1000) [Path('/home/innom-dt/.fastai/data/biwi_head_pose/01/depth.cal'),Path('/home/innom-dt/.fastai/data/biwi_head_pose/01/frame_00003_pose.txt'),Path('/home/innom-dt/.fastai/data/biwi_head_pose/01/frame_00003_rgb.jpg'),Path('/home/innom-dt/.fastai/data/biwi_head_pose/01/frame_00004_pose.txt'),Path('/home/innom-dt/.fastai/data/biwi_head_pose/01/frame_00004_rgb.jpg'),Path('/home/innom-dt/.fastai/data/biwi_head_pose/01/frame_00005_pose.txt'),Path('/home/innom-dt/.fastai/data/biwi_head_pose/01/frame_00005_rgb.jpg'),Path('/home/innom-dt/.fastai/data/biwi_head_pose/01/frame_00006_pose.txt'),Path('/home/innom-dt/.fastai/data/biwi_head_pose/01/frame_00006_rgb.jpg'),Path('/home/innom-dt/.fastai/data/biwi_head_pose/01/frame_00007_pose.txt')...]

# recursivley get all images in the 24 subdirectories
img_files = get_image_files(path)
# get the file names for the corresponding pose.txt files
def img2pose(x): return Path(f'{str(x)[:-7]}pose.txt')
pose_file = img2pose(img_files[0])
pose_file
Path('/home/innom-dt/.fastai/data/biwi_head_pose/22/frame_00304_pose.txt')

!cat $pose_file
0.999485 -0.00797222 -0.031067 
-0.00416483 0.928156 -0.372168 
0.031802 0.372106 0.927645 

62.3638 96.2159 979.839 

im = PILImage.create(img_files[0])
im.shape
(480, 640)

im.to_thumb(160)

np.genfromtxt


np.genfromtxt
<function numpy.genfromtxt(fname, dtype=<class 'float'>, comments='#', delimiter=None, skip_header=0, skip_footer=0, converters=None, missing_values=None, filling_values=None, usecols=None, names=None, excludelist=None, deletechars=" !#$%&'()*+,-./:;<=>?@[\\]^{|}~", replace_space='_', autostrip=False, case_sensitive=True, defaultfmt='f%i', unpack=None, usemask=False, loose=True, invalid_raise=True, max_rows=None, encoding='bytes', *, like=None)>

# Contains the calibration values for this folder's rgb camera
# Skip the last six lines in the file 
cal = np.genfromtxt(path/'01'/'rgb.cal', skip_footer=6)
cal
array([[517.679,   0.   , 320.   ],
       [  0.   , 517.679, 240.5  ],
       [  0.   ,   0.   ,   1.   ]])

# Extract the 2D coordinates for the center of a head
# Serves as the get_y function for a DataBlock
def get_ctr(f):
    # Skip the last 3 lines in the file
    ctr = np.genfromtxt(img2pose(f), skip_header=3)
    c1 = ctr[0] * cal[0][0]/ctr[2] + cal[0][2]
    c2 = ctr[1] * cal[1][1]/ctr[2] + cal[1][2]
    return tensor([c1,c2])

np.genfromtxt(img2pose(img_files[0]), skip_header=3)
array([ 62.3638,  96.2159, 979.839 ])

get_ctr(img_files[0])
tensor([352.9487, 291.3338])

PointBlock


PointBlock
<fastai.data.block.TransformBlock at 0x7fb65aaf5490>

# Construct custom DataBlock
biwi = DataBlock(
    blocks=(ImageBlock, PointBlock),
    get_items=get_image_files,
    get_y=get_ctr,
    # Have the validation set contain images for a single person
    splitter=FuncSplitter(lambda o: o.parent.name=='13'),
    batch_tfms=[*aug_transforms(size=(240,320)), 
                Normalize.from_stats(*imagenet_stats)]
)

dls = biwi.dataloaders(path)
dls.show_batch(max_n=9, figsize=(8,6))


xb,yb = dls.one_batch()
xb.shape,yb.shape
(torch.Size([64, 3, 240, 320]), torch.Size([64, 1, 2]))

yb[0]
TensorPoint([[-0.1246,  0.0960]], device='cuda:0')

Training a Model

# Set range of coordinate values for the model output to [-1,1]
learn = cnn_learner(dls, resnet18, y_range=(-1,1))

def sigmoid_range(x, lo, hi): return torch.sigmoid(x) * (hi-lo) + lo

plot_function(partial(sigmoid_range,lo=-1,hi=1), min=-4, max=4)


dls.loss_func
FlattenedLoss of MSELoss()

min_lr, steep_lr, valley = learn.lr_find(suggest_funcs=(minimum, steep, valley))


min_lr
0.006918309628963471

steep_lr
2.0892961401841603e-05

valley
0.0010000000474974513

lr = 1e-2
learn.fine_tune(3, lr)
epoch train_loss valid_loss time
0 0.048689 0.026659 00:36
epoch train_loss valid_loss time
0 0.007270 0.002140 00:47
1 0.002966 0.000160 00:47
2 0.001556 0.000042 00:47

# Calculate the Root Mean Squared Error
math.sqrt(0.000042)
0.00648074069840786

learn.show_results(ds_idx=1, nrows=3, figsize=(6,8))

References

Previous: Notes on fastai Book Ch. 5

Next: Notes on fastai Book Ch. 7