from fastai.vision.all import *
Below you will find the exact imports for everything we use today
from fastcore.xtras import Path
from fastai.callback.hook import summary
from fastai.callback.progress import ProgressCallback
from fastai.callback.schedule import lr_find, fit_flat_cos
from fastai.data.block import DataBlock
from fastai.data.external import untar_data, URLs
from fastai.data.transforms import get_image_files, FuncSplitter, Normalize
from fastai.layers import Mish
from fastai.losses import BaseLoss
from fastai.optimizer import ranger
from fastai.torch_core import tensor
from fastai.vision.augment import aug_transforms
from fastai.vision.core import PILImage, PILMask
from fastai.vision.data import ImageBlock, MaskBlock, imagenet_stats
from fastai.vision.learner import unet_learner
from PIL import Image
import numpy as np
from torch import nn
from torchvision.models.resnet import resnet34
import torch
import torch.nn.functional as F
path = untar_data(URLs.CAMVID)
Our validation set is inside a text document called valid.txt and split by new lines. Let's read it in:
valid_fnames = (path/'valid.txt').read_text().split('\n')
valid_fnames[:5]
Let's look at an image and see how everything aligns up
path_im = path/'images'
path_lbl = path/'labels'
First we need our filenames
fnames = get_image_files(path_im)
lbl_names = get_image_files(path_lbl)
And now let's work with one of them
img_fn = fnames[10]
img = PILImage.create(img_fn)
img.show(figsize=(5,5))
Now let's grab our y's. They live in the labels folder and are denoted by a _P
get_msk = lambda o: path/'labels'/f'{o.stem}_P{o.suffix}'
The stem and suffix grab everything before and after the period respectively.
Our masks are of type PILMask and we will make our gradient percentage (alpha) equal to 1 as we are not overlaying this on anything yet
msk = PILMask.create(get_msk(img_fn))
msk.show(figsize=(5,5), alpha=1)
Now if we look at what our mask actually is, we can see it's a giant array of pixels:
tensor(msk)
Where each one represents a class that we can find in codes.txt. Let's make a vocabulary with it
codes = np.loadtxt(path/'codes.txt', dtype=str); codes
We need a split function that will split from our list of valid filenames we grabbed earlier. Let's try making our own.
def FileSplitter(fname):
"Split `items` depending on the value of `mask`."
valid = Path(fname).read_text().split('\n')
def _func(x): return x.name in valid
def _inner(o, **kwargs): return FuncSplitter(_func)(o)
return _inner
This takes in our filenames, and checks for all of our filenames in all of our items in our validation filenames
This first round we will train at half the image size
sz = msk.shape; sz
half = tuple(int(x/2) for x in sz); half
camvid = DataBlock(blocks=(ImageBlock, MaskBlock(codes)),
get_items=get_image_files,
splitter=FileSplitter(path/'valid.txt'),
get_y=get_msk,
batch_tfms=[*aug_transforms(size=half), Normalize.from_stats(*imagenet_stats)])
dls = camvid.dataloaders(path/'images', bs=8)
Let's look at a batch, and look at all the classes between codes 1 and 30 (ignoring Animal and Wall)
dls.show_batch(max_n=4, vmin=1, vmax=30, figsize=(14,10))
Lastly let's make our vocabulary a part of our DataLoaders, as our loss function needs to deal with the Void label
dls.vocab = codes
Now we need a methodology for grabbing that particular code from our output of numbers. Let's make everything into a dictionary
name2id = {v:k for k,v in enumerate(codes)}
name2id
Awesome! Let's make an accuracy function
void_code = name2id['Void']
For segmentation, we want to squeeze all the outputted values to have it as a matrix of digits for our segmentation mask. From there, we want to match their argmax to the target's mask for each pixel and take the average
def acc_camvid(inp, targ):
targ = targ.squeeze(1)
mask = targ != void_code
return (inp.argmax(dim=1)[mask]==targ[mask]).float().mean()
U-Net allows us to look at pixel-wise representations of our images through sizing it down and then blowing it bck up into a high resolution image. The first part we call an "encoder" and the second a "decoder"
On the image, the authors of the UNET paper describe the arrows as "denotions of different operations"
We have a special unet_learner. Something new is we can pass in some model configurations where we can declare a few things to customize it with!
- Blur/blur final: avoid checkerboard artifacts
- Self attention: A self-attention layer
- y_range: Last activations go through a sigmoid for rescaling
- Last cross - Cross-connection with the direct model input
- Bottle - Bottlenck or not on that cross
- Activation function
- Norm type
Let's make a unet_learner that uses some of the new state of the art techniques. Specifically:
- Self-attention layers:
self_attention = True - Mish activation function:
act_cls = Mish
Along with this we will use the Ranger as optimizer function.
opt = ranger
learn = unet_learner(dls, resnet34, metrics=acc_camvid, self_attention=True, act_cls=Mish, opt_func=opt)
learn.summary()
If we do a learn.summary we can see this blow-up trend, and see that our model came in frozen. Let's find a learning rate
learn.lr_find()
lr = 1e-3
With our new optimizer, we will also want to use a different fit function, called fit_flat_cos
learn.fit_flat_cos(10, slice(lr))
learn.save('stage-1')
learn.load('stage-1');
learn.show_results(max_n=4, figsize=(12,6))