Chapter 5, MNIST + Cross Entropy Loss

Part 1 (2020)
1

Hi! I’ve totally rebuilt my MNIST model with Cross Entropy Loss, however I’ve experience something really strange… I’m getting an 89 percent accuracy after the first epoch.
(Also I can’t use predict, I’m getting an

AttributeError: 'list' object has no attribute 'decode_batch'
error, I have no idea what that is…
Here’s my code

from fastai.vision.all import *
from fastbook import *

path = untar_data(URLs.MNIST)

path.ls()

(#2) [Path('/storage/data/mnist_png/training'),Path('/storage/data/mnist_png/testing')]

path_train = path/'training'
path_valid = path/'testing'

train_x = get_image_files(path_train).sorted()

train_x = [(tensor(Image.open(element)).float()) / 255 for element in train_x ]

train_x = torch.stack(train_x).view(-1,28*28)

train_y = [int(element.parent.name) for element in get_image_files(path_train).sorted()]

train_y = tensor(train_y)

dl = DataLoader(list(zip(train_x,train_y)),batch_size=256,shuffle=True)

valid_x = get_image_files(path_valid).sorted()

valid_x = [(tensor(Image.open(element)).float()) / 255 for element in valid_x ]

valid_x = torch.stack(valid_x).view(-1,28*28)

valid_y = [int(element.parent.name) for element in get_image_files(path_valid).sorted()]

valid_y = tensor(valid_y)

dl_valid = DataLoader(list(zip(valid_x,valid_y)),batch_size=256,shuffle=True)

dls = DataLoaders(dl,dl_valid)

simple_net = nn.Sequential(
nn.Linear(28*28,50),
nn.ReLU(),
nn.Linear(50,10)
)

learner = Learner(dls,simple_net,opt_func=SGD,loss_func=nn.CrossEntropyLoss(),metrics=accuracy,lr=0.001)

learner.fit(10,0.1)
epoch train_loss valid_loss accuracy time
0 0.461113 0.376712 0.896700 00:01
1 0.337488 0.320396 0.906000 00:01
2 0.304718 0.276135 0.921400 00:01
3 0.275385 0.254352 0.927000 00:01
4 0.240319 0.235078 0.931800 00:01
5 0.225280 0.219778 0.936700 00:01
6 0.218840 0.209284 0.939000 00:01
7 0.201014 0.194134 0.942900 00:01
8 0.190363 0.185128 0.944700 00:01
9 0.184855 0.178727 0.948300 00:01
2 Likes