Bad performance with CNN for basic image classification task

Question

how are you doing?

I'm playing around with CNN in FastAI.

My model with 2 millions parameters only has around 80% accuracy. I also tried with Data normalization, Batch normalization, Label smoothing, Mixup but the results are still capped at 80-81% (they did converged faster with those techniques).

I assumed bigger model would make better predictions so I increase the parameters (from 2 millions to 182 millions). However, the result is still at 82% after 40 epochs

imagenette = DataBlock(blocks = (ImageBlock, CategoryBlock),
                   get_items = get_image_files,
                   get_y = parent_label,
                   splitter = GrandparentSplitter(train_name='train', valid_name='val'),
                   item_tfms = Resize(460),
                   batch_tfms = aug_transforms(size=244)
                  )
dls = imagenette.dataloaders(path)



def conv_manual(ni, nf, ks=3, act=True):
conv = nn.Conv2d(ni, nf, stride=2, kernel_size=ks, padding = ks//2)
if act: conv = nn.Sequential(
    conv, 
    nn.ReLU(),
    #nn.AvgPool2d(2,stride=2, padding=1),
    BatchNorm(nf)
)
return conv

test_cnn = sequential(
    conv_manual(3,8, ks=5),            #122x122
    conv_manual(8,64),                 #61x61
    conv_manual(64,100),                #31x31
    conv_manual(100,250),                #16x16
    conv_manual(250,800),               #8x8
    conv_manual(800,3000),              #4x4
    conv_manual(3000, 6000),             #2x2
    conv_manual(6000,10, act=False),   #1x1  
    Flatten(),
)

test_model.fit_one_cycle(10, 0.00015)

I hope you could give me some thoughts about it

Thank you

Answer 1

1. You can try more architectures such as ResNets, U-Nets, Inception Networks.
2. You also need to perform some more hyperparameter tuning check out some tools for tuning your models