Super-resolution

Using the Convolutional ResNet architecture to achieve image super-resolution, instead of classification

Adapted from

The challenge of this module is to: (a) upscale a 32x32 image to a 64x64 image while (b) un-erasing randomly cropped portions of the training data.

source

preprocess

 preprocess (examples, pipe, erase)

Factory to preprocess images for training and testing

source

get_imagenet_super_rez_dls

 get_imagenet_super_rez_dls (bs=512)

Get the Imagenet Super resolution data

dls = get_imagenet_super_rez_dls(1024)
xb, yb = dls.peek()
CPU times: user 55.2 s, sys: 8.12 s, total: 1min 3s
Wall time: 25.3 s
show_images(denorm(xb[:6, ...]))

First, let’s clean up the ResidualConvBlock. It seems that there was an issue where the original Conv helper did activation, then normalization. It seems it should be the other way around..

Here’s the shared model code.

source

Conv

 Conv (c_in, c_out, stride=2, ks=3, **kwargs)

Convolutional neural network wrapper

source

ResidualConvBlock

 ResidualConvBlock (c_in, c_out, stride=2, ks=3,
                    with_final_activation=True)

Convolutional block with residual links without a final activation

source

UpsamplingResidualConvBlock

 UpsamplingResidualConvBlock (*args, **kwargs)

Conv block with an upsampling pass

source

KaimingMixin

 KaimingMixin ()

Helper to initialize the network using Kaiming

Here’s the training code

source

train

 train (model, dls, lr=0.004, n_epochs=25,
        extra_cbs=[<slowai.learner.MetricsCB object at 0x7faa60e4b370>],
        loss_fn=<function mse_loss>)

source

viz

 viz (xb, yb, yp, n=3)

Different approaches to super-resolution

Let’s try a few different methods to invent higher-resolution details.

Approach #1: Autoencoder

Recall that we had poor results in FashionMNIST, so temper your expectations.

source

AutoEncoder

 AutoEncoder (nfs:list[int]=(32, 64, 128, 256, 512, 1024))

Project into a hidden space and reproject into the original space

lr_find(AutoEncoder.kaiming(), dls, gamma=1.1, start_lr=1e-5, loss_fn=F.mse_loss)
0.00% [0/2 00:00<?]
80.61% [79/98 00:51<00:12 1.099] 

ae = train(AutoEncoder.kaiming(), dls)
loss epoch train
0.845 0 train
0.558 0 eval
0.487 1 train
0.432 1 eval
0.387 2 train
0.354 2 eval
0.325 3 train
0.325 3 eval
0.286 4 train
0.270 4 eval
0.258 5 train
0.240 5 eval
0.241 6 train
0.236 6 eval
0.226 7 train
0.228 7 eval
0.216 8 train
0.240 8 eval
0.206 9 train
0.201 9 eval
0.197 10 train
0.196 10 eval
0.190 11 train
0.186 11 eval
0.183 12 train
0.184 12 eval
0.177 13 train
0.171 13 eval
0.172 14 train
0.170 14 eval
0.168 15 train
0.168 15 eval
0.164 16 train
0.162 16 eval
0.161 17 train
0.155 17 eval
0.160 18 train
0.154 18 eval
0.156 19 train
0.148 19 eval
0.154 20 train
0.148 20 eval
0.153 21 train
0.146 21 eval
0.152 22 train
0.144 22 eval
0.152 23 train
0.143 23 eval
0.151 24 train
0.142 24 eval 

with torch.no_grad():
    yp = ae(xb.to(def_device)).cpu()
viz(xb, yb, yp)

Not much better than the original.

Approach #2: U-net

This is nearly identical except that, for a particular up-sampling block, the logits from the same scale down-sampling are added or concatenated.

source

TinyUnet

 TinyUnet (nfs:list[int]=(32, 64, 128, 256, 512, 1024), n_blocks=(3, 2, 2,
           1, 1))

U-net

lr_find(TinyUnet.kaiming(), dls, gamma=1.1, start_lr=1e-5, loss_fn=F.mse_loss)
0.00% [0/2 00:00<?]
58.16% [57/98 00:38<00:27 1.355] 

un = train(TinyUnet.kaiming(), dls)
loss epoch train
1.265 0 train
0.377 0 eval
0.249 1 train
0.196 1 eval
0.173 2 train
0.149 2 eval
0.142 3 train
0.124 3 eval
0.128 4 train
0.119 4 eval
0.120 5 train
0.107 5 eval
0.115 6 train
0.105 6 eval
0.110 7 train
0.099 7 eval
0.107 8 train
0.098 8 eval
0.105 9 train
0.097 9 eval
0.103 10 train
0.095 10 eval
0.102 11 train
0.094 11 eval
0.101 12 train
0.093 12 eval
0.099 13 train
0.092 13 eval
0.099 14 train
0.092 14 eval
0.098 15 train
0.091 15 eval
0.097 16 train
0.091 16 eval
0.097 17 train
0.090 17 eval
0.096 18 train
0.090 18 eval
0.096 19 train
0.090 19 eval
0.096 20 train
0.090 20 eval
0.096 21 train
0.089 21 eval
0.096 22 train
0.089 22 eval
0.096 23 train
0.089 23 eval
0.096 24 train
0.089 24 eval 

with torch.no_grad():
    yp = un(xb.to(def_device)).cpu()
viz(xb, yb, yp)

del un

Approach #3: U-net with perceptual loss

First, let’s train a model with an identical downsampling circuit and a classification head

source

TinyImageResNet4

 TinyImageResNet4 (nfs:list[int]=(32, 64, 128, 256, 512, 1024))

Convolutional classification model

dls_clf = DataLoaders.from_dsd(dls.splits).listify(["image_high_rez", "idx"])
dls_clf.bs = 1024
m = MetricsCB(MulticlassAccuracy(num_classes=200))
clf = train(
    TinyImageResNet4.kaiming(),
    dls_clf,
    loss_fn=F.cross_entropy,
    lr=1e-2,
    n_epochs=25,
    extra_cbs=[m],
)
MulticlassAccuracy loss epoch train
0.064 4.714 0 train
0.133 4.083 0 eval
0.138 4.059 1 train
0.200 3.607 1 eval
0.197 3.648 2 train
0.223 3.540 2 eval
0.243 3.372 3 train
0.289 3.102 3 eval
0.284 3.136 4 train
0.292 3.084 4 eval
0.316 2.957 5 train
0.353 2.740 5 eval
0.349 2.783 6 train
0.360 2.755 6 eval
0.379 2.628 7 train
0.391 2.617 7 eval
0.404 2.498 8 train
0.376 2.737 8 eval
0.434 2.352 9 train
0.407 2.626 9 eval
0.457 2.234 10 train
0.419 2.598 10 eval
0.483 2.110 11 train
0.445 2.428 11 eval
0.509 1.987 12 train
0.455 2.427 12 eval
0.535 1.871 13 train
0.473 2.370 13 eval
0.564 1.730 14 train
0.476 2.352 14 eval
0.592 1.612 15 train
0.505 2.244 15 eval
0.624 1.475 16 train
0.496 2.339 16 eval
0.656 1.355 17 train
0.513 2.276 17 eval
0.685 1.230 18 train
0.535 2.194 18 eval
0.712 1.123 19 train
0.537 2.167 19 eval
0.737 1.038 20 train
0.551 2.166 20 eval
0.756 0.955 21 train
0.555 2.140 21 eval
0.773 0.897 22 train
0.560 2.123 22 eval
0.783 0.861 23 train
0.563 2.114 23 eval
0.788 0.837 24 train
0.564 2.115 24 eval 

torch.save(clf, "../models/tiny-image-resnet-4.pt")
clf = torch.load("../models/tiny-image-resnet-4.pt")
@torch.no_grad()
def get_clf_features(x):
    x = xb.cuda()
    for l in clf.conv_layers[:4]:
        x = l(x)
    x = torch.flatten(x)
    return x

Now, we can train using these “perceptual” features

def custom_loss(y, y_pred):
    """Compositive MSE and perceputal loss function"""
    perceptual_loss = F.mse_loss(
        get_clf_features(y),
        get_clf_features(y_pred),
    )
    return F.mse_loss(y, y_pred) + (0.1 * perceptual_loss)

Now, we can train with this loss function

unp = train(TinyUnet.kaiming(), dls, loss_fn=custom_loss)
loss epoch train
2.163 0 train
0.460 0 eval
0.315 1 train
0.238 1 eval
0.204 2 train
0.169 2 eval
0.162 3 train
0.139 3 eval
0.141 4 train
0.123 4 eval
0.128 5 train
0.115 5 eval
0.119 6 train
0.107 6 eval
0.115 7 train
0.104 7 eval
0.111 8 train
0.102 8 eval
0.108 9 train
0.100 9 eval
0.106 10 train
0.098 10 eval
0.105 11 train
0.096 11 eval
0.103 12 train
0.096 12 eval
0.102 13 train
0.094 13 eval
0.101 14 train
0.094 14 eval
0.101 15 train
0.093 15 eval
0.100 16 train
0.092 16 eval
0.099 17 train
0.092 17 eval
0.099 18 train
0.092 18 eval
0.098 19 train
0.091 19 eval
0.098 20 train
0.091 20 eval
0.098 21 train
0.091 21 eval
0.098 22 train
0.091 22 eval
0.098 23 train
0.091 23 eval
0.098 24 train
0.091 24 eval 

with torch.no_grad():
    yp = unp(xb.to(def_device)).cpu()
viz(xb, yb, yp)

torch.save(unp, "../models/tiny-image-resnet-un-p.pt")
unp = torch.load("../models/tiny-image-resnet-un-p.pt")

And, finally, train with the same features from the classifier in the downsampler path

source

initialize_unet_weights_with_clf_weights

 initialize_unet_weights_with_clf_weights (c, unet)
unpp = train(
    initialize_unet_weights_with_clf_weights(clf, TinyUnet.kaiming()),
    dls,
    loss_fn=custom_loss,
)
loss epoch train
1.419 0 train
0.420 0 eval
0.266 1 train
0.183 1 eval
0.159 2 train
0.130 2 eval
0.130 3 train
0.110 3 eval
0.146 4 train
0.117 4 eval
0.110 5 train
0.097 5 eval
0.103 6 train
0.093 6 eval
0.100 7 train
0.091 7 eval
0.098 8 train
0.090 8 eval
0.096 9 train
0.089 9 eval
0.096 10 train
0.088 10 eval
0.095 11 train
0.088 11 eval
0.094 12 train
0.087 12 eval
0.093 13 train
0.087 13 eval
0.093 14 train
0.086 14 eval
0.093 15 train
0.086 15 eval
0.092 16 train
0.086 16 eval
0.092 17 train
0.086 17 eval
0.092 18 train
0.086 18 eval
0.092 19 train
0.085 19 eval
0.091 20 train
0.085 20 eval
0.091 21 train
0.085 21 eval
0.091 22 train
0.085 22 eval
0.091 23 train
0.085 23 eval
0.091 24 train
0.085 24 eval 

with torch.no_grad():
    yp = unpp(xb.to(def_device)).cpu()
viz(xb, yb, yp)

del unpp