Updated super-res for v2?

fastai
1

Hi all,

Just starting out trying the learn all the changes brought about in fastai v2.

One of my favorite projects was the superres notebooks from the v1 course. I had made my own models to try to deblur using the same techniques, but with v2 I don’t even know where to begin to load and use the model on a new photo. I know I need to change my thinking from giving the unet_learner ‘data’ to giving it a data-loader, but I’m really stuck.

Has anyone already tried to convert to fastai v2 the super-res notebook, and in particular this one: lesson7-superres.ipynb? Will I have to retrain a new model using fastai v2?

Any help or pointers would be greatly appreciated!

2

The entire v3 course ported to v2 (part 1) can be found here:

superres is included https://github.com/fastai/fastai/blob/master/dev_nbs/course/lesson7-superres.ipynb

3

Oh wow! I didn’t even realize, thanks!

4

Sorry, back to my 2nd question - can we use a model trained with v1 for v2? When trying to load the model, I’m getting a lot of errors:
learn = unet_learner(dls, arch, loss_func=F.l1_loss, blur=True, norm_type=NormType.Weight)
learn.load('super_res_5b');

RuntimeError: Error(s) in loading state_dict for DynamicUnet:
	Missing key(s) in state_dict: "layers.3.0.0.weight", "layers.3.1.0.weight", "layers.4.shuf.0.0.weight", "layers.4.shuf.0.0.bias", "layers.4.conv1.0.weight", "layers.4.conv2.0.weight", "layers.5.shuf.0.0.weight", "layers.5.shuf.0.0.bias", "layers.5.conv1.0.weight", "layers.5.conv2.0.weight", "layers.6.shuf.0.0.weight", "layers.6.shuf.0.0.bias", "layers.6.conv1.0.weight", "layers.6.conv2.0.weight", "layers.7.shuf.0.0.weight", "layers.7.shuf.0.0.bias", "layers.7.conv1.0.weight", "layers.7.conv2.0.weight", "layers.8.0.0.weight", "layers.8.0.0.bias", "layers.11.convpath.0.0.weight", "layers.11.convpath.0.0.bias", "layers.11.convpath.1.0.weight", "layers.11.convpath.1.0.bias", "layers.12.0.weight", "layers.12.0.bias". 
	Unexpected key(s) in state_dict: "layers.3.0.0.weight_g", "layers.3.0.0.weight_v", "layers.3.1.0.weight_g", "layers.3.1.0.weight_v", "layers.4.shuf.conv.0.bias", "layers.4.shuf.conv.0.weight_g", "layers.4.shuf.conv.0.weight_v", "layers.4.conv1.0.weight_g", "layers.4.conv1.0.weight_v", "layers.4.conv2.0.weight_g", "layers.4.conv2.0.weight_v", "layers.5.shuf.conv.0.bias", "layers.5.shuf.conv.0.weight_g", "layers.5.shuf.conv.0.weight_v", "layers.5.conv1.0.weight_g", "layers.5.conv1.0.weight_v", "layers.5.conv2.0.weight_g", "layers.5.conv2.0.weight_v", "layers.6.shuf.conv.0.bias", "layers.6.shuf.conv.0.weight_g", "layers.6.shuf.conv.0.weight_v", "layers.6.conv1.0.weight_g", "layers.6.conv1.0.weight_v", "layers.6.conv2.0.weight_g", "layers.6.conv2.0.weight_v", "layers.7.shuf.conv.0.bias", "layers.7.shuf.conv.0.weight_g", "layers.7.shuf.conv.0.weight_v", "layers.7.conv1.0.weight_g", "layers.7.conv1.0.weight_v", "layers.7.conv2.0.weight_g", "layers.7.conv2.0.weight_v", "layers.8.conv.0.bias", "layers.8.conv.0.weight_g", "layers.8.conv.0.weight_v", "layers.10.layers.0.0.bias", "layers.10.layers.0.0.weight_g", "layers.10.layers.0.0.weight_v", "layers.10.layers.1.0.bias", "layers.10.layers.1.0.weight_g", "layers.10.layers.1.0.weight_v", "layers.11.0.bias", "layers.11.0.weight_g", "layers.11.0.weight_v".
5

Not if the model definition changed. V2 is a complete rewrite, which includes the unet, as you’ve seen :slight_smile:

You’d need to make a model with the old unet code to make it work.

6

Hello!

I’m revisiting this notebook and rebuilding my models now that fastai changed the way models were saved.

Unfortunately, the following code is providing errors for unet_config: not defined. Is there a workaround for this? I’m not familiar with what this bit of code is doing so not sure how to replace it.

learn = unet_learner(dls, arch, loss_func=feat_loss, metrics=LossMetrics(feat_loss.metric_names),
                     config=unet_config(blur=True, norm_type=NormType.Weight))

I double checked and this is the same in the master, here: fastai/lesson7-superres.ipynb at master · fastai/fastai · GitHub

7

I solved my own problem - looks like the arguments are passed directly like so:

learn = unet_learner(dls, arch, loss_func=feat_loss, metrics=LossMetrics(feat_loss.metric_names),
                     blur=True, norm_type=NormType.Weight)