Training classifier on bottleneck

[tsaxena]

@alexandre01 I am trying to train a classifier on the icons using the bottleneck embeddings that I get by model inference using pretrained model. In some cases though it doesnt seem to work. I used the code in your latent_ops notebook and encode each icon. In some cases I get this error
The size of tensor a (10) must match the size of tensor b (8) at non-singleton dimension 0
Am I missing something?

[tsaxena]

I am using the following methods to encode icon, but does seem to work on some of the icon indices.

`
def encode(data):
model_args = batchify((data[key] for key in cfg.model_args), device)
with torch.no_grad():
z = model(*model_args, encode_mode=True)
return z

def encode_icon(idx):
data = dataset.get(id=idx, random_aug=False)
return encode(data)
`

[alexandre01]

Hello @tsaxena!
Yes unfortunately this is due to the fact that the pretrained model was trained with a maximum of 8 paths per SVG. Since an index embedding is used in the model, this means one cannot perform inference using a larger amount.

You'd need to train a model with more paths or filter your classification dataset to eight paths.

[tsaxena]

Thanks for the prompt reply @alexandre01 . So does that mean you did not use all 100k icons for training the pretrained model?

[alexandre01]

Yes, the dataset is about 100k icons, but because of time constraints the pretrained model was only trained on a filtered subset. And I don't have access to the GPU server I used anymore.

[tsaxena]

I eventually want to fine tune the network on svgs that will have more than 8 paths. Do you suggest training from scratch? From what I understand, the max number of SVG paths is a configuration parameter that can be changed.

[pwichmann]

I am not @alexandre01. But what you say is correct, I think.

The max number of paths is a config parameter and can be changed in deepsvg/model/confg.py:

self.max_num_groups = 8          # Number of paths (N_P)

There are also individual configs in configs/deepsvg/ that overwrite this variable. You may need to increase the number of paths here as well.

You would need to retrain to cope with a larger number of paths. If you have a beefy GPU, retraining does not take very long. On my RTX 3090, I can retrain from scratch within hours.

What is your use case?