Why view_embedding is not trainable?

[AlisherBlack]

This is the first time I've seen positional encoding defined as nn.Parameter but set to not be trainable, even though it's initialized with random weights. Could this be a mistake? If I misunderstood something, I apologize.

LiDAR-Diffusion/lidm/modules/encoders/modules.py

Line 255 in 8416ddb

x = torch.cat(img_feats, 1).float() + self.view_embedding

with torch.no_grad():
    img_feats = [self.model.encode_image(self.preprocess(img))[:, None] for img in x]
    x = torch.cat(img_feats, 1).float() + self.view_embedding
    x = self.linear(x)

P.S. self.linear is not trainable as well