File size: 2,602 Bytes
a3f6fc6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 | import torch
import einops
import torch.nn as nn
from transformers import AutoImageProcessor, Dinov2Model
class WebSSLEncoder(nn.Module):
def __init__(self, name, feature_key, plus=False, output_dim=None, postprocess=None, n_patches=196):
super().__init__()
print("Encoder feature_key:", feature_key)
self.name = name
repo = "facebook/webssl-dino300m-full2b-224"
self.processor = AutoImageProcessor.from_pretrained(repo, do_rescale=False)
self.base_model = Dinov2Model.from_pretrained(repo)
self.feature_key = feature_key
self.emb_dim = self.base_model.config.hidden_size
self.output_dim = self.emb_dim # for compatibility
if feature_key == "x_norm_patchtokens":
self.latent_ndim = 2
elif feature_key == "x_norm_clstoken":
self.latent_ndim = 1
else:
raise ValueError(f"Invalid feature key: {feature_key}")
self.patch_size = self.base_model.config.patch_size
self.postprocess = postprocess
if postprocess is not None:
if postprocess == 'avg_pool':
self.latent_ndim = 1
def forward(self, x):
# Accept arbitrary number of leading dimensions before (C, H, W)
# and preserve them on return.
# Example: input shape (...prefix, C, H, W)
prefix_shape = x.shape[:-3]
c, h, w = x.shape[-3:]
# Collapse all leading dims into a single batch dimension for the base model
prod_prefix = 1
for d in prefix_shape:
prod_prefix *= d
x = x.reshape(prod_prefix, c, h, w)
inputs = self.processor(images=x, return_tensors="pt")
device = next(self.base_model.parameters()).device
target_dtype = next(self.base_model.parameters()).dtype
inputs = {
k: v.to(device=device, dtype=target_dtype) if isinstance(v, torch.Tensor) else v
for k, v in inputs.items()
}
outputs = self.base_model(**inputs)
if self.feature_key == "x_norm_clstoken":
emb = outputs.last_hidden_state[:, 0, :] # CLS token
elif self.feature_key == "x_norm_patchtokens":
emb = outputs.last_hidden_state[:, 1:, :] # Patch tokens (skip 4 register tokens)
emb = emb.reshape(*prefix_shape, *emb.shape[1:])
if self.postprocess == 'avg_pool':
emb = torch.mean(emb, dim=-2) # (...prefix, e)
if self.latent_ndim == 1:
emb = emb.unsqueeze(len(prefix_shape)) # dummy patch dim, b v 1 e
return emb |