File size: 2,105 Bytes
b579854 |
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 |
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import torch.nn as nn
from transformers import ViTImageProcessor, ViTModel, AutoImageProcessor, AutoModel, Dinov2Model
class DinoWrapper(nn.Module):
"""
Dino v1 wrapper using huggingface transformer implementation.
"""
def __init__(self, model_name: str, freeze: bool = True):
super().__init__()
self.model, self.processor = self._build_dino(model_name)
if freeze:
self._freeze()
def forward(self, image):
# image: [N, C, H, W], on cpu
# RGB image with [0,1] scale and properly sized
inputs = self.processor(images=image.float(), return_tensors="pt", do_rescale=False, do_resize=False).to(self.model.device)
# This resampling of positional embedding uses bicubic interpolation
outputs = self.model(**inputs)
last_hidden_states = outputs.last_hidden_state
return last_hidden_states
def _freeze(self):
print(f"======== Freezing DinoWrapper ========")
self.model.eval()
for name, param in self.model.named_parameters():
param.requires_grad = False
@staticmethod
def _build_dino(model_name: str, proxy_error_retries: int = 3, proxy_error_cooldown: int = 5):
import requests
try:
processor = AutoImageProcessor.from_pretrained('facebook/dinov2-base')
processor.do_center_crop = False
model = AutoModel.from_pretrained('facebook/dinov2-base')
return model, processor
except requests.exceptions.ProxyError as err:
if proxy_error_retries > 0:
print(f"Huggingface ProxyError: Retrying in {proxy_error_cooldown} seconds...")
import time
time.sleep(proxy_error_cooldown)
return DinoWrapper._build_dino(model_name, proxy_error_retries - 1, proxy_error_cooldown)
else:
raise err |