difference between training and inference of omniaid

[garychan22]

I have noticed that there are some differences between the training and inference of omniaid.

During training, using

RealGen/RealGen_v2/flow_grpo/aigidetect_model/OmniAID.py

Lines 250 to 258 in 35de0a2

	def forward(self, images) -> dict:
	batch_size = images.size(0)
	with torch.no_grad():
	# Use the pooler_output as the routing feature
	routing_features = self.feature_extractor(images, output_hidden_states=False).pooler_output
	hidden_states = self.embeddings(images)
	hidden_states = self.ln_pre(hidden_states)

During inference, using

RealGen/RealGen_v1/flow_grpo/reward-server/reward_server/omniaid_scorer.py

Lines 86 to 89 in 35de0a2

	def forward(self, images) -> dict:
	batch_size = images.size(0)
	hidden_states = self.embeddings(images)
	routing_features = torch.mean(hidden_states[:, 1:, :], dim=1)

what is the reason for this? I want to implement the offline OmniAiD-DINO reward without GAN-style training.

Thanks~

[yunncheng]

Since RealGen v1 utilizes an early architectural implementation of OmniAID, whereas v2 employs the latest version, we recommend directly copying the implementation from OmniAID_DINO.py if you intend to perform offline training.

[garychan22]

Since RealGen v1 utilizes an early architectural implementation of OmniAID, whereas v2 employs the latest version, we recommend directly copying the implementation from OmniAID_DINO.py if you intend to perform offline training.

Thanks for your reply!

So, I just need to directly import the model class from the models/OmniAID.py and OmniAID_DINO.py, and load their pre-trained weights, and using

prob = torch.softmax(pred, dim=1)[:, 0]

as the image quality score?

Did I understand correctly?

Thanks~