From d69cc8f46f238e91e2f597cd301cc53b1d4b8bec Mon Sep 17 00:00:00 2001 From: Volpeon Date: Sat, 25 Mar 2023 17:49:30 +0100 Subject: Fix training with guidance --- training/functional.py | 12 +++++++++--- 1 file changed, 9 insertions(+), 3 deletions(-) (limited to 'training') diff --git a/training/functional.py b/training/functional.py index d285366..109845b 100644 --- a/training/functional.py +++ b/training/functional.py @@ -344,9 +344,15 @@ def loss_step( raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}") if guidance_scale != 0: - # Chunk the noise and model_pred into two parts and compute the loss on each part separately. - model_pred_uncond, model_pred_text = torch.chunk(model_pred, 2, dim=0) - model_pred = model_pred_uncond + guidance_scale * (model_pred_text - model_pred_uncond) + uncond_encoder_hidden_states = get_extended_embeddings( + text_encoder, + batch["negative_input_ids"], + batch["negative_attention_mask"] + ) + uncond_encoder_hidden_states = uncond_encoder_hidden_states.to(dtype=unet.dtype) + + model_pred_uncond = unet(noisy_latents, timesteps, uncond_encoder_hidden_states).sample + model_pred = model_pred_uncond + guidance_scale * (model_pred - model_pred_uncond) loss = F.mse_loss(model_pred.float(), target.float(), reduction="none") elif prior_loss_weight != 0: -- cgit v1.2.3-54-g00ecf