Method save_samples

scripts/diffusers_textual_inversion_2.py:515–612

(
        self,
        step,
        text_encoder,
        height,
        width,
        guidance_scale,
        eta,
        num_inference_steps,
    )

    @torch.no_grad()
    def save_samples(
        self,
        step,
        text_encoder,
        height,
        width,
        guidance_scale,
        eta,
        num_inference_steps,
    ):
        samples_path = f"{self.output_dir}/samples"
        os.makedirs(samples_path, exist_ok=True)
        checker = NoCheck()

        unwrapped = self.accelerator.unwrap_model(text_encoder)
        # Save a sample image
        pipeline = StableDiffusionPipeline(
            text_encoder=unwrapped,
            vae=self.vae,
            unet=self.unet,
            tokenizer=self.tokenizer,
            scheduler=LMSDiscreteScheduler(
                beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear"
            ),
            safety_checker=NoCheck(),
            feature_extractor=CLIPFeatureExtractor.from_pretrained(
                "openai/clip-vit-base-patch32"
            ),
        ).to("cuda")
        pipeline.enable_attention_slicing()

        if self.stable_sample_batches > 0:
            stable_latents = torch.randn(
                (
                    self.sample_batch_size,
                    pipeline.unet.in_channels,
                    height // 8,
                    width // 8,
                ),
                device=pipeline.device,
                generator=torch.Generator(device=pipeline.device).manual_seed(
                    self.seed
                ),
            )

            stable_prompts = [
                choice.format(self.placeholder_token)
                for choice in (self.templates * self.sample_batch_size)[
                    : self.sample_batch_size
                ]
            ]

            # Generate and save stable samples
            for i in range(0, self.stable_sample_batches):
                samples = pipeline(
                    prompt=stable_prompts,
                    height=384,
                    latents=stable_latents,