Moved common training code into separate module

1 files changed, 16 insertions, 110 deletions

diff --git a/train_dreambooth.py b/train_dreambooth.py
index 0f8fece..9749c62 100644
--- a/train_dreambooth.py
+++ b/train_dreambooth.py
@@ -16,7 +16,6 @@ from accelerate.utils import LoggerType, set_seed
 from diffusers import AutoencoderKL, DDPMScheduler, DPMSolverMultistepScheduler, UNet2DConditionModel
 from diffusers.optimization import get_scheduler, get_cosine_with_hard_restarts_schedule_with_warmup
 from diffusers.training_utils import EMAModel
-from PIL import Image
 from tqdm.auto import tqdm
 from transformers import CLIPTextModel, CLIPTokenizer
 from slugify import slugify
@@ -25,6 +24,7 @@ from common import load_text_embeddings
 from pipelines.stable_diffusion.vlpn_stable_diffusion import VlpnStableDiffusion
 from data.csv import CSVDataModule
 from training.optimization import get_one_cycle_schedule
+from training.util import AverageMeter, CheckpointerBase, freeze_params, save_args
 from models.clip.prompt import PromptProcessor
 
 logger = get_logger(__name__)
@@ -385,41 +385,7 @@ def parse_args():
     return args
 
 
-def save_args(basepath: Path, args, extra={}):
-    info = {"args": vars(args)}
-    info["args"].update(extra)
-    with open(basepath.joinpath("args.json"), "w") as f:
-        json.dump(info, f, indent=4)
-
-
-def freeze_params(params):
-    for param in params:
-        param.requires_grad = False
-
-
-def make_grid(images, rows, cols):
-    w, h = images[0].size
-    grid = Image.new('RGB', size=(cols*w, rows*h))
-    for i, image in enumerate(images):
-        grid.paste(image, box=(i % cols*w, i//cols*h))
-    return grid
-
-
-class AverageMeter:
-    def __init__(self, name=None):
-        self.name = name
-        self.reset()
-
-    def reset(self):
-        self.sum = self.count = self.avg = 0
-
-    def update(self, val, n=1):
-        self.sum += val * n
-        self.count += n
-        self.avg = self.sum / self.count
-
-
-class Checkpointer:
+class Checkpointer(CheckpointerBase):
     def __init__(
         self,
         datamodule,
@@ -437,9 +403,20 @@ class Checkpointer:
         sample_image_size,
         sample_batches,
         sample_batch_size,
-        seed
+        seed,
     ):
-        self.datamodule = datamodule
+        super().__init__(
+            datamodule=datamodule,
+            output_dir=output_dir,
+            instance_identifier=instance_identifier,
+            placeholder_token=placeholder_token,
+            placeholder_token_id=placeholder_token_id,
+            sample_image_size=sample_image_size,
+            seed=seed or torch.random.seed(),
+            sample_batches=sample_batches,
+            sample_batch_size=sample_batch_size
+        )
+
         self.accelerator = accelerator
         self.vae = vae
         self.unet = unet
@@ -447,14 +424,6 @@ class Checkpointer:
         self.tokenizer = tokenizer
         self.text_encoder = text_encoder
         self.scheduler = scheduler
-        self.output_dir = output_dir
-        self.instance_identifier = instance_identifier
-        self.placeholder_token = placeholder_token
-        self.placeholder_token_id = placeholder_token_id
-        self.sample_image_size = sample_image_size
-        self.seed = seed or torch.random.seed()
-        self.sample_batches = sample_batches
-        self.sample_batch_size = sample_batch_size
 
     @torch.no_grad()
     def save_model(self):
@@ -481,8 +450,6 @@ class Checkpointer:
 
     @torch.no_grad()
     def save_samples(self, step, num_inference_steps, guidance_scale=7.5, eta=0.0):
-        samples_path = Path(self.output_dir).joinpath("samples")
-
         unet = self.ema_unet.averaged_model if self.ema_unet is not None else self.accelerator.unwrap_model(self.unet)
         text_encoder = self.accelerator.unwrap_model(self.text_encoder)
 
@@ -495,72 +462,11 @@ class Checkpointer:
         ).to(self.accelerator.device)
         pipeline.set_progress_bar_config(dynamic_ncols=True)
 
-        train_data = self.datamodule.train_dataloader()
-        val_data = self.datamodule.val_dataloader()
-
-        generator = torch.Generator(device=pipeline.device).manual_seed(self.seed)
-        stable_latents = torch.randn(
-            (self.sample_batch_size, pipeline.unet.in_channels, self.sample_image_size // 8, self.sample_image_size // 8),
-            device=pipeline.device,
-            generator=generator,
-        )
-
-        with torch.autocast("cuda"), torch.inference_mode():
-            for pool, data, latents in [("stable", val_data, stable_latents), ("val", val_data, None), ("train", train_data, None)]:
-                all_samples = []
-                file_path = samples_path.joinpath(pool, f"step_{step}.jpg")
-                file_path.parent.mkdir(parents=True, exist_ok=True)
-
-                data_enum = enumerate(data)
-
-                batches = [
-                    batch
-                    for j, batch in data_enum
-                    if j * data.batch_size < self.sample_batch_size * self.sample_batches
-                ]
-                prompts = [
-                    prompt.format(identifier=self.instance_identifier)
-                    for batch in batches
-                    for prompt in batch["prompts"]
-                ]
-                nprompts = [
-                    prompt
-                    for batch in batches
-                    for prompt in batch["nprompts"]
-                ]
-
-                for i in range(self.sample_batches):
-                    prompt = prompts[i * self.sample_batch_size:(i + 1) * self.sample_batch_size]
-                    nprompt = nprompts[i * self.sample_batch_size:(i + 1) * self.sample_batch_size]
-
-                    samples = pipeline(
-                        prompt=prompt,
-                        negative_prompt=nprompt,
-                        height=self.sample_image_size,
-                        width=self.sample_image_size,
-                        image=latents[:len(prompt)] if latents is not None else None,
-                        generator=generator if latents is not None else None,
-                        guidance_scale=guidance_scale,
-                        eta=eta,
-                        num_inference_steps=num_inference_steps,
-                        output_type='pil'
-                    ).images
-
-                    all_samples += samples
-
-                    del samples
-
-                image_grid = make_grid(all_samples, self.sample_batches, self.sample_batch_size)
-                image_grid.save(file_path, quality=85)
-
-                del all_samples
-                del image_grid
+        super().save_samples(pipeline, step, num_inference_steps, guidance_scale, eta)
 
         del unet
         del text_encoder
         del pipeline
-        del generator
-        del stable_latents
 
         if torch.cuda.is_available():
             torch.cuda.empty_cache()