Additions to fine-tune TPDE dataset

README.md

@@ -1,6 +1,8 @@
 # Experiments with Stable Diffusion
 
-This repository extends and adds to the [original training repo](https://github.com/pesser/stable-diffusion) for Stable Diffusion.
+Fine-tune stable diffusion with images from Paris. Part of the _This Place Does Exist_ experiment.
+
+This repository extends and adds to the [fine-tune repo by Justin Pinkney](https//github.com/justinpinkney/stable-diffusion.git) which in turn extends the [original training repo](https://github.com/pesser/stable-diffusion) for Stable Diffusion.
 
 Currently it adds:
 

configs/stable-diffusion/paris.yaml

@@ -0,0 +1,138 @@
+model:
+  base_learning_rate: 1.0e-04
+  target: ldm.models.diffusion.ddpm.LatentDiffusion
+  params:
+    linear_start: 0.00085
+    linear_end: 0.0120
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: "image"
+    cond_stage_key: "txt"
+    image_size: 64
+    channels: 4
+    cond_stage_trainable: false   # Note: different from the one we trained before
+    conditioning_key: crossattn
+    scale_factor: 0.18215
+
+    scheduler_config: # 10000 warmup steps
+      target: ldm.lr_scheduler.LambdaLinearScheduler
+      params:
+        warm_up_steps: [ 1 ] # NOTE for resuming. use 10000 if starting from scratch
+        cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
+        f_start: [ 1.e-6 ]
+        f_max: [ 1. ]
+        f_min: [ 1. ]
+
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        image_size: 32 # unused
+        in_channels: 4
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions: [ 4, 2, 1 ]
+        num_res_blocks: 2
+        channel_mult: [ 1, 2, 4, 4 ]
+        num_heads: 8
+        use_spatial_transformer: True
+        transformer_depth: 1
+        context_dim: 768
+        use_checkpoint: True
+        legacy: False
+
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      ckpt_path: "models/first_stage_models/kl-f8/model.ckpt"
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: ldm.modules.encoders.modules.FrozenCLIPEmbedder
+
+
+data:
+  target: main.DataModuleFromConfig
+  params:
+    batch_size: 4
+    num_workers: 4
+    use_worker_init_fn: false
+    # num_val_workers: 0 # Avoid a weird val dataloader issue
+    train:
+      target: paris_dataloader.ParisDataset #implements https://pytorch.org/docs/stable/data.html#torch.utils.data.IterableDataset
+      params:
+        image_folder: ../VLoD/
+        width: 512
+        height: 512
+        max_images: 50
+        image_transforms:
+        - target: torchvision.transforms.Resize
+          params:
+            size: 512
+            interpolation: 3
+        - target: torchvision.transforms.RandomCrop
+          params:
+            size: 512
+        - target: torchvision.transforms.RandomHorizontalFlip
+    validation:
+      target: ldm.data.simple.TextOnly
+      params:
+        captions:
+        - "The 1st arrondisement of Paris"
+        - "The 14th arrondisement of Paris"
+        - "A shop front"
+        - "Boulevard de Magenta"
+        output_size: 512
+        n_gpus: 2 # small hack to sure we see all our samples
+
+
+lightning:
+  find_unused_parameters: False
+
+  modelcheckpoint:
+    params:
+      every_n_train_steps: 2000
+      save_top_k: -1
+      monitor: null
+
+  callbacks:
+    image_logger:
+      target: main.ImageLogger
+      params:
+        batch_frequency: 2000
+        max_images: 4
+        increase_log_steps: False
+        log_first_step: True
+        log_all_val: True
+        log_images_kwargs:
+          use_ema_scope: True
+          inpaint: False
+          plot_progressive_rows: False
+          plot_diffusion_rows: False
+          N: 4
+          unconditional_guidance_scale: 3.0
+          unconditional_guidance_label: [""]
+
+  trainer:
+    benchmark: True
+    num_sanity_val_steps: 0
+    accumulate_grad_batches: 1
+    # gpus: "1," # for some reason triggers erro

main.py

@@ -215,7 +215,7 @@ class DataModuleFromConfig(pl.LightningDataModule):
         else:
             init_fn = None
         return DataLoader(self.datasets["train"], batch_size=self.batch_size,
-                          num_workers=self.num_workers, shuffle=False if is_iterable_dataset else True,
+                          num_workers=self.num_workers, # shuffle=False if is_iterable_dataset else True,
                           worker_init_fn=init_fn)
 
     def _val_dataloader(self, shuffle=False):

paris_dataloader.py

@@ -0,0 +1,97 @@
+import math
+from os import PathLike
+import os
+from pathlib import Path
+from typing import Iterator, Optional
+from ldm.data.base import Txt2ImgIterableBaseDataset
+from PIL import Image, ImageOps
+import numpy as np
+from torch.utils.data import Dataset, ConcatDataset, ChainDataset, IterableDataset
+from torch.utils.data import get_worker_info
+from ldm.util import instantiate_from_config
+from torchvision import transforms
+from einops import rearrange
+
+Image.init() # required to initialise PIL.Image.EXTENSION
+
+def extract_labels(f: Path) -> dict:
+    # get the labels for the image path
+    arr = int(f.parent.parent.name.split("E")[0])
+    street = " ".join(f.parent.name.split(' ')[:-1]).replace('_', '\'')
+    return {
+        "arrondisement": arr,
+        "street": street,
+    }
+
+# adapted from stylegan dataset_tool
+def is_image_ext(f: Path) -> bool:
+    return f.suffix.lower() in Image.EXTENSION
+
+def open_image_folder(source_dir, *, max_images: Optional[int] = None):
+    image_files = [f for f in sorted(Path(source_dir).rglob('*')) if is_image_ext(f) and os.path.isfile(f) and f.parent.name != "BU"]
+    image_files = image_files[:max_images] 
+
+    labeled_images = [{
+        "abspath": f.resolve(),
+        "relpath": f.relative_to(source_dir),
+        "labels":  extract_labels(f.relative_to(source_dir))
+        } for f in image_files]
+
+    return labeled_images
+
+def make_ordinal(n):
+    '''
+    Convert an integer into its ordinal representation::
+
+        make_ordinal(0)   => '0th'
+        make_ordinal(3)   => '3rd'
+        make_ordinal(122) => '122nd'
+        make_ordinal(213) => '213th'
+    '''
+    n = int(n)
+    if 11 <= (n % 100) <= 13:
+        suffix = 'th'
+    else:
+        suffix = ['th', 'st', 'nd', 'rd', 'th'][min(n % 10, 4)]
+    return str(n) + suffix
+
+
+class ParisDataset(IterableDataset):
+    # see also ldm.utils.data.simple.FolderData
+    def __init__(self, image_folder: os.PathLike, width, height, max_images, image_transforms):
+        super(ParisDataset).__init__()
+        assert os.path.exists(image_folder), "image_folder does not exist"
+        self.labeled_images = open_image_folder(image_folder, max_images=max_images)
+        self.width = width
+        self.height = height
+
+        image_transforms = [instantiate_from_config(tt) for tt in image_transforms]
+        image_transforms.extend([transforms.ToTensor(),
+                                 transforms.Lambda(lambda x: rearrange(x * 2. - 1., 'c h w -> h w c'))])
+        image_transforms = transforms.Compose(image_transforms)
+        self.tform = image_transforms
+
+    def __iter__(self):
+        worker_info = get_worker_info()
+        if worker_info is None:  # single-process data loading, return the full iterator
+            iter_start = 0
+            iter_end = len(self.labeled_images)
+        else:  # in a worker process
+            # split workload
+            per_worker = int(math.ceil(len(self.labeled_images) / float(worker_info.num_workers)))
+            worker_id = worker_info.id
+            iter_start = worker_id * per_worker
+            iter_end = min(iter_start + per_worker, len(self.labeled_images))
+
+        for image in self.labeled_images[iter_start:iter_end]:
+            yield {
+                #with tform, the scaling is superfluous
+                'image': self.tform(ImageOps.fit(Image.open(image['abspath']), (self.width, self.height)).convert("RGB")),
+                'txt': f"Shop front on the {image['labels']['street']} in the {make_ordinal(image['labels']['arrondisement'])} arrondissement of Paris",
+            }
+
+
+if __name__ == "__main__":
+    d = ParisDataset("../VLoD/", 512, 512)
+    for i in d:
+        print(i)