first (in)stable steps

configs/stable-diffusion/txt2img-ldm-vae-f8.yaml

@@ -0,0 +1,130 @@
+model:
+  base_learning_rate: 1.0e-04  # TODO: run with scale_lr False
+  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: "jpg"
+    cond_stage_key: "txt"
+    image_size: 32
+    channels: 4
+    cond_stage_trainable: true
+    conditioning_key: crossattn
+    monitor: val/loss_simple_ema
+    scale_factor: 0.18215
+
+    scheduler_config: # 10000 warmup steps
+      target: ldm.lr_scheduler.LambdaLinearScheduler
+      params:
+        warm_up_steps: [ 10000 ]
+        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
+        in_channels: 4
+        out_channels: 4
+        model_channels: 128 # 320   # TODO increase
+        attention_resolutions: [ 4, 2, 1 ]  # is equal to fixed spatial resolution: 32 , 16 , 8
+        num_res_blocks: 2
+        channel_mult: [ 1,2,4,4 ]
+        #num_head_channels: 32
+        num_heads: 8
+        use_spatial_transformer: True
+        transformer_depth: 1
+        context_dim: 1280
+        use_checkpoint: True
+
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ckpt_path: "/home/robin/projects/latent-diffusion/models/first_stage_models/kl-f8/model.ckpt"
+        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.BERTEmbedder
+      params:
+        n_embed: 1280
+        n_layer: 3 #32 # TODO: increase
+
+
+data:
+  target: ldm.data.laion.WebDataModuleFromConfig
+  params:
+    tar_base: "pipe:aws s3 cp s3://s-datasets/laion5b/laion2B-data/"
+    batch_size: 60
+    num_workers: 4
+    n_nodes: 2  # TODO: runs with two gpus
+    train:
+      shards: '{000000..000010}.tar -'  # TODO: wild guess, change
+      image_key: jpg
+      image_transforms:
+      - target: torchvision.transforms.Resize
+        params:
+          size: 512
+          interpolation: 3
+      - target: torchvision.transforms.RandomCrop
+        params:
+          size: 512
+
+      shuffle: 5000
+      n_examples: 16519100  # TODO: find out
+    validation:
+      shards: '{000011..000012}.tar -'  # TODO: wild guess, change
+      image_key: jpg
+      image_transforms:
+      - target: torchvision.transforms.Resize
+        params:
+          size: 512
+          interpolation: 3
+      - target: torchvision.transforms.CenterCrop
+        params:
+          size: 512
+
+      shuffle: 0
+      n_examples: 60000  # TODO: find out
+    val_num_workers: 2
+
+
+
+lightning:
+  callbacks:
+    image_logger:
+      target: main.ImageLogger
+      params:
+        batch_frequency: 5000 # 5000
+        max_images: 8
+        increase_log_steps: False
+        log_first_step: True
+
+
+  trainer:
+    replace_sampler_ddp: False
+    benchmark: True
+    val_check_interval: 20000  # every 20k training steps
+    num_sanity_val_steps: 0

ldm/data/laion.py

@@ -2,7 +2,178 @@ import webdataset as wds
 from PIL import Image
 import io
 import os
+import torchvision
+from PIL import Image
+import glob
+import random
+import numpy as np
+import pytorch_lightning as pl
 from tqdm import tqdm
+from omegaconf import OmegaConf
+from einops import rearrange
+import torch
+
+
+from ldm.util import instantiate_from_config
+
+
+def dict_collation_fn(samples, combine_tensors=True, combine_scalars=True):
+    """Take a list  of samples (as dictionary) and create a batch, preserving the keys.
+    If `tensors` is True, `ndarray` objects are combined into
+    tensor batches.
+    :param dict samples: list of samples
+    :param bool tensors: whether to turn lists of ndarrays into a single ndarray
+    :returns: single sample consisting of a batch
+    :rtype: dict
+    """
+    batched = {key: [] for key in samples[0]}
+    # assert isinstance(samples[0][first_key], (list, tuple)), type(samples[first_key])
+
+    for s in samples:
+        [batched[key].append(s[key]) for key in batched]
+
+
+    result = {}
+    for key in batched:
+        if isinstance(batched[key][0], (int, float)):
+            if combine_scalars:
+                result[key] = np.array(list(batched[key]))
+        elif isinstance(batched[key][0], torch.Tensor):
+            if combine_tensors:
+                # import torch
+
+                result[key] = torch.stack(list(batched[key]))
+        elif isinstance(batched[key][0], np.ndarray):
+            if combine_tensors:
+                result[key] = np.array(list(batched[key]))
+        else:
+            result[key] = list(batched[key])
+        # result.append(b)
+    return result
+
+
+class WebDataModuleFromConfig(pl.LightningDataModule):
+    def __init__(self, tar_base, batch_size, train=None, validation=None,
+                 test=None, num_workers=4, load_ddp=True, n_nodes=1,
+                 **kwargs):
+        super().__init__(self)
+        print(f'Setting tar base to {tar_base}')
+        self.tar_base = tar_base
+        self.batch_size = batch_size
+        self.num_workers = num_workers
+        self.train = train
+        self.validation = validation
+        self.test = test
+        self.load_ddp = load_ddp
+        self.multinode = n_nodes > 1
+        self.n_nodes = n_nodes  # n gpu ??
+
+    def make_loader(self, dataset_config, train=True):
+        if 'image_transforms' in dataset_config:
+            image_transforms = [instantiate_from_config(tt) for tt in dataset_config.image_transforms]
+        else:
+            image_transforms = []
+
+        image_transforms.extend([torchvision.transforms.ToTensor(),
+                                 torchvision.transforms.Lambda(lambda x: rearrange(x * 2. - 1., 'c h w -> h w c'))])
+        image_transforms = torchvision.transforms.Compose(image_transforms)
+
+        if 'transforms' in dataset_config:
+            transforms_config = OmegaConf.to_container(dataset_config.transforms)
+        else:
+            transforms_config = dict()
+
+        transform_dict = {dkey: load_partial_from_config(transforms_config[dkey]) if transforms_config[
+                                                                                         dkey] != 'identity' else identity
+                          for dkey in transforms_config}
+        img_key = dataset_config.get('image_key', 'jpeg')
+        transform_dict.update({img_key: image_transforms})
+
+        shuffle = dataset_config.get('shuffle', 0)
+
+        # TODO fid strategy when n exmples not known beforehand
+        n_examples = dataset_config.get('n_examples', 1e6) // self.n_nodes
+
+        shards_to_load = dataset_config.shards
+        dset_name = 'unknown'
+        if isinstance(shards_to_load, str):
+            print(f'Loading tars based on the string {shards_to_load}')
+            tars = os.path.join(self.tar_base, shards_to_load)
+            start_shard_id, end_shard_id = dataset_config.shards.split('{')[-1].split('}')[0].split('..')
+            n_shards = int(end_shard_id) - int(start_shard_id) + 1
+            dset_name = dataset_config.shards.split('-')[0]
+        elif isinstance(shards_to_load, int):
+            print(f'Creating tar list, max shard is {shards_to_load}')
+            try:
+                tars = [tf for tf in natsorted(glob(os.path.join(self.tar_base, '*.tar'))) if
+                        int(tf.split('/')[-1].split('.')[0]) < shards_to_load]
+                n_shards = len(tars)
+                random.shuffle(tars)
+
+            except ValueError as e:
+                print('tarfile names should follow the pattern <zero_padded_number>.tar . Check names of the files')
+                raise e
+        else:
+            raise ValueError(
+                'shards should be either a string containing consecutive shards or an int defining the max shard number')
+
+        print(f'Got {n_shards} shard files in datafolder for {"training" if train else "validation"}')
+
+        # if self.num_workers > 0:
+        #     assert n_shards % self.num_workers == 0 , f'Number of workers which is {self.num_workers} does not evenly divide number of shards which is {n_shards}'
+        print(f'Loading webdataset based dataloader based on {n_shards} of {dset_name} dataset.')
+
+        # start creating the dataset
+        nodesplitter = wds.shardlists.split_by_node if self.multinode else wds.shardlists.single_node_only
+        epoch_length = n_examples // (self.batch_size)
+
+        dset = wds.WebDataset(tars, nodesplitter=nodesplitter).shuffle(shuffle)
+
+        with_epoch_args = {'nsamples': n_examples, 'nbatches': epoch_length}
+
+        if 'filters' in dataset_config:
+            for stage in tqdm(dataset_config.filters,
+                              desc=f'Applying the following filters: {[f for f in dataset_config.filters]}'):
+                f = getattr(dset, stage)
+                dset = f(dset, *dataset_config.filters[stage].args,
+                         **dataset_config.filters[stage].get('kwargs', dict()))
+
+        print(f'Dataset holding {len(dset.pipeline[0].urls)} shards')
+
+        dset = (dset
+                .decode('pil')
+                # .to_tuple("jpg;png;jpeg pickle cls hls")
+                # .map_tuple(image_transforms,load_partial_from_config(nns_transform) if 'target' in nns_transform else identity,identity,identity)
+                .map_dict(**transform_dict)
+                .repeat()
+                .batched(self.batch_size, partial=False,
+                         collation_fn=dict_collation_fn)
+                .with_length(n_examples)
+                .with_epoch(**with_epoch_args)
+                )
+
+        loader = wds.WebLoader(dset, batch_size=None, shuffle=False,
+                               num_workers=self.num_workers)
+
+        return loader, n_examples
+
+    def train_dataloader(self):
+        assert self.train is not None
+        loader, dset_size = self.make_loader(self.train)
+        # if self.load_ddp:
+        #     loader = loader.ddp_equalize(dset_size // self.batch_size)
+        return loader
+
+    def val_dataloader(self):
+        assert self.train is not None
+        loader, _ = self.make_loader(self.validation, train=False)
+        return loader
+
+    def test_dataloader(self):
+        assert self.train is not None
+        loader, _ = self.make_loader(self.test, train=False)
+        return loader
+
 
 if __name__ == "__main__":
     url = "pipe:aws s3 cp s3://s-datasets/laion5b/laion2B-data/000000.tar -"
@@ -18,7 +189,7 @@ if __name__ == "__main__":
     image = Image.open(io.BytesIO(example["jpg"]))
     outdir = "tmp"
     os.makedirs(outdir, exist_ok=True)
-    image.save(os.path.join(outdir, example["__key__"]+".png"))
+    image.save(os.path.join(outdir, example["__key__"] + ".png"))
 
 
     def load_example(example):

ldm/models/diffusion/ddpm.py

@@ -664,7 +664,7 @@ class LatentDiffusion(DDPM):
             if cond_key is None:
                 cond_key = self.cond_stage_key
             if cond_key != self.first_stage_key:
-                if cond_key in ['caption', 'coordinates_bbox']:
+                if cond_key in ['caption', 'coordinates_bbox', "txt"]:
                     xc = batch[cond_key]
                 elif cond_key == 'class_label':
                     xc = batch
@@ -762,66 +762,6 @@ class LatentDiffusion(DDPM):
             else:
                 return self.first_stage_model.decode(z)
 
-    # same as above but without decorator
-    def differentiable_decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
-        if predict_cids:
-            if z.dim() == 4:
-                z = torch.argmax(z.exp(), dim=1).long()
-            z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
-            z = rearrange(z, 'b h w c -> b c h w').contiguous()
-
-        z = 1. / self.scale_factor * z
-
-        if hasattr(self, "split_input_params"):
-            if self.split_input_params["patch_distributed_vq"]:
-                ks = self.split_input_params["ks"]  # eg. (128, 128)
-                stride = self.split_input_params["stride"]  # eg. (64, 64)
-                uf = self.split_input_params["vqf"]
-                bs, nc, h, w = z.shape
-                if ks[0] > h or ks[1] > w:
-                    ks = (min(ks[0], h), min(ks[1], w))
-                    print("reducing Kernel")
-
-                if stride[0] > h or stride[1] > w:
-                    stride = (min(stride[0], h), min(stride[1], w))
-                    print("reducing stride")
-
-                fold, unfold, normalization, weighting = self.get_fold_unfold(z, ks, stride, uf=uf)
-
-                z = unfold(z)  # (bn, nc * prod(**ks), L)
-                # 1. Reshape to img shape
-                z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
-
-                # 2. apply model loop over last dim
-                if isinstance(self.first_stage_model, VQModelInterface):  
-                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i],
-                                                                 force_not_quantize=predict_cids or force_not_quantize)
-                                   for i in range(z.shape[-1])]
-                else:
-
-                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i])
-                                   for i in range(z.shape[-1])]
-
-                o = torch.stack(output_list, axis=-1)  # # (bn, nc, ks[0], ks[1], L)
-                o = o * weighting
-                # Reverse 1. reshape to img shape
-                o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
-                # stitch crops together
-                decoded = fold(o)
-                decoded = decoded / normalization  # norm is shape (1, 1, h, w)
-                return decoded
-            else:
-                if isinstance(self.first_stage_model, VQModelInterface):
-                    return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
-                else:
-                    return self.first_stage_model.decode(z)
-
-        else:
-            if isinstance(self.first_stage_model, VQModelInterface):
-                return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
-            else:
-                return self.first_stage_model.decode(z)
-
     @torch.no_grad()
     def encode_first_stage(self, x):
         if hasattr(self, "split_input_params"):
@@ -1268,8 +1208,8 @@ class LatentDiffusion(DDPM):
             if hasattr(self.cond_stage_model, "decode"):
                 xc = self.cond_stage_model.decode(c)
                 log["conditioning"] = xc
-            elif self.cond_stage_key in ["caption"]:
+            elif self.cond_stage_key in ["caption", "txt"]:
-                xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["caption"])
+                xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key])
                 log["conditioning"] = xc
             elif self.cond_stage_key == 'class_label':
                 xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"])

main.py

@@ -667,8 +667,11 @@ if __name__ == "__main__":
         data.prepare_data()
         data.setup()
         print("#### Data #####")
-        for k in data.datasets:
+        try:
-            print(f"{k}, {data.datasets[k].__class__.__name__}, {len(data.datasets[k])}")
+            for k in data.datasets:
+                print(f"{k}, {data.datasets[k].__class__.__name__}, {len(data.datasets[k])}")
+        except:
+            print("datasets not yet initialized.")
 
         # configure learning rate
         bs, base_lr = config.data.params.batch_size, config.model.base_learning_rate