bring in the dynamic thresholding technique from the Imagen paper, which purportedly improves classifier free guidance for the cascading ddpm

block in a section that links to available pre-trained models for those who are interested

README.md

@@ -12,7 +12,7 @@ This model is SOTA for text-to-image for now.
 
 Please join <a href="https://discord.gg/xBPBXfcFHd"><img alt="Join us on Discord" src="https://img.shields.io/discord/823813159592001537?color=5865F2&logo=discord&logoColor=white"></a> if you are interested in helping out with the replication with the <a href="https://laion.ai/">LAION</a> community | <a href="https://www.youtube.com/watch?v=AIOE1l1W0Tw">Yannic Interview</a>
 
-There was enough interest for a <a href="https://github.com/lucidrains/dalle2-jax">Jax version</a>. I will also eventually extend this to <a href="https://github.com/lucidrains/dalle2-video">text to video</a>, once the repository is in a good place.
+As of 5/23/22, it is no longer SOTA. SOTA will be <a href="https://github.com/lucidrains/imagen-pytorch">here</a>. Jax versions as well as text-to-video project will be shifted towards the Imagen architecture, as it is way simpler.
 
 ## Status
 
@@ -24,6 +24,11 @@ There was enough interest for a <a href="https://github.com/lucidrains/dalle2-ja
 
 *ongoing at 21k steps*
 
+## Pre-Trained Models
+- LAION is training prior models. Checkpoints are available on <a href="https://huggingface.co/zenglishuci/conditioned-prior">🤗huggingface</a> and the training statistics are available on <a href="https://wandb.ai/nousr_laion/conditioned-prior/reports/LAION-DALLE2-PyTorch-Prior--VmlldzoyMDI2OTIx">🐝WANDB</a>.
+- Decoder - <a href="https://wandb.ai/veldrovive/dalle2_train_decoder/runs/jkrtg0so?workspace=user-veldrovive">In-progress test run</a> 🚧
+- DALL-E 2 🚧
+
 ## Install
 
 ```bash
@@ -1079,6 +1084,7 @@ This library would not have gotten to this working state without the help of
 - [x] use pydantic for config drive training
 - [x] for both diffusion prior and decoder, all exponential moving averaged models needs to be saved and restored as well (as well as the step number)
 - [x] offer save / load methods on the trainer classes to automatically take care of state dicts for scalers / optimizers / saving versions and checking for breaking changes
+- [x] allow for creation of diffusion prior model off pydantic config classes - consider the same for tracker configs
 - [ ] become an expert with unets, cleanup unet code, make it fully configurable, port all learnings over to https://github.com/lucidrains/x-unet (test out unet² in ddpm repo) - consider https://github.com/lucidrains/uformer-pytorch attention-based unet
 - [ ] transcribe code to Jax, which lowers the activation energy for distributed training, given access to TPUs
 - [ ] train on a toy task, offer in colab
@@ -1189,4 +1195,12 @@ This library would not have gotten to this working state without the help of
 }
 ```
 
+```bibtex
+@misc{Saharia2022,
+    title   = {Imagen: unprecedented photorealism × deep level of language understanding},
+    author  = {Chitwan Saharia*, William Chan*, Saurabh Saxena†, Lala Li†, Jay Whang†, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho†, David Fleet†, Mohammad Norouzi*},
+    year    = {2022}
+}
+```
+
 *Creating noise from data is easy; creating data from noise is generative modeling.* - <a href="https://arxiv.org/abs/2011.13456">Yang Song's paper</a>

configs/README.md

@@ -6,9 +6,10 @@ For more complex configuration, we provide the option of using a configuration f
 
 The decoder trainer has 7 main configuration options. A full example of their use can be found in the [example decoder configuration](train_decoder_config.example.json).
 
-**<ins>Unets</ins>:**
+**<ins>Unet</ins>:**
+
+This is a single unet config, which belongs as an array nested under the decoder config as a list of `unets`
 
-Each member of this array defines a single unet that will be added to the decoder.
 | Option | Required | Default | Description |
 | ------ | -------- | ------- | ----------- |
 | `dim`  | Yes      | N/A     | The starting channels of the unet. |
@@ -22,6 +23,7 @@ Any parameter from the `Unet` constructor can also be given here.
 Defines the configuration options for the decoder model. The unets defined above will automatically be inserted.
 | Option | Required | Default | Description |
 | ------ | -------- | ------- | ----------- |
+| `unets` | Yes | N/A | A list of unets, using the configuration above |
 | `image_sizes` | Yes | N/A | The resolution of the image after each upsampling step. The length of this array should be the number of unets defined. |
 | `image_size` | Yes | N/A | Not used. Can be any number. |
 | `timesteps` | No | `1000` | The number of diffusion timesteps used for generation. |

configs/train_decoder_config.example.json

@@ -1,16 +1,16 @@
 {
-    "unets": [
-        {
-            "dim": 128,
-            "image_embed_dim": 768,
-            "cond_dim": 64,
-            "channels": 3,
-            "dim_mults": [1, 2, 4, 8],
-            "attn_dim_head": 32,
-            "attn_heads": 16
-        }
-    ],
     "decoder": {
+        "unets": [
+            {
+                "dim": 128,
+                "image_embed_dim": 768,
+                "cond_dim": 64,
+                "channels": 3,
+                "dim_mults": [1, 2, 4, 8],
+                "attn_dim_head": 32,
+                "attn_heads": 16
+            }
+        ],
         "image_sizes": [64],
         "channels": 3,
         "timesteps": 1000,

dalle2_pytorch/dalle2_pytorch.py

@@ -890,6 +890,8 @@ class DiffusionPrior(BaseGaussianDiffusion):
         )
 
         if exists(clip):
+            assert image_channels == clip.image_channels, f'channels of image ({image_channels}) should be equal to the channels that CLIP accepts ({clip.image_channels})'
+
             if isinstance(clip, CLIP):
                 clip = XClipAdapter(clip, **clip_adapter_overrides)
             elif isinstance(clip, CoCa):
@@ -1702,6 +1704,8 @@ class Decoder(BaseGaussianDiffusion):
         vb_loss_weight = 0.001,
         unconditional = False,
         auto_normalize_img = True,                  # whether to take care of normalizing the image from [0, 1] to [-1, 1] and back automatically - you can turn this off if you want to pass in the [-1, 1] ranged image yourself from the dataloader
+        use_dynamic_thres = False,                  # from the Imagen paper
+        dynamic_thres_percentile = 0.9
     ):
         super().__init__(
             beta_schedule = beta_schedule,
@@ -1710,12 +1714,19 @@ class Decoder(BaseGaussianDiffusion):
         )
 
         self.unconditional = unconditional
-        assert not (condition_on_text_encodings and unconditional), 'unconditional decoder image generation cannot be set to True if conditioning on text is present'
 
-        assert self.unconditional or (exists(clip) ^ exists(image_size)), 'either CLIP is supplied, or you must give the image_size and channels (usually 3 for RGB)'
+        # text conditioning
+
+        assert not (condition_on_text_encodings and unconditional), 'unconditional decoder image generation cannot be set to True if conditioning on text is present'
+        self.condition_on_text_encodings = condition_on_text_encodings
+
+        # clip
 
         self.clip = None
         if exists(clip):
+            assert not unconditional, 'clip must not be given if doing unconditional image training'
+            assert channels == clip.image_channels, f'channels of image ({channels}) should be equal to the channels that CLIP accepts ({clip.image_channels})'
+
             if isinstance(clip, CLIP):
                 clip = XClipAdapter(clip, **clip_adapter_overrides)
             elif isinstance(clip, CoCa):
@@ -1725,13 +1736,20 @@ class Decoder(BaseGaussianDiffusion):
             assert isinstance(clip, BaseClipAdapter)
 
             self.clip = clip
-            self.clip_image_size = clip.image_size
-            self.channels = clip.image_channels
-        else:
-            self.clip_image_size = image_size
-            self.channels = channels
 
-        self.condition_on_text_encodings = condition_on_text_encodings
+        # determine image size, with image_size and image_sizes taking precedence
+
+        if exists(image_size) or exists(image_sizes):
+            assert exists(image_size) ^ exists(image_sizes), 'only one of image_size or image_sizes must be given'
+            image_size = default(image_size, lambda: image_sizes[-1])
+        elif exists(clip):
+            image_size = clip.image_size
+        else:
+            raise Error('either image_size, image_sizes, or clip must be given to decoder')
+
+        # channels
+
+        self.channels = channels
 
         # automatically take care of ensuring that first unet is unconditional
         # while the rest of the unets are conditioned on the low resolution image produced by previous unet
@@ -1773,7 +1791,7 @@ class Decoder(BaseGaussianDiffusion):
 
         # unet image sizes
 
-        image_sizes = default(image_sizes, (self.clip_image_size,))
+        image_sizes = default(image_sizes, (image_size,))
         image_sizes = tuple(sorted(set(image_sizes)))
 
         assert len(self.unets) == len(image_sizes), f'you did not supply the correct number of u-nets ({len(self.unets)}) for resolutions {image_sizes}'
@@ -1810,7 +1828,13 @@ class Decoder(BaseGaussianDiffusion):
         self.clip_denoised = clip_denoised
         self.clip_x_start = clip_x_start
 
+        # dynamic thresholding settings, if clipping denoised during sampling
+
+        self.use_dynamic_thres = use_dynamic_thres
+        self.dynamic_thres_percentile = dynamic_thres_percentile
+
         # normalize and unnormalize image functions
+
         self.normalize_img = normalize_neg_one_to_one if auto_normalize_img else identity
         self.unnormalize_img = unnormalize_zero_to_one if auto_normalize_img else identity
 
@@ -1851,7 +1875,21 @@ class Decoder(BaseGaussianDiffusion):
             x_recon = self.predict_start_from_noise(x, t = t, noise = pred)
 
         if clip_denoised:
-            x_recon.clamp_(-1., 1.)
+            # s is the threshold amount
+            # static thresholding would just be s = 1
+            s = 1.
+            if self.use_dynamic_thres:
+                s = torch.quantile(
+                    rearrange(x_recon, 'b ... -> b (...)').abs(),
+                    self.dynamic_thres_percentile,
+                    dim = -1
+                )
+
+                s.clamp_(min = 1.)
+                s = s.view(-1, *((1,) * (x_recon.ndim - 1)))
+
+            # clip by threshold, depending on whether static or dynamic
+            x_recon = x_recon.clamp(-s, s) / s
 
         model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
 

dalle2_pytorch/train_configs.py

@@ -3,15 +3,61 @@ from torchvision import transforms as T
 from pydantic import BaseModel, validator, root_validator
 from typing import List, Iterable, Optional, Union, Tuple, Dict, Any
 
+from dalle2_pytorch.dalle2_pytorch import Unet, Decoder, DiffusionPrior, DiffusionPriorNetwork
+
+# helper functions
+
 def exists(val):
     return val is not None
 
 def default(val, d):
     return val if exists(val) else d
 
+def ListOrTuple(inner_type):
+    return Union[List[inner_type], Tuple[inner_type]]
+
+# pydantic classes
+
+class DiffusionPriorNetworkConfig(BaseModel):
+    dim: int
+    depth: int
+    num_timesteps: int = None
+    num_time_embeds: int = 1
+    num_image_embeds: int = 1
+    num_text_embeds: int = 1
+    dim_head: int = 64
+    heads: int = 8
+    ff_mult: int = 4
+    norm_out: bool = True
+    attn_dropout: float = 0.
+    ff_dropout: float = 0.
+    final_proj: bool = True
+    normformer: bool = False
+    rotary_emb: bool = True
+
+class DiffusionPriorConfig(BaseModel):
+    # only clip-less diffusion prior config for now
+    net: DiffusionPriorNetworkConfig
+    image_embed_dim: int
+    image_size: int
+    image_channels: int = 3
+    timesteps: int = 1000
+    cond_drop_prob: float = 0.
+    loss_type: str = 'l2'
+    predict_x_start: bool = True
+    beta_schedule: str = 'cosine'
+
+    def create(self):
+        kwargs = self.dict()
+        diffusion_prior_network = DiffusionPriorNetwork(**kwargs.pop('net'))
+        return DiffusionPrior(net = diffusion_prior_network, **kwargs)
+
+    class Config:
+        extra = "allow"
+
 class UnetConfig(BaseModel):
     dim: int
-    dim_mults: List[int]
+    dim_mults: ListOrTuple(int)
     image_embed_dim: int = None
     cond_dim: int = None
     channels: int = 3
@@ -22,13 +68,22 @@ class UnetConfig(BaseModel):
         extra = "allow"
 
 class DecoderConfig(BaseModel):
+    unets: ListOrTuple(UnetConfig)
     image_size: int = None
-    image_sizes: Union[List[int], Tuple[int]] = None
+    image_sizes: ListOrTuple(int) = None
     channels: int = 3
     timesteps: int = 1000
     loss_type: str = 'l2'
     beta_schedule: str = 'cosine'
     learned_variance: bool = True
+    image_cond_drop_prob: float = 0.1
+    text_cond_drop_prob: float = 0.5
+
+    def create(self):
+        decoder_kwargs = self.dict()
+        unet_configs = decoder_kwargs.pop('unets')
+        unets = [Unet(**config) for config in unet_configs]
+        return Decoder(unets, **decoder_kwargs)
 
     @validator('image_sizes')
     def check_image_sizes(cls, image_sizes, values):
@@ -86,17 +141,17 @@ class DecoderTrainConfig(BaseModel):
     wd: float = 0.01
     max_grad_norm: float = 0.5
     save_every_n_samples: int = 100000
-    n_sample_images: int = 6                # The number of example images to produce when sampling the train and test dataset
+    n_sample_images: int = 6                       # The number of example images to produce when sampling the train and test dataset
     device: str = 'cuda:0'
-    epoch_samples: int = None               # Limits the number of samples per epoch. None means no limit. Required if resample_train is true as otherwise the number of samples per epoch is infinite.
+    epoch_samples: int = None                      # Limits the number of samples per epoch. None means no limit. Required if resample_train is true as otherwise the number of samples per epoch is infinite.
-    validation_samples: int = None          # Same as above but for validation.
+    validation_samples: int = None                 # Same as above but for validation.
     use_ema: bool = True
     ema_beta: float = 0.99
     amp: bool = False
-    save_all: bool = False                  # Whether to preserve all checkpoints
+    save_all: bool = False                         # Whether to preserve all checkpoints
-    save_latest: bool = True                # Whether to always save the latest checkpoint
+    save_latest: bool = True                       # Whether to always save the latest checkpoint
-    save_best: bool = True                  # Whether to save the best checkpoint
+    save_best: bool = True                         # Whether to save the best checkpoint
-    unet_training_mask: List[bool] = None   # If None, use all unets
+    unet_training_mask: ListOrTuple(bool) = None   # If None, use all unets
 
 class DecoderEvaluateConfig(BaseModel):
     n_evaluation_samples: int = 1000
@@ -120,7 +175,6 @@ class DecoderLoadConfig(BaseModel):
     resume: bool = False                    # If using wandb, whether to resume the run
 
 class TrainDecoderConfig(BaseModel):
-    unets: List[UnetConfig]
     decoder: DecoderConfig
     data: DecoderDataConfig
     train: DecoderTrainConfig

dalle2_pytorch/trainer.py

@@ -288,7 +288,7 @@ class DiffusionPriorTrainer(nn.Module):
 
         self.register_buffer('step', torch.tensor([0]))
 
-    def save(self, path, overwrite = True):
+    def save(self, path, overwrite = True, **kwargs):
         path = Path(path)
         assert not (path.exists() and not overwrite)
         path.parent.mkdir(parents = True, exist_ok = True)
@@ -298,7 +298,8 @@ class DiffusionPriorTrainer(nn.Module):
             optimizer = self.optimizer.state_dict(),
             model = self.diffusion_prior.state_dict(),
             version = get_pkg_version(),
-            step = self.step.item()
+            step = self.step.item(),
+            **kwargs
         )
 
         if self.use_ema:
@@ -319,7 +320,7 @@ class DiffusionPriorTrainer(nn.Module):
         self.step.copy_(torch.ones_like(self.step) * loaded_obj['step'])
 
         if only_model:
-            return
+            return loaded_obj
 
         self.scaler.load_state_dict(loaded_obj['scaler'])
         self.optimizer.load_state_dict(loaded_obj['optimizer'])
@@ -328,6 +329,8 @@ class DiffusionPriorTrainer(nn.Module):
             assert 'ema' in loaded_obj
             self.ema_diffusion_prior.load_state_dict(loaded_obj['ema'], strict = strict)
 
+        return loaded_obj
+
     def update(self):
         if exists(self.max_grad_norm):
             self.scaler.unscale_(self.optimizer)
@@ -449,7 +452,7 @@ class DecoderTrainer(nn.Module):
 
         self.register_buffer('step', torch.tensor([0.]))
 
-    def save(self, path, overwrite = True):
+    def save(self, path, overwrite = True, **kwargs):
         path = Path(path)
         assert not (path.exists() and not overwrite)
         path.parent.mkdir(parents = True, exist_ok = True)
@@ -457,7 +460,8 @@ class DecoderTrainer(nn.Module):
         save_obj = dict(
             model = self.decoder.state_dict(),
             version = get_pkg_version(),
-            step = self.step.item()
+            step = self.step.item(),
+            **kwargs
         )
 
         for ind in range(0, self.num_unets):
@@ -485,7 +489,7 @@ class DecoderTrainer(nn.Module):
         self.step.copy_(torch.ones_like(self.step) * loaded_obj['step'])
 
         if only_model:
-            return
+            return loaded_obj
 
         for ind in range(0, self.num_unets):
             scaler_key = f'scaler{ind}'
@@ -500,6 +504,8 @@ class DecoderTrainer(nn.Module):
             assert 'ema' in loaded_obj
             self.ema_unets.load_state_dict(loaded_obj['ema'], strict = strict)
 
+        return loaded_obj
+
     @property
     def unets(self):
         return nn.ModuleList([ema.ema_model for ema in self.ema_unets])

setup.py

@@ -10,7 +10,7 @@ setup(
       'dream = dalle2_pytorch.cli:dream'
     ],
   },
-  version = '0.4.7',
+  version = '0.5.0',
   license='MIT',
   description = 'DALL-E 2',
   author = 'Phil Wang',

train_decoder.py

@@ -85,20 +85,6 @@ def create_dataloaders(
         "test_sampling": test_sampling_dataloader
     }
 
-
-def create_decoder(device, decoder_config, unets_config):
-    """Creates a sample decoder"""
-
-    unets = [Unet(**config.dict()) for config in unets_config]
-
-    decoder = Decoder(
-        unet=unets,
-        **decoder_config.dict()
-    )
-
-    decoder.to(device=device)
-    return decoder
-
 def get_dataset_keys(dataloader):
     """
     It is sometimes neccesary to get the keys the dataloader is returning. Since the dataset is burried in the dataloader, we need to do a process to recover it.
@@ -428,7 +414,7 @@ def initialize_training(config):
         **config.data.dict()
     )
 
-    decoder = create_decoder(device, config.decoder, config.unets)
+    decoder = config.decoder.create().to(device = device)
     num_parameters = sum(p.numel() for p in decoder.parameters())
     print(print_ribbon("Loaded Config", repeat=40))
     print(f"Number of parameters: {num_parameters}")