make sure diffusion prior trainer can operate with no warmup

README.md

@@ -371,6 +371,7 @@ loss.backward()
 unet1 = Unet(
     dim = 128,
     image_embed_dim = 512,
+    text_embed_dim = 512,
     cond_dim = 128,
     channels = 3,
     dim_mults=(1, 2, 4, 8),
@@ -395,7 +396,7 @@ decoder = Decoder(
 ).cuda()
 
 for unet_number in (1, 2):
-    loss = decoder(images, unet_number = unet_number) # this can optionally be decoder(images, text) if you wish to condition on the text encodings as well, though it was hinted in the paper it didn't do much
+    loss = decoder(images, text = text, unet_number = unet_number) # this can optionally be decoder(images, text) if you wish to condition on the text encodings as well, though it was hinted in the paper it didn't do much
     loss.backward()
 
 # do above for many steps
@@ -626,6 +627,18 @@ images = dalle2(
 # save your image (in this example, of size 256x256)
 ```
 
+Alternatively, you can also use <a href="https://github.com/mlfoundations/open_clip">Open Clip</a>
+
+```bash
+$ pip install open-clip-torch
+```
+
+```python
+from dalle2_pytorch import OpenClipAdapter
+
+clip = OpenClipAdapter()
+```
+
 Now you'll just have to worry about training the Prior and the Decoder!
 
 ## Inpainting
@@ -860,25 +873,23 @@ unet1 = Unet(
     text_embed_dim = 512,
     cond_dim = 128,
     channels = 3,
-    dim_mults=(1, 2, 4, 8)
+    dim_mults=(1, 2, 4, 8),
+    cond_on_text_encodings = True,
 ).cuda()
 
 unet2 = Unet(
     dim = 16,
     image_embed_dim = 512,
-    text_embed_dim = 512,
     cond_dim = 128,
     channels = 3,
     dim_mults = (1, 2, 4, 8, 16),
-    cond_on_text_encodings = True
 ).cuda()
 
 decoder = Decoder(
     unet = (unet1, unet2),
     image_sizes = (128, 256),
     clip = clip,
-    timesteps = 1000,
+    timesteps = 1000
-    condition_on_text_encodings = True
 ).cuda()
 
 decoder_trainer = DecoderTrainer(
@@ -903,8 +914,8 @@ for unet_number in (1, 2):
 # after much training
 # you can sample from the exponentially moving averaged unets as so
 
-mock_image_embed = torch.randn(4, 512).cuda()
+mock_image_embed = torch.randn(32, 512).cuda()
-images = decoder_trainer.sample(mock_image_embed, text = text) # (4, 3, 256, 256)
+images = decoder_trainer.sample(image_embed = mock_image_embed, text = text) # (4, 3, 256, 256)
 ```
 
 ### Diffusion Prior Training
@@ -1112,7 +1123,8 @@ For detailed information on training the diffusion prior, please refer to the [d
 - [x] allow for unet to be able to condition non-cross attention style as well
 - [x] speed up inference, read up on papers (ddim)
 - [x] add inpainting ability using resampler from repaint paper https://arxiv.org/abs/2201.09865
-- [ ] try out the nested unet from https://arxiv.org/abs/2005.09007 after hearing several positive testimonies from researchers, for segmentation anyhow
+- [x] add the final combination of upsample feature maps, used in unet squared, seems to have an effect in local experiments
+- [ ] consider elucidated dalle2 https://arxiv.org/abs/2206.00364
 - [ ] interface out the vqgan-vae so a pretrained one can be pulled off the shelf to validate latent diffusion + DALL-E2
 
 ## Citations
@@ -1241,4 +1253,25 @@ For detailed information on training the diffusion prior, please refer to the [d
 }
 ```
 
+```bibtex
+@misc{chen2022analog,
+    title   = {Analog Bits: Generating Discrete Data using Diffusion Models with Self-Conditioning},
+    author  = {Ting Chen and Ruixiang Zhang and Geoffrey Hinton},
+    year    = {2022},
+    eprint  = {2208.04202},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.CV}
+}
+```
+
+```bibtex
+@article{Qiao2019WeightS,
+    title   = {Weight Standardization},
+    author  = {Siyuan Qiao and Huiyu Wang and Chenxi Liu and Wei Shen and Alan Loddon Yuille},
+    journal = {ArXiv},
+    year    = {2019},
+    volume  = {abs/1903.10520}
+}
+```
+
 *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>

dalle2_pytorch/trainer.py

@@ -9,7 +9,7 @@ from collections.abc import Iterable
 import torch
 import torch.nn.functional as F
 from torch import nn
-from torch.optim.lr_scheduler import LambdaLR
+from torch.optim.lr_scheduler import LambdaLR, CosineAnnealingLR
 from torch.cuda.amp import autocast, GradScaler
 
 from dalle2_pytorch.dalle2_pytorch import Decoder, DiffusionPrior
@@ -174,20 +174,25 @@ class DiffusionPriorTrainer(nn.Module):
     def __init__(
         self,
         diffusion_prior,
-        accelerator,
+        accelerator = None,
         use_ema = True,
         lr = 3e-4,
         wd = 1e-2,
         eps = 1e-6,
         max_grad_norm = None,
         group_wd_params = True,
-        warmup_steps = 1,
+        warmup_steps = None,
+        cosine_decay_max_steps = None,
         **kwargs
     ):
         super().__init__()
         assert isinstance(diffusion_prior, DiffusionPrior)
-        assert isinstance(accelerator, Accelerator)
+
         ema_kwargs, kwargs = groupby_prefix_and_trim('ema_', kwargs)
+        accelerator_kwargs, kwargs = groupby_prefix_and_trim('accelerator_', kwargs)
+
+        if not exists(accelerator):
+            accelerator = Accelerator(**accelerator_kwargs)
 
         # assign some helpful member vars
 
@@ -229,8 +234,11 @@ class DiffusionPriorTrainer(nn.Module):
             **self.optim_kwargs,
             **kwargs
         )
-        
+
-        self.scheduler = LambdaLR(self.optimizer, lr_lambda = lambda _: 1.0)
+        if exists(cosine_decay_max_steps):
+            self.scheduler = CosineAnnealingLR(optimizer, T_max = cosine_decay_max_steps)
+        else:
+            self.scheduler = LambdaLR(self.optimizer, lr_lambda = lambda _: 1.0)
         
         self.warmup_scheduler = warmup.LinearWarmup(self.optimizer, warmup_period = warmup_steps) if exists(warmup_steps) else None
 
@@ -267,6 +275,7 @@ class DiffusionPriorTrainer(nn.Module):
             # FIXME: LambdaLR can't be saved due to pickling issues
             save_obj = dict(
                 optimizer = self.optimizer.state_dict(),
+                scheduler = self.scheduler.state_dict(),
                 warmup_scheduler = self.warmup_scheduler,
                 model = self.accelerator.unwrap_model(self.diffusion_prior).state_dict(),
                 version = version.parse(__version__),
@@ -300,7 +309,7 @@ class DiffusionPriorTrainer(nn.Module):
 
         # all processes need to load checkpoint. no restriction here
         if isinstance(path_or_state, str):
-            path = Path(path)
+            path = Path(path_or_state)
             assert path.exists()
             loaded_obj = torch.load(str(path), map_location=self.device)
 
@@ -313,7 +322,9 @@ class DiffusionPriorTrainer(nn.Module):
         # unwrap the model when loading from checkpoint
         self.accelerator.unwrap_model(self.diffusion_prior).load_state_dict(loaded_obj['model'], strict = strict)
         self.step.copy_(torch.ones_like(self.step, device=self.device) * loaded_obj['step'].to(self.device))
+
         self.optimizer.load_state_dict(loaded_obj['optimizer'])
+        self.scheduler.load_state_dict(loaded_obj['scheduler'])
 
         # set warmupstep
         if exists(self.warmup_scheduler):
@@ -346,7 +357,8 @@ class DiffusionPriorTrainer(nn.Module):
 
         # accelerator will ocassionally skip optimizer steps in a "dynamic loss scaling strategy"
         if not self.accelerator.optimizer_step_was_skipped:
-            with self.warmup_scheduler.dampening():
+            sched_context = self.warmup_scheduler.dampening if exists(self.warmup_scheduler) else nullcontext
+            with sched_context():
                 self.scheduler.step()
 
         if self.use_ema:
@@ -429,6 +441,7 @@ class DecoderTrainer(nn.Module):
         wd = 1e-2,
         eps = 1e-8,
         warmup_steps = None,
+        cosine_decay_max_steps = None,
         max_grad_norm = 0.5,
         amp = False,
         group_wd_params = True,
@@ -450,7 +463,7 @@ class DecoderTrainer(nn.Module):
         # be able to finely customize learning rate, weight decay
         # per unet
 
-        lr, wd, eps, warmup_steps = map(partial(cast_tuple, length = self.num_unets), (lr, wd, eps, warmup_steps))
+        lr, wd, eps, warmup_steps, cosine_decay_max_steps = map(partial(cast_tuple, length = self.num_unets), (lr, wd, eps, warmup_steps, cosine_decay_max_steps))
 
         assert all([unet_lr <= 1e-2 for unet_lr in lr]), 'your learning rate is too high, recommend sticking with 1e-4, at most 5e-4'
 
@@ -458,7 +471,7 @@ class DecoderTrainer(nn.Module):
         schedulers = []
         warmup_schedulers = []
 
-        for unet, unet_lr, unet_wd, unet_eps, unet_warmup_steps in zip(decoder.unets, lr, wd, eps, warmup_steps):
+        for unet, unet_lr, unet_wd, unet_eps, unet_warmup_steps, unet_cosine_decay_max_steps in zip(decoder.unets, lr, wd, eps, warmup_steps, cosine_decay_max_steps):
             if isinstance(unet, nn.Identity):
                 optimizers.append(None)
                 schedulers.append(None)
@@ -474,7 +487,11 @@ class DecoderTrainer(nn.Module):
                 )
 
                 optimizers.append(optimizer)
-                scheduler = LambdaLR(optimizer, lr_lambda = lambda step: 1.0)
+
+                if exists(unet_cosine_decay_max_steps):
+                    scheduler = CosineAnnealingLR(optimizer, T_max = unet_cosine_decay_max_steps)
+                else:
+                    scheduler = LambdaLR(optimizer, lr_lambda = lambda step: 1.0)
 
                 warmup_scheduler = warmup.LinearWarmup(optimizer, warmup_period = unet_warmup_steps) if exists(unet_warmup_steps) else None
                 warmup_schedulers.append(warmup_scheduler)
@@ -554,9 +571,15 @@ class DecoderTrainer(nn.Module):
 
         for ind in range(0, self.num_unets):
             optimizer_key = f'optim{ind}'
+            scheduler_key = f'sched{ind}'
+
             optimizer = getattr(self, optimizer_key)
-            state_dict = optimizer.state_dict() if optimizer is not None else None
+            scheduler = getattr(self, scheduler_key)
-            save_obj = {**save_obj, optimizer_key: state_dict}
+
+            optimizer_state_dict = optimizer.state_dict() if exists(optimizer) else None
+            scheduler_state_dict = scheduler.state_dict() if exists(scheduler) else None
+
+            save_obj = {**save_obj, optimizer_key: optimizer_state_dict, scheduler_key: scheduler_state_dict}
 
         if self.use_ema:
             save_obj = {**save_obj, 'ema': self.ema_unets.state_dict()}
@@ -577,10 +600,18 @@ class DecoderTrainer(nn.Module):
 
             optimizer_key = f'optim{ind}'
             optimizer = getattr(self, optimizer_key)
+
+            scheduler_key = f'sched{ind}'
+            scheduler = getattr(self, scheduler_key)
+
             warmup_scheduler = self.warmup_schedulers[ind]
-            if optimizer is not None:
+
+            if exists(optimizer):
                 optimizer.load_state_dict(loaded_obj[optimizer_key])
 
+            if exists(scheduler):
+                scheduler.load_state_dict(loaded_obj[scheduler_key])
+
             if exists(warmup_scheduler):
                 warmup_scheduler.last_step = last_step
 

dalle2_pytorch/version.py

@@ -1 +1 @@
-__version__ = '1.0.3'
+__version__ = '1.8.2'

setup.py

@@ -26,7 +26,7 @@ setup(
   install_requires=[
     'accelerate',
     'click',
-    'clip-anytorch',
+    'clip-anytorch>=2.4.0',
     'coca-pytorch>=0.0.5',
     'ema-pytorch>=0.0.7',
     'einops>=0.4',