starling.models.diffusion.DiffusionModel

class DiffusionModel

Bases: LightningModule

Denoising diffusion probabilistic model for latent space generation.

Implements the diffusion process described in: - Sohl-Dickstein et al. (2015): Nonequilibrium Thermodynamics - Ho et al. (2020): Denoising Diffusion Probabilistic Models - Rombach et al. (2021): High-resolution image synthesis with latent diffusion

Methods

__init__	A discrete-time denoising-diffusion model framework for latent space diffusion models.
add_module	Add a child module to the current module.
all_gather	Gather tensors or collections of tensors from multiple processes.
apply	Apply fn recursively to every submodule (as returned by .children()) as well as self.
backward	Called to perform backward on the loss returned in training_step().
bfloat16	Casts all floating point parameters and buffers to bfloat16 datatype.
buffers	Return an iterator over module buffers.
children	Return an iterator over immediate children modules.
clip_gradients	Handles gradient clipping internally.
compile	Compile this Module's forward using torch.compile().
compute_snr	Computes SNR as per https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849
configure_callbacks	Configure model-specific callbacks.
configure_gradient_clipping	Perform gradient clipping for the optimizer parameters.
configure_model	Hook to create modules in a strategy and precision aware context.
configure_optimizers	Configure the optimizer and the learning rate scheduler for the model.
configure_sharded_model	Deprecated.
cpu	See torch.nn.Module.cpu().
cuda	Moves all model parameters and buffers to the GPU.
double	See torch.nn.Module.double().
eval	Set the module in evaluation mode.
extra_repr	Return the extra representation of the module.
float	See torch.nn.Module.float().
forward	Forward pass of the model, calculates the loss based on the predicted noise and the actual noise.
freeze	Freeze all params for inference.
get_buffer	Return the buffer given by target if it exists, otherwise throw an error.
get_extra_state	Return any extra state to include in the module's state_dict.
get_parameter	Return the parameter given by target if it exists, otherwise throw an error.
get_submodule	Return the submodule given by target if it exists, otherwise throw an error.
half	See torch.nn.Module.half().
ipu	Move all model parameters and buffers to the IPU.
load_from_checkpoint	Primary way of loading a model from a checkpoint.
load_state_dict	Copy parameters and buffers from state_dict into this module and its descendants.
log	Log a key, value pair.
log_dict	Log a dictionary of values at once.
lr_scheduler_step	Override this method to adjust the default way the Trainer calls each scheduler.
lr_schedulers	Returns the learning rate scheduler(s) that are being used during training.
manual_backward	Call this directly from your training_step() when doing optimizations manually.
modules	Return an iterator over all modules in the network.
mtia	Move all model parameters and buffers to the MTIA.
named_buffers	Return an iterator over module buffers, yielding both the name of the buffer as well as the buffer itself.
named_children	Return an iterator over immediate children modules, yielding both the name of the module as well as the module itself.
named_modules	Return an iterator over all modules in the network, yielding both the name of the module as well as the module itself.
named_parameters	Return an iterator over module parameters, yielding both the name of the parameter as well as the parameter itself.
on_after_backward	Called after loss.backward() and before optimizers are stepped.
on_after_batch_transfer	Override to alter or apply batch augmentations to your batch after it is transferred to the device.
on_before_backward	Called before loss.backward().
on_before_batch_transfer	Override to alter or apply batch augmentations to your batch before it is transferred to the device.
on_before_optimizer_step	Called before optimizer.step().
on_before_zero_grad	Called after training_step() and before optimizer.zero_grad().
on_fit_end	Called at the very end of fit.
on_fit_start	Called at the very beginning of fit.
on_load_checkpoint	Called by Lightning to restore your model.
on_predict_batch_end	Called in the predict loop after the batch.
on_predict_batch_start	Called in the predict loop before anything happens for that batch.
on_predict_end	Called at the end of predicting.
on_predict_epoch_end	Called at the end of predicting.
on_predict_epoch_start	Called at the beginning of predicting.
on_predict_model_eval	Called when the predict loop starts.
on_predict_start	Called at the beginning of predicting.
on_save_checkpoint	Called by Lightning when saving a checkpoint to give you a chance to store anything else you might want to save.
on_test_batch_end	Called in the test loop after the batch.
on_test_batch_start	Called in the test loop before anything happens for that batch.
on_test_end	Called at the end of testing.
on_test_epoch_end	Called in the test loop at the very end of the epoch.
on_test_epoch_start	Called in the test loop at the very beginning of the epoch.
on_test_model_eval	Called when the test loop starts.
on_test_model_train	Called when the test loop ends.
on_test_start	Called at the beginning of testing.
on_train_batch_end	Called in the training loop after the batch.
on_train_batch_start	Called in the training loop before anything happens for that batch.
on_train_end	Called at the end of training before logger experiment is closed.
on_train_epoch_end	Called in the training loop at the very end of the epoch.
on_train_epoch_start	Called in the training loop at the very beginning of the epoch.
on_train_start	Called at the beginning of training after sanity check.
on_validation_batch_end	Called in the validation loop after the batch.
on_validation_batch_start	Called in the validation loop before anything happens for that batch.
on_validation_end	Called at the end of validation.
on_validation_epoch_end	Called in the validation loop at the very end of the epoch.
on_validation_epoch_start	Called in the validation loop at the very beginning of the epoch.
on_validation_model_eval	Called when the validation loop starts.
on_validation_model_train	Called when the validation loop ends.
on_validation_model_zero_grad	Called by the training loop to release gradients before entering the validation loop.
on_validation_start	Called at the beginning of validation.
optimizer_step	Override this method to adjust the default way the Trainer calls the optimizer.
optimizer_zero_grad	Override this method to change the default behaviour of optimizer.zero_grad().
optimizers	Returns the optimizer(s) that are being used during training.
p_loss	A function that runs the model and calculates the loss based on the predicted noise and the actual noise.
parameters	Return an iterator over module parameters.
predict_dataloader	An iterable or collection of iterables specifying prediction samples.
predict_step	Step function called during predict().
prepare_data	Use this to download and prepare data.
print	Prints only from process 0.
q_sample	Add the noise to x_start tensor based on the timestamp t
register_backward_hook	Register a backward hook on the module.
register_buffer	Add a buffer to the module.
register_forward_hook	Register a forward hook on the module.
register_forward_pre_hook	Register a forward pre-hook on the module.
register_full_backward_hook	Register a backward hook on the module.
register_full_backward_pre_hook	Register a backward pre-hook on the module.
register_load_state_dict_post_hook	Register a post-hook to be run after module's load_state_dict() is called.
register_load_state_dict_pre_hook	Register a pre-hook to be run before module's load_state_dict() is called.
register_module	Alias for add_module().
register_parameter	Add a parameter to the module.
register_state_dict_post_hook	Register a post-hook for the state_dict() method.
register_state_dict_pre_hook	Register a pre-hook for the state_dict() method.
requires_grad_	Change if autograd should record operations on parameters in this module.
save_hyperparameters	Save arguments to hparams attribute.
sequence2labels	Converts sequences to labels based on user defined models,
set_extra_state	Set extra state contained in the loaded state_dict.
set_submodule	Set the submodule given by target if it exists, otherwise throw an error.
setup	Called at the beginning of fit (train + validate), validate, test, or predict.
share_memory	See torch.Tensor.share_memory_().
state_dict	Return a dictionary containing references to the whole state of the module.
teardown	Called at the end of fit (train + validate), validate, test, or predict.
test_dataloader	An iterable or collection of iterables specifying test samples.
test_step	Operates on a single batch of data from the test set.
to	See torch.nn.Module.to().
to_empty	Move the parameters and buffers to the specified device without copying storage.
to_onnx	Saves the model in ONNX format.
to_torchscript	By default compiles the whole model to a torch.jit.ScriptModule.
toggle_optimizer	Makes sure only the gradients of the current optimizer's parameters are calculated in the training step to prevent dangling gradients in multiple-optimizer setup.
toggled_optimizer	Makes sure only the gradients of the current optimizer's parameters are calculated in the training step to prevent dangling gradients in multiple-optimizer setup.
train	Set the module in training mode.
train_dataloader	An iterable or collection of iterables specifying training samples.
training_step	Training step that computes diffusion loss on a batch.
transfer_batch_to_device	Override this hook if your DataLoader returns tensors wrapped in a custom data structure.
type	See torch.nn.Module.type().
unfreeze	Unfreeze all parameters for training.
untoggle_optimizer	Resets the state of required gradients that were toggled with toggle_optimizer().
val_dataloader	An iterable or collection of iterables specifying validation samples.
validation_step	Validation step that evaluates diffusion loss on a batch.
xpu	Move all model parameters and buffers to the XPU.
zero_grad	Reset gradients of all model parameters.

Attributes

CHECKPOINT_HYPER_PARAMS_KEY
CHECKPOINT_HYPER_PARAMS_NAME
CHECKPOINT_HYPER_PARAMS_TYPE
SCHEDULER_MAPPING
T_destination
automatic_optimization	If set to False you are responsible for calling .backward(), .step(), .zero_grad().
call_super_init
current_epoch	The current epoch in the Trainer, or 0 if not attached.
device
device_mesh	Strategies like ModelParallelStrategy will create a device mesh that can be accessed in the configure_model() hook to parallelize the LightningModule.
dtype
dump_patches
example_input_array	The example input array is a specification of what the module can consume in the forward() method.
fabric
global_rank	The index of the current process across all nodes and devices.
global_step	Total training batches seen across all epochs.
hparams	The collection of hyperparameters saved with save_hyperparameters().
hparams_initial	The collection of hyperparameters saved with save_hyperparameters().
local_rank	The index of the current process within a single node.
logger	Reference to the logger object in the Trainer.
loggers	Reference to the list of loggers in the Trainer.
on_gpu	Returns True if this model is currently located on a GPU.
strict_loading	Determines how Lightning loads this model using .load_state_dict(..., strict=model.strict_loading).
trainer
training

SCHEDULER_MAPPING = {'cosine': <function cosine_beta_schedule>, 'linear': <function linear_beta_schedule>, 'sigmoid': <function sigmoid_beta_schedule>}

__init__(model: Module, sequence_encoder: Module, distance_map_encoder: Module, beta_scheduler: str = 'cosine', timesteps: int = 1000, set_lr: float = 0.0001, min_snr_loss: bool = False, min_snr_gamma: float = 5.0, config_scheduler: str = 'LinearWarmupCosineAnnealingLR') → None

A discrete-time denoising-diffusion model framework for latent space diffusion models. The model is based on the work of Sohl-Dickstein et al. [1], Ho et al. [2], and Rombach et al. [3].

References

1) Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N. & Ganguli, S. Deep Unsupervised Learning using Nonequilibrium Thermodynamics. in Proceedings of the 32nd International Conference on Machine Learning (eds. Bach, F. & Blei, D.) vol. 37 2256–2265 (PMLR, Lille, France, 07–09 Jul 2015).

2. Ho, J., Jain, A. & Abbeel, P. Denoising Diffusion Probabilistic Models. arXiv [cs.LG] (2020).

3) Rombach, R., Blattmann, A., Lorenz, D., Esser, P. & Ommer, B. High-resolution image synthesis with latent diffusion models. arXiv [cs.CV] (2021).

Parameters:

• model (nn.Module) – A neural network model that takes in an image, a timestamp, and optionally labels to condition on and outputs the predicted noise

• encoder_model (nn.Module) – A VAE model that takes in the data (e.g., a distance map) and outputs the compressed representation of the data (e.g., a latent space). The denoising-diffusion model is then trained to denoise the latent space.

• image_size (int) – The size of the latent space (height and width)

• beta_scheduler (str, optional) – The name of the beta scheduler to use, by default “cosine”

• timesteps (int, optional) – The number of timesteps to run the diffusion process, by default 1000

• schedule_fn_kwargs (Union[dict, None], optional) – Additional arguments to pass to the beta scheduler function, by default None

• labels (str, optional) – The type of labels to condition the model on, by default “learned-embeddings”

• set_lr (float, optional) – The initial learning rate for the optimizer, by default 1e-4

• config_scheduler (str, optional) – The name of the learning rate scheduler to use, by default “CosineAnnealingLR”

Raises:

ValueError – If the beta scheduler is not implemented

q_sample(x_start: Tensor, t: int, noise: Tensor = None) → Tensor

Add the noise to x_start tensor based on the timestamp t

Parameters:

• x_start (torch.Tensor) – The starting image tensor

• t (int) – The timestep of the denoising-diffusion process

• noise (torch.Tensor, optional) – Sampled noise to add, by default None

Returns:

Returns the properly (according to the timestamp) noised tensor

Return type:

torch.Tensor

sequence2labels(sequences: List, sequence_mask, ionic_strength) → Tensor

Converts sequences to labels based on user defined models,

Parameters:

sequences (List) – A list of sequences to convert to labels

Returns:

Returns the labels for the decoder

Return type:

torch.Tensor

Raises:

ValueError – If the labels are not one of the three options

p_loss(x_start: Tensor, t: int, labels: Tensor, mask: Tensor, ionic_strengths: Tensor, noise: Tensor = None) → Tensor

A function that runs the model and calculates the loss based on the predicted noise and the actual noise. The loss can either be L1 or L2.

Parameters:

• x_start (torch.Tensor) – The starting image tensor

• t (int) – The timestep along the denoising-diffusion process

• labels (torch.Tensor, optional) – Labels to condition the model on, by default None

• noise (torch.Tensor, optional) – Sampled noise from N(0,I), by default None

Returns:

Returns the loss

Return type:

torch.Tensor

Raises:

ValueError – If the loss type is not one of the two options (l1, l2)

forward(x: Tensor, labels: Tensor, mask, ionic_strengths) → Tensor

Forward pass of the model, calculates the loss based on the predicted noise and the actual noise.

Parameters:

• x (torch.Tensor) – The starting tensor to noise/denoise

• labels (torch.Tensor, optional) – Sequences to condition the model on, by default None

Returns:

Returns the loss

Return type:

torch.Tensor

training_step(batch: Tensor, batch_idx: int) → Tensor

Training step that computes diffusion loss on a batch.

validation_step(batch: Tensor, batch_idx: int) → Tensor

Validation step that evaluates diffusion loss on a batch.

compute_snr(timesteps)

Computes SNR as per https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849

configure_optimizers()

Configure the optimizer and the learning rate scheduler for the model. Here I am using NVIDIA suggested settings for learning rate and weight decay. For ResNet50 they have seen best performance with CosineAnnealingLR, initial learning rate of 0.256 for batch size of 256 and linearly scaling it down/up for other batch sizes. The weight decay is set to 1/32768 for all parameters except the batch normalization layers. For further information check: https://catalog.ngc.nvidia.com/orgs/nvidia/resources/resnet_50_v1_5_for_pytorch

Returns:

Returns the optimizer and the learning rate scheduler

Return type:

List

Raises:

ValueError – If the scheduler is not implemented