# Copyright 2022 Cerebras Systems.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import logging
from abc import abstractmethod
import numpy as np
import torch
from torch.utils.data.dataloader import default_collate
import cerebras.pytorch as cstorch
from cerebras.modelzoo.common.input_utils import get_streaming_batch_size
from cerebras.modelzoo.data.vision.diffusion.config import (
DiffusionBaseProcessorConfig,
)
from cerebras.modelzoo.data.vision.diffusion.dit_transforms import (
LabelDropout,
NoiseGenerator,
)
from cerebras.modelzoo.data.vision.preprocessing import get_preprocess_transform
from cerebras.modelzoo.data.vision.transforms import LambdaWithParam
from cerebras.modelzoo.data.vision.utils import (
ShardedSampler,
is_gpu_distributed,
)
from cerebras.modelzoo.models.vision.dit.layers.vae.utils import (
DiagonalGaussianDistribution,
)
[docs]class DiffusionBaseProcessor:
def __init__(self, config: DiffusionBaseProcessorConfig):
if isinstance(config, dict):
config = DiffusionBaseProcessorConfig(**config)
self.data_dir = config.data_dir
self.mp_type = cstorch.amp.get_floating_point_dtype()
self.allowable_split = ["train", "val"]
# Preprocessing params
self.pp_params = dict(
noaugment=config.noaugment,
transforms=config.transforms,
)
# params for data loader
self.batch_size = get_streaming_batch_size(config.batch_size)
self.shuffle = config.shuffle and config.split == "train"
self.shuffle_seed = config.shuffle_seed
self.drop_last = config.drop_last
self.split = config.split
self.use_worker_cache = config.use_worker_cache
# multi-processing params.
self.num_workers = config.num_workers
self.prefetch_factor = config.prefetch_factor
self.persistent_workers = config.persistent_workers
self.distributed = is_gpu_distributed()
# DiT related
# copied to train/eval params in utils.py
self.vae_scaling_factor = config.vae_scaling_factor
self.dropout_rate = config.label_dropout_rate
self.latent_height = config.latent_size[0]
self.latent_width = config.latent_size[1]
self.latent_channels = config.latent_channels
self.num_diffusion_steps = config.num_diffusion_steps
self.schedule_name = config.schedule_name
def _passthrough(self, x):
return x
[docs] def create_dataloader(self):
"""
Dataloader returns a dict with keys:
"input": Tensor of shape (batch_size, latent_channels, latent_height, latent_width)
"label": Tensor of shape (batch_size, ) with dropout applied with `label_dropout_rate`
"diffusion_noise": Tensor of shape (batch_size, latent_channels, latent_height, latent_width)
represents diffusion noise to be applied
"timestep": Tensor of shape (batch_size, ) that
indicates the timesteps for each diffusion sample
"""
dataset = self.create_dataset()
shuffle = self.shuffle
generator = torch.Generator(device="cpu")
if self.shuffle_seed is not None:
generator.manual_seed(self.shuffle_seed)
if self.distributed:
data_sampler = torch.utils.data.distributed.DistributedSampler(
dataset,
shuffle=shuffle,
seed=self.shuffle_seed,
)
else:
data_sampler = ShardedSampler(
dataset, shuffle, self.shuffle_seed, self.drop_last
)
if self.num_workers:
dataloader = torch.utils.data.DataLoader(
dataset,
batch_size=self.batch_size,
sampler=data_sampler,
num_workers=self.num_workers,
pin_memory=self.distributed,
drop_last=self.drop_last,
prefetch_factor=self.prefetch_factor,
persistent_workers=self.persistent_workers,
generator=generator,
)
else:
dataloader = torch.utils.data.DataLoader(
dataset,
batch_size=self.batch_size,
sampler=data_sampler,
pin_memory=self.distributed,
drop_last=self.drop_last,
generator=generator,
)
# Intialize classes
self.label_dropout = LabelDropout(self.dropout_rate, self.num_classes)
self.noise_generator = NoiseGenerator(
self.latent_width,
self.latent_height,
self.latent_channels,
self.num_diffusion_steps,
)
self.latent_dist_fn = DiagonalGaussianDistribution
dataloader.collate_fn = self.custom_collate_fn
return dataloader
def custom_collate_fn(self, batch):
batch = default_collate(batch)
input, label = batch
data = self.noise_generator(*self.label_dropout(input, label))
# Pop `vae_noise` tensor, not needed if not used
vae_noise = data.pop("vae_noise")
# torch.clamp in this class does not support half dtypes
latent_dist = self.latent_dist_fn(input.to(torch.float32))
# overwrite with latent
data["input"] = (
latent_dist.sample(vae_noise)
.mul_(self.vae_scaling_factor)
.to(data["diffusion_noise"].dtype)
)
return data
def check_split_valid(self, split):
if split not in self.allowable_split:
raise ValueError(
f"Dataset split {split} is invalid. Only values in "
f"{self.allowable_split} are allowed."
)
def _get_target_transform(self, x, *args, **kwargs):
return np.int32(x)
def process_transform(self):
if self.pp_params["noaugment"]:
transform_specs = [
{"name": "to_tensor"},
]
logging.warning(
"User specified `noaugment=True`. "
"The input data will only be converted to tensor."
)
self.pp_params["transforms"] = transform_specs
transform = get_preprocess_transform(self.pp_params)
target_transform = LambdaWithParam(self._get_target_transform)
return transform, target_transform
@abstractmethod
def create_dataset(self):
raise NotImplementedError(
"create_dataset must be implemented in a child class!!"
)