Files
SHARK-Studio/apps/stable_diffusion/scripts/train_lora_word.py
2023-08-18 14:06:06 -04:00

694 lines
22 KiB
Python

# Install the required libs
# pip install -U git+https://github.com/huggingface/diffusers.git
# pip install accelerate transformers ftfy
# HuggingFace Token
# YOUR_TOKEN = "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"
# Import required libraries
import itertools
import math
import os
from typing import List
import random
import torch_mlir
import numpy as np
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
from torch.utils.data import Dataset
import PIL
import logging
from diffusers import (
AutoencoderKL,
DDPMScheduler,
PNDMScheduler,
StableDiffusionPipeline,
UNet2DConditionModel,
)
from PIL import Image
from tqdm.auto import tqdm
from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
from diffusers.loaders import AttnProcsLayers
from diffusers.models.attention_processor import LoRAXFormersAttnProcessor
import torch_mlir
from torch_mlir.dynamo import make_simple_dynamo_backend
import torch._dynamo as dynamo
from torch.fx.experimental.proxy_tensor import make_fx
from torch_mlir_e2e_test.linalg_on_tensors_backends import refbackend
from shark.shark_inference import SharkInference
torch._dynamo.config.verbose = True
from diffusers import (
AutoencoderKL,
DDPMScheduler,
PNDMScheduler,
StableDiffusionPipeline,
UNet2DConditionModel,
)
from diffusers.optimization import get_scheduler
from diffusers.pipelines.stable_diffusion import (
StableDiffusionSafetyChecker,
)
from PIL import Image
from tqdm.auto import tqdm
from transformers import (
CLIPFeatureExtractor,
CLIPTextModel,
CLIPTokenizer,
)
from io import BytesIO
from dataclasses import dataclass
from apps.stable_diffusion.src import (
args,
get_schedulers,
set_init_device_flags,
clear_all,
)
from apps.stable_diffusion.src.utils import update_lora_weight
# Setup the dataset
class LoraDataset(Dataset):
def __init__(
self,
data_root,
tokenizer,
size=512,
repeats=100,
interpolation="bicubic",
set="train",
prompt="myloraprompt",
center_crop=False,
):
self.data_root = data_root
self.tokenizer = tokenizer
self.size = size
self.center_crop = center_crop
self.prompt = prompt
self.image_paths = [
os.path.join(self.data_root, file_path)
for file_path in os.listdir(self.data_root)
]
self.num_images = len(self.image_paths)
self._length = self.num_images
if set == "train":
self._length = self.num_images * repeats
self.interpolation = {
"linear": PIL.Image.LINEAR,
"bilinear": PIL.Image.BILINEAR,
"bicubic": PIL.Image.BICUBIC,
"lanczos": PIL.Image.LANCZOS,
}[interpolation]
def __len__(self):
return self._length
def __getitem__(self, i):
example = {}
image = Image.open(self.image_paths[i % self.num_images])
if not image.mode == "RGB":
image = image.convert("RGB")
example["input_ids"] = self.tokenizer(
self.prompt,
padding="max_length",
truncation=True,
max_length=self.tokenizer.model_max_length,
return_tensors="pt",
).input_ids[0]
# default to score-sde preprocessing
img = np.array(image).astype(np.uint8)
if self.center_crop:
crop = min(img.shape[0], img.shape[1])
(
h,
w,
) = (
img.shape[0],
img.shape[1],
)
img = img[
(h - crop) // 2 : (h + crop) // 2,
(w - crop) // 2 : (w + crop) // 2,
]
image = Image.fromarray(img)
image = image.resize(
(self.size, self.size), resample=self.interpolation
)
image = np.array(image).astype(np.uint8)
image = (image / 127.5 - 1.0).astype(np.float32)
example["pixel_values"] = torch.from_numpy(image).permute(2, 0, 1)
return example
def torch_device(device):
device_tokens = device.split("=>")
if len(device_tokens) == 1:
device_str = device_tokens[0].strip()
else:
device_str = device_tokens[1].strip()
device_type_tokens = device_str.split("://")
if device_type_tokens[0] == "metal":
device_type_tokens[0] = "vulkan"
if len(device_type_tokens) > 1:
return device_type_tokens[0] + ":" + device_type_tokens[1]
else:
return device_type_tokens[0]
########## Setting up the model ##########
def lora_train(
prompt: str,
height: int,
width: int,
steps: int,
guidance_scale: float,
seed: int,
batch_count: int,
batch_size: int,
scheduler: str,
custom_model: str,
hf_model_id: str,
precision: str,
device: str,
max_length: int,
training_images_dir: str,
lora_save_dir: str,
use_lora: str,
):
from apps.stable_diffusion.web.ui.utils import (
get_custom_model_pathfile,
Config,
)
import apps.stable_diffusion.web.utils.global_obj as global_obj
print(
"Note LoRA training is not compatible with the latest torch-mlir branch"
)
print(
"To run LoRA training you'll need this to follow this guide for the torch-mlir branch: https://github.com/nod-ai/SHARK/tree/main/shark/examples/shark_training/stable_diffusion"
)
torch.manual_seed(seed)
args.prompts = [prompt]
args.steps = steps
# set ckpt_loc and hf_model_id.
types = (
".ckpt",
".safetensors",
) # the tuple of file types
args.ckpt_loc = ""
args.hf_model_id = ""
if custom_model == "None":
if not hf_model_id:
return (
None,
"Please provide either custom model or huggingface model ID, both must not be "
"empty.",
)
args.hf_model_id = hf_model_id
elif ".ckpt" in custom_model or ".safetensors" in custom_model:
args.ckpt_loc = custom_model
else:
args.hf_model_id = custom_model
args.training_images_dir = training_images_dir
args.lora_save_dir = lora_save_dir
args.precision = precision
args.batch_size = batch_size
args.max_length = max_length
args.height = height
args.width = width
args.device = torch_device(device)
args.use_lora = use_lora
# Load the Stable Diffusion model
text_encoder = CLIPTextModel.from_pretrained(
args.hf_model_id, subfolder="text_encoder"
)
vae = AutoencoderKL.from_pretrained(args.hf_model_id, subfolder="vae")
unet = UNet2DConditionModel.from_pretrained(
args.hf_model_id, subfolder="unet"
)
def freeze_params(params):
for param in params:
param.requires_grad = False
# Freeze everything but LoRA
freeze_params(vae.parameters())
freeze_params(unet.parameters())
freeze_params(text_encoder.parameters())
# Move vae and unet to device
vae.to(args.device)
unet.to(args.device)
text_encoder.to(args.device)
if use_lora != "":
update_lora_weight(unet, args.use_lora, "unet")
else:
lora_attn_procs = {}
for name in unet.attn_processors.keys():
cross_attention_dim = (
None
if name.endswith("attn1.processor")
else unet.config.cross_attention_dim
)
if name.startswith("mid_block"):
hidden_size = unet.config.block_out_channels[-1]
elif name.startswith("up_blocks"):
block_id = int(name[len("up_blocks.")])
hidden_size = list(reversed(unet.config.block_out_channels))[
block_id
]
elif name.startswith("down_blocks"):
block_id = int(name[len("down_blocks.")])
hidden_size = unet.config.block_out_channels[block_id]
lora_attn_procs[name] = LoRAXFormersAttnProcessor(
hidden_size=hidden_size,
cross_attention_dim=cross_attention_dim,
)
unet.set_attn_processor(lora_attn_procs)
lora_layers = AttnProcsLayers(unet.attn_processors)
class VaeModel(torch.nn.Module):
def __init__(self):
super().__init__()
self.vae = vae
def forward(self, input):
x = self.vae.encode(input, return_dict=False)[0]
return x
class UnetModel(torch.nn.Module):
def __init__(self):
super().__init__()
self.unet = unet
def forward(self, x, y, z):
return self.unet.forward(x, y, z, return_dict=False)[0]
shark_vae = VaeModel()
shark_unet = UnetModel()
####### Creating our training data ########
tokenizer = CLIPTokenizer.from_pretrained(
args.hf_model_id,
subfolder="tokenizer",
)
# Let's create the Dataset and Dataloader
train_dataset = LoraDataset(
data_root=args.training_images_dir,
tokenizer=tokenizer,
size=vae.sample_size,
prompt=args.prompts[0],
repeats=100,
center_crop=False,
set="train",
)
def create_dataloader(train_batch_size=1):
return torch.utils.data.DataLoader(
train_dataset, batch_size=train_batch_size, shuffle=True
)
# Create noise_scheduler for training
noise_scheduler = DDPMScheduler.from_config(
args.hf_model_id, subfolder="scheduler"
)
######## Training ###########
# Define hyperparameters for our training. If you are not happy with your results,
# you can tune the `learning_rate` and the `max_train_steps`
# Setting up all training args
hyperparameters = {
"learning_rate": 5e-04,
"scale_lr": True,
"max_train_steps": steps,
"train_batch_size": batch_size,
"gradient_accumulation_steps": 1,
"gradient_checkpointing": True,
"mixed_precision": "fp16",
"seed": 42,
"output_dir": "sd-concept-output",
}
# creating output directory
cwd = os.getcwd()
out_dir = os.path.join(cwd, hyperparameters["output_dir"])
while not os.path.exists(str(out_dir)):
try:
os.mkdir(out_dir)
except OSError as error:
print("Output directory not created")
###### Torch-MLIR Compilation ######
def _remove_nones(fx_g: torch.fx.GraphModule) -> List[int]:
removed_indexes = []
for node in fx_g.graph.nodes:
if node.op == "output":
assert (
len(node.args) == 1
), "Output node must have a single argument"
node_arg = node.args[0]
if isinstance(node_arg, (list, tuple)):
node_arg = list(node_arg)
node_args_len = len(node_arg)
for i in range(node_args_len):
curr_index = node_args_len - (i + 1)
if node_arg[curr_index] is None:
removed_indexes.append(curr_index)
node_arg.pop(curr_index)
node.args = (tuple(node_arg),)
break
if len(removed_indexes) > 0:
fx_g.graph.lint()
fx_g.graph.eliminate_dead_code()
fx_g.recompile()
removed_indexes.sort()
return removed_indexes
def _unwrap_single_tuple_return(fx_g: torch.fx.GraphModule) -> bool:
"""
Replace tuple with tuple element in functions that return one-element tuples.
Returns true if an unwrapping took place, and false otherwise.
"""
unwrapped_tuple = False
for node in fx_g.graph.nodes:
if node.op == "output":
assert (
len(node.args) == 1
), "Output node must have a single argument"
node_arg = node.args[0]
if isinstance(node_arg, tuple):
if len(node_arg) == 1:
node.args = (node_arg[0],)
unwrapped_tuple = True
break
if unwrapped_tuple:
fx_g.graph.lint()
fx_g.recompile()
return unwrapped_tuple
def _returns_nothing(fx_g: torch.fx.GraphModule) -> bool:
for node in fx_g.graph.nodes:
if node.op == "output":
assert (
len(node.args) == 1
), "Output node must have a single argument"
node_arg = node.args[0]
if isinstance(node_arg, tuple):
return len(node_arg) == 0
return False
def transform_fx(fx_g):
for node in fx_g.graph.nodes:
if node.op == "call_function":
if node.target in [
torch.ops.aten.empty,
]:
# aten.empty should be filled with zeros.
if node.target in [torch.ops.aten.empty]:
with fx_g.graph.inserting_after(node):
new_node = fx_g.graph.call_function(
torch.ops.aten.zero_,
args=(node,),
)
node.append(new_node)
node.replace_all_uses_with(new_node)
new_node.args = (node,)
fx_g.graph.lint()
@make_simple_dynamo_backend
def refbackend_torchdynamo_backend(
fx_graph: torch.fx.GraphModule, example_inputs: List[torch.Tensor]
):
# handling usage of empty tensor without initializing
transform_fx(fx_graph)
fx_graph.recompile()
if _returns_nothing(fx_graph):
return fx_graph
removed_none_indexes = _remove_nones(fx_graph)
was_unwrapped = _unwrap_single_tuple_return(fx_graph)
mlir_module = torch_mlir.compile(
fx_graph, example_inputs, output_type="linalg-on-tensors"
)
bytecode_stream = BytesIO()
mlir_module.operation.write_bytecode(bytecode_stream)
bytecode = bytecode_stream.getvalue()
shark_module = SharkInference(
mlir_module=bytecode, device=args.device, mlir_dialect="tm_tensor"
)
shark_module.compile()
def compiled_callable(*inputs):
inputs = [x.numpy() for x in inputs]
result = shark_module("forward", inputs)
if was_unwrapped:
result = [
result,
]
if not isinstance(result, list):
result = torch.from_numpy(result)
else:
result = tuple(torch.from_numpy(x) for x in result)
result = list(result)
for removed_index in removed_none_indexes:
result.insert(removed_index, None)
result = tuple(result)
return result
return compiled_callable
def predictions(torch_func, jit_func, batchA, batchB):
res = jit_func(batchA.numpy(), batchB.numpy())
if res is not None:
# prediction = torch.from_numpy(res)
prediction = res
else:
prediction = None
return prediction
logger = logging.getLogger(__name__)
train_batch_size = hyperparameters["train_batch_size"]
gradient_accumulation_steps = hyperparameters[
"gradient_accumulation_steps"
]
learning_rate = hyperparameters["learning_rate"]
if hyperparameters["scale_lr"]:
learning_rate = (
learning_rate
* gradient_accumulation_steps
* train_batch_size
# * accelerator.num_processes
)
# Initialize the optimizer
optimizer = torch.optim.AdamW(
lora_layers.parameters(), # only optimize the embeddings
lr=learning_rate,
)
# Training function
def train_func(batch_pixel_values, batch_input_ids):
# Convert images to latent space
latents = shark_vae(batch_pixel_values).sample().detach()
latents = latents * 0.18215
# Sample noise that we'll add to the latents
noise = torch.randn_like(latents)
bsz = latents.shape[0]
# Sample a random timestep for each image
timesteps = torch.randint(
0,
noise_scheduler.num_train_timesteps,
(bsz,),
device=latents.device,
).long()
# Add noise to the latents according to the noise magnitude at each timestep
# (this is the forward diffusion process)
noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
# Get the text embedding for conditioning
encoder_hidden_states = text_encoder(batch_input_ids)[0]
# Predict the noise residual
noise_pred = shark_unet(
noisy_latents,
timesteps,
encoder_hidden_states,
)
# Get the target for loss depending on the prediction type
if noise_scheduler.config.prediction_type == "epsilon":
target = noise
elif noise_scheduler.config.prediction_type == "v_prediction":
target = noise_scheduler.get_velocity(latents, noise, timesteps)
else:
raise ValueError(
f"Unknown prediction type {noise_scheduler.config.prediction_type}"
)
loss = (
F.mse_loss(noise_pred, target, reduction="none")
.mean([1, 2, 3])
.mean()
)
loss.backward()
optimizer.step()
optimizer.zero_grad()
return loss
def training_function():
max_train_steps = hyperparameters["max_train_steps"]
output_dir = hyperparameters["output_dir"]
gradient_checkpointing = hyperparameters["gradient_checkpointing"]
train_dataloader = create_dataloader(train_batch_size)
# We need to recalculate our total training steps as the size of the training dataloader may have changed.
num_update_steps_per_epoch = math.ceil(
len(train_dataloader) / gradient_accumulation_steps
)
num_train_epochs = math.ceil(
max_train_steps / num_update_steps_per_epoch
)
# Train!
total_batch_size = (
train_batch_size
* gradient_accumulation_steps
# train_batch_size * accelerator.num_processes * gradient_accumulation_steps
)
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(
f" Instantaneous batch size per device = {train_batch_size}"
)
logger.info(
f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}"
)
logger.info(
f" Gradient Accumulation steps = {gradient_accumulation_steps}"
)
logger.info(f" Total optimization steps = {max_train_steps}")
# Only show the progress bar once on each machine.
progress_bar = tqdm(
# range(max_train_steps), disable=not accelerator.is_local_main_process
range(max_train_steps)
)
progress_bar.set_description("Steps")
global_step = 0
params__ = [
i for i in text_encoder.get_input_embeddings().parameters()
]
for epoch in range(num_train_epochs):
unet.train()
for step, batch in enumerate(train_dataloader):
dynamo_callable = dynamo.optimize(
refbackend_torchdynamo_backend
)(train_func)
lam_func = lambda x, y: dynamo_callable(
torch.from_numpy(x), torch.from_numpy(y)
)
loss = predictions(
train_func,
lam_func,
batch["pixel_values"],
batch["input_ids"],
)
# Checks if the accelerator has performed an optimization step behind the scenes
progress_bar.update(1)
global_step += 1
logs = {"loss": loss.detach().item()}
progress_bar.set_postfix(**logs)
if global_step >= max_train_steps:
break
training_function()
# Save the lora weights
unet.save_attn_procs(args.lora_save_dir)
for param in itertools.chain(unet.parameters(), text_encoder.parameters()):
if param.grad is not None:
del param.grad # free some memory
torch.cuda.empty_cache()
if __name__ == "__main__":
if args.clear_all:
clear_all()
dtype = torch.float32 if args.precision == "fp32" else torch.half
cpu_scheduling = not args.scheduler.startswith("Shark")
set_init_device_flags()
schedulers = get_schedulers(args.hf_model_id)
scheduler_obj = schedulers[args.scheduler]
seed = args.seed
if len(args.prompts) != 1:
print("Need exactly one prompt for the LoRA word")
lora_train(
args.prompts[0],
args.height,
args.width,
args.training_steps,
args.guidance_scale,
args.seed,
args.batch_count,
args.batch_size,
args.scheduler,
"None",
args.hf_model_id,
args.precision,
args.device,
args.max_length,
args.training_images_dir,
args.lora_save_dir,
args.use_lora,
)