stable-diffusion-webui/modules/hypernetworks/hypernetwork.py

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

783 lines
35 KiB
Python
Raw Normal View History

2022-10-07 14:22:22 -06:00
import datetime
import glob
import html
import os
import inspect
from contextlib import closing
2022-10-18 10:00:01 -06:00
2022-10-22 05:07:00 -06:00
import modules.textual_inversion.dataset
2022-10-18 10:00:01 -06:00
import torch
2022-10-22 05:07:00 -06:00
import tqdm
2022-10-18 10:00:01 -06:00
from einops import rearrange, repeat
2022-10-22 05:07:00 -06:00
from ldm.util import default
from modules import devices, sd_models, shared, sd_samplers, hashes, sd_hijack_checkpoint, errors
from modules.textual_inversion import textual_inversion, logging
from modules.textual_inversion.learn_schedule import LearnRateScheduler
2022-10-22 05:07:00 -06:00
from torch import einsum
from torch.nn.init import normal_, xavier_normal_, xavier_uniform_, kaiming_normal_, kaiming_uniform_, zeros_
2023-05-09 23:43:42 -06:00
from collections import deque
2022-10-22 10:57:58 -06:00
from statistics import stdev, mean
2022-11-04 00:57:17 -06:00
optimizer_dict = {optim_name : cls_obj for optim_name, cls_obj in inspect.getmembers(torch.optim, inspect.isclass) if optim_name != "Optimizer"}
2022-10-07 14:22:22 -06:00
class HypernetworkModule(torch.nn.Module):
2022-10-22 05:07:00 -06:00
activation_dict = {
"linear": torch.nn.Identity,
2022-10-22 05:07:00 -06:00
"relu": torch.nn.ReLU,
"leakyrelu": torch.nn.LeakyReLU,
"elu": torch.nn.ELU,
"swish": torch.nn.Hardswish,
"tanh": torch.nn.Tanh,
"sigmoid": torch.nn.Sigmoid,
2022-10-22 05:07:00 -06:00
}
2022-10-25 00:24:59 -06:00
activation_dict.update({cls_name.lower(): cls_obj for cls_name, cls_obj in inspect.getmembers(torch.nn.modules.activation) if inspect.isclass(cls_obj) and cls_obj.__module__ == 'torch.nn.modules.activation'})
2022-10-22 05:07:00 -06:00
def __init__(self, dim, state_dict=None, layer_structure=None, activation_func=None, weight_init='Normal',
add_layer_norm=False, activate_output=False, dropout_structure=None):
2022-10-07 14:22:22 -06:00
super().__init__()
self.multiplier = 1.0
2022-10-19 18:10:45 -06:00
assert layer_structure is not None, "layer_structure must not be None"
assert layer_structure[0] == 1, "Multiplier Sequence should start with size 1!"
assert layer_structure[-1] == 1, "Multiplier Sequence should end with size 1!"
2022-10-07 14:22:22 -06:00
linears = []
for i in range(len(layer_structure) - 1):
2022-10-22 05:07:00 -06:00
# Add a fully-connected layer
linears.append(torch.nn.Linear(int(dim * layer_structure[i]), int(dim * layer_structure[i+1])))
2022-10-22 05:07:00 -06:00
# Add an activation func except last layer
2022-10-26 00:50:58 -06:00
if activation_func == "linear" or activation_func is None or (i >= len(layer_structure) - 2 and not activate_output):
2022-10-22 05:07:00 -06:00
pass
elif activation_func in self.activation_dict:
linears.append(self.activation_dict[activation_func]())
else:
2022-10-22 06:02:41 -06:00
raise RuntimeError(f'hypernetwork uses an unsupported activation function: {activation_func}')
2022-10-22 05:07:00 -06:00
# Add layer normalization
if add_layer_norm:
linears.append(torch.nn.LayerNorm(int(dim * layer_structure[i+1])))
# Everything should be now parsed into dropout structure, and applied here.
# Since we only have dropouts after layers, dropout structure should start with 0 and end with 0.
if dropout_structure is not None and dropout_structure[i+1] > 0:
assert 0 < dropout_structure[i+1] < 1, "Dropout probability should be 0 or float between 0 and 1!"
linears.append(torch.nn.Dropout(p=dropout_structure[i+1]))
# Code explanation : [1, 2, 1] -> dropout is missing when last_layer_dropout is false. [1, 2, 2, 1] -> [0, 0.3, 0, 0], when its True, [0, 0.3, 0.3, 0].
2022-10-22 07:00:44 -06:00
self.linear = torch.nn.Sequential(*linears)
2022-10-07 14:22:22 -06:00
if state_dict is not None:
self.fix_old_state_dict(state_dict)
self.load_state_dict(state_dict)
2022-10-07 14:22:22 -06:00
else:
for layer in self.linear:
if type(layer) == torch.nn.Linear or type(layer) == torch.nn.LayerNorm:
w, b = layer.weight.data, layer.bias.data
if weight_init == "Normal" or type(layer) == torch.nn.LayerNorm:
normal_(w, mean=0.0, std=0.01)
normal_(b, mean=0.0, std=0)
elif weight_init == 'XavierUniform':
xavier_uniform_(w)
zeros_(b)
elif weight_init == 'XavierNormal':
xavier_normal_(w)
zeros_(b)
elif weight_init == 'KaimingUniform':
kaiming_uniform_(w, nonlinearity='leaky_relu' if 'leakyrelu' == activation_func else 'relu')
zeros_(b)
elif weight_init == 'KaimingNormal':
kaiming_normal_(w, nonlinearity='leaky_relu' if 'leakyrelu' == activation_func else 'relu')
zeros_(b)
else:
raise KeyError(f"Key {weight_init} is not defined as initialization!")
2022-10-07 14:22:22 -06:00
self.to(devices.device)
def fix_old_state_dict(self, state_dict):
changes = {
'linear1.bias': 'linear.0.bias',
'linear1.weight': 'linear.0.weight',
'linear2.bias': 'linear.1.bias',
'linear2.weight': 'linear.1.weight',
}
for fr, to in changes.items():
x = state_dict.get(fr, None)
if x is None:
continue
del state_dict[fr]
state_dict[to] = x
2022-10-07 14:22:22 -06:00
def forward(self, x):
return x + self.linear(x) * (self.multiplier if not self.training else 1)
def trainables(self):
layer_structure = []
for layer in self.linear:
if type(layer) == torch.nn.Linear or type(layer) == torch.nn.LayerNorm:
2022-10-19 18:10:45 -06:00
layer_structure += [layer.weight, layer.bias]
return layer_structure
2022-10-13 11:12:37 -06:00
#param layer_structure : sequence used for length, use_dropout : controlling boolean, last_layer_dropout : for compatibility check.
def parse_dropout_structure(layer_structure, use_dropout, last_layer_dropout):
if layer_structure is None:
layer_structure = [1, 2, 1]
if not use_dropout:
return [0] * len(layer_structure)
dropout_values = [0]
dropout_values.extend([0.3] * (len(layer_structure) - 3))
if last_layer_dropout:
dropout_values.append(0.3)
else:
dropout_values.append(0)
dropout_values.append(0)
return dropout_values
2022-10-07 14:22:22 -06:00
class Hypernetwork:
filename = None
name = None
def __init__(self, name=None, enable_sizes=None, layer_structure=None, activation_func=None, weight_init=None, add_layer_norm=False, use_dropout=False, activate_output=False, **kwargs):
2022-10-07 14:22:22 -06:00
self.filename = None
self.name = name
self.layers = {}
self.step = 0
self.sd_checkpoint = None
self.sd_checkpoint_name = None
self.layer_structure = layer_structure
2022-10-19 18:10:45 -06:00
self.activation_func = activation_func
self.weight_init = weight_init
self.add_layer_norm = add_layer_norm
2022-10-22 05:07:00 -06:00
self.use_dropout = use_dropout
2022-10-26 00:50:58 -06:00
self.activate_output = activate_output
self.last_layer_dropout = kwargs.get('last_layer_dropout', True)
self.dropout_structure = kwargs.get('dropout_structure', None)
if self.dropout_structure is None:
self.dropout_structure = parse_dropout_structure(self.layer_structure, self.use_dropout, self.last_layer_dropout)
2022-11-04 00:57:17 -06:00
self.optimizer_name = None
self.optimizer_state_dict = None
self.optional_info = None
2022-10-07 14:22:22 -06:00
for size in enable_sizes or []:
self.layers[size] = (
HypernetworkModule(size, None, self.layer_structure, self.activation_func, self.weight_init,
self.add_layer_norm, self.activate_output, dropout_structure=self.dropout_structure),
HypernetworkModule(size, None, self.layer_structure, self.activation_func, self.weight_init,
self.add_layer_norm, self.activate_output, dropout_structure=self.dropout_structure),
)
self.eval()
2022-10-07 14:22:22 -06:00
def weights(self):
res = []
2023-05-10 02:37:18 -06:00
for layers in self.layers.values():
2022-11-23 04:00:00 -07:00
for layer in layers:
res += layer.parameters()
return res
2022-10-07 14:22:22 -06:00
def train(self, mode=True):
2023-05-10 02:37:18 -06:00
for layers in self.layers.values():
2022-10-07 14:22:22 -06:00
for layer in layers:
layer.train(mode=mode)
2022-11-23 04:00:00 -07:00
for param in layer.parameters():
param.requires_grad = mode
2022-10-07 14:22:22 -06:00
def to(self, device):
2023-05-10 02:37:18 -06:00
for layers in self.layers.values():
for layer in layers:
layer.to(device)
return self
def set_multiplier(self, multiplier):
2023-05-10 02:37:18 -06:00
for layers in self.layers.values():
for layer in layers:
layer.multiplier = multiplier
return self
def eval(self):
2023-05-10 02:37:18 -06:00
for layers in self.layers.values():
2022-11-23 04:00:00 -07:00
for layer in layers:
layer.eval()
for param in layer.parameters():
param.requires_grad = False
2022-10-07 14:22:22 -06:00
def save(self, filename):
state_dict = {}
2022-11-04 00:57:17 -06:00
optimizer_saved_dict = {}
2022-10-07 14:22:22 -06:00
for k, v in self.layers.items():
state_dict[k] = (v[0].state_dict(), v[1].state_dict())
state_dict['step'] = self.step
state_dict['name'] = self.name
state_dict['layer_structure'] = self.layer_structure
2022-10-19 18:10:45 -06:00
state_dict['activation_func'] = self.activation_func
state_dict['is_layer_norm'] = self.add_layer_norm
state_dict['weight_initialization'] = self.weight_init
2022-10-07 14:22:22 -06:00
state_dict['sd_checkpoint'] = self.sd_checkpoint
state_dict['sd_checkpoint_name'] = self.sd_checkpoint_name
2022-10-26 00:50:58 -06:00
state_dict['activate_output'] = self.activate_output
state_dict['use_dropout'] = self.use_dropout
state_dict['dropout_structure'] = self.dropout_structure
state_dict['last_layer_dropout'] = (self.dropout_structure[-2] != 0) if self.dropout_structure is not None else self.last_layer_dropout
state_dict['optional_info'] = self.optional_info if self.optional_info else None
2022-11-04 00:57:17 -06:00
if self.optimizer_name is not None:
optimizer_saved_dict['optimizer_name'] = self.optimizer_name
2022-10-07 14:22:22 -06:00
torch.save(state_dict, filename)
2022-11-04 01:09:19 -06:00
if shared.opts.save_optimizer_state and self.optimizer_state_dict:
2023-01-14 00:25:21 -07:00
optimizer_saved_dict['hash'] = self.shorthash()
2022-11-04 00:57:17 -06:00
optimizer_saved_dict['optimizer_state_dict'] = self.optimizer_state_dict
torch.save(optimizer_saved_dict, filename + '.optim')
2022-10-07 14:22:22 -06:00
def load(self, filename):
self.filename = filename
if self.name is None:
self.name = os.path.splitext(os.path.basename(filename))[0]
state_dict = torch.load(filename, map_location='cpu')
self.layer_structure = state_dict.get('layer_structure', [1, 2, 1])
2023-01-14 00:25:21 -07:00
self.optional_info = state_dict.get('optional_info', None)
2022-10-19 18:10:45 -06:00
self.activation_func = state_dict.get('activation_func', None)
self.weight_init = state_dict.get('weight_initialization', 'Normal')
self.add_layer_norm = state_dict.get('is_layer_norm', False)
self.dropout_structure = state_dict.get('dropout_structure', None)
self.use_dropout = True if self.dropout_structure is not None and any(self.dropout_structure) else state_dict.get('use_dropout', False)
2022-10-26 00:50:58 -06:00
self.activate_output = state_dict.get('activate_output', True)
self.last_layer_dropout = state_dict.get('last_layer_dropout', False)
# Dropout structure should have same length as layer structure, Every digits should be in [0,1), and last digit must be 0.
if self.dropout_structure is None:
self.dropout_structure = parse_dropout_structure(self.layer_structure, self.use_dropout, self.last_layer_dropout)
2023-01-14 00:25:21 -07:00
if shared.opts.print_hypernet_extra:
if self.optional_info is not None:
print(f" INFO:\n {self.optional_info}\n")
2023-01-14 00:25:21 -07:00
print(f" Layer structure: {self.layer_structure}")
print(f" Activation function: {self.activation_func}")
print(f" Weight initialization: {self.weight_init}")
print(f" Layer norm: {self.add_layer_norm}")
print(f" Dropout usage: {self.use_dropout}" )
print(f" Activate last layer: {self.activate_output}")
print(f" Dropout structure: {self.dropout_structure}")
optimizer_saved_dict = torch.load(self.filename + '.optim', map_location='cpu') if os.path.exists(self.filename + '.optim') else {}
if self.shorthash() == optimizer_saved_dict.get('hash', None):
2022-11-04 00:57:17 -06:00
self.optimizer_state_dict = optimizer_saved_dict.get('optimizer_state_dict', None)
else:
self.optimizer_state_dict = None
if self.optimizer_state_dict:
self.optimizer_name = optimizer_saved_dict.get('optimizer_name', 'AdamW')
if shared.opts.print_hypernet_extra:
print("Loaded existing optimizer from checkpoint")
print(f"Optimizer name is {self.optimizer_name}")
2022-11-04 00:57:17 -06:00
else:
self.optimizer_name = "AdamW"
if shared.opts.print_hypernet_extra:
print("No saved optimizer exists in checkpoint")
2022-10-07 14:22:22 -06:00
for size, sd in state_dict.items():
if type(size) == int:
self.layers[size] = (
HypernetworkModule(size, sd[0], self.layer_structure, self.activation_func, self.weight_init,
self.add_layer_norm, self.activate_output, self.dropout_structure),
HypernetworkModule(size, sd[1], self.layer_structure, self.activation_func, self.weight_init,
self.add_layer_norm, self.activate_output, self.dropout_structure),
)
2022-10-07 14:22:22 -06:00
self.name = state_dict.get('name', self.name)
self.step = state_dict.get('step', 0)
self.sd_checkpoint = state_dict.get('sd_checkpoint', None)
self.sd_checkpoint_name = state_dict.get('sd_checkpoint_name', None)
self.eval()
2022-10-07 14:22:22 -06:00
2023-01-14 00:25:21 -07:00
def shorthash(self):
sha256 = hashes.sha256(self.filename, f'hypernet/{self.name}')
2023-02-04 01:38:56 -07:00
return sha256[0:10] if sha256 else None
2023-01-14 00:25:21 -07:00
2022-10-07 14:22:22 -06:00
2022-10-11 05:53:02 -06:00
def list_hypernetworks(path):
2022-10-07 14:22:22 -06:00
res = {}
for filename in sorted(glob.iglob(os.path.join(path, '**/*.pt'), recursive=True), key=str.lower):
2022-10-11 05:53:02 -06:00
name = os.path.splitext(os.path.basename(filename))[0]
# Prevent a hypothetical "None.pt" from being listed.
if name != "None":
2023-01-14 00:25:21 -07:00
res[name] = filename
2022-10-11 05:53:02 -06:00
return res
2022-10-07 14:22:22 -06:00
2022-10-11 05:53:02 -06:00
def load_hypernetwork(name):
path = shared.hypernetworks.get(name, None)
2022-10-11 05:53:02 -06:00
if path is None:
return None
2022-10-07 14:22:22 -06:00
try:
hypernetwork = Hypernetwork()
hypernetwork.load(path)
return hypernetwork
except Exception:
errors.report(f"Error loading hypernetwork {path}", exc_info=True)
return None
def load_hypernetworks(names, multipliers=None):
already_loaded = {}
for hypernetwork in shared.loaded_hypernetworks:
if hypernetwork.name in names:
already_loaded[hypernetwork.name] = hypernetwork
shared.loaded_hypernetworks.clear()
for i, name in enumerate(names):
hypernetwork = already_loaded.get(name, None)
if hypernetwork is None:
hypernetwork = load_hypernetwork(name)
if hypernetwork is None:
continue
hypernetwork.set_multiplier(multipliers[i] if multipliers else 1.0)
shared.loaded_hypernetworks.append(hypernetwork)
2022-10-07 14:22:22 -06:00
def apply_single_hypernetwork(hypernetwork, context_k, context_v, layer=None):
hypernetwork_layers = (hypernetwork.layers if hypernetwork is not None else {}).get(context_k.shape[2], None)
2022-10-07 14:22:22 -06:00
2022-10-11 05:53:02 -06:00
if hypernetwork_layers is None:
return context_k, context_v
2022-10-07 14:22:22 -06:00
2022-10-11 05:53:02 -06:00
if layer is not None:
layer.hyper_k = hypernetwork_layers[0]
layer.hyper_v = hypernetwork_layers[1]
2022-10-07 14:22:22 -06:00
context_k = devices.cond_cast_unet(hypernetwork_layers[0](devices.cond_cast_float(context_k)))
context_v = devices.cond_cast_unet(hypernetwork_layers[1](devices.cond_cast_float(context_v)))
return context_k, context_v
def apply_hypernetworks(hypernetworks, context, layer=None):
context_k = context
context_v = context
for hypernetwork in hypernetworks:
context_k, context_v = apply_single_hypernetwork(hypernetwork, context_k, context_v, layer)
2022-10-11 05:53:02 -06:00
return context_k, context_v
2022-10-07 14:22:22 -06:00
2023-07-13 00:30:33 -06:00
def attention_CrossAttention_forward(self, x, context=None, mask=None, **kwargs):
2022-10-11 05:53:02 -06:00
h = self.heads
q = self.to_q(x)
context = default(context, x)
2022-10-07 14:22:22 -06:00
context_k, context_v = apply_hypernetworks(shared.loaded_hypernetworks, context, self)
2022-10-07 14:22:22 -06:00
k = self.to_k(context_k)
v = self.to_v(context_v)
2023-05-10 02:05:02 -06:00
q, k, v = (rearrange(t, 'b n (h d) -> (b h) n d', h=h) for t in (q, k, v))
2022-10-07 14:22:22 -06:00
sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
if mask is not None:
mask = rearrange(mask, 'b ... -> b (...)')
max_neg_value = -torch.finfo(sim.dtype).max
mask = repeat(mask, 'b j -> (b h) () j', h=h)
sim.masked_fill_(~mask, max_neg_value)
# attention, what we cannot get enough of
attn = sim.softmax(dim=-1)
out = einsum('b i j, b j d -> b i d', attn, v)
out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
return self.to_out(out)
def stack_conds(conds):
if len(conds) == 1:
return torch.stack(conds)
# same as in reconstruct_multicond_batch
token_count = max([x.shape[0] for x in conds])
for i in range(len(conds)):
if conds[i].shape[0] != token_count:
last_vector = conds[i][-1:]
last_vector_repeated = last_vector.repeat([token_count - conds[i].shape[0], 1])
conds[i] = torch.vstack([conds[i], last_vector_repeated])
return torch.stack(conds)
2022-10-22 10:57:58 -06:00
def statistics(data):
2022-10-23 06:47:39 -06:00
if len(data) < 2:
std = 0
else:
std = stdev(data)
total_information = f"loss:{mean(data):.3f}" + u"\u00B1" + f"({std/ (len(data) ** 0.5):.3f})"
2022-10-22 10:57:58 -06:00
recent_data = data[-32:]
2022-10-23 06:47:39 -06:00
if len(recent_data) < 2:
std = 0
else:
std = stdev(recent_data)
recent_information = f"recent 32 loss:{mean(recent_data):.3f}" + u"\u00B1" + f"({std / (len(recent_data) ** 0.5):.3f})"
2022-10-22 10:57:58 -06:00
return total_information, recent_information
def create_hypernetwork(name, enable_sizes, overwrite_old, layer_structure=None, activation_func=None, weight_init=None, add_layer_norm=False, use_dropout=False, dropout_structure=None):
2022-12-24 16:02:22 -07:00
# Remove illegal characters from name.
name = "".join( x for x in name if (x.isalnum() or x in "._- "))
assert name, "Name cannot be empty!"
2022-12-24 16:02:22 -07:00
fn = os.path.join(shared.cmd_opts.hypernetwork_dir, f"{name}.pt")
if not overwrite_old:
assert not os.path.exists(fn), f"file {fn} already exists"
if type(layer_structure) == str:
layer_structure = [float(x.strip()) for x in layer_structure.split(",")]
if use_dropout and dropout_structure and type(dropout_structure) == str:
dropout_structure = [float(x.strip()) for x in dropout_structure.split(",")]
else:
dropout_structure = [0] * len(layer_structure)
2022-12-24 16:02:22 -07:00
hypernet = modules.hypernetworks.hypernetwork.Hypernetwork(
name=name,
enable_sizes=[int(x) for x in enable_sizes],
layer_structure=layer_structure,
activation_func=activation_func,
weight_init=weight_init,
add_layer_norm=add_layer_norm,
use_dropout=use_dropout,
dropout_structure=dropout_structure
2022-12-24 16:02:22 -07:00
)
hypernet.save(fn)
shared.reload_hypernetworks()
2022-12-24 16:02:22 -07:00
def train_hypernetwork(id_task, hypernetwork_name, learn_rate, batch_size, gradient_step, data_root, log_directory, training_width, training_height, varsize, steps, clip_grad_mode, clip_grad_value, shuffle_tags, tag_drop_out, latent_sampling_method, use_weight, create_image_every, save_hypernetwork_every, template_filename, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height):
from modules import images, processing
save_hypernetwork_every = save_hypernetwork_every or 0
create_image_every = create_image_every or 0
template_file = textual_inversion.textual_inversion_templates.get(template_filename, None)
textual_inversion.validate_train_inputs(hypernetwork_name, learn_rate, batch_size, gradient_step, data_root, template_file, template_filename, steps, save_hypernetwork_every, create_image_every, log_directory, name="hypernetwork")
template_file = template_file.path
2022-10-07 14:22:22 -06:00
2022-10-11 05:53:02 -06:00
path = shared.hypernetworks.get(hypernetwork_name, None)
hypernetwork = Hypernetwork()
hypernetwork.load(path)
shared.loaded_hypernetworks = [hypernetwork]
2022-10-07 14:22:22 -06:00
2023-01-03 08:34:51 -07:00
shared.state.job = "train-hypernetwork"
2022-10-07 14:22:22 -06:00
shared.state.textinfo = "Initializing hypernetwork training..."
shared.state.job_count = steps
2022-11-04 01:04:03 -06:00
hypernetwork_name = hypernetwork_name.rsplit('(', 1)[0]
2022-10-07 14:22:22 -06:00
filename = os.path.join(shared.cmd_opts.hypernetwork_dir, f'{hypernetwork_name}.pt')
log_directory = os.path.join(log_directory, datetime.datetime.now().strftime("%Y-%m-%d"), hypernetwork_name)
unload = shared.opts.unload_models_when_training
2022-10-07 14:22:22 -06:00
if save_hypernetwork_every > 0:
hypernetwork_dir = os.path.join(log_directory, "hypernetworks")
os.makedirs(hypernetwork_dir, exist_ok=True)
else:
hypernetwork_dir = None
if create_image_every > 0:
images_dir = os.path.join(log_directory, "images")
os.makedirs(images_dir, exist_ok=True)
else:
images_dir = None
checkpoint = sd_models.select_checkpoint()
2022-10-07 14:22:22 -06:00
2022-10-20 14:43:08 -06:00
initial_step = hypernetwork.step or 0
if initial_step >= steps:
shared.state.textinfo = "Model has already been trained beyond specified max steps"
2022-10-07 14:22:22 -06:00
return hypernetwork, filename
2022-10-20 14:43:08 -06:00
scheduler = LearnRateScheduler(learn_rate, steps, initial_step)
clip_grad = torch.nn.utils.clip_grad_value_ if clip_grad_mode == "value" else torch.nn.utils.clip_grad_norm_ if clip_grad_mode == "norm" else None
2022-11-04 22:48:38 -06:00
if clip_grad:
clip_grad_sched = LearnRateScheduler(clip_grad_value, steps, initial_step, verbose=False)
2022-10-11 13:03:05 -06:00
if shared.opts.training_enable_tensorboard:
tensorboard_writer = textual_inversion.tensorboard_setup(log_directory)
# dataset loading may take a while, so input validations and early returns should be done before this
2022-10-07 14:22:22 -06:00
shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..."
pin_memory = shared.opts.pin_memory
ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=hypernetwork_name, model=shared.sd_model, cond_model=shared.sd_model.cond_stage_model, device=devices.device, template_file=template_file, include_cond=True, batch_size=batch_size, gradient_step=gradient_step, shuffle_tags=shuffle_tags, tag_drop_out=tag_drop_out, latent_sampling_method=latent_sampling_method, varsize=varsize, use_weight=use_weight)
if shared.opts.save_training_settings_to_txt:
saved_params = dict(
2023-01-13 23:56:59 -07:00
model_name=checkpoint.model_name, model_hash=checkpoint.shorthash, num_of_dataset_images=len(ds),
**{field: getattr(hypernetwork, field) for field in ['layer_structure', 'activation_func', 'weight_init', 'add_layer_norm', 'use_dropout', ]}
)
logging.save_settings_to_file(log_directory, {**saved_params, **locals()})
2022-10-07 14:22:22 -06:00
latent_sampling_method = ds.latent_sampling_method
dl = modules.textual_inversion.dataset.PersonalizedDataLoader(ds, latent_sampling_method=latent_sampling_method, batch_size=ds.batch_size, pin_memory=pin_memory)
old_parallel_processing_allowed = shared.parallel_processing_allowed
2022-10-07 14:22:22 -06:00
if unload:
shared.parallel_processing_allowed = False
shared.sd_model.cond_stage_model.to(devices.cpu)
shared.sd_model.first_stage_model.to(devices.cpu)
2022-10-07 14:22:22 -06:00
weights = hypernetwork.weights()
hypernetwork.train()
2022-10-07 14:22:22 -06:00
2022-11-04 00:57:17 -06:00
# Here we use optimizer from saved HN, or we can specify as UI option.
if hypernetwork.optimizer_name in optimizer_dict:
2022-11-04 00:57:17 -06:00
optimizer = optimizer_dict[hypernetwork.optimizer_name](params=weights, lr=scheduler.learn_rate)
optimizer_name = hypernetwork.optimizer_name
2022-11-04 00:57:17 -06:00
else:
print(f"Optimizer type {hypernetwork.optimizer_name} is not defined!")
2022-11-04 00:57:17 -06:00
optimizer = torch.optim.AdamW(params=weights, lr=scheduler.learn_rate)
optimizer_name = 'AdamW'
2022-11-04 00:57:17 -06:00
if hypernetwork.optimizer_state_dict: # This line must be changed if Optimizer type can be different from saved optimizer.
try:
optimizer.load_state_dict(hypernetwork.optimizer_state_dict)
except RuntimeError as e:
print("Cannot resume from saved optimizer!")
print(e)
2022-10-11 13:03:05 -06:00
scaler = torch.cuda.amp.GradScaler()
batch_size = ds.batch_size
gradient_step = ds.gradient_step
# n steps = batch_size * gradient_step * n image processed
steps_per_epoch = len(ds) // batch_size // gradient_step
max_steps_per_epoch = len(ds) // batch_size - (len(ds) // batch_size) % gradient_step
loss_step = 0
_loss_step = 0 #internal
# size = len(ds.indexes)
# loss_dict = defaultdict(lambda : deque(maxlen = 1024))
2023-01-15 11:02:54 -07:00
loss_logging = deque(maxlen=len(ds) * 3) # this should be configurable parameter, this is 3 * epoch(dataset size)
# losses = torch.zeros((size,))
# previous_mean_losses = [0]
# previous_mean_loss = 0
# print("Mean loss of {} elements".format(size))
2022-10-11 13:03:05 -06:00
steps_without_grad = 0
last_saved_file = "<none>"
last_saved_image = "<none>"
forced_filename = "<none>"
pbar = tqdm.tqdm(total=steps - initial_step)
try:
sd_hijack_checkpoint.add()
2023-05-10 02:37:18 -06:00
for _ in range((steps-initial_step) * gradient_step):
if scheduler.finished:
break
if shared.state.interrupted:
break
for j, batch in enumerate(dl):
# works as a drop_last=True for gradient accumulation
if j == max_steps_per_epoch:
break
scheduler.apply(optimizer, hypernetwork.step)
if scheduler.finished:
break
if shared.state.interrupted:
break
if clip_grad:
clip_grad_sched.step(hypernetwork.step)
with devices.autocast():
x = batch.latent_sample.to(devices.device, non_blocking=pin_memory)
if use_weight:
w = batch.weight.to(devices.device, non_blocking=pin_memory)
if tag_drop_out != 0 or shuffle_tags:
shared.sd_model.cond_stage_model.to(devices.device)
c = shared.sd_model.cond_stage_model(batch.cond_text).to(devices.device, non_blocking=pin_memory)
shared.sd_model.cond_stage_model.to(devices.cpu)
else:
c = stack_conds(batch.cond).to(devices.device, non_blocking=pin_memory)
if use_weight:
loss = shared.sd_model.weighted_forward(x, c, w)[0] / gradient_step
del w
else:
loss = shared.sd_model.forward(x, c)[0] / gradient_step
del x
del c
_loss_step += loss.item()
scaler.scale(loss).backward()
# go back until we reach gradient accumulation steps
if (j + 1) % gradient_step != 0:
continue
2023-01-15 11:02:54 -07:00
loss_logging.append(_loss_step)
if clip_grad:
clip_grad(weights, clip_grad_sched.learn_rate)
scaler.step(optimizer)
scaler.update()
hypernetwork.step += 1
pbar.update()
optimizer.zero_grad(set_to_none=True)
loss_step = _loss_step
_loss_step = 0
steps_done = hypernetwork.step + 1
epoch_num = hypernetwork.step // steps_per_epoch
epoch_step = hypernetwork.step % steps_per_epoch
2023-01-11 08:28:55 -07:00
description = f"Training hypernetwork [Epoch {epoch_num}: {epoch_step+1}/{steps_per_epoch}]loss: {loss_step:.7f}"
pbar.set_description(description)
if hypernetwork_dir is not None and steps_done % save_hypernetwork_every == 0:
# Before saving, change name to match current checkpoint.
hypernetwork_name_every = f'{hypernetwork_name}-{steps_done}'
last_saved_file = os.path.join(hypernetwork_dir, f'{hypernetwork_name_every}.pt')
hypernetwork.optimizer_name = optimizer_name
if shared.opts.save_optimizer_state:
hypernetwork.optimizer_state_dict = optimizer.state_dict()
save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, last_saved_file)
hypernetwork.optimizer_state_dict = None # dereference it after saving, to save memory.
2023-01-13 04:57:38 -07:00
if shared.opts.training_enable_tensorboard:
epoch_num = hypernetwork.step // len(ds)
epoch_step = hypernetwork.step - (epoch_num * len(ds)) + 1
2023-01-15 10:46:21 -07:00
mean_loss = sum(loss_logging) / len(loss_logging)
2023-01-13 04:57:38 -07:00
textual_inversion.tensorboard_add(tensorboard_writer, loss=mean_loss, global_step=hypernetwork.step, step=epoch_step, learn_rate=scheduler.learn_rate, epoch_num=epoch_num)
textual_inversion.write_loss(log_directory, "hypernetwork_loss.csv", hypernetwork.step, steps_per_epoch, {
"loss": f"{loss_step:.7f}",
"learn_rate": scheduler.learn_rate
})
if images_dir is not None and steps_done % create_image_every == 0:
forced_filename = f'{hypernetwork_name}-{steps_done}'
last_saved_image = os.path.join(images_dir, forced_filename)
hypernetwork.eval()
rng_state = torch.get_rng_state()
cuda_rng_state = None
if torch.cuda.is_available():
cuda_rng_state = torch.cuda.get_rng_state_all()
shared.sd_model.cond_stage_model.to(devices.device)
shared.sd_model.first_stage_model.to(devices.device)
p = processing.StableDiffusionProcessingTxt2Img(
sd_model=shared.sd_model,
do_not_save_grid=True,
do_not_save_samples=True,
)
p.disable_extra_networks = True
if preview_from_txt2img:
p.prompt = preview_prompt
p.negative_prompt = preview_negative_prompt
p.steps = preview_steps
p.sampler_name = sd_samplers.samplers[preview_sampler_index].name
p.cfg_scale = preview_cfg_scale
p.seed = preview_seed
p.width = preview_width
p.height = preview_height
else:
p.prompt = batch.cond_text[0]
p.steps = 20
p.width = training_width
p.height = training_height
preview_text = p.prompt
with closing(p):
processed = processing.process_images(p)
image = processed.images[0] if len(processed.images) > 0 else None
if unload:
shared.sd_model.cond_stage_model.to(devices.cpu)
shared.sd_model.first_stage_model.to(devices.cpu)
torch.set_rng_state(rng_state)
if torch.cuda.is_available():
torch.cuda.set_rng_state_all(cuda_rng_state)
hypernetwork.train()
if image is not None:
shared.state.assign_current_image(image)
2023-01-15 11:02:54 -07:00
if shared.opts.training_enable_tensorboard and shared.opts.training_tensorboard_save_images:
textual_inversion.tensorboard_add_image(tensorboard_writer,
f"Validation at epoch {epoch_num}", image,
hypernetwork.step)
last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, shared.opts.samples_format, processed.infotexts[0], p=p, forced_filename=forced_filename, save_to_dirs=False)
last_saved_image += f", prompt: {preview_text}"
2022-10-07 14:22:22 -06:00
shared.state.job_no = hypernetwork.step
2022-10-07 14:22:22 -06:00
shared.state.textinfo = f"""
2022-10-07 14:22:22 -06:00
<p>
Loss: {loss_step:.7f}<br/>
2022-11-22 10:49:01 -07:00
Step: {steps_done}<br/>
Last prompt: {html.escape(batch.cond_text[0])}<br/>
2022-10-19 17:07:57 -06:00
Last saved hypernetwork: {html.escape(last_saved_file)}<br/>
2022-10-07 14:22:22 -06:00
Last saved image: {html.escape(last_saved_image)}<br/>
</p>
"""
except Exception:
errors.report("Exception in training hypernetwork", exc_info=True)
finally:
pbar.leave = False
pbar.close()
hypernetwork.eval()
sd_hijack_checkpoint.remove()
2022-10-07 14:22:22 -06:00
filename = os.path.join(shared.cmd_opts.hypernetwork_dir, f'{hypernetwork_name}.pt')
2022-11-04 00:57:17 -06:00
hypernetwork.optimizer_name = optimizer_name
if shared.opts.save_optimizer_state:
hypernetwork.optimizer_state_dict = optimizer.state_dict()
save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, filename)
2022-10-07 14:22:22 -06:00
2022-11-04 00:57:17 -06:00
del optimizer
hypernetwork.optimizer_state_dict = None # dereference it after saving, to save memory.
shared.sd_model.cond_stage_model.to(devices.device)
shared.sd_model.first_stage_model.to(devices.device)
shared.parallel_processing_allowed = old_parallel_processing_allowed
2022-10-07 14:22:22 -06:00
return hypernetwork, filename
def save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, filename):
old_hypernetwork_name = hypernetwork.name
old_sd_checkpoint = hypernetwork.sd_checkpoint if hasattr(hypernetwork, "sd_checkpoint") else None
old_sd_checkpoint_name = hypernetwork.sd_checkpoint_name if hasattr(hypernetwork, "sd_checkpoint_name") else None
try:
2023-01-13 23:56:59 -07:00
hypernetwork.sd_checkpoint = checkpoint.shorthash
hypernetwork.sd_checkpoint_name = checkpoint.model_name
hypernetwork.name = hypernetwork_name
hypernetwork.save(filename)
except:
hypernetwork.sd_checkpoint = old_sd_checkpoint
hypernetwork.sd_checkpoint_name = old_sd_checkpoint_name
hypernetwork.name = old_hypernetwork_name
raise