2022-10-22 19:07:00 +08:00
|
|
|
import csv
|
2022-10-08 04:22:22 +08:00
|
|
|
import datetime
|
|
|
|
import glob
|
|
|
|
import html
|
|
|
|
import os
|
|
|
|
import sys
|
|
|
|
import traceback
|
2022-10-25 13:48:49 +08:00
|
|
|
import inspect
|
2022-10-08 04:22:22 +08:00
|
|
|
|
2022-10-22 19:07:00 +08:00
|
|
|
import modules.textual_inversion.dataset
|
2022-10-08 04:22:22 +08:00
|
|
|
import torch
|
2022-10-22 19:07:00 +08:00
|
|
|
import tqdm
|
2022-10-19 00:00:01 +08:00
|
|
|
from einops import rearrange, repeat
|
2022-10-22 19:07:00 +08:00
|
|
|
from ldm.util import default
|
2022-11-19 17:01:51 +08:00
|
|
|
from modules import devices, processing, sd_models, shared, sd_samplers
|
2022-10-15 03:43:55 +08:00
|
|
|
from modules.textual_inversion import textual_inversion
|
2022-10-13 01:49:47 +08:00
|
|
|
from modules.textual_inversion.learn_schedule import LearnRateScheduler
|
2022-10-22 19:07:00 +08:00
|
|
|
from torch import einsum
|
2022-10-25 13:48:49 +08:00
|
|
|
from torch.nn.init import normal_, xavier_normal_, xavier_uniform_, kaiming_normal_, kaiming_uniform_, zeros_
|
2022-10-18 23:51:36 +08:00
|
|
|
|
2022-10-23 20:07:07 +08:00
|
|
|
from collections import defaultdict, deque
|
2022-10-23 00:57:58 +08:00
|
|
|
from statistics import stdev, mean
|
2022-10-18 23:51:36 +08:00
|
|
|
|
2022-10-25 13:48:49 +08:00
|
|
|
|
2022-11-04 14:57:17 +08:00
|
|
|
optimizer_dict = {optim_name : cls_obj for optim_name, cls_obj in inspect.getmembers(torch.optim, inspect.isclass) if optim_name != "Optimizer"}
|
|
|
|
|
2022-10-08 04:22:22 +08:00
|
|
|
class HypernetworkModule(torch.nn.Module):
|
2022-10-14 01:12:37 +08:00
|
|
|
multiplier = 1.0
|
2022-10-22 19:07:00 +08:00
|
|
|
activation_dict = {
|
2022-10-27 13:00:47 +08:00
|
|
|
"linear": torch.nn.Identity,
|
2022-10-22 19:07:00 +08:00
|
|
|
"relu": torch.nn.ReLU,
|
|
|
|
"leakyrelu": torch.nn.LeakyReLU,
|
|
|
|
"elu": torch.nn.ELU,
|
|
|
|
"swish": torch.nn.Hardswish,
|
2022-10-25 13:48:49 +08:00
|
|
|
"tanh": torch.nn.Tanh,
|
|
|
|
"sigmoid": torch.nn.Sigmoid,
|
2022-10-22 19:07:00 +08:00
|
|
|
}
|
2022-10-25 14:24:59 +08:00
|
|
|
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 19:07:00 +08:00
|
|
|
|
2022-10-26 21:24:33 +08:00
|
|
|
def __init__(self, dim, state_dict=None, layer_structure=None, activation_func=None, weight_init='Normal',
|
2022-10-26 23:51:51 +08:00
|
|
|
add_layer_norm=False, use_dropout=False, activate_output=False, last_layer_dropout=True):
|
2022-10-08 04:22:22 +08:00
|
|
|
super().__init__()
|
2022-10-20 00:31:12 +08:00
|
|
|
|
2022-10-20 08:10:45 +08:00
|
|
|
assert layer_structure is not None, "layer_structure must not be None"
|
2022-10-20 00:31:12 +08:00
|
|
|
assert layer_structure[0] == 1, "Multiplier Sequence should start with size 1!"
|
|
|
|
assert layer_structure[-1] == 1, "Multiplier Sequence should end with size 1!"
|
2022-10-22 19:07:00 +08:00
|
|
|
|
2022-10-18 23:51:36 +08:00
|
|
|
linears = []
|
|
|
|
for i in range(len(layer_structure) - 1):
|
2022-10-22 19:07:00 +08:00
|
|
|
|
|
|
|
# Add a fully-connected layer
|
2022-10-18 23:51:36 +08:00
|
|
|
linears.append(torch.nn.Linear(int(dim * layer_structure[i]), int(dim * layer_structure[i+1])))
|
2022-10-22 19:07:00 +08:00
|
|
|
|
2022-10-26 12:10:30 +08:00
|
|
|
# Add an activation func except last layer
|
2022-10-26 14:50:58 +08:00
|
|
|
if activation_func == "linear" or activation_func is None or (i >= len(layer_structure) - 2 and not activate_output):
|
2022-10-22 19:07:00 +08:00
|
|
|
pass
|
|
|
|
elif activation_func in self.activation_dict:
|
|
|
|
linears.append(self.activation_dict[activation_func]())
|
2022-10-20 22:43:03 +08:00
|
|
|
else:
|
2022-10-22 20:02:41 +08:00
|
|
|
raise RuntimeError(f'hypernetwork uses an unsupported activation function: {activation_func}')
|
2022-10-22 19:07:00 +08:00
|
|
|
|
|
|
|
# Add layer normalization
|
2022-10-21 00:37:11 +08:00
|
|
|
if add_layer_norm:
|
|
|
|
linears.append(torch.nn.LayerNorm(int(dim * layer_structure[i+1])))
|
2022-10-18 23:51:36 +08:00
|
|
|
|
2022-10-26 12:10:30 +08:00
|
|
|
# Add dropout except last layer
|
2022-10-26 23:51:51 +08:00
|
|
|
if use_dropout and (i < len(layer_structure) - 3 or last_layer_dropout and i < len(layer_structure) - 2):
|
2022-10-22 21:44:39 +08:00
|
|
|
linears.append(torch.nn.Dropout(p=0.3))
|
2022-10-22 21:00:44 +08:00
|
|
|
|
2022-10-18 23:51:36 +08:00
|
|
|
self.linear = torch.nn.Sequential(*linears)
|
2022-10-08 04:22:22 +08:00
|
|
|
|
|
|
|
if state_dict is not None:
|
2022-10-20 00:31:12 +08:00
|
|
|
self.fix_old_state_dict(state_dict)
|
|
|
|
self.load_state_dict(state_dict)
|
2022-10-08 04:22:22 +08:00
|
|
|
else:
|
2022-10-18 23:51:36 +08:00
|
|
|
for layer in self.linear:
|
2022-10-21 15:13:24 +08:00
|
|
|
if type(layer) == torch.nn.Linear or type(layer) == torch.nn.LayerNorm:
|
2022-10-25 13:48:49 +08:00
|
|
|
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)
|
2022-10-27 13:44:53 +08:00
|
|
|
normal_(b, mean=0.0, std=0)
|
2022-10-25 13:48:49 +08:00
|
|
|
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!")
|
2022-10-08 04:22:22 +08:00
|
|
|
self.to(devices.device)
|
|
|
|
|
2022-10-20 00:31:12 +08:00
|
|
|
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
|
2022-10-18 23:51:36 +08:00
|
|
|
|
2022-10-08 04:22:22 +08:00
|
|
|
def forward(self, x):
|
2022-10-18 23:51:36 +08:00
|
|
|
return x + self.linear(x) * self.multiplier
|
|
|
|
|
|
|
|
def trainables(self):
|
2022-10-19 22:30:33 +08:00
|
|
|
layer_structure = []
|
2022-10-18 23:51:36 +08:00
|
|
|
for layer in self.linear:
|
2022-10-21 15:13:24 +08:00
|
|
|
if type(layer) == torch.nn.Linear or type(layer) == torch.nn.LayerNorm:
|
2022-10-20 08:10:45 +08:00
|
|
|
layer_structure += [layer.weight, layer.bias]
|
2022-10-19 22:30:33 +08:00
|
|
|
return layer_structure
|
2022-10-14 01:12:37 +08:00
|
|
|
|
|
|
|
|
|
|
|
def apply_strength(value=None):
|
|
|
|
HypernetworkModule.multiplier = value if value is not None else shared.opts.sd_hypernetwork_strength
|
2022-10-08 04:22:22 +08:00
|
|
|
|
|
|
|
|
|
|
|
class Hypernetwork:
|
|
|
|
filename = None
|
|
|
|
name = None
|
|
|
|
|
2022-10-26 21:24:33 +08:00
|
|
|
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):
|
2022-10-08 04:22:22 +08:00
|
|
|
self.filename = None
|
|
|
|
self.name = name
|
|
|
|
self.layers = {}
|
|
|
|
self.step = 0
|
|
|
|
self.sd_checkpoint = None
|
|
|
|
self.sd_checkpoint_name = None
|
2022-10-19 22:30:33 +08:00
|
|
|
self.layer_structure = layer_structure
|
2022-10-20 08:10:45 +08:00
|
|
|
self.activation_func = activation_func
|
2022-10-25 13:48:49 +08:00
|
|
|
self.weight_init = weight_init
|
2022-10-22 19:07:00 +08:00
|
|
|
self.add_layer_norm = add_layer_norm
|
|
|
|
self.use_dropout = use_dropout
|
2022-10-26 14:50:58 +08:00
|
|
|
self.activate_output = activate_output
|
2022-10-26 21:24:33 +08:00
|
|
|
self.last_layer_dropout = kwargs['last_layer_dropout'] if 'last_layer_dropout' in kwargs else True
|
2022-11-04 14:57:17 +08:00
|
|
|
self.optimizer_name = None
|
|
|
|
self.optimizer_state_dict = None
|
2022-10-08 04:22:22 +08:00
|
|
|
|
2022-10-12 00:22:30 +08:00
|
|
|
for size in enable_sizes or []:
|
2022-10-19 22:30:33 +08:00
|
|
|
self.layers[size] = (
|
2022-10-26 21:24:33 +08:00
|
|
|
HypernetworkModule(size, None, self.layer_structure, self.activation_func, self.weight_init,
|
|
|
|
self.add_layer_norm, self.use_dropout, self.activate_output, last_layer_dropout=self.last_layer_dropout),
|
|
|
|
HypernetworkModule(size, None, self.layer_structure, self.activation_func, self.weight_init,
|
|
|
|
self.add_layer_norm, self.use_dropout, self.activate_output, last_layer_dropout=self.last_layer_dropout),
|
2022-10-19 22:30:33 +08:00
|
|
|
)
|
2022-11-23 19:00:00 +08:00
|
|
|
self.eval_mode()
|
2022-10-08 04:22:22 +08:00
|
|
|
|
|
|
|
def weights(self):
|
|
|
|
res = []
|
2022-11-23 19:00:00 +08:00
|
|
|
for k, layers in self.layers.items():
|
|
|
|
for layer in layers:
|
|
|
|
res += layer.parameters()
|
|
|
|
return res
|
2022-10-08 04:22:22 +08:00
|
|
|
|
2022-11-23 19:00:00 +08:00
|
|
|
def train_mode(self):
|
2022-10-08 04:22:22 +08:00
|
|
|
for k, layers in self.layers.items():
|
|
|
|
for layer in layers:
|
2022-10-21 00:37:11 +08:00
|
|
|
layer.train()
|
2022-11-23 19:00:00 +08:00
|
|
|
for param in layer.parameters():
|
|
|
|
param.requires_grad = True
|
2022-10-08 04:22:22 +08:00
|
|
|
|
2022-11-23 19:00:00 +08:00
|
|
|
def eval_mode(self):
|
|
|
|
for k, layers in self.layers.items():
|
|
|
|
for layer in layers:
|
|
|
|
layer.eval()
|
|
|
|
for param in layer.parameters():
|
|
|
|
param.requires_grad = False
|
2022-10-08 04:22:22 +08:00
|
|
|
|
|
|
|
def save(self, filename):
|
|
|
|
state_dict = {}
|
2022-11-04 14:57:17 +08:00
|
|
|
optimizer_saved_dict = {}
|
2022-10-08 04:22:22 +08:00
|
|
|
|
|
|
|
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
|
2022-10-19 22:30:33 +08:00
|
|
|
state_dict['layer_structure'] = self.layer_structure
|
2022-10-20 08:10:45 +08:00
|
|
|
state_dict['activation_func'] = self.activation_func
|
2022-10-22 19:07:00 +08:00
|
|
|
state_dict['is_layer_norm'] = self.add_layer_norm
|
2022-10-25 13:48:49 +08:00
|
|
|
state_dict['weight_initialization'] = self.weight_init
|
2022-10-22 19:07:00 +08:00
|
|
|
state_dict['use_dropout'] = self.use_dropout
|
2022-10-08 04:22:22 +08:00
|
|
|
state_dict['sd_checkpoint'] = self.sd_checkpoint
|
|
|
|
state_dict['sd_checkpoint_name'] = self.sd_checkpoint_name
|
2022-10-26 14:50:58 +08:00
|
|
|
state_dict['activate_output'] = self.activate_output
|
2022-10-26 21:24:33 +08:00
|
|
|
state_dict['last_layer_dropout'] = self.last_layer_dropout
|
2022-11-04 14:57:17 +08:00
|
|
|
|
|
|
|
if self.optimizer_name is not None:
|
|
|
|
optimizer_saved_dict['optimizer_name'] = self.optimizer_name
|
|
|
|
|
2022-10-08 04:22:22 +08:00
|
|
|
torch.save(state_dict, filename)
|
2022-11-04 15:09:19 +08:00
|
|
|
if shared.opts.save_optimizer_state and self.optimizer_state_dict:
|
2022-11-04 14:57:17 +08:00
|
|
|
optimizer_saved_dict['hash'] = sd_models.model_hash(filename)
|
|
|
|
optimizer_saved_dict['optimizer_state_dict'] = self.optimizer_state_dict
|
|
|
|
torch.save(optimizer_saved_dict, filename + '.optim')
|
2022-10-08 04:22:22 +08:00
|
|
|
|
|
|
|
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')
|
|
|
|
|
2022-10-20 00:31:12 +08:00
|
|
|
self.layer_structure = state_dict.get('layer_structure', [1, 2, 1])
|
2022-10-25 13:48:49 +08:00
|
|
|
print(self.layer_structure)
|
2022-10-20 08:10:45 +08:00
|
|
|
self.activation_func = state_dict.get('activation_func', None)
|
2022-10-25 13:48:49 +08:00
|
|
|
print(f"Activation function is {self.activation_func}")
|
|
|
|
self.weight_init = state_dict.get('weight_initialization', 'Normal')
|
|
|
|
print(f"Weight initialization is {self.weight_init}")
|
2022-10-22 19:07:00 +08:00
|
|
|
self.add_layer_norm = state_dict.get('is_layer_norm', False)
|
2022-10-25 13:48:49 +08:00
|
|
|
print(f"Layer norm is set to {self.add_layer_norm}")
|
2022-10-26 15:00:03 +08:00
|
|
|
self.use_dropout = state_dict.get('use_dropout', False)
|
2022-10-25 13:48:49 +08:00
|
|
|
print(f"Dropout usage is set to {self.use_dropout}" )
|
2022-10-26 14:50:58 +08:00
|
|
|
self.activate_output = state_dict.get('activate_output', True)
|
2022-10-26 21:24:33 +08:00
|
|
|
print(f"Activate last layer is set to {self.activate_output}")
|
|
|
|
self.last_layer_dropout = state_dict.get('last_layer_dropout', False)
|
2022-10-20 00:31:12 +08:00
|
|
|
|
2022-11-04 14:57:17 +08:00
|
|
|
optimizer_saved_dict = torch.load(self.filename + '.optim', map_location = 'cpu') if os.path.exists(self.filename + '.optim') else {}
|
|
|
|
self.optimizer_name = optimizer_saved_dict.get('optimizer_name', 'AdamW')
|
|
|
|
print(f"Optimizer name is {self.optimizer_name}")
|
|
|
|
if sd_models.model_hash(filename) == optimizer_saved_dict.get('hash', None):
|
|
|
|
self.optimizer_state_dict = optimizer_saved_dict.get('optimizer_state_dict', None)
|
|
|
|
else:
|
|
|
|
self.optimizer_state_dict = None
|
|
|
|
if self.optimizer_state_dict:
|
|
|
|
print("Loaded existing optimizer from checkpoint")
|
|
|
|
else:
|
|
|
|
print("No saved optimizer exists in checkpoint")
|
|
|
|
|
2022-10-08 04:22:22 +08:00
|
|
|
for size, sd in state_dict.items():
|
|
|
|
if type(size) == int:
|
2022-10-19 22:30:33 +08:00
|
|
|
self.layers[size] = (
|
2022-10-26 21:24:33 +08:00
|
|
|
HypernetworkModule(size, sd[0], self.layer_structure, self.activation_func, self.weight_init,
|
|
|
|
self.add_layer_norm, self.use_dropout, self.activate_output, last_layer_dropout=self.last_layer_dropout),
|
|
|
|
HypernetworkModule(size, sd[1], self.layer_structure, self.activation_func, self.weight_init,
|
|
|
|
self.add_layer_norm, self.use_dropout, self.activate_output, last_layer_dropout=self.last_layer_dropout),
|
2022-10-19 22:30:33 +08:00
|
|
|
)
|
2022-10-08 04:22:22 +08:00
|
|
|
|
|
|
|
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)
|
|
|
|
|
|
|
|
|
2022-10-11 19:53:02 +08:00
|
|
|
def list_hypernetworks(path):
|
2022-10-08 04:22:22 +08:00
|
|
|
res = {}
|
2022-11-04 21:53:11 +08:00
|
|
|
for filename in sorted(glob.iglob(os.path.join(path, '**/*.pt'), recursive=True)):
|
2022-10-11 19:53:02 +08:00
|
|
|
name = os.path.splitext(os.path.basename(filename))[0]
|
2022-10-28 16:41:57 +08:00
|
|
|
# Prevent a hypothetical "None.pt" from being listed.
|
|
|
|
if name != "None":
|
2022-11-04 14:57:17 +08:00
|
|
|
res[name + f"({sd_models.model_hash(filename)})"] = filename
|
2022-10-11 19:53:02 +08:00
|
|
|
return res
|
2022-10-08 04:22:22 +08:00
|
|
|
|
2022-10-11 19:53:02 +08:00
|
|
|
|
|
|
|
def load_hypernetwork(filename):
|
|
|
|
path = shared.hypernetworks.get(filename, None)
|
2022-10-28 16:41:57 +08:00
|
|
|
# Prevent any file named "None.pt" from being loaded.
|
|
|
|
if path is not None and filename != "None":
|
2022-10-11 19:53:02 +08:00
|
|
|
print(f"Loading hypernetwork {filename}")
|
2022-10-08 04:22:22 +08:00
|
|
|
try:
|
2022-10-11 19:53:02 +08:00
|
|
|
shared.loaded_hypernetwork = Hypernetwork()
|
|
|
|
shared.loaded_hypernetwork.load(path)
|
|
|
|
|
2022-10-08 04:22:22 +08:00
|
|
|
except Exception:
|
2022-10-11 19:53:02 +08:00
|
|
|
print(f"Error loading hypernetwork {path}", file=sys.stderr)
|
2022-10-08 04:22:22 +08:00
|
|
|
print(traceback.format_exc(), file=sys.stderr)
|
2022-10-11 19:53:02 +08:00
|
|
|
else:
|
|
|
|
if shared.loaded_hypernetwork is not None:
|
|
|
|
print(f"Unloading hypernetwork")
|
2022-10-08 04:22:22 +08:00
|
|
|
|
2022-10-11 19:53:02 +08:00
|
|
|
shared.loaded_hypernetwork = None
|
2022-10-08 04:22:22 +08:00
|
|
|
|
|
|
|
|
2022-10-09 23:56:36 +08:00
|
|
|
def find_closest_hypernetwork_name(search: str):
|
|
|
|
if not search:
|
|
|
|
return None
|
|
|
|
search = search.lower()
|
|
|
|
applicable = [name for name in shared.hypernetworks if search in name.lower()]
|
|
|
|
if not applicable:
|
|
|
|
return None
|
|
|
|
applicable = sorted(applicable, key=lambda name: len(name))
|
|
|
|
return applicable[0]
|
|
|
|
|
|
|
|
|
2022-10-11 19:53:02 +08:00
|
|
|
def apply_hypernetwork(hypernetwork, context, layer=None):
|
|
|
|
hypernetwork_layers = (hypernetwork.layers if hypernetwork is not None else {}).get(context.shape[2], None)
|
2022-10-08 04:22:22 +08:00
|
|
|
|
2022-10-11 19:53:02 +08:00
|
|
|
if hypernetwork_layers is None:
|
|
|
|
return context, context
|
2022-10-08 04:22:22 +08:00
|
|
|
|
2022-10-11 19:53:02 +08:00
|
|
|
if layer is not None:
|
|
|
|
layer.hyper_k = hypernetwork_layers[0]
|
|
|
|
layer.hyper_v = hypernetwork_layers[1]
|
2022-10-08 04:22:22 +08:00
|
|
|
|
2022-10-11 19:53:02 +08:00
|
|
|
context_k = hypernetwork_layers[0](context)
|
|
|
|
context_v = hypernetwork_layers[1](context)
|
|
|
|
return context_k, context_v
|
2022-10-08 04:22:22 +08:00
|
|
|
|
|
|
|
|
2022-10-11 19:53:02 +08:00
|
|
|
def attention_CrossAttention_forward(self, x, context=None, mask=None):
|
|
|
|
h = self.heads
|
|
|
|
|
|
|
|
q = self.to_q(x)
|
|
|
|
context = default(context, x)
|
2022-10-08 04:22:22 +08:00
|
|
|
|
2022-10-11 19:53:02 +08:00
|
|
|
context_k, context_v = apply_hypernetwork(shared.loaded_hypernetwork, context, self)
|
2022-10-08 04:22:22 +08:00
|
|
|
k = self.to_k(context_k)
|
|
|
|
v = self.to_v(context_v)
|
|
|
|
|
|
|
|
q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
|
|
|
|
|
|
|
|
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)
|
|
|
|
|
|
|
|
|
2022-10-15 14:24:59 +08:00
|
|
|
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)
|
|
|
|
|
2022-10-20 00:31:12 +08:00
|
|
|
|
2022-10-23 00:57:58 +08:00
|
|
|
def statistics(data):
|
2022-10-23 20:47:39 +08:00
|
|
|
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})"
|
2022-10-23 00:57:58 +08:00
|
|
|
recent_data = data[-32:]
|
2022-10-23 20:47:39 +08:00
|
|
|
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})"
|
2022-10-23 00:57:58 +08:00
|
|
|
return total_information, recent_information
|
|
|
|
|
|
|
|
|
|
|
|
def report_statistics(loss_info:dict):
|
|
|
|
keys = sorted(loss_info.keys(), key=lambda x: sum(loss_info[x]) / len(loss_info[x]))
|
|
|
|
for key in keys:
|
2022-10-23 12:43:34 +08:00
|
|
|
try:
|
|
|
|
print("Loss statistics for file " + key)
|
2022-10-23 20:41:39 +08:00
|
|
|
info, recent = statistics(list(loss_info[key]))
|
2022-10-23 12:43:34 +08:00
|
|
|
print(info)
|
|
|
|
print(recent)
|
|
|
|
except Exception as e:
|
|
|
|
print(e)
|
2022-10-23 00:57:58 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
2022-11-20 11:35:26 +08:00
|
|
|
def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_step, data_root, log_directory, training_width, training_height, steps, shuffle_tags, tag_drop_out, latent_sampling_method, create_image_every, save_hypernetwork_every, template_file, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height):
|
2022-10-20 03:32:22 +08:00
|
|
|
# images allows training previews to have infotext. Importing it at the top causes a circular import problem.
|
2022-10-20 03:21:32 +08:00
|
|
|
from modules import images
|
|
|
|
|
2022-10-29 19:09:17 +08:00
|
|
|
save_hypernetwork_every = save_hypernetwork_every or 0
|
|
|
|
create_image_every = create_image_every or 0
|
2022-11-20 11:35:26 +08:00
|
|
|
textual_inversion.validate_train_inputs(hypernetwork_name, learn_rate, batch_size, gradient_step, data_root, template_file, steps, save_hypernetwork_every, create_image_every, log_directory, name="hypernetwork")
|
2022-10-08 04:22:22 +08:00
|
|
|
|
2022-10-11 19:53:02 +08:00
|
|
|
path = shared.hypernetworks.get(hypernetwork_name, None)
|
|
|
|
shared.loaded_hypernetwork = Hypernetwork()
|
|
|
|
shared.loaded_hypernetwork.load(path)
|
2022-10-08 04:22:22 +08:00
|
|
|
|
|
|
|
shared.state.textinfo = "Initializing hypernetwork training..."
|
|
|
|
shared.state.job_count = steps
|
|
|
|
|
2022-11-04 15:04:03 +08:00
|
|
|
hypernetwork_name = hypernetwork_name.rsplit('(', 1)[0]
|
2022-10-08 04:22:22 +08:00
|
|
|
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)
|
2022-10-12 00:03:08 +08:00
|
|
|
unload = shared.opts.unload_models_when_training
|
2022-10-08 04:22:22 +08:00
|
|
|
|
|
|
|
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
|
|
|
|
|
2022-10-29 19:09:17 +08:00
|
|
|
hypernetwork = shared.loaded_hypernetwork
|
2022-10-30 01:49:29 +08:00
|
|
|
checkpoint = sd_models.select_checkpoint()
|
2022-10-29 19:09:17 +08:00
|
|
|
|
2022-11-20 11:35:26 +08:00
|
|
|
initial_step = hypernetwork.step or 0
|
|
|
|
if initial_step >= steps:
|
2022-10-29 19:09:17 +08:00
|
|
|
shared.state.textinfo = f"Model has already been trained beyond specified max steps"
|
|
|
|
return hypernetwork, filename
|
|
|
|
|
2022-11-20 11:35:26 +08:00
|
|
|
scheduler = LearnRateScheduler(learn_rate, steps, initial_step)
|
|
|
|
|
2022-10-29 19:09:17 +08:00
|
|
|
# dataset loading may take a while, so input validations and early returns should be done before this
|
2022-10-08 04:22:22 +08:00
|
|
|
shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..."
|
2022-11-20 11:35:26 +08:00
|
|
|
|
|
|
|
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)
|
2022-11-21 09:15:46 +08:00
|
|
|
|
|
|
|
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)
|
2022-10-29 19:09:17 +08:00
|
|
|
|
2022-10-12 00:03:08 +08:00
|
|
|
if unload:
|
|
|
|
shared.sd_model.cond_stage_model.to(devices.cpu)
|
|
|
|
shared.sd_model.first_stage_model.to(devices.cpu)
|
2022-10-29 19:09:17 +08:00
|
|
|
|
|
|
|
weights = hypernetwork.weights()
|
2022-11-23 19:00:00 +08:00
|
|
|
hypernetwork.train_mode()
|
2022-11-05 22:09:42 +08:00
|
|
|
|
2022-11-04 14:57:17 +08:00
|
|
|
# Here we use optimizer from saved HN, or we can specify as UI option.
|
2022-11-05 22:09:42 +08:00
|
|
|
if hypernetwork.optimizer_name in optimizer_dict:
|
2022-11-04 14:57:17 +08:00
|
|
|
optimizer = optimizer_dict[hypernetwork.optimizer_name](params=weights, lr=scheduler.learn_rate)
|
2022-11-05 22:09:42 +08:00
|
|
|
optimizer_name = hypernetwork.optimizer_name
|
2022-11-04 14:57:17 +08:00
|
|
|
else:
|
2022-11-05 22:09:42 +08:00
|
|
|
print(f"Optimizer type {hypernetwork.optimizer_name} is not defined!")
|
2022-11-23 01:49:01 +08:00
|
|
|
optimizer = torch.optim.AdamW(params=weights, lr=scheduler.learn_rate)
|
|
|
|
optimizer_name = 'AdamW'
|
2022-11-05 22:09:42 +08:00
|
|
|
|
2022-11-04 14:57:17 +08:00
|
|
|
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)
|
2022-10-12 03:03:05 +08:00
|
|
|
|
2022-11-20 11:35:26 +08:00
|
|
|
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))
|
|
|
|
# losses = torch.zeros((size,))
|
|
|
|
# previous_mean_losses = [0]
|
|
|
|
# previous_mean_loss = 0
|
|
|
|
# print("Mean loss of {} elements".format(size))
|
|
|
|
|
2022-10-22 19:48:43 +08:00
|
|
|
steps_without_grad = 0
|
|
|
|
|
2022-10-29 19:09:17 +08:00
|
|
|
last_saved_file = "<none>"
|
|
|
|
last_saved_image = "<none>"
|
|
|
|
forced_filename = "<none>"
|
|
|
|
|
2022-11-20 11:35:26 +08:00
|
|
|
pbar = tqdm.tqdm(total=steps - initial_step)
|
|
|
|
try:
|
|
|
|
for i 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
|
|
|
|
|
|
|
|
with torch.autocast("cuda"):
|
|
|
|
x = batch.latent_sample.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)
|
|
|
|
loss = shared.sd_model(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
|
|
|
|
# print(f"grad:{weights[0].grad.detach().cpu().abs().mean().item():.7f}")
|
|
|
|
# scaler.unscale_(optimizer)
|
|
|
|
# print(f"grad:{weights[0].grad.detach().cpu().abs().mean().item():.15f}")
|
|
|
|
# torch.nn.utils.clip_grad_norm_(weights, max_norm=1.0)
|
|
|
|
# print(f"grad:{weights[0].grad.detach().cpu().abs().mean().item():.15f}")
|
|
|
|
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
|
2022-10-23 00:57:58 +08:00
|
|
|
|
2022-11-20 11:35:26 +08:00
|
|
|
epoch_num = hypernetwork.step // steps_per_epoch
|
|
|
|
epoch_step = hypernetwork.step % steps_per_epoch
|
|
|
|
|
|
|
|
pbar.set_description(f"[Epoch {epoch_num}: {epoch_step+1}/{steps_per_epoch}]loss: {loss_step:.7f}")
|
|
|
|
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.
|
|
|
|
|
|
|
|
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)
|
2022-11-23 19:00:00 +08:00
|
|
|
hypernetwork.eval_mode()
|
2022-11-20 11:35:26 +08:00
|
|
|
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,
|
|
|
|
)
|
|
|
|
|
|
|
|
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
|
2022-10-21 00:37:11 +08:00
|
|
|
|
2022-11-20 11:35:26 +08:00
|
|
|
preview_text = p.prompt
|
2022-10-21 00:37:11 +08:00
|
|
|
|
2022-11-20 11:35:26 +08:00
|
|
|
processed = processing.process_images(p)
|
|
|
|
image = processed.images[0] if len(processed.images) > 0 else None
|
2022-10-21 00:37:11 +08:00
|
|
|
|
2022-11-20 11:35:26 +08:00
|
|
|
if unload:
|
|
|
|
shared.sd_model.cond_stage_model.to(devices.cpu)
|
|
|
|
shared.sd_model.first_stage_model.to(devices.cpu)
|
2022-11-23 19:00:00 +08:00
|
|
|
hypernetwork.train_mode()
|
2022-11-20 11:35:26 +08:00
|
|
|
if image is not None:
|
|
|
|
shared.state.current_image = image
|
|
|
|
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}"
|
2022-10-08 04:22:22 +08:00
|
|
|
|
2022-11-20 11:35:26 +08:00
|
|
|
shared.state.job_no = hypernetwork.step
|
2022-10-08 04:22:22 +08:00
|
|
|
|
2022-11-20 11:35:26 +08:00
|
|
|
shared.state.textinfo = f"""
|
2022-10-08 04:22:22 +08:00
|
|
|
<p>
|
2022-11-20 11:35:26 +08:00
|
|
|
Loss: {loss_step:.7f}<br/>
|
2022-11-23 01:49:01 +08:00
|
|
|
Step: {steps_done}<br/>
|
2022-11-20 11:35:26 +08:00
|
|
|
Last prompt: {html.escape(batch.cond_text[0])}<br/>
|
2022-10-20 07:07:57 +08:00
|
|
|
Last saved hypernetwork: {html.escape(last_saved_file)}<br/>
|
2022-10-08 04:22:22 +08:00
|
|
|
Last saved image: {html.escape(last_saved_image)}<br/>
|
|
|
|
</p>
|
|
|
|
"""
|
2022-11-20 11:35:26 +08:00
|
|
|
except Exception:
|
|
|
|
print(traceback.format_exc(), file=sys.stderr)
|
|
|
|
finally:
|
|
|
|
pbar.leave = False
|
|
|
|
pbar.close()
|
2022-11-23 19:00:00 +08:00
|
|
|
hypernetwork.eval_mode()
|
2022-11-20 11:35:26 +08:00
|
|
|
#report_statistics(loss_dict)
|
2022-10-08 04:22:22 +08:00
|
|
|
|
2022-10-30 01:49:29 +08:00
|
|
|
filename = os.path.join(shared.cmd_opts.hypernetwork_dir, f'{hypernetwork_name}.pt')
|
2022-11-04 14:57:17 +08:00
|
|
|
hypernetwork.optimizer_name = optimizer_name
|
|
|
|
if shared.opts.save_optimizer_state:
|
|
|
|
hypernetwork.optimizer_state_dict = optimizer.state_dict()
|
2022-10-30 01:49:29 +08:00
|
|
|
save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, filename)
|
2022-11-04 14:57:17 +08:00
|
|
|
del optimizer
|
|
|
|
hypernetwork.optimizer_state_dict = None # dereference it after saving, to save memory.
|
2022-11-20 11:35:26 +08:00
|
|
|
shared.sd_model.cond_stage_model.to(devices.device)
|
|
|
|
shared.sd_model.first_stage_model.to(devices.device)
|
|
|
|
|
2022-10-08 04:22:22 +08:00
|
|
|
return hypernetwork, filename
|
2022-10-30 01:49:29 +08:00
|
|
|
|
|
|
|
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:
|
|
|
|
hypernetwork.sd_checkpoint = checkpoint.hash
|
|
|
|
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
|