File size: 10,035 Bytes
f7009b3 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 |
import torch
from abc import ABC
from torch import nn
from torch.nn import functional as F
from .diffusion import DiffusionLoss, DDIMSampler, DDPMSampler
from .transformer import TransformerModel
from .mamba import MambaModel
from .lstm import LstmModel
from .gatemlp import GMLPModel
class ModelDiffusion(nn.Module, ABC):
config = {}
def __init__(self, sequence_length):
super().__init__()
DiffusionLoss.config = self.config
self.criteria = DiffusionLoss()
if self.config.get("post_d_model") is None:
assert self.config["d_model"] == self.config["condition_dim"]
self.sequence_length = sequence_length
# to define model after this function
self.to_condition = nn.Linear(self.config["d_condition"], self.config["d_model"])
self.to_permutation_state = nn.Embedding(self.config["num_permutation"], self.config["d_model"])
self.to_permutation_state.weight = \
nn.Parameter(torch.ones_like(self.to_permutation_state.weight) / self.config["d_model"])
def forward(self, output_shape=None, x_0=None, condition=None, permutation_state=None, **kwargs):
# condition
if condition is not None:
assert len(condition.shape) == 2
assert condition.shape[-1] == self.config["d_condition"]
condition = self.to_condition(condition.to(self.device)[:, None, :])
else: # not use condition
condition = self.to_condition(torch.zeros(size=(1, 1, 1), device=self.device))
# process
if kwargs.get("sample"):
if permutation_state is not False:
permutation_state = torch.randint(0, self.to_permutation_state.num_embeddings, (1,), device=self.device)
permutation_state = self.to_permutation_state(permutation_state)[:, None, :]
else: # permutation state == False
permutation_state = 0.
return self.sample(x=None, condition=condition+permutation_state)
else: # train
if permutation_state is not None:
permutation_state = self.to_permutation_state(permutation_state)[:, None, :]
else: # not use permutation state
permutation_state = 0.
# Given condition c and ground truth token x, compute loss
c = self.model(output_shape, condition+permutation_state)
loss = self.criteria(x=x_0, c=c, **kwargs)
return loss
@torch.no_grad()
def sample(self, x=None, condition=None):
z = self.model([1, self.sequence_length, self.config["d_model"]], condition)
if x is None:
x = torch.randn((1, self.sequence_length, self.config["model_dim"]), device=z.device)
x = self.criteria.sample(x, z)
return x
@property
def device(self):
return next(self.parameters()).device
class ModelMSELoss(nn.Module, ABC):
config = {}
def __init__(self, sequence_length):
super().__init__()
if self.config.get("post_d_model") is None:
assert self.config["d_model"] == self.config["condition_dim"]
self.sequence_length = sequence_length
# to define model after this function
self.to_condition = nn.Linear(self.config["d_condition"], self.config["d_model"])
self.to_permutation_state = nn.Embedding(self.config["num_permutation"], self.config["d_model"])
self.to_permutation_state.weight = \
nn.Parameter(torch.ones_like(self.to_permutation_state.weight) / self.config["d_model"])
def forward(self, output_shape=None, x_0=None, condition=None, permutation_state=None, **kwargs):
# condition
if condition is not None:
assert len(condition.shape) == 2
assert condition.shape[-1] == self.config["d_condition"]
condition = self.to_condition(condition.to(self.device)[:, None, :])
else: # not use condition
condition = self.to_condition(torch.zeros(size=(1, 1, 1), device=self.device))
# process
if kwargs.get("sample"):
if permutation_state is not False:
permutation_state = torch.randint(0, self.to_permutation_state.num_embeddings, (1,), device=self.device)
permutation_state = self.to_permutation_state(permutation_state)[:, None, :]
else: # permutation state == False
permutation_state = 0.
return self.sample(x=None, condition=condition+permutation_state)
else: # train
if permutation_state is not None:
permutation_state = self.to_permutation_state(permutation_state)[:, None, :]
else: # not use permutation state
permutation_state = 0.
# Given condition c and ground truth token x, compute loss
c = self.model(output_shape, condition+permutation_state)
assert c.shape[-1] == x_0.shape[-1], "d_model should be equal to dim_per_token"
# preprocess nan to zero
mask = torch.isnan(x_0)
x_0 = torch.nan_to_num(x_0, 0.)
# get the gradient
loss = F.mse_loss(c, x_0, reduction="none")
loss[mask] = torch.nan
return loss.nanmean()
@torch.no_grad()
def sample(self, x=None, condition=None):
z = self.model([1, self.sequence_length, self.config["d_model"]], condition)
return z
@property
def device(self):
return next(self.parameters()).device
class MambaDiffusion(ModelDiffusion):
def __init__(self, sequence_length, positional_embedding):
super().__init__(sequence_length=sequence_length)
MambaModel.config = self.config
self.model = MambaModel(positional_embedding=positional_embedding)
class TransformerDiffusion(ModelDiffusion):
def __init__(self, sequence_length, positional_embedding):
super().__init__(sequence_length=sequence_length)
TransformerModel.config = self.config
self.model = TransformerModel(positional_embedding=positional_embedding)
class LstmDiffusion(ModelDiffusion):
def __init__(self, sequence_length, positional_embedding):
super().__init__(sequence_length=sequence_length)
LstmModel.config = self.config
self.model = LstmModel(positional_embedding=positional_embedding)
class GMLPDiffusion(ModelDiffusion):
def __init__(self, sequence_length, positional_embedding):
super().__init__(sequence_length=sequence_length)
GMLPModel.config = self.config
self.model = GMLPModel(positional_embedding=positional_embedding)
class MambaMSELoss(ModelMSELoss):
def __init__(self, sequence_length, positional_embedding):
super().__init__(sequence_length=sequence_length)
MambaModel.config = self.config
self.model = MambaModel(positional_embedding=positional_embedding)
class ClassConditionMambaDiffusion(MambaDiffusion):
def __init__(self, sequence_length, positional_embedding, input_class=10):
super().__init__(sequence_length, positional_embedding)
self.get_condition = nn.Sequential(
nn.Linear(input_class, self.config["d_condition"]),
nn.SiLU(),
) # to condition
self.to_permutation_state = nn.Embedding(self.config["num_permutation"], self.config["d_model"])
# condition module
self.to_condition_linear = nn.Linear(self.config["d_condition"], self.config["d_model"])
to_condition_gate = torch.zeros(size=(1, sequence_length, 1))
to_condition_gate[:, -8:, :] = 1.
self.register_buffer("to_condition_gate", to_condition_gate)
# reset to_condition
del self.to_condition
self.to_condition = self._to_condition
def forward(self, output_shape=None, x_0=None, condition=None, **kwargs):
condition = self.get_condition(condition.to(self.device))
return super().forward(output_shape=output_shape, x_0=x_0, condition=condition, **kwargs)
def _to_condition(self, x):
assert len(x.shape) == 3
x = self.to_condition_linear(x)
x = x * self.to_condition_gate
return x
class ClassConditionMambaDiffusionFull(MambaDiffusion):
def __init__(self, sequence_length, positional_embedding, input_class=10, init_noise_intensity=1e-4):
super().__init__(sequence_length, positional_embedding)
self.get_condition = nn.Sequential(
nn.Linear(input_class, self.config["d_condition"]),
nn.LayerNorm(self.config["d_condition"]),
) # to condition
self.to_permutation_state = nn.Embedding(self.config["num_permutation"], self.config["d_model"])
# condition module
self.to_condition_linear = nn.Linear(self.config["d_condition"], self.config["d_model"])
self.to_condition_conv = nn.Sequential(
nn.Conv1d(1, sequence_length, 9, 1, 4),
nn.GroupNorm(num_groups=1, num_channels=sequence_length),
nn.Conv1d(sequence_length, sequence_length, 9, 1, 4),
) # [batch_size, sequence_length, d_model]
# reset to_condition
del self.to_condition
def forward(self, output_shape=None, x_0=None, condition=None, **kwargs):
if kwargs.get("pre_training"):
self.to_condition = self._zero_condition
condition = None
else: # train with condition
self.to_condition = self._to_condition
condition = self.get_condition(condition.to(self.device))
return super().forward(output_shape=output_shape, x_0=x_0, condition=condition, **kwargs)
def _to_condition(self, x):
assert len(x.shape) == 3
x = self.to_condition_linear(x) # [batch_size, 1, d_model]
x = self.to_condition_conv(x) # [batch_size, sequence_length, d_model]
return x
def _zero_condition(self, x):
return torch.zeros(size=(x.shape[0], self.sequence_length, self.config["d_model"]), device=x.device) |