Update modeling_harmon.py
Browse files- modeling_harmon.py +288 -288
modeling_harmon.py
CHANGED
@@ -1,288 +1,288 @@
|
|
1 |
-
import torch
|
2 |
-
import math
|
3 |
-
import numpy as np
|
4 |
-
import torch.nn as nn
|
5 |
-
import copy
|
6 |
-
from einops import rearrange
|
7 |
-
from torch.nn.modules.module import T
|
8 |
-
from transformers.cache_utils import DynamicCache
|
9 |
-
|
10 |
-
from tqdm import tqdm
|
11 |
-
from transformers import Qwen2ForCausalLM, Qwen2Config, PreTrainedModel
|
12 |
-
|
13 |
-
from .diffusion_utils import *
|
14 |
-
from .gaussian_diffusion import *
|
15 |
-
from .respace import *
|
16 |
-
from .misc import *
|
17 |
-
from .diffloss import *
|
18 |
-
|
19 |
-
|
20 |
-
from .configuration_harmon import HarmonConfig
|
21 |
-
from .vae import AutoencoderKL
|
22 |
-
from .mar import mar_base, mar_large, mar_huge
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
def build_mlp(hidden_size, projector_dim, z_dim):
|
27 |
-
return nn.Sequential(
|
28 |
-
nn.Linear(hidden_size, projector_dim),
|
29 |
-
nn.SiLU(),
|
30 |
-
nn.Linear(projector_dim, z_dim),)
|
31 |
-
|
32 |
-
|
33 |
-
def mask_by_order(mask_len, order, bsz, seq_len):
|
34 |
-
masking = torch.zeros(bsz, seq_len, device=order.device)
|
35 |
-
masking = torch.scatter(masking, dim=-1, index=order[:, :mask_len.long()],
|
36 |
-
src=torch.ones(bsz, seq_len, device=order.device)).bool()
|
37 |
-
return masking
|
38 |
-
|
39 |
-
|
40 |
-
class HarmonModel(PreTrainedModel):
|
41 |
-
config_class = HarmonConfig
|
42 |
-
|
43 |
-
def __init__(self, config: HarmonConfig):
|
44 |
-
super().__init__(config)
|
45 |
-
# VAE
|
46 |
-
self.vae = AutoencoderKL(
|
47 |
-
embed_dim=16,
|
48 |
-
ch_mult=(1, 1, 2, 2, 4)
|
49 |
-
)
|
50 |
-
self.vae_scale = 0.2325
|
51 |
-
|
52 |
-
# LLM
|
53 |
-
self.llm = Qwen2ForCausalLM(config=Qwen2Config.from_dict(config.llm))
|
54 |
-
|
55 |
-
# MAR
|
56 |
-
mar_config = copy.deepcopy(config.mar)
|
57 |
-
mar_type = mar_config.pop('type')
|
58 |
-
if mar_type == 'mar_base':
|
59 |
-
self.mar = mar_base(**mar_config)
|
60 |
-
elif mar_type == 'mar_large':
|
61 |
-
self.mar = mar_large(**mar_config)
|
62 |
-
elif mar_type == 'mar_huge':
|
63 |
-
self.mar = mar_huge(**mar_config)
|
64 |
-
else:
|
65 |
-
raise ValueError
|
66 |
-
|
67 |
-
# projection layers
|
68 |
-
self.proj_in = build_mlp(hidden_size=self.mar.encoder_embed_dim,
|
69 |
-
projector_dim=self.llm.config.hidden_size,
|
70 |
-
z_dim=self.llm.config.hidden_size)
|
71 |
-
self.proj_out = build_mlp(hidden_size=self.llm.config.hidden_size,
|
72 |
-
projector_dim=self.llm.config.hidden_size,
|
73 |
-
z_dim=self.mar.encoder_embed_dim)
|
74 |
-
|
75 |
-
@property
|
76 |
-
def llm_model(self):
|
77 |
-
return self.llm.model
|
78 |
-
|
79 |
-
@property
|
80 |
-
def device(self):
|
81 |
-
return self.llm.device
|
82 |
-
|
83 |
-
@property
|
84 |
-
def dtype(self):
|
85 |
-
return self.llm.dtype
|
86 |
-
|
87 |
-
@property
|
88 |
-
def gen_seq_len(self):
|
89 |
-
return self.mar.seq_len
|
90 |
-
|
91 |
-
@property
|
92 |
-
def token_embed_dim(self):
|
93 |
-
return self.vae.embed_dim * (self.mar.patch_size ** 2)
|
94 |
-
|
95 |
-
@torch.no_grad()
|
96 |
-
def encode(self, x):
|
97 |
-
posterior = self.vae.encode(x)
|
98 |
-
z = posterior.
|
99 |
-
z = rearrange(z, 'b c (m p) (n q) -> b m n (c p q)',
|
100 |
-
p=self.mar.patch_size, q=self.mar.patch_size)
|
101 |
-
|
102 |
-
return z
|
103 |
-
|
104 |
-
@torch.no_grad()
|
105 |
-
def decode(self, z):
|
106 |
-
z /= self.vae_scale
|
107 |
-
z = rearrange(z, 'b m n (c p q) -> b c (m p) (n q)',
|
108 |
-
p=self.mar.patch_size, q=self.mar.patch_size)
|
109 |
-
|
110 |
-
x = self.vae.decode(z)
|
111 |
-
return x
|
112 |
-
|
113 |
-
def prepare_forward_input(self,
|
114 |
-
x,
|
115 |
-
inputs_embeds=None,
|
116 |
-
input_ids=None,
|
117 |
-
attention_mask=None,
|
118 |
-
past_key_values=None):
|
119 |
-
b, l, _ = x.shape
|
120 |
-
attention_mask = attention_mask.to(device=self.device, dtype=torch.bool)
|
121 |
-
attention_mask = torch.cat([
|
122 |
-
attention_mask, attention_mask.new_ones(b, l)
|
123 |
-
], dim=1)
|
124 |
-
position_ids = torch.cumsum(attention_mask, dim=1) - 1
|
125 |
-
position_ids[position_ids < 0] = 0
|
126 |
-
|
127 |
-
# import pdb; pdb.set_trace()
|
128 |
-
|
129 |
-
# prepare context
|
130 |
-
if past_key_values is not None:
|
131 |
-
inputs_embeds = x
|
132 |
-
position_ids = position_ids[:, -l:]
|
133 |
-
else:
|
134 |
-
if inputs_embeds is None:
|
135 |
-
input_ids = input_ids.to(self.device)
|
136 |
-
inputs_embeds = self.llm.get_input_embeddings()(input_ids)
|
137 |
-
inputs_embeds = torch.cat([inputs_embeds, x], dim=1)
|
138 |
-
|
139 |
-
return dict(inputs_embeds=inputs_embeds,
|
140 |
-
attention_mask=attention_mask,
|
141 |
-
position_ids=position_ids,
|
142 |
-
past_key_values=past_key_values)
|
143 |
-
|
144 |
-
def extract_visual_feature(self, x, mask=None, detach=False):
|
145 |
-
b, m, n, _ = x.shape
|
146 |
-
x = x.view(b, m*n, -1)
|
147 |
-
# x: b mn c
|
148 |
-
if mask is None:
|
149 |
-
mask = torch.zeros_like(x[..., 0])
|
150 |
-
null_embeds = self.mar.fake_latent.expand(x.shape[0], -1)
|
151 |
-
x_enc = self.mar.forward_mae_encoder(x, mask, null_embeds, image_shape=(m, n))
|
152 |
-
|
153 |
-
z_enc = self.proj_in(x_enc)
|
154 |
-
# Move buffers to the end of the image sequence
|
155 |
-
z_enc = torch.cat([
|
156 |
-
z_enc[:, self.mar.buffer_size:],
|
157 |
-
z_enc[:, :self.mar.buffer_size]], dim=1)
|
158 |
-
|
159 |
-
if detach:
|
160 |
-
x_enc = x_enc.detach()
|
161 |
-
z_enc = z_enc.detach()
|
162 |
-
|
163 |
-
return x_enc, z_enc
|
164 |
-
|
165 |
-
def forward_mae_encoder(self, x, mask, detach=False, **context):
|
166 |
-
b, m, n, _ = x.shape
|
167 |
-
x_enc, z_enc = self.extract_visual_feature(x, mask=mask, detach=detach)
|
168 |
-
inputs = self.prepare_forward_input(x=z_enc, **context)
|
169 |
-
output = self.llm_model(**inputs, return_dict=True)
|
170 |
-
|
171 |
-
z_llm = output.last_hidden_state[:, -z_enc.shape[1]:]
|
172 |
-
|
173 |
-
# move buffers back to the start of the image sequence
|
174 |
-
z_llm = torch.cat([
|
175 |
-
z_llm[:, -self.mar.buffer_size:],
|
176 |
-
z_llm[:, :-self.mar.buffer_size]], dim=1)
|
177 |
-
|
178 |
-
# residual learning
|
179 |
-
x_enc = x_enc + self.proj_out(z_llm)
|
180 |
-
|
181 |
-
return x_enc
|
182 |
-
|
183 |
-
@staticmethod
|
184 |
-
def curtail_cache(past_key_values, cur_len):
|
185 |
-
for past_key_values_ in past_key_values:
|
186 |
-
keys, values = past_key_values_
|
187 |
-
keys.data = keys.data[:, :, :cur_len]
|
188 |
-
values.data = values.data[:, :, :cur_len]
|
189 |
-
|
190 |
-
@torch.no_grad()
|
191 |
-
def sample(self,
|
192 |
-
input_ids=None, inputs_embeds=None,
|
193 |
-
attention_mask=None, num_iter=64, cfg=1.0, cfg_schedule="constant", temperature=1.0,
|
194 |
-
progress=False, mask=None, past_key_values=None, image_shape=None, x_con=None, **kwargs):
|
195 |
-
if inputs_embeds is None and input_ids is not None:
|
196 |
-
inputs_embeds = self.llm.get_input_embeddings()(input_ids)
|
197 |
-
|
198 |
-
bsz = attention_mask.shape[0]
|
199 |
-
if cfg != 1.0:
|
200 |
-
assert bsz % 2 == 0
|
201 |
-
|
202 |
-
if image_shape is None:
|
203 |
-
m = n = int(self.gen_seq_len ** 0.5)
|
204 |
-
else:
|
205 |
-
m, n = image_shape
|
206 |
-
|
207 |
-
if mask is None:
|
208 |
-
mask = torch.ones(bsz, m*n, device=self.device, dtype=self.dtype)
|
209 |
-
else:
|
210 |
-
mask = mask.view(bsz, m*n)
|
211 |
-
tokens = torch.zeros(bsz, m*n, self.token_embed_dim,
|
212 |
-
device=self.device, dtype=self.dtype)
|
213 |
-
orders = self.mar.sample_orders(bsz, seq_len=m*n)
|
214 |
-
if cfg != 1.0:
|
215 |
-
orders[bsz//2:] = orders[:bsz//2]
|
216 |
-
|
217 |
-
indices = list(range(num_iter))
|
218 |
-
if progress:
|
219 |
-
indices = tqdm(indices)
|
220 |
-
|
221 |
-
# past key values can be prepared outside (usually in multi-turn editing)
|
222 |
-
if past_key_values is None:
|
223 |
-
output = self.llm_model(inputs_embeds=inputs_embeds,
|
224 |
-
attention_mask=None,
|
225 |
-
position_ids=None,
|
226 |
-
past_key_values=DynamicCache.from_legacy_cache(),
|
227 |
-
return_dict=True,
|
228 |
-
use_cache=True)
|
229 |
-
past_key_values = output.past_key_values
|
230 |
-
|
231 |
-
# generate latents
|
232 |
-
for step in indices:
|
233 |
-
cur_tokens = tokens.clone()
|
234 |
-
x_enc = self.forward_mae_encoder(tokens.view(bsz, m, n, -1),
|
235 |
-
mask.to(self.dtype),
|
236 |
-
past_key_values=past_key_values,
|
237 |
-
# inputs_embeds=inputs_embeds,
|
238 |
-
attention_mask=attention_mask)
|
239 |
-
# import pdb; pdb.set_trace()
|
240 |
-
self.curtail_cache(past_key_values, inputs_embeds.shape[1])
|
241 |
-
# import pdb; pdb.set_trace()
|
242 |
-
|
243 |
-
z = self.mar.forward_mae_decoder(x_enc, mask.to(self.dtype), image_shape=(m, n), x_con=x_con)
|
244 |
-
|
245 |
-
# mask ratio for the next round, following MaskGIT and MAGE.
|
246 |
-
mask_ratio = np.cos(math.pi / 2. * (step + 1) / num_iter)
|
247 |
-
mask_len = torch.Tensor([np.floor(m*n * mask_ratio)]).to(self.device)
|
248 |
-
|
249 |
-
# masks out at least one for the next iteration
|
250 |
-
mask_len = torch.maximum(torch.Tensor([1]).to(self.device),
|
251 |
-
torch.minimum(torch.sum(mask, dim=-1, keepdims=True) - 1, mask_len))
|
252 |
-
|
253 |
-
# get masking for next iteration and locations to be predicted in this iteration
|
254 |
-
mask_next = mask_by_order(mask_len[0], orders, bsz, m*n).to(self.device)
|
255 |
-
if cfg != 1.0:
|
256 |
-
mask_next[bsz//2:] = mask_next[:bsz//2]
|
257 |
-
if step >= num_iter - 1:
|
258 |
-
mask_to_pred = mask[:bsz].bool()
|
259 |
-
else:
|
260 |
-
mask_to_pred = torch.logical_xor(mask[:bsz].bool(), mask_next.bool())
|
261 |
-
mask = mask_next
|
262 |
-
# if not cfg == 1.0:
|
263 |
-
# mask_to_pred = torch.cat([mask_to_pred, mask_to_pred], dim=0)
|
264 |
-
|
265 |
-
# sample token latents for this step
|
266 |
-
z = z[mask_to_pred.nonzero(as_tuple=True)]
|
267 |
-
# cfg schedule follow Muse
|
268 |
-
if cfg_schedule == "linear":
|
269 |
-
cfg_iter = 1 + (cfg - 1) * (m*n - mask_len[0]) / (m*n)
|
270 |
-
elif cfg_schedule == "constant":
|
271 |
-
cfg_iter = cfg
|
272 |
-
else:
|
273 |
-
raise NotImplementedError
|
274 |
-
sampled_token_latent = self.mar.diffloss.sample(z, temperature, cfg_iter).to(self.dtype)
|
275 |
-
# if not cfg == 1.0:
|
276 |
-
# sampled_token_latent, _ = sampled_token_latent.chunk(2, dim=0) # Remove null class samples
|
277 |
-
# mask_to_pred, _ = mask_to_pred.chunk(2, dim=0)
|
278 |
-
|
279 |
-
cur_tokens[mask_to_pred.nonzero(as_tuple=True)] = sampled_token_latent
|
280 |
-
if cfg != 1.0:
|
281 |
-
cur_tokens[bsz//2:] = cur_tokens[:bsz//2]
|
282 |
-
tokens = cur_tokens.clone()
|
283 |
-
|
284 |
-
pred = self.decode(tokens.view(bsz, m, n, -1))
|
285 |
-
|
286 |
-
if cfg != 1.0:
|
287 |
-
pred = pred[:bsz//2]
|
288 |
-
return pred
|
|
|
1 |
+
import torch
|
2 |
+
import math
|
3 |
+
import numpy as np
|
4 |
+
import torch.nn as nn
|
5 |
+
import copy
|
6 |
+
from einops import rearrange
|
7 |
+
from torch.nn.modules.module import T
|
8 |
+
from transformers.cache_utils import DynamicCache
|
9 |
+
|
10 |
+
from tqdm import tqdm
|
11 |
+
from transformers import Qwen2ForCausalLM, Qwen2Config, PreTrainedModel
|
12 |
+
|
13 |
+
from .diffusion_utils import *
|
14 |
+
from .gaussian_diffusion import *
|
15 |
+
from .respace import *
|
16 |
+
from .misc import *
|
17 |
+
from .diffloss import *
|
18 |
+
|
19 |
+
|
20 |
+
from .configuration_harmon import HarmonConfig
|
21 |
+
from .vae import AutoencoderKL
|
22 |
+
from .mar import mar_base, mar_large, mar_huge
|
23 |
+
|
24 |
+
|
25 |
+
|
26 |
+
def build_mlp(hidden_size, projector_dim, z_dim):
|
27 |
+
return nn.Sequential(
|
28 |
+
nn.Linear(hidden_size, projector_dim),
|
29 |
+
nn.SiLU(),
|
30 |
+
nn.Linear(projector_dim, z_dim),)
|
31 |
+
|
32 |
+
|
33 |
+
def mask_by_order(mask_len, order, bsz, seq_len):
|
34 |
+
masking = torch.zeros(bsz, seq_len, device=order.device)
|
35 |
+
masking = torch.scatter(masking, dim=-1, index=order[:, :mask_len.long()],
|
36 |
+
src=torch.ones(bsz, seq_len, device=order.device)).bool()
|
37 |
+
return masking
|
38 |
+
|
39 |
+
|
40 |
+
class HarmonModel(PreTrainedModel):
|
41 |
+
config_class = HarmonConfig
|
42 |
+
|
43 |
+
def __init__(self, config: HarmonConfig):
|
44 |
+
super().__init__(config)
|
45 |
+
# VAE
|
46 |
+
self.vae = AutoencoderKL(
|
47 |
+
embed_dim=16,
|
48 |
+
ch_mult=(1, 1, 2, 2, 4)
|
49 |
+
)
|
50 |
+
self.vae_scale = 0.2325
|
51 |
+
|
52 |
+
# LLM
|
53 |
+
self.llm = Qwen2ForCausalLM(config=Qwen2Config.from_dict(config.llm))
|
54 |
+
|
55 |
+
# MAR
|
56 |
+
mar_config = copy.deepcopy(config.mar)
|
57 |
+
mar_type = mar_config.pop('type')
|
58 |
+
if mar_type == 'mar_base':
|
59 |
+
self.mar = mar_base(**mar_config)
|
60 |
+
elif mar_type == 'mar_large':
|
61 |
+
self.mar = mar_large(**mar_config)
|
62 |
+
elif mar_type == 'mar_huge':
|
63 |
+
self.mar = mar_huge(**mar_config)
|
64 |
+
else:
|
65 |
+
raise ValueError
|
66 |
+
|
67 |
+
# projection layers
|
68 |
+
self.proj_in = build_mlp(hidden_size=self.mar.encoder_embed_dim,
|
69 |
+
projector_dim=self.llm.config.hidden_size,
|
70 |
+
z_dim=self.llm.config.hidden_size)
|
71 |
+
self.proj_out = build_mlp(hidden_size=self.llm.config.hidden_size,
|
72 |
+
projector_dim=self.llm.config.hidden_size,
|
73 |
+
z_dim=self.mar.encoder_embed_dim)
|
74 |
+
|
75 |
+
@property
|
76 |
+
def llm_model(self):
|
77 |
+
return self.llm.model
|
78 |
+
|
79 |
+
@property
|
80 |
+
def device(self):
|
81 |
+
return self.llm.device
|
82 |
+
|
83 |
+
@property
|
84 |
+
def dtype(self):
|
85 |
+
return self.llm.dtype
|
86 |
+
|
87 |
+
@property
|
88 |
+
def gen_seq_len(self):
|
89 |
+
return self.mar.seq_len
|
90 |
+
|
91 |
+
@property
|
92 |
+
def token_embed_dim(self):
|
93 |
+
return self.vae.embed_dim * (self.mar.patch_size ** 2)
|
94 |
+
|
95 |
+
@torch.no_grad()
|
96 |
+
def encode(self, x):
|
97 |
+
posterior = self.vae.encode(x)
|
98 |
+
z = posterior.mode().mul_(self.vae_scale)
|
99 |
+
z = rearrange(z, 'b c (m p) (n q) -> b m n (c p q)',
|
100 |
+
p=self.mar.patch_size, q=self.mar.patch_size)
|
101 |
+
|
102 |
+
return z
|
103 |
+
|
104 |
+
@torch.no_grad()
|
105 |
+
def decode(self, z):
|
106 |
+
z /= self.vae_scale
|
107 |
+
z = rearrange(z, 'b m n (c p q) -> b c (m p) (n q)',
|
108 |
+
p=self.mar.patch_size, q=self.mar.patch_size)
|
109 |
+
|
110 |
+
x = self.vae.decode(z)
|
111 |
+
return x
|
112 |
+
|
113 |
+
def prepare_forward_input(self,
|
114 |
+
x,
|
115 |
+
inputs_embeds=None,
|
116 |
+
input_ids=None,
|
117 |
+
attention_mask=None,
|
118 |
+
past_key_values=None):
|
119 |
+
b, l, _ = x.shape
|
120 |
+
attention_mask = attention_mask.to(device=self.device, dtype=torch.bool)
|
121 |
+
attention_mask = torch.cat([
|
122 |
+
attention_mask, attention_mask.new_ones(b, l)
|
123 |
+
], dim=1)
|
124 |
+
position_ids = torch.cumsum(attention_mask, dim=1) - 1
|
125 |
+
position_ids[position_ids < 0] = 0
|
126 |
+
|
127 |
+
# import pdb; pdb.set_trace()
|
128 |
+
|
129 |
+
# prepare context
|
130 |
+
if past_key_values is not None:
|
131 |
+
inputs_embeds = x
|
132 |
+
position_ids = position_ids[:, -l:]
|
133 |
+
else:
|
134 |
+
if inputs_embeds is None:
|
135 |
+
input_ids = input_ids.to(self.device)
|
136 |
+
inputs_embeds = self.llm.get_input_embeddings()(input_ids)
|
137 |
+
inputs_embeds = torch.cat([inputs_embeds, x], dim=1)
|
138 |
+
|
139 |
+
return dict(inputs_embeds=inputs_embeds,
|
140 |
+
attention_mask=attention_mask,
|
141 |
+
position_ids=position_ids,
|
142 |
+
past_key_values=past_key_values)
|
143 |
+
|
144 |
+
def extract_visual_feature(self, x, mask=None, detach=False):
|
145 |
+
b, m, n, _ = x.shape
|
146 |
+
x = x.view(b, m*n, -1)
|
147 |
+
# x: b mn c
|
148 |
+
if mask is None:
|
149 |
+
mask = torch.zeros_like(x[..., 0])
|
150 |
+
null_embeds = self.mar.fake_latent.expand(x.shape[0], -1)
|
151 |
+
x_enc = self.mar.forward_mae_encoder(x, mask, null_embeds, image_shape=(m, n))
|
152 |
+
|
153 |
+
z_enc = self.proj_in(x_enc)
|
154 |
+
# Move buffers to the end of the image sequence
|
155 |
+
z_enc = torch.cat([
|
156 |
+
z_enc[:, self.mar.buffer_size:],
|
157 |
+
z_enc[:, :self.mar.buffer_size]], dim=1)
|
158 |
+
|
159 |
+
if detach:
|
160 |
+
x_enc = x_enc.detach()
|
161 |
+
z_enc = z_enc.detach()
|
162 |
+
|
163 |
+
return x_enc, z_enc
|
164 |
+
|
165 |
+
def forward_mae_encoder(self, x, mask, detach=False, **context):
|
166 |
+
b, m, n, _ = x.shape
|
167 |
+
x_enc, z_enc = self.extract_visual_feature(x, mask=mask, detach=detach)
|
168 |
+
inputs = self.prepare_forward_input(x=z_enc, **context)
|
169 |
+
output = self.llm_model(**inputs, return_dict=True)
|
170 |
+
|
171 |
+
z_llm = output.last_hidden_state[:, -z_enc.shape[1]:]
|
172 |
+
|
173 |
+
# move buffers back to the start of the image sequence
|
174 |
+
z_llm = torch.cat([
|
175 |
+
z_llm[:, -self.mar.buffer_size:],
|
176 |
+
z_llm[:, :-self.mar.buffer_size]], dim=1)
|
177 |
+
|
178 |
+
# residual learning
|
179 |
+
x_enc = x_enc + self.proj_out(z_llm)
|
180 |
+
|
181 |
+
return x_enc
|
182 |
+
|
183 |
+
@staticmethod
|
184 |
+
def curtail_cache(past_key_values, cur_len):
|
185 |
+
for past_key_values_ in past_key_values:
|
186 |
+
keys, values = past_key_values_
|
187 |
+
keys.data = keys.data[:, :, :cur_len]
|
188 |
+
values.data = values.data[:, :, :cur_len]
|
189 |
+
|
190 |
+
@torch.no_grad()
|
191 |
+
def sample(self,
|
192 |
+
input_ids=None, inputs_embeds=None,
|
193 |
+
attention_mask=None, num_iter=64, cfg=1.0, cfg_schedule="constant", temperature=1.0,
|
194 |
+
progress=False, mask=None, past_key_values=None, image_shape=None, x_con=None, **kwargs):
|
195 |
+
if inputs_embeds is None and input_ids is not None:
|
196 |
+
inputs_embeds = self.llm.get_input_embeddings()(input_ids)
|
197 |
+
|
198 |
+
bsz = attention_mask.shape[0]
|
199 |
+
if cfg != 1.0:
|
200 |
+
assert bsz % 2 == 0
|
201 |
+
|
202 |
+
if image_shape is None:
|
203 |
+
m = n = int(self.gen_seq_len ** 0.5)
|
204 |
+
else:
|
205 |
+
m, n = image_shape
|
206 |
+
|
207 |
+
if mask is None:
|
208 |
+
mask = torch.ones(bsz, m*n, device=self.device, dtype=self.dtype)
|
209 |
+
else:
|
210 |
+
mask = mask.view(bsz, m*n)
|
211 |
+
tokens = torch.zeros(bsz, m*n, self.token_embed_dim,
|
212 |
+
device=self.device, dtype=self.dtype)
|
213 |
+
orders = self.mar.sample_orders(bsz, seq_len=m*n)
|
214 |
+
if cfg != 1.0:
|
215 |
+
orders[bsz//2:] = orders[:bsz//2]
|
216 |
+
|
217 |
+
indices = list(range(num_iter))
|
218 |
+
if progress:
|
219 |
+
indices = tqdm(indices)
|
220 |
+
|
221 |
+
# past key values can be prepared outside (usually in multi-turn editing)
|
222 |
+
if past_key_values is None:
|
223 |
+
output = self.llm_model(inputs_embeds=inputs_embeds,
|
224 |
+
attention_mask=None,
|
225 |
+
position_ids=None,
|
226 |
+
past_key_values=DynamicCache.from_legacy_cache(),
|
227 |
+
return_dict=True,
|
228 |
+
use_cache=True)
|
229 |
+
past_key_values = output.past_key_values
|
230 |
+
|
231 |
+
# generate latents
|
232 |
+
for step in indices:
|
233 |
+
cur_tokens = tokens.clone()
|
234 |
+
x_enc = self.forward_mae_encoder(tokens.view(bsz, m, n, -1),
|
235 |
+
mask.to(self.dtype),
|
236 |
+
past_key_values=past_key_values,
|
237 |
+
# inputs_embeds=inputs_embeds,
|
238 |
+
attention_mask=attention_mask)
|
239 |
+
# import pdb; pdb.set_trace()
|
240 |
+
self.curtail_cache(past_key_values, inputs_embeds.shape[1])
|
241 |
+
# import pdb; pdb.set_trace()
|
242 |
+
|
243 |
+
z = self.mar.forward_mae_decoder(x_enc, mask.to(self.dtype), image_shape=(m, n), x_con=x_con)
|
244 |
+
|
245 |
+
# mask ratio for the next round, following MaskGIT and MAGE.
|
246 |
+
mask_ratio = np.cos(math.pi / 2. * (step + 1) / num_iter)
|
247 |
+
mask_len = torch.Tensor([np.floor(m*n * mask_ratio)]).to(self.device)
|
248 |
+
|
249 |
+
# masks out at least one for the next iteration
|
250 |
+
mask_len = torch.maximum(torch.Tensor([1]).to(self.device),
|
251 |
+
torch.minimum(torch.sum(mask, dim=-1, keepdims=True) - 1, mask_len))
|
252 |
+
|
253 |
+
# get masking for next iteration and locations to be predicted in this iteration
|
254 |
+
mask_next = mask_by_order(mask_len[0], orders, bsz, m*n).to(self.device)
|
255 |
+
if cfg != 1.0:
|
256 |
+
mask_next[bsz//2:] = mask_next[:bsz//2]
|
257 |
+
if step >= num_iter - 1:
|
258 |
+
mask_to_pred = mask[:bsz].bool()
|
259 |
+
else:
|
260 |
+
mask_to_pred = torch.logical_xor(mask[:bsz].bool(), mask_next.bool())
|
261 |
+
mask = mask_next
|
262 |
+
# if not cfg == 1.0:
|
263 |
+
# mask_to_pred = torch.cat([mask_to_pred, mask_to_pred], dim=0)
|
264 |
+
|
265 |
+
# sample token latents for this step
|
266 |
+
z = z[mask_to_pred.nonzero(as_tuple=True)]
|
267 |
+
# cfg schedule follow Muse
|
268 |
+
if cfg_schedule == "linear":
|
269 |
+
cfg_iter = 1 + (cfg - 1) * (m*n - mask_len[0]) / (m*n)
|
270 |
+
elif cfg_schedule == "constant":
|
271 |
+
cfg_iter = cfg
|
272 |
+
else:
|
273 |
+
raise NotImplementedError
|
274 |
+
sampled_token_latent = self.mar.diffloss.sample(z, temperature, cfg_iter).to(self.dtype)
|
275 |
+
# if not cfg == 1.0:
|
276 |
+
# sampled_token_latent, _ = sampled_token_latent.chunk(2, dim=0) # Remove null class samples
|
277 |
+
# mask_to_pred, _ = mask_to_pred.chunk(2, dim=0)
|
278 |
+
|
279 |
+
cur_tokens[mask_to_pred.nonzero(as_tuple=True)] = sampled_token_latent
|
280 |
+
if cfg != 1.0:
|
281 |
+
cur_tokens[bsz//2:] = cur_tokens[:bsz//2]
|
282 |
+
tokens = cur_tokens.clone()
|
283 |
+
|
284 |
+
pred = self.decode(tokens.view(bsz, m, n, -1))
|
285 |
+
|
286 |
+
if cfg != 1.0:
|
287 |
+
pred = pred[:bsz//2]
|
288 |
+
return pred
|