Spaces:
Runtime error
Runtime error
Singularity666
commited on
Commit
•
eae59cc
1
Parent(s):
8724709
Update app.py
Browse files
app.py
CHANGED
@@ -1,4 +1,3 @@
|
|
1 |
-
#mgie_llava.py:
|
2 |
from typing import List, Optional, Tuple, Union
|
3 |
|
4 |
import torch
|
@@ -19,11 +18,9 @@ DEFAULT_IMAGE_PATCH_TOKEN = "<im_patch>"
|
|
19 |
DEFAULT_IM_START_TOKEN = "<im_start>"
|
20 |
DEFAULT_IM_END_TOKEN = "<im_end>"
|
21 |
|
22 |
-
|
23 |
class LlavaConfig(LlamaConfig):
|
24 |
model_type = "llava"
|
25 |
|
26 |
-
|
27 |
class LlavaLlamaModel(LlamaModel):
|
28 |
config_class = LlavaConfig
|
29 |
|
@@ -31,9 +28,7 @@ class LlavaLlamaModel(LlamaModel):
|
|
31 |
super(LlavaLlamaModel, self).__init__(config)
|
32 |
|
33 |
if hasattr(config, "mm_vision_tower"):
|
34 |
-
# HACK: for FSDP
|
35 |
self.vision_tower = [CLIPVisionModel.from_pretrained(config.mm_vision_tower)]
|
36 |
-
# self.vision_tower = CLIPVisionModel.from_pretrained(config.mm_vision_tower)
|
37 |
|
38 |
if hasattr(config, "use_mm_proj"):
|
39 |
self.mm_projector = nn.Linear(config.mm_hidden_size, config.hidden_size)
|
@@ -94,22 +89,15 @@ class LlavaLlamaModel(LlamaModel):
|
|
94 |
return_dict: Optional[bool] = None,
|
95 |
) -> Union[Tuple, BaseModelOutputWithPast]:
|
96 |
|
97 |
-
# HACK: replace back original embeddings for LLaVA pretraining
|
98 |
orig_embeds_params = getattr(self, 'orig_embeds_params', None)
|
99 |
-
# if orig_embeds_params is not None:
|
100 |
-
# orig_embeds_params = orig_embeds_params[0]
|
101 |
-
# with torch.no_grad():
|
102 |
-
# self.get_input_embeddings().weight.data[:-2] = orig_embeds_params[:-2].data
|
103 |
|
104 |
if inputs_embeds is None:
|
105 |
inputs_embeds = self.embed_tokens(input_ids)
|
106 |
|
107 |
vision_tower = self.get_vision_tower()
|
108 |
if vision_tower is not None and (input_ids.shape[1] != 1 or self.training) and images is not None:
|
109 |
-
# TODO: this is a modified multimodal LLM -- Haotian Liu
|
110 |
with torch.no_grad():
|
111 |
if type(images) is list:
|
112 |
-
# variable length images
|
113 |
image_features = []
|
114 |
for image in images:
|
115 |
image_forward_out = vision_tower(image.unsqueeze(0), output_hidden_states=True)
|
@@ -133,7 +121,6 @@ class LlavaLlamaModel(LlamaModel):
|
|
133 |
cur_image_idx = 0
|
134 |
for cur_input_ids, cur_input_embeds in zip(input_ids, inputs_embeds):
|
135 |
if (cur_input_ids == vision_tower.config.im_patch_token).sum() == 0:
|
136 |
-
# multimodal LLM, but the current sample is not multimodal
|
137 |
cur_input_embeds = cur_input_embeds + (0. * dummy_image_features).sum()
|
138 |
new_input_embeds.append(cur_input_embeds)
|
139 |
cur_image_idx += 1
|
@@ -191,7 +178,7 @@ class EditMapper(nn.Module):
|
|
191 |
self.hid2feat = nn.Linear(512, 768)
|
192 |
|
193 |
def forward(self, llm, emb):
|
194 |
-
hid = self.llm2hid(llm+emb)
|
195 |
hid = self.mapper(hid, self.query.repeat(llm.shape[0], 1, 1))
|
196 |
feat = self.hid2feat(hid)
|
197 |
|
@@ -208,18 +195,19 @@ class LlavaLlamaForCausalLM(LlamaForCausalLM):
|
|
208 |
|
209 |
self.edit_head = EditMapper()
|
210 |
|
211 |
-
|
212 |
-
|
213 |
-
|
|
|
|
|
214 |
self.vae.requires_grad_(False)
|
215 |
self.unet.register_to_config(in_channels=8)
|
216 |
with torch.no_grad():
|
217 |
conv = torch.nn.Conv2d(8, self.unet.conv_in.out_channels, self.unet.conv_in.kernel_size, self.unet.conv_in.stride, self.unet.conv_in.padding)
|
218 |
conv.weight.zero_()
|
219 |
conv.weight[:, :4, :, :].copy_(self.unet.conv_in.weight)
|
220 |
-
self.unet.conv_in = conv
|
221 |
|
222 |
-
# Initialize weights and apply final processing
|
223 |
self.post_init()
|
224 |
|
225 |
def get_model(self):
|
@@ -228,13 +216,6 @@ class LlavaLlamaForCausalLM(LlamaForCausalLM):
|
|
228 |
def get_vision_tower(self):
|
229 |
return self.get_model().get_vision_tower()
|
230 |
|
231 |
-
def get_vision_tower(self):
|
232 |
-
model = self.get_model()
|
233 |
-
vision_tower = model.vision_tower
|
234 |
-
if type(vision_tower) is list:
|
235 |
-
vision_tower = vision_tower[0]
|
236 |
-
return vision_tower
|
237 |
-
|
238 |
def forward(
|
239 |
self,
|
240 |
input_ids: torch.LongTensor = None,
|
@@ -250,12 +231,9 @@ class LlavaLlamaForCausalLM(LlamaForCausalLM):
|
|
250 |
p2p_inp=None, p2p_ans=None
|
251 |
) -> Union[Tuple, CausalLMOutputWithPast]:
|
252 |
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
253 |
-
output_hidden_states =
|
254 |
-
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
255 |
-
)
|
256 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
257 |
|
258 |
-
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
|
259 |
outputs = self.model(
|
260 |
input_ids=input_ids,
|
261 |
attention_mask=attention_mask,
|
@@ -273,24 +251,23 @@ class LlavaLlamaForCausalLM(LlamaForCausalLM):
|
|
273 |
|
274 |
loss = None
|
275 |
if labels is not None:
|
276 |
-
# Shift so that tokens < n predict n
|
277 |
shift_logits = logits[..., :-1, :].contiguous()
|
278 |
shift_labels = labels[..., 1:].contiguous()
|
279 |
-
# Flatten the tokens
|
280 |
loss_fct = CrossEntropyLoss()
|
281 |
shift_logits = shift_logits.view(-1, self.config.vocab_size)
|
282 |
shift_labels = shift_labels.view(-1)
|
283 |
-
# Enable model/pipeline parallelism
|
284 |
shift_labels = shift_labels.to(shift_logits.device)
|
285 |
loss = loss_fct(shift_logits, shift_labels)
|
286 |
|
287 |
if labels is not None:
|
288 |
llm = []
|
289 |
for i in range(labels.shape[0]):
|
290 |
-
try:
|
291 |
-
|
292 |
-
|
293 |
-
|
|
|
|
|
294 |
llm = torch.cat(llm, dim=0)
|
295 |
hid_edit = self.edit_head(llm, self.model.embed_tokens.weight[-8:].unsqueeze(dim=0).repeat(labels.shape[0], 1, 1))
|
296 |
|
@@ -300,25 +277,41 @@ class LlavaLlamaForCausalLM(LlamaForCausalLM):
|
|
300 |
self.model.embed_tokens.weight[-8:].unsqueeze(dim=0).repeat(labels.shape[0], 1, 1))
|
301 |
|
302 |
with torch.no_grad():
|
303 |
-
lat_ans, lat_inp = self.vae.encode(p2p_ans).latent_dist.sample()*self.vae.config.scaling_factor, self.vae.encode(p2p_inp).latent_dist.mode()
|
304 |
lat_ans, lat_inp = [torch.from_numpy(lat_ans.data.cpu().float().numpy()).to(lat_ans.device),
|
305 |
torch.from_numpy(lat_inp.data.cpu().float().numpy()).to(lat_inp.device)]
|
306 |
|
307 |
noise = torch.randn_like(lat_ans)
|
308 |
-
ts = torch.randint(0, self.scheduler.config.num_train_timesteps, (B,
|
309 |
lat_noise = self.scheduler.add_noise(lat_ans, noise, ts)
|
310 |
|
311 |
prob = torch.rand(B, device=lat_ans.device)
|
312 |
-
mask = (prob<(DROP*2)).reshape(B, 1, 1)
|
313 |
hid_edit = torch.where(mask, hid_null, hid_edit)
|
314 |
-
mask = (1.0-((prob>=DROP).to(lat_inp.dtype)*(prob<(DROP*3)).to(lat_inp.dtype))).reshape(B, 1, 1, 1)
|
315 |
lat_inp *= mask
|
316 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
317 |
out = self.unet(torch.cat([lat_noise, lat_inp], dim=1), ts, hid_edit).sample
|
318 |
|
319 |
loss_ce, loss_edit = loss, nn.functional.mse_loss(out, noise, reduction='mean')
|
320 |
-
if int(os.environ
|
321 |
-
loss = loss_ce+loss_edit*0.5
|
322 |
|
323 |
if not return_dict:
|
324 |
output = (logits,) + outputs[1:]
|
@@ -338,7 +331,6 @@ class LlavaLlamaForCausalLM(LlamaForCausalLM):
|
|
338 |
if past_key_values:
|
339 |
input_ids = input_ids[:, -1:]
|
340 |
|
341 |
-
# if `inputs_embeds` are passed, we only want to use them in the 1st generation step
|
342 |
if inputs_embeds is not None and past_key_values is None:
|
343 |
model_inputs = {"inputs_embeds": inputs_embeds}
|
344 |
else:
|
@@ -370,10 +362,8 @@ class LlavaLlamaForCausalLM(LlamaForCausalLM):
|
|
370 |
input_embeddings = self.get_input_embeddings().weight.data
|
371 |
output_embeddings = self.get_output_embeddings().weight.data
|
372 |
|
373 |
-
input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(
|
374 |
-
|
375 |
-
output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(
|
376 |
-
dim=0, keepdim=True)
|
377 |
|
378 |
input_embeddings[-num_new_tokens:] = input_embeddings_avg
|
379 |
output_embeddings[-num_new_tokens:] = output_embeddings_avg
|
@@ -394,9 +384,9 @@ class LlavaLlamaForCausalLM(LlamaForCausalLM):
|
|
394 |
elif embed_tokens_weight.shape[0] == num_new_tokens:
|
395 |
input_embeddings[-num_new_tokens:] = embed_tokens_weight
|
396 |
else:
|
397 |
-
raise ValueError(f"Unexpected embed_tokens_weight shape. Pretrained: {embed_tokens_weight.shape}. Current: {input_embeddings.shape}.
|
398 |
|
399 |
vision_config.im_patch_token = tokenizer.convert_tokens_to_ids([DEFAULT_IMAGE_PATCH_TOKEN])[0]
|
400 |
|
401 |
AutoConfig.register("llava", LlavaConfig)
|
402 |
-
AutoModelForCausalLM.register(LlavaConfig, LlavaLlamaForCausalLM)
|
|
|
|
|
1 |
from typing import List, Optional, Tuple, Union
|
2 |
|
3 |
import torch
|
|
|
18 |
DEFAULT_IM_START_TOKEN = "<im_start>"
|
19 |
DEFAULT_IM_END_TOKEN = "<im_end>"
|
20 |
|
|
|
21 |
class LlavaConfig(LlamaConfig):
|
22 |
model_type = "llava"
|
23 |
|
|
|
24 |
class LlavaLlamaModel(LlamaModel):
|
25 |
config_class = LlavaConfig
|
26 |
|
|
|
28 |
super(LlavaLlamaModel, self).__init__(config)
|
29 |
|
30 |
if hasattr(config, "mm_vision_tower"):
|
|
|
31 |
self.vision_tower = [CLIPVisionModel.from_pretrained(config.mm_vision_tower)]
|
|
|
32 |
|
33 |
if hasattr(config, "use_mm_proj"):
|
34 |
self.mm_projector = nn.Linear(config.mm_hidden_size, config.hidden_size)
|
|
|
89 |
return_dict: Optional[bool] = None,
|
90 |
) -> Union[Tuple, BaseModelOutputWithPast]:
|
91 |
|
|
|
92 |
orig_embeds_params = getattr(self, 'orig_embeds_params', None)
|
|
|
|
|
|
|
|
|
93 |
|
94 |
if inputs_embeds is None:
|
95 |
inputs_embeds = self.embed_tokens(input_ids)
|
96 |
|
97 |
vision_tower = self.get_vision_tower()
|
98 |
if vision_tower is not None and (input_ids.shape[1] != 1 or self.training) and images is not None:
|
|
|
99 |
with torch.no_grad():
|
100 |
if type(images) is list:
|
|
|
101 |
image_features = []
|
102 |
for image in images:
|
103 |
image_forward_out = vision_tower(image.unsqueeze(0), output_hidden_states=True)
|
|
|
121 |
cur_image_idx = 0
|
122 |
for cur_input_ids, cur_input_embeds in zip(input_ids, inputs_embeds):
|
123 |
if (cur_input_ids == vision_tower.config.im_patch_token).sum() == 0:
|
|
|
124 |
cur_input_embeds = cur_input_embeds + (0. * dummy_image_features).sum()
|
125 |
new_input_embeds.append(cur_input_embeds)
|
126 |
cur_image_idx += 1
|
|
|
178 |
self.hid2feat = nn.Linear(512, 768)
|
179 |
|
180 |
def forward(self, llm, emb):
|
181 |
+
hid = self.llm2hid(llm + emb)
|
182 |
hid = self.mapper(hid, self.query.repeat(llm.shape[0], 1, 1))
|
183 |
feat = self.hid2feat(hid)
|
184 |
|
|
|
195 |
|
196 |
self.edit_head = EditMapper()
|
197 |
|
198 |
+
self.scheduler, self.vae, self.unet = [
|
199 |
+
diffusers.DDPMScheduler.from_pretrained('runwayml/stable-diffusion-v1-5', subfolder='scheduler'),
|
200 |
+
diffusers.AutoencoderKL.from_pretrained('runwayml/stable-diffusion-v1-5', subfolder='vae'),
|
201 |
+
diffusers.UNet2DConditionModel.from_pretrained('runwayml/stable-diffusion-v1-5', subfolder='unet')
|
202 |
+
]
|
203 |
self.vae.requires_grad_(False)
|
204 |
self.unet.register_to_config(in_channels=8)
|
205 |
with torch.no_grad():
|
206 |
conv = torch.nn.Conv2d(8, self.unet.conv_in.out_channels, self.unet.conv_in.kernel_size, self.unet.conv_in.stride, self.unet.conv_in.padding)
|
207 |
conv.weight.zero_()
|
208 |
conv.weight[:, :4, :, :].copy_(self.unet.conv_in.weight)
|
209 |
+
self.unet.conv_in = conv
|
210 |
|
|
|
211 |
self.post_init()
|
212 |
|
213 |
def get_model(self):
|
|
|
216 |
def get_vision_tower(self):
|
217 |
return self.get_model().get_vision_tower()
|
218 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
219 |
def forward(
|
220 |
self,
|
221 |
input_ids: torch.LongTensor = None,
|
|
|
231 |
p2p_inp=None, p2p_ans=None
|
232 |
) -> Union[Tuple, CausalLMOutputWithPast]:
|
233 |
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
234 |
+
output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
|
|
|
|
235 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
236 |
|
|
|
237 |
outputs = self.model(
|
238 |
input_ids=input_ids,
|
239 |
attention_mask=attention_mask,
|
|
|
251 |
|
252 |
loss = None
|
253 |
if labels is not None:
|
|
|
254 |
shift_logits = logits[..., :-1, :].contiguous()
|
255 |
shift_labels = labels[..., 1:].contiguous()
|
|
|
256 |
loss_fct = CrossEntropyLoss()
|
257 |
shift_logits = shift_logits.view(-1, self.config.vocab_size)
|
258 |
shift_labels = shift_labels.view(-1)
|
|
|
259 |
shift_labels = shift_labels.to(shift_logits.device)
|
260 |
loss = loss_fct(shift_logits, shift_labels)
|
261 |
|
262 |
if labels is not None:
|
263 |
llm = []
|
264 |
for i in range(labels.shape[0]):
|
265 |
+
try:
|
266 |
+
p = labels[i].data.cpu().tolist().index(32003) - 1
|
267 |
+
except:
|
268 |
+
p = len(labels[i]) - 9
|
269 |
+
p = min(len(hidden_states[i]) - 9, p)
|
270 |
+
llm.append(hidden_states[i][p:p + 8].unsqueeze(0))
|
271 |
llm = torch.cat(llm, dim=0)
|
272 |
hid_edit = self.edit_head(llm, self.model.embed_tokens.weight[-8:].unsqueeze(dim=0).repeat(labels.shape[0], 1, 1))
|
273 |
|
|
|
277 |
self.model.embed_tokens.weight[-8:].unsqueeze(dim=0).repeat(labels.shape[0], 1, 1))
|
278 |
|
279 |
with torch.no_grad():
|
280 |
+
lat_ans, lat_inp = self.vae.encode(p2p_ans).latent_dist.sample() * self.vae.config.scaling_factor, self.vae.encode(p2p_inp).latent_dist.mode()
|
281 |
lat_ans, lat_inp = [torch.from_numpy(lat_ans.data.cpu().float().numpy()).to(lat_ans.device),
|
282 |
torch.from_numpy(lat_inp.data.cpu().float().numpy()).to(lat_inp.device)]
|
283 |
|
284 |
noise = torch.randn_like(lat_ans)
|
285 |
+
ts = torch.randint(0, self.scheduler.config.num_train_timesteps, (B,), device=noise.device).long()
|
286 |
lat_noise = self.scheduler.add_noise(lat_ans, noise, ts)
|
287 |
|
288 |
prob = torch.rand(B, device=lat_ans.device)
|
289 |
+
mask = (prob < (DROP * 2)).reshape(B, 1, 1)
|
290 |
hid_edit = torch.where(mask, hid_null, hid_edit)
|
291 |
+
mask = (1.0 - ((prob >= DROP).to(lat_inp.dtype) * (prob < (DROP * 3)).to(lat_inp.dtype))).reshape(B, 1, 1, 1)
|
292 |
lat_inp *= mask
|
293 |
|
294 |
+
# Progressive Feature Blending
|
295 |
+
beta_1, beta_2 = 0.7, 0.3
|
296 |
+
visual_features = lat_inp # Assuming lat_inp represents the visual features
|
297 |
+
B_1 = beta_1 * hid_edit + (1 - beta_1) * visual_features
|
298 |
+
B_2 = beta_2 * hid_edit + (1 - beta_2) * visual_features
|
299 |
+
|
300 |
+
# Cross-Attention Masking
|
301 |
+
attention_scores = torch.matmul(hid_edit, hid_edit.transpose(-1, -2))
|
302 |
+
mask = torch.zeros_like(hid_edit)
|
303 |
+
mask[:, 3:5] = 1.0 # Emphasize central elements (e.g., "hat", "blue")
|
304 |
+
masked_attention_scores = attention_scores * mask
|
305 |
+
hid_edit = torch.matmul(F.softmax(masked_attention_scores, dim=-1), hid_edit)
|
306 |
+
|
307 |
+
# Use blended features in subsequent processing
|
308 |
+
hid_edit = B_1 + B_2
|
309 |
+
|
310 |
out = self.unet(torch.cat([lat_noise, lat_inp], dim=1), ts, hid_edit).sample
|
311 |
|
312 |
loss_ce, loss_edit = loss, nn.functional.mse_loss(out, noise, reduction='mean')
|
313 |
+
if int(os.environ.get('LOCAL_RANK', 0)) == 0: print('loss_ce:', loss_ce, '/', 'loss_edit:', loss_edit)
|
314 |
+
loss = loss_ce + loss_edit * 0.5
|
315 |
|
316 |
if not return_dict:
|
317 |
output = (logits,) + outputs[1:]
|
|
|
331 |
if past_key_values:
|
332 |
input_ids = input_ids[:, -1:]
|
333 |
|
|
|
334 |
if inputs_embeds is not None and past_key_values is None:
|
335 |
model_inputs = {"inputs_embeds": inputs_embeds}
|
336 |
else:
|
|
|
362 |
input_embeddings = self.get_input_embeddings().weight.data
|
363 |
output_embeddings = self.get_output_embeddings().weight.data
|
364 |
|
365 |
+
input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(dim=0, keepdim=True)
|
366 |
+
output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(dim=0, keepdim=True)
|
|
|
|
|
367 |
|
368 |
input_embeddings[-num_new_tokens:] = input_embeddings_avg
|
369 |
output_embeddings[-num_new_tokens:] = output_embeddings_avg
|
|
|
384 |
elif embed_tokens_weight.shape[0] == num_new_tokens:
|
385 |
input_embeddings[-num_new_tokens:] = embed_tokens_weight
|
386 |
else:
|
387 |
+
raise ValueError(f"Unexpected embed_tokens_weight shape. Pretrained: {embed_tokens_weight.shape}. Current: {input_embeddings.shape}. Number of new tokens: {num_new_tokens}.")
|
388 |
|
389 |
vision_config.im_patch_token = tokenizer.convert_tokens_to_ids([DEFAULT_IMAGE_PATCH_TOKEN])[0]
|
390 |
|
391 |
AutoConfig.register("llava", LlavaConfig)
|
392 |
+
AutoModelForCausalLM.register(LlavaConfig, LlavaLlamaForCausalLM)
|