koukyo1994
commited on
Commit
•
21e5dd0
1
Parent(s):
e999941
Upload folder using huggingface_hub
Browse files- configuration_llama_action.py +13 -0
- modeling_llama_action.py +237 -0
configuration_llama_action.py
ADDED
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import LlamaConfig
|
2 |
+
|
3 |
+
|
4 |
+
class LlamaActionConfig(LlamaConfig):
|
5 |
+
model_type = "llama_action"
|
6 |
+
|
7 |
+
def __init__(self, **kwargs):
|
8 |
+
super().__init__(**kwargs)
|
9 |
+
self.num_spatio_embeddings = kwargs.get("num_spatio_embeddings", 582)
|
10 |
+
self.num_temporal_embeddings = kwargs.get("num_temporal_embeddings", 25)
|
11 |
+
self.num_action_embeddings = kwargs.get("num_action_tokens", 5)
|
12 |
+
self.num_image_patches = kwargs.get("num_image_patches", 576)
|
13 |
+
self.action_dim = kwargs.get("action_dim", 3)
|
modeling_llama_action.py
ADDED
@@ -0,0 +1,237 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import List, Optional, Tuple, Union
|
2 |
+
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
from transformers import LlamaForCausalLM
|
6 |
+
from transformers.modeling_outputs import CausalLMOutputWithPast
|
7 |
+
|
8 |
+
from .configuration_llama_action import LlamaActionConfig
|
9 |
+
|
10 |
+
|
11 |
+
class LearnableFactorizedSpatioTemporalPositionalEmbedding(nn.Module):
|
12 |
+
def __init__(self, num_spatio_embeddings: int, num_temporal_embeddings: int, embedding_dim: int):
|
13 |
+
super().__init__()
|
14 |
+
self.spatio_embeddings = nn.Embedding(num_spatio_embeddings, embedding_dim)
|
15 |
+
self.temporal_embeddings = nn.Embedding(num_temporal_embeddings, embedding_dim)
|
16 |
+
self.num_spatio_embeddings = num_spatio_embeddings
|
17 |
+
self.num_temporal_embeddings = num_temporal_embeddings
|
18 |
+
|
19 |
+
def forward(self, attention_mask: torch.LongTensor, past_key_values_length: int):
|
20 |
+
seq_length = attention_mask.size(1)
|
21 |
+
batch_size = attention_mask.size(0)
|
22 |
+
|
23 |
+
if past_key_values_length == 0:
|
24 |
+
# create a tensor of the form [0, 1, 2, ..., num_spatio_embeddings-1]
|
25 |
+
spatio_indices = torch.arange(
|
26 |
+
self.num_spatio_embeddings,
|
27 |
+
device=attention_mask.device
|
28 |
+
).repeat(self.num_temporal_embeddings).unsqueeze(0).repeat((batch_size, 1))
|
29 |
+
|
30 |
+
# create a tensor of the form [0, 0, 0, ..., 1, 1, 1, ..., 2, 2, 2, ...]
|
31 |
+
temporal_indices = torch.arange(
|
32 |
+
self.num_temporal_embeddings,
|
33 |
+
device=attention_mask.device
|
34 |
+
).repeat_interleave(self.num_spatio_embeddings).unsqueeze(0).repeat((batch_size, 1))
|
35 |
+
|
36 |
+
spatio_indices = spatio_indices[:, :seq_length]
|
37 |
+
temporal_indices = temporal_indices[:, :seq_length]
|
38 |
+
|
39 |
+
else:
|
40 |
+
temporal_index = past_key_values_length // self.num_spatio_embeddings
|
41 |
+
spatio_index = past_key_values_length % self.num_spatio_embeddings
|
42 |
+
spatio_indices = torch.tensor([[spatio_index]], device=attention_mask.device).repeat((batch_size, 1))
|
43 |
+
temporal_indices = torch.tensor([[temporal_index]], device=attention_mask.device).repeat((batch_size, 1))
|
44 |
+
|
45 |
+
return self.spatio_embeddings(spatio_indices) + self.temporal_embeddings(temporal_indices)
|
46 |
+
|
47 |
+
|
48 |
+
class LlamaActionForCausalLM(LlamaForCausalLM):
|
49 |
+
config_class = LlamaActionConfig
|
50 |
+
|
51 |
+
def __init__(self, config: LlamaActionConfig):
|
52 |
+
super().__init__(config)
|
53 |
+
|
54 |
+
self.num_spatio_embeddings = config.num_spatio_embeddings
|
55 |
+
self.num_temporal_embeddings = config.num_temporal_embeddings
|
56 |
+
self.num_image_patches = config.num_image_patches
|
57 |
+
self.num_action_embeddings = config.num_action_embeddings
|
58 |
+
|
59 |
+
self.pos_embedding_spatio_temporal = LearnableFactorizedSpatioTemporalPositionalEmbedding(
|
60 |
+
config.num_spatio_embeddings, config.num_temporal_embeddings, config.hidden_size,
|
61 |
+
)
|
62 |
+
|
63 |
+
self.action_projection = nn.Linear(config.action_dim, config.hidden_size)
|
64 |
+
|
65 |
+
self.post_init()
|
66 |
+
|
67 |
+
def forward(
|
68 |
+
self,
|
69 |
+
input_ids: Optional[torch.Tensor] = None,
|
70 |
+
actions: Optional[torch.Tensor] = None,
|
71 |
+
attention_mask: Optional[torch.Tensor] = None,
|
72 |
+
position_ids: Optional[torch.Tensor] = None,
|
73 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
74 |
+
labels: Optional[torch.Tensor] = None,
|
75 |
+
past_key_values: Optional[List[torch.Tensor]] = None,
|
76 |
+
use_cache: Optional[bool] = None,
|
77 |
+
output_attentions: Optional[bool] = None,
|
78 |
+
output_hidden_states: Optional[bool] = None,
|
79 |
+
return_dict: Optional[bool] = None,
|
80 |
+
) -> Union[Tuple[torch.Tensor], CausalLMOutputWithPast]:
|
81 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
82 |
+
if labels is not None:
|
83 |
+
use_cache = False
|
84 |
+
|
85 |
+
if input_ids is not None and inputs_embeds is not None:
|
86 |
+
raise ValueError(
|
87 |
+
"You cannot specify both input_ids and inputs_embeds at the same time"
|
88 |
+
)
|
89 |
+
elif input_ids is not None:
|
90 |
+
pass
|
91 |
+
elif inputs_embeds is not None:
|
92 |
+
pass
|
93 |
+
else:
|
94 |
+
raise ValueError("You have to specify either input_ids or inputs_embeds")
|
95 |
+
|
96 |
+
inputs_embeds = self.model.get_input_embeddings()(input_ids)
|
97 |
+
if past_key_values is None or len(past_key_values) == 0:
|
98 |
+
inputs_embeds_list = torch.split(
|
99 |
+
inputs_embeds,
|
100 |
+
split_size_or_sections=self.num_image_patches,
|
101 |
+
dim=1
|
102 |
+
)
|
103 |
+
actions_list = torch.split(
|
104 |
+
actions,
|
105 |
+
split_size_or_sections=self.num_action_embeddings,
|
106 |
+
dim=1
|
107 |
+
)
|
108 |
+
|
109 |
+
embeddings = []
|
110 |
+
if len(inputs_embeds_list) == len(actions_list):
|
111 |
+
# mostly used in training phase
|
112 |
+
for inputs_embeds, action_embeds in zip(inputs_embeds_list, actions_list):
|
113 |
+
action_features = self.action_projection(action_embeds)
|
114 |
+
embeddings.append(inputs_embeds)
|
115 |
+
embeddings.append(action_features)
|
116 |
+
elif len(inputs_embeds_list) < len(actions_list):
|
117 |
+
# used in inference phase (mostly)
|
118 |
+
for i, inputs_embeds in enumerate(inputs_embeds_list):
|
119 |
+
embeddings.append(inputs_embeds)
|
120 |
+
if i < len(inputs_embeds_list) - 1:
|
121 |
+
# the last frame might be generating image tokens, so we don't add action embedding
|
122 |
+
action_embeds = self.action_projection(actions_list[i])
|
123 |
+
embeddings.append(action_embeds)
|
124 |
+
if inputs_embeds_list[-1].size(1) == self.num_image_patches:
|
125 |
+
# if the last frame has generated all image tokens, we add action embedding
|
126 |
+
action_embeds = self.action_projection(actions_list[len(inputs_embeds_list) - 1])
|
127 |
+
embeddings.append(action_embeds)
|
128 |
+
else:
|
129 |
+
if isinstance(past_key_values, tuple):
|
130 |
+
past_key_values_length = past_key_values[0][0].size(2)
|
131 |
+
else:
|
132 |
+
past_key_values_length = past_key_values.get_seq_length()
|
133 |
+
embeddings = []
|
134 |
+
# create an interleaved sequence of image and action embeddings like image, image, ..., image, action, action, ..., action
|
135 |
+
# we only generate image tokens, so we add action tokens after generating one frame
|
136 |
+
if past_key_values_length % self.num_spatio_embeddings == (self.num_spatio_embeddings - self.num_action_embeddings):
|
137 |
+
seq_index = past_key_values_length // self.num_spatio_embeddings + 1
|
138 |
+
actions_list = torch.split(
|
139 |
+
actions,
|
140 |
+
split_size_or_sections=self.num_action_embeddings,
|
141 |
+
dim=1
|
142 |
+
)
|
143 |
+
action_features = self.action_projection(actions_list[seq_index - 1])
|
144 |
+
embeddings.append(action_features)
|
145 |
+
embeddings.append(inputs_embeds)
|
146 |
+
else:
|
147 |
+
pass
|
148 |
+
|
149 |
+
if len(embeddings) > 0:
|
150 |
+
inputs_embeds = torch.cat(embeddings, dim=1)
|
151 |
+
|
152 |
+
# insert spatio-temporal positional embedding
|
153 |
+
if past_key_values is not None:
|
154 |
+
if isinstance(past_key_values, tuple):
|
155 |
+
past_key_values_length = past_key_values[0][0].size(2)
|
156 |
+
else:
|
157 |
+
past_key_values_length = past_key_values.get_seq_length()
|
158 |
+
else:
|
159 |
+
past_key_values_length = 0
|
160 |
+
inputs_embeds += self.pos_embedding_spatio_temporal(inputs_embeds, past_key_values_length)
|
161 |
+
|
162 |
+
outputs = self.model(
|
163 |
+
input_ids=None,
|
164 |
+
attention_mask=attention_mask,
|
165 |
+
position_ids=position_ids,
|
166 |
+
past_key_values=past_key_values,
|
167 |
+
inputs_embeds=inputs_embeds,
|
168 |
+
use_cache=use_cache,
|
169 |
+
output_attentions=output_attentions,
|
170 |
+
output_hidden_states=output_hidden_states,
|
171 |
+
return_dict=return_dict,
|
172 |
+
)
|
173 |
+
|
174 |
+
sequence_output = outputs[0]
|
175 |
+
logits = self.lm_head(sequence_output).contiguous()
|
176 |
+
|
177 |
+
loss = None
|
178 |
+
if labels is not None:
|
179 |
+
shift_logits = logits[..., :-1, :].contiguous()
|
180 |
+
shift_labels = labels[..., 1:].contiguous()
|
181 |
+
loss_fct = nn.CrossEntropyLoss()
|
182 |
+
loss = loss_fct(shift_logits.view(-1, self.config.vocab_size), shift_labels.view(-1))
|
183 |
+
|
184 |
+
if not return_dict:
|
185 |
+
output = (logits,) + outputs[1:]
|
186 |
+
return ((loss,) + output) if loss is not None else output
|
187 |
+
|
188 |
+
return CausalLMOutputWithPast(
|
189 |
+
loss=loss,
|
190 |
+
logits=logits,
|
191 |
+
past_key_values=outputs.past_key_values,
|
192 |
+
hidden_states=outputs.hidden_states,
|
193 |
+
attentions=outputs.attentions,
|
194 |
+
)
|
195 |
+
|
196 |
+
def prepare_inputs_for_generation(
|
197 |
+
self,
|
198 |
+
input_ids,
|
199 |
+
past_key_values=None,
|
200 |
+
attention_mask=None,
|
201 |
+
use_cache=None,
|
202 |
+
**kwargs):
|
203 |
+
batch_size = input_ids.size(0)
|
204 |
+
seq_length = input_ids.size(1)
|
205 |
+
n_frames = seq_length // self.num_image_patches
|
206 |
+
attention_mask_length = n_frames * (self.num_image_patches + self.num_action_embeddings)
|
207 |
+
if seq_length % self.num_image_patches != 0:
|
208 |
+
n_last_frame_tokens = seq_length % self.num_image_patches
|
209 |
+
attention_mask_length += n_last_frame_tokens
|
210 |
+
else:
|
211 |
+
print(f"attempting to generate new frame - frame no: {n_frames + 1}")
|
212 |
+
attention_mask = torch.ones((batch_size, attention_mask_length), device=input_ids.device, dtype=torch.long)
|
213 |
+
# cut decoder_input_ids if past_key_values is used
|
214 |
+
if past_key_values is not None and len(past_key_values) > 0:
|
215 |
+
if isinstance(past_key_values, tuple):
|
216 |
+
past_length = past_key_values[0][0].size(2)
|
217 |
+
else:
|
218 |
+
past_length = past_key_values.get_seq_length()
|
219 |
+
if input_ids.size(1) > past_length:
|
220 |
+
remove_prefix_length = past_length
|
221 |
+
else:
|
222 |
+
remove_prefix_length = input_ids.size(1) - 1
|
223 |
+
input_ids = input_ids[:, remove_prefix_length:]
|
224 |
+
seq_length = input_ids.size(1)
|
225 |
+
past_key_values_length = past_length
|
226 |
+
mask_seq_length = seq_length + past_key_values_length
|
227 |
+
if past_key_values_length % self.num_spatio_embeddings == (self.num_spatio_embeddings - self.num_action_embeddings):
|
228 |
+
mask_seq_length += self.num_action_embeddings
|
229 |
+
attention_mask = torch.ones((batch_size, mask_seq_length), device=input_ids.device, dtype=torch.long)
|
230 |
+
|
231 |
+
return {
|
232 |
+
"input_ids": input_ids,
|
233 |
+
"attention_mask": attention_mask,
|
234 |
+
"actions": kwargs.get("actions"),
|
235 |
+
"past_key_values": past_key_values,
|
236 |
+
"use_cache": use_cache,
|
237 |
+
}
|