File size: 10,877 Bytes
21e5dd0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
229
230
231
232
233
234
235
236
237
238
from typing import List, Optional, Tuple, Union

import torch
import torch.nn as nn
from transformers import LlamaForCausalLM
from transformers.modeling_outputs import CausalLMOutputWithPast

from .configuration_llama_action import LlamaActionConfig


class LearnableFactorizedSpatioTemporalPositionalEmbedding(nn.Module):
    def __init__(self, num_spatio_embeddings: int, num_temporal_embeddings: int, embedding_dim: int):
        super().__init__()
        self.spatio_embeddings = nn.Embedding(num_spatio_embeddings, embedding_dim)
        self.temporal_embeddings = nn.Embedding(num_temporal_embeddings, embedding_dim)
        self.num_spatio_embeddings = num_spatio_embeddings
        self.num_temporal_embeddings = num_temporal_embeddings

    def forward(self, attention_mask: torch.LongTensor, past_key_values_length: int):
        seq_length = attention_mask.size(1)
        batch_size = attention_mask.size(0)

        if past_key_values_length == 0:
            # create a tensor of the form [0, 1, 2, ..., num_spatio_embeddings-1]
            spatio_indices = torch.arange(
                self.num_spatio_embeddings,
                device=attention_mask.device
            ).repeat(self.num_temporal_embeddings).unsqueeze(0).repeat((batch_size, 1))

            # create a tensor of the form [0, 0, 0, ..., 1, 1, 1, ..., 2, 2, 2, ...]
            temporal_indices = torch.arange(
                self.num_temporal_embeddings,
                device=attention_mask.device
            ).repeat_interleave(self.num_spatio_embeddings).unsqueeze(0).repeat((batch_size, 1))

            spatio_indices = spatio_indices[:, :seq_length]
            temporal_indices = temporal_indices[:, :seq_length]
            
        else:
            temporal_index = past_key_values_length // self.num_spatio_embeddings
            spatio_index = past_key_values_length % self.num_spatio_embeddings
            spatio_indices = torch.tensor([[spatio_index]], device=attention_mask.device).repeat((batch_size, 1))
            temporal_indices = torch.tensor([[temporal_index]], device=attention_mask.device).repeat((batch_size, 1))

        return self.spatio_embeddings(spatio_indices) + self.temporal_embeddings(temporal_indices)


class LlamaActionForCausalLM(LlamaForCausalLM):
    config_class = LlamaActionConfig

    def __init__(self, config: LlamaActionConfig):
        super().__init__(config)

        self.num_spatio_embeddings = config.num_spatio_embeddings
        self.num_temporal_embeddings = config.num_temporal_embeddings
        self.num_image_patches = config.num_image_patches
        self.num_action_embeddings = config.num_action_embeddings

        self.pos_embedding_spatio_temporal = LearnableFactorizedSpatioTemporalPositionalEmbedding(
            config.num_spatio_embeddings, config.num_temporal_embeddings, config.hidden_size,
        )

        self.action_projection = nn.Linear(config.action_dim, config.hidden_size)

        self.post_init()

    def forward(
        self,
        input_ids: Optional[torch.Tensor] = None,
        actions: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        labels: Optional[torch.Tensor] = None,
        past_key_values: Optional[List[torch.Tensor]] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple[torch.Tensor], CausalLMOutputWithPast]:
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        if labels is not None:
            use_cache = False

        if input_ids is not None and inputs_embeds is not None:
            raise ValueError(
                "You cannot specify both input_ids and inputs_embeds at the same time"
            )
        elif input_ids is not None:
            pass
        elif inputs_embeds is not None:
            pass
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        inputs_embeds = self.model.get_input_embeddings()(input_ids)
        if past_key_values is None or len(past_key_values) == 0:
            inputs_embeds_list = torch.split(
                inputs_embeds,
                split_size_or_sections=self.num_image_patches,
                dim=1
            )
            actions_list = torch.split(
                actions,
                split_size_or_sections=self.num_action_embeddings,
                dim=1
            )

            embeddings = []
            if len(inputs_embeds_list) == len(actions_list):
                # mostly used in training phase
                for inputs_embeds, action_embeds in zip(inputs_embeds_list, actions_list):
                    action_features = self.action_projection(action_embeds)
                    embeddings.append(inputs_embeds)
                    embeddings.append(action_features)
            elif len(inputs_embeds_list) < len(actions_list):
                # used in inference phase (mostly)
                for i, inputs_embeds in enumerate(inputs_embeds_list):
                    embeddings.append(inputs_embeds)
                    if i < len(inputs_embeds_list) - 1:
                        # the last frame might be generating image tokens, so we don't add action embedding
                        action_embeds = self.action_projection(actions_list[i])
                        embeddings.append(action_embeds)
                if inputs_embeds_list[-1].size(1) == self.num_image_patches:
                    # if the last frame has generated all image tokens, we add action embedding
                    action_embeds = self.action_projection(actions_list[len(inputs_embeds_list) - 1])
                    embeddings.append(action_embeds)
        else:
            if isinstance(past_key_values, tuple):
                past_key_values_length = past_key_values[0][0].size(2)
            else:
                past_key_values_length = past_key_values.get_seq_length()
            embeddings = []
            # create an interleaved sequence of image and action embeddings like image, image, ..., image, action, action, ..., action
            # we only generate image tokens, so we add action tokens after generating one frame
            if past_key_values_length % self.num_spatio_embeddings == (self.num_spatio_embeddings - self.num_action_embeddings):
                seq_index = past_key_values_length // self.num_spatio_embeddings + 1
                actions_list = torch.split(
                    actions,
                    split_size_or_sections=self.num_action_embeddings,
                    dim=1
                )
                action_features = self.action_projection(actions_list[seq_index - 1])
                embeddings.append(action_features)
                embeddings.append(inputs_embeds)
            else:
                pass

        if len(embeddings) > 0:
            inputs_embeds = torch.cat(embeddings, dim=1)

        # insert spatio-temporal positional embedding
        if past_key_values is not None:
            if isinstance(past_key_values, tuple):
                past_key_values_length = past_key_values[0][0].size(2)
            else:
                past_key_values_length = past_key_values.get_seq_length()
        else:
            past_key_values_length = 0
        inputs_embeds += self.pos_embedding_spatio_temporal(inputs_embeds, past_key_values_length)

        outputs = self.model(
            input_ids=None,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        sequence_output = outputs[0]
        logits = self.lm_head(sequence_output).contiguous()

        loss = None
        if labels is not None:
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            loss_fct = nn.CrossEntropyLoss()
            loss = loss_fct(shift_logits.view(-1, self.config.vocab_size), shift_labels.view(-1))

        if not return_dict:
            output = (logits,) + outputs[1:]
            return ((loss,) + output) if loss is not None else output

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        attention_mask=None,
        use_cache=None,
        **kwargs):
        batch_size = input_ids.size(0)
        seq_length = input_ids.size(1)
        n_frames = seq_length // self.num_image_patches
        attention_mask_length = n_frames * (self.num_image_patches + self.num_action_embeddings)
        if seq_length % self.num_image_patches != 0:
            n_last_frame_tokens = seq_length % self.num_image_patches
            attention_mask_length += n_last_frame_tokens
        else:
            print(f"attempting to generate new frame - frame no: {n_frames + 1}")
        attention_mask = torch.ones((batch_size, attention_mask_length), device=input_ids.device, dtype=torch.long)
        # cut decoder_input_ids if past_key_values is used
        if past_key_values is not None and len(past_key_values) > 0:
            if isinstance(past_key_values, tuple):
                past_length = past_key_values[0][0].size(2)
            else:
                past_length = past_key_values.get_seq_length()
            if input_ids.size(1) > past_length:
                remove_prefix_length = past_length
            else:
                remove_prefix_length = input_ids.size(1) - 1
            input_ids = input_ids[:, remove_prefix_length:]
            seq_length = input_ids.size(1)
            past_key_values_length = past_length
            mask_seq_length = seq_length + past_key_values_length
            if past_key_values_length % self.num_spatio_embeddings == (self.num_spatio_embeddings - self.num_action_embeddings):
                mask_seq_length += self.num_action_embeddings
            attention_mask = torch.ones((batch_size, mask_seq_length), device=input_ids.device, dtype=torch.long)

        return {
            "input_ids": input_ids,
            "attention_mask": attention_mask,
            "actions": kwargs.get("actions"),
            "past_key_values": past_key_values,
            "use_cache": use_cache,
        }