File size: 8,497 Bytes
19d8873
19f124c
 
 
69ab272
19d8873
19f124c
 
19d8873
19f124c
 
19d8873
 
 
19f124c
19d8873
19f124c
 
19d8873
19f124c
 
 
 
 
 
 
0021de3
19f124c
 
 
 
 
19d8873
19f124c
 
 
19d8873
19f124c
 
 
19d8873
19f124c
19d8873
 
19f124c
 
69ab272
19d8873
19f124c
 
 
 
19d8873
 
 
 
 
 
 
 
0021de3
 
 
 
 
 
 
 
69ab272
19d8873
 
 
 
 
 
 
 
 
935404c
19d8873
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
69ab272
 
19d8873
69ab272
 
 
19d8873
 
69ab272
 
 
 
19d8873
69ab272
 
 
19d8873
69ab272
 
 
19d8873
69ab272
 
19d8873
 
 
 
 
69ab272
19d8873
 
69ab272
19d8873
 
 
69ab272
19d8873
 
0021de3
 
 
 
 
 
 
 
69ab272
0021de3
 
 
 
69ab272
19d8873
69ab272
 
0021de3
19d8873
 
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
# modeling_emuru.py
import torch
import torch.nn as nn
from transformers import PreTrainedModel, T5ForConditionalGeneration, T5Config, AutoTokenizer
from configuration_emuru import EmuruConfig
# from .configuration_emuru import EmuruConfig
from diffusers import AutoencoderKL
from einops.layers.torch import Rearrange
from einops import rearrange, repeat

class Emuru(PreTrainedModel):
    config_class = EmuruConfig  # Link to your configuration

    def __init__(self, config):
        super().__init__(config)
        # Initialize the tokenizer (if you want it as part of your model)
        self.tokenizer = AutoTokenizer.from_pretrained(config.tokenizer_config)

        # Load T5 using the provided filename from config
        t5_config = T5Config.from_pretrained(config.t5_config)
        t5_config.vocab_size = len(self.tokenizer)
        self.T5 = T5ForConditionalGeneration(t5_config)
        self.T5.lm_head = nn.Identity()
        self.sos = nn.Embedding(1, t5_config.d_model)

        vae_latent_size = 8 * config.vae_channels * config.slices_per_query
        self.vae_to_t5 = nn.Linear(vae_latent_size, t5_config.d_model)
        self.t5_to_vae = nn.Linear(t5_config.d_model, vae_latent_size, bias=False)

        self.padding_token = nn.Parameter(torch.empty((1, vae_latent_size)), requires_grad=False)
        self.padding_token_threshold = nn.Parameter(torch.empty(1), requires_grad=False)

        # Load VAE
        self.vae = AutoencoderKL.from_pretrained(config.vae_config)
        self.set_training(self.vae, False)

        # Define the rearrange layers
        self.query_rearrange = Rearrange('b c h (w q) -> b w (q c h)', q=config.slices_per_query)
        self.z_rearrange = Rearrange('b w (q c h) -> b c h (w q)', c=config.vae_channels, q=config.slices_per_query)

        # Define your loss functions
        self.mse_criterion = nn.MSELoss()

        # Initialize weights following Hugging Face conventions (if needed)
        self.init_weights()


    def set_training(self, model, training):
        model.train() if training else model.eval()
        for param in model.parameters():
            param.requires_grad = training

    # --- Implement the rest of your methods ---
    # For example, _img_encode, forward, generate, etc.
    # You can largely port your existing code here, making sure that:
    #  - The forward method returns a dictionary with your losses and outputs.
    #  - You use the Hugging Face methods for saving/loading weights.


    def forward(self, img=None, input_ids=None, attention_mask=None, noise=0, **kwargs):
        decoder_inputs_embeds, z_sequence, z = self._img_encode(img, noise)

        output = self.T5(input_ids, attention_mask=attention_mask, decoder_inputs_embeds=decoder_inputs_embeds)
        vae_latent = self.t5_to_vae(output.logits[:, :-1])
        pred_latent = self.z_rearrange(vae_latent)

        mse_loss = self.mse_criterion(vae_latent, z_sequence)
        return mse_loss, pred_latent, z


    def old_generate(self, text=None, img=None, z_sequence=None, input_ids=None, max_new_tokens=256,
                 stopping_criteria='latent', stopping_after=10, stopping_errors=1):
        assert text is not None or input_ids is not None, 'Either text or input_ids must be provided'
        assert img is not None or z_sequence is not None, 'Either img or z_sequence must be provided'

        if input_ids is None:
            input_ids = self.tokenizer(text, return_tensors='pt', padding=True).input_ids
            input_ids = input_ids.to(next(self.T5.parameters()).device)
        
        if z_sequence is None:
            _, z_sequence, _ = self._img_encode(img)
        z_sequence = [z_sequence]

        sos = repeat(self.sos.weight, '1 d -> b 1 d', b=input_ids.size(0))
        for _ in range(max_new_tokens):
            if len(z_sequence) == 0:
                decoder_inputs_embeds = sos
            else:
                decoder_inputs_embeds = self.vae_to_t5(torch.cat(z_sequence, dim=1))
                decoder_inputs_embeds = torch.cat([sos, decoder_inputs_embeds], dim=1)
            output = self.T5(input_ids, decoder_inputs_embeds=decoder_inputs_embeds)
            vae_latent = self.t5_to_vae(output.logits[:, -1:])
            z_sequence.append(vae_latent)

            if stopping_criteria == 'latent':
                curr_z_sequence = torch.cat(z_sequence, dim=1)
                pad_token = repeat(self.padding_token, '1 d -> b 1 d', b=input_ids.size(0)).to(decoder_inputs_embeds.device)
                similarity = torch.nn.functional.cosine_similarity(curr_z_sequence, pad_token, dim=-1)
                similarity = similarity[:, -stopping_after:] > self.padding_token_threshold
                if torch.all(similarity.sum(-1) >= (stopping_after - stopping_errors)):
                    # z_sequence = [curr_z_sequence[:, :-stopping_after]]
                    z_sequence = [curr_z_sequence]
                    break
            elif stopping_criteria == 'pixel':
                raise NotImplementedError

        z_sequence = torch.cat(z_sequence, dim=1)
        img = torch.clamp(self.vae.decode(self.z_rearrange(z_sequence)).sample, -1, 1)
        return img


    def generate(self,
                     style_text=None,
                     gen_text=None,
                     style_img=None,
                     input_ids=None,
                     z_sequence=None,
                     max_new_tokens=256,
                     stopping_criteria='latent',
                     stopping_after=10,
                     stopping_patience=1,
                     trim_image=True):
        assert (gen_text is not None and style_text is not None) or input_ids is not None, 'Either gen_text and style_text or input_ids must be provided'
        assert style_img is not None or z_sequence is not None, 'Either style_img or z_sequence must be provided'

        if input_ids is None:
            input_ids = self.tokenizer(gen_text + ' ' + style_text, return_tensors='pt', padding=True).input_ids
            input_ids = input_ids.to(self.device)

        if z_sequence is None:
            _, z_sequence, _ = self._img_encode(style_img)
        z_sequence = [z_sequence]

        sos = repeat(self.sos.weight, '1 d -> b 1 d', b=input_ids.size(0))
        pad_token = repeat(self.padding_token, '1 d -> b 1 d', b=input_ids.size(0))

        for _ in range(max_new_tokens):
            if len(z_sequence) == 0:
                decoder_inputs_embeds = sos
            else:
                decoder_inputs_embeds = self.vae_to_t5(torch.cat(z_sequence, dim=1))
                decoder_inputs_embeds = torch.cat([sos, decoder_inputs_embeds], dim=1)
            output = self.T5(input_ids, decoder_inputs_embeds=decoder_inputs_embeds)
            vae_latent = self.t5_to_vae(output.logits[:, -1:])
            z_sequence.append(vae_latent)

            if stopping_criteria == 'latent':
                curr_z_sequence = torch.cat(z_sequence, dim=1)
                similarity = torch.nn.functional.cosine_similarity(curr_z_sequence, pad_token, dim=-1)
                similarity = similarity[:, -stopping_after:] > self.padding_token_threshold
                if torch.all(similarity.sum(-1) >= (stopping_after - stopping_patience)):
                    z_sequence = [curr_z_sequence[:, :-similarity.sum(-1)]] if trim_image else [curr_z_sequence]
                    break
            elif stopping_criteria == 'pixel':
                raise NotImplementedError

        z_sequence = torch.cat(z_sequence, dim=1)
        img = torch.clamp(self.vae.decode(self.z_rearrange(z_sequence)).sample, -1, 1)
        return img, z_sequence
    

    def _img_encode(self, img, noise=0):
        posterior = self.vae.encode(img.float())
        z = posterior.latent_dist.sample()
        z_sequence = self.query_rearrange(z)

        noise_sequence = z_sequence
        if noise > 0:
            noise_sequence = z_sequence + torch.randn_like(z_sequence) * noise

        decoder_inputs_embeds = self.vae_to_t5(noise_sequence)
        sos = repeat(self.sos.weight, '1 d -> b 1 d', b=decoder_inputs_embeds.size(0))
        decoder_inputs_embeds = torch.cat([sos, decoder_inputs_embeds], dim=1)
        return decoder_inputs_embeds, z_sequence, z


    def compute_padding_token(self):
        raise NotImplementedError("compute_padding_token not implemented")


    def compute_padding_token_threshold(self):
        raise NotImplementedError("compute_padding_token_threshold not implemented")