File size: 8,525 Bytes
9e31d55
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os

import torch
import torch.nn as nn
from transformers import PreTrainedModel, PretrainedConfig

os.environ["TOKENIZERS_PARALLELISM"] = "false"
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")

# ===================== OPTIMIZED EMG MODEL =====================


class OptimizedEMGCell(nn.Module):
    def __init__(self, input_size, hidden_size, dropout_rate=0.1, use_layer_norm=False):
        super(OptimizedEMGCell, self).__init__()
        self.input_size = input_size
        self.hidden_size = hidden_size
        self.use_layer_norm = use_layer_norm
        self.clamp_min = -1
        self.clamp_max = 1

        # Fused linear transformations for better efficiency
        self.input_transform_linear = nn.Linear(input_size, hidden_size * 2)
        self.hidden_transform_linear = nn.Linear(hidden_size, hidden_size * 2)
        
        # SIMPLIFIED: Use standard dropout instead of variational
        self.dropout = nn.Dropout(dropout_rate) if dropout_rate > 0 else None
        
        # Layer normalization for training stability
        if use_layer_norm:
            self.input_norm = nn.LayerNorm(hidden_size)
            self.hidden_norm = nn.LayerNorm(hidden_size)
            self.cell_norm = nn.LayerNorm(hidden_size)
        
        self.init_weights()
        
    def init_weights(self):
        for linear in [self.input_transform_linear, self.hidden_transform_linear]:
            # Use smaller initialization for RNN stability
            nn.init.uniform_(linear.weight, -0.1, 0.1)
            nn.init.zeros_(linear.bias)
            
    def forward(self, input, hidden):
        h_prev, c_prev = hidden
        
        # Project input and hidden states
        input_connections = self.input_transform_linear(input)
        hidden_connections = self.hidden_transform_linear(h_prev)
        
        # Split projections
        i_move, i_merge = torch.chunk(input_connections, 2, dim=-1)
        h_move, h_merge = torch.chunk(hidden_connections, 2, dim=-1)

        # EMG computation
        # merge_gate = torch.clamp(i_merge, self.clamp_min, self.clamp_max) * torch.sigmoid(torch.clamp(h_merge, self.clamp_min, self.clamp_max))
        merge_gate = torch.clamp(i_merge * torch.sigmoid(h_merge), self.clamp_min, self.clamp_max)
        move_gate = torch.clamp(torch.sigmoid(i_move) * h_move, self.clamp_min, self.clamp_max)

        if self.use_layer_norm:
            c_prev = self.cell_norm(c_prev)

        context_gate = torch.tanh(torch.clamp(c_prev + merge_gate, self.clamp_min, self.clamp_max))

        if self.use_layer_norm:
            context_gate = self.input_norm(context_gate)

        c_next = context_gate
        
        if self.use_layer_norm:
            c_next = self.hidden_norm(c_next)
        
        # Apply dropout to output instead of complex variational dropout
        m_next = (1 - move_gate) * merge_gate + move_gate * c_next
        if self.dropout is not None:
            m_next = self.dropout(m_next)
        
        return m_next, c_next


class OptimizedEMG(nn.Module):
    """Enhanced EMG with gradient checkpointing and other optimizations"""
    def __init__(self, input_size, hidden_size, num_layers, dropout_rate=0.1, 
                 use_gradient_checkpointing=False):
        super(OptimizedEMG, self).__init__()
        self.input_size = input_size
        self.hidden_size = hidden_size
        self.num_layers = num_layers
        self.use_gradient_checkpointing = use_gradient_checkpointing
        
        self.cells = nn.ModuleList([
            OptimizedEMGCell(
                input_size if i == 0 else hidden_size, 
                hidden_size, 
                dropout_rate
            ) for i in range(num_layers)
        ])

    def forward(self, x, hidden=None):
        batch_size, seq_len, _ = x.size()

        if hidden is None:
            hidden = [(torch.zeros(batch_size, self.hidden_size, device=x.device), 
                      torch.zeros(batch_size, self.hidden_size, device=x.device)) 
                     for _ in range(self.num_layers)]

        outputs = []

        for t in range(seq_len):
            layer_input = x[:, t, :]
            
            for layer_idx, cell in enumerate(self.cells):
                m_prev, c_prev = hidden[layer_idx]
                
                if self.use_gradient_checkpointing and self.training:
                    m_next, c_next = torch.utils.checkpoint.checkpoint(
                        cell, layer_input, (m_prev, c_prev), use_reentrant=False
                    )
                else:
                    m_next, c_next = cell(layer_input, (m_prev, c_prev))
                
                hidden[layer_idx] = (m_next, c_next)
                layer_input = m_next
            
            outputs.append(layer_input)

        output = torch.stack(outputs, dim=1)
        return output, hidden


# ===================== HUGGING FACE COMPATIBLE MODEL =====================

class EMGConfig(PretrainedConfig):
    """Configuration class for EMG model"""
    model_type = "emg"
    
    def __init__(
        self,
        vocab_size=50000,
        embedding_dim=512,
        hidden_dim=512,
        num_layers=2,
        dropout=0.1,
        use_layer_norm=True,
        use_gradient_checkpointing=False,
        tie_word_embeddings=True,
        **kwargs
    ):
        super().__init__(**kwargs)
        self.vocab_size = vocab_size
        self.embedding_dim = embedding_dim
        self.hidden_dim = hidden_dim
        self.num_layers = num_layers
        self.dropout = dropout
        self.use_layer_norm = use_layer_norm
        self.use_gradient_checkpointing = use_gradient_checkpointing
        self.tie_word_embeddings = tie_word_embeddings


class EMGLanguageModel(PreTrainedModel):
    """Hugging Face compatible EMG Language Model"""
    config_class = EMGConfig
    
    def __init__(self, config):
        super().__init__(config)
        self.config = config
        
        self.embedding = nn.Embedding(config.vocab_size, config.embedding_dim)
        self.emg = OptimizedEMG(
            config.embedding_dim, 
            config.hidden_dim, 
            config.num_layers, 
            config.dropout,
            config.use_gradient_checkpointing
        )
        self.output_projection = nn.Linear(config.hidden_dim, config.vocab_size)
        
        # Tie embedding and output weights if dimensions match
        if config.tie_word_embeddings and config.embedding_dim == config.hidden_dim:
            self.output_projection.weight = self.embedding.weight
            
        # Initialize weights
        self.apply(self._init_weights)

    def _init_weights(self, module):
        """Initialize the weights"""
        if isinstance(module, nn.Linear):
            nn.init.xavier_uniform_(module.weight)
            if module.bias is not None:
                nn.init.zeros_(module.bias)
        elif isinstance(module, nn.Embedding):
            nn.init.normal_(module.weight, mean=0.0, std=0.02)

    def forward(self, input_ids, hidden=None, labels=None, **kwargs):
        embedded = self.embedding(input_ids)
        output, hidden = self.emg(embedded, hidden)
        logits = self.output_projection(output)
        
        loss = None
        if labels is not None:
            # Shift so that tokens < n predict n
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            
            # Flatten the tokens
            loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), 
                          shift_labels.view(-1))
        
        return {'loss': loss, 'logits': logits, 'hidden_states': hidden}
        
    def generate(self, input_ids, max_length=50, temperature=1.0, top_k=50):
        self.eval()
        generated = input_ids
        hidden = None

        for _ in range(max_length):
            outputs = self.forward(generated[:, -1:], hidden)
            logits = outputs['logits'][:, -1, :] / temperature

            # Top-k sampling
            top_k_logits, top_k_indices = torch.topk(logits, top_k)
            probs = F.softmax(top_k_logits, dim=-1)
            next_token = top_k_indices.gather(1, torch.multinomial(probs, num_samples=1))

            generated = torch.cat([generated, next_token], dim=1)

        return generated