File size: 11,494 Bytes
6ac124b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
"""HuggingFace compatible Namer model."""

from __future__ import annotations

import math
from typing import Optional, Union

import torch
import torch.nn as nn
from transformers import PreTrainedModel, PretrainedConfig
from transformers.modeling_outputs import CausalLMOutputWithCrossAttentions
from transformers.generation import GenerationMixin


class NamerConfig(PretrainedConfig):
    """Configuration class for NamerModel."""
    
    model_type = "custom"
    
    def __init__(
        self,
        vocab_size: int = 41,
        max_output_len: int = 20,
        d_model: int = 128,
        nhead: int = 4,
        num_encoder_layers: int = 4,
        dim_feedforward: int = 512,
        dropout: float = 0.1,
        pad_token_id: int = 10,
        eos_token_id: int = 40,  # <EOS> token index
        **kwargs,
    ):
        self.vocab_size = vocab_size
        self.max_output_len = max_output_len
        self.d_model = d_model
        self.nhead = nhead
        self.num_encoder_layers = num_encoder_layers
        self.dim_feedforward = dim_feedforward
        self.dropout = dropout
        
        super().__init__(
            pad_token_id=pad_token_id,
            eos_token_id=eos_token_id,
            **kwargs,
        )


class PositionalEncoding(nn.Module):
    """Sinusoidal positional encoding for transformer."""

    def __init__(self, d_model: int, max_len: int = 5000) -> None:
        super().__init__()

        pe = torch.zeros(max_len, d_model)
        position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
        div_term = torch.exp(
            torch.arange(0, d_model, 2).float()
            * (-math.log(10000.0) / d_model)
        )

        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)

        self.register_buffer("pe", pe)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """Add positional encoding to input."""
        return x + self.pe[: x.size(1)]


class NamerModel(PreTrainedModel, GenerationMixin):
    """HuggingFace compatible Namer transformer model.
    
    Converts integer digit sequences to English number names.
    """
    
    config_class = NamerConfig
    base_model_prefix = "namer"
    
    def __init__(self, config: NamerConfig):
        super().__init__(config)
        
        self.vocab_size = config.vocab_size
        self.max_output_len = config.max_output_len
        self.d_model = config.d_model
        
        # Digit embedding (10 digits + 1 padding token = 11)
        self.digit_embedding = nn.Embedding(11, config.d_model, padding_idx=config.pad_token_id)
        
        # Positional encoding
        self.pos_encoder = PositionalEncoding(config.d_model, max_len=100)
        
        # Transformer encoder
        encoder_layer = nn.TransformerEncoderLayer(
            d_model=config.d_model,
            nhead=config.nhead,
            dim_feedforward=config.dim_feedforward,
            dropout=config.dropout,
            batch_first=True,
        )
        self.transformer_encoder = nn.TransformerEncoder(
            encoder_layer, num_layers=config.num_encoder_layers
        )
        
        # Output projection
        self.output_projection = nn.Linear(config.d_model, config.vocab_size)
        
        # Learned queries for each output position
        self.output_queries = nn.Parameter(torch.randn(config.max_output_len, config.d_model))
        
        # Cross-attention from output positions to encoded input
        self.cross_attention = nn.MultiheadAttention(
            config.d_model, config.nhead, dropout=config.dropout, batch_first=True
        )
        
        # Final output layers
        self.output_norm = nn.LayerNorm(config.d_model)
        
        self.post_init()
    
    def forward(
        self,
        input_ids: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        labels: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> CausalLMOutputWithCrossAttentions:
        """Forward pass for HF compatibility.
        
        Args:
            input_ids: (batch_size, seq_len) tensor of digit indices (0-9), padding=10
            attention_mask: Optional mask for padding
            labels: Optional target labels for training
            
        Returns:
            CausalLMOutputWithCrossAttentions with logits
        """
        if input_ids is None:
            raise ValueError("input_ids must be provided")
        
        batch_size, seq_len = input_ids.shape
        
        # Handle padding: convert -1 padding to 10 (our padding index)
        digits = input_ids.clone()
        digits[digits == -1] = self.config.pad_token_id
        
        # Create padding mask for transformer (True = padding)
        if attention_mask is None:
            src_key_padding_mask = digits == self.config.pad_token_id
        else:
            src_key_padding_mask = ~attention_mask.bool()
        
        # Embed digits: (batch, seq_len, d_model)
        embedded = self.digit_embedding(digits)
        
        # Add positional encoding
        embedded = self.pos_encoder(embedded)
        
        # Transformer encoder: (batch, seq_len, d_model)
        memory = self.transformer_encoder(
            embedded, src_key_padding_mask=src_key_padding_mask
        )
        
        # Expand queries for batch: (batch, max_output_len, d_model)
        queries = self.output_queries.unsqueeze(0).expand(batch_size, -1, -1)
        
        # Cross-attention from queries to encoded input
        attn_output, _ = self.cross_attention(
            queries, memory, memory, key_padding_mask=src_key_padding_mask
        )
        
        # Normalize and project to vocab
        output = self.output_norm(attn_output)
        logits = self.output_projection(output)
        
        loss = None
        if labels is not None:
            loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
            loss = loss_fct(logits.view(-1, self.vocab_size), labels.view(-1))
        
        return CausalLMOutputWithCrossAttentions(
            loss=loss,
            logits=logits,
            hidden_states=None,
            attentions=None,
            cross_attentions=None,
        )
    
    def prepare_inputs_for_generation(self, input_ids, **kwargs):
        """Prepare inputs for text generation."""
        return {"input_ids": input_ids}
    
    def _reorder_cache(self, past_key_values, beam_idx):
        """Reorder cache for beam search."""
        return past_key_values


class NamerPipeline:
    """Simple pipeline for Namer model inference.
    
    Usage:
        from transformers import AutoModel
        
        # Load model
        model = AutoModel.from_pretrained(
            "edwinhere/namer", 
            trust_remote_code=True
        )
        
        # Create pipeline
        pipe = NamerPipeline(model)
        
        # Generate
        result = pipe.generate(42)  # "forty two"
        result = pipe(42)  # {"generated_text": "forty two"}
    """
    
    def __init__(self, model: NamerModel, tokenizer=None, device: str = None):
        if device is None:
            device = "cuda" if torch.cuda.is_available() else "cpu"
        self.model = model.to(device)
        self.model.eval()
        self.device = device
        self.tokenizer = tokenizer  # Placeholder if we add a tokenizer later
        
        # Vocabulary mapping (index -> word)
        # Must match utils.py vocabulary exactly
        self.id2word = {
            0: "zero", 1: "one", 2: "two", 3: "three", 4: "four",
            5: "five", 6: "six", 7: "seven", 8: "eight", 9: "nine",
            10: "ten", 11: "eleven", 12: "twelve", 13: "thirteen", 14: "fourteen",
            15: "fifteen", 16: "sixteen", 17: "seventeen", 18: "eighteen", 19: "nineteen",
            20: "twenty", 21: "thirty", 22: "forty", 23: "fifty",
            24: "sixty", 25: "seventy", 26: "eighty", 27: "ninety",
            28: "hundred",
            29: "thousand", 30: "million", 31: "billion", 32: "trillion",
            33: "quadrillion", 34: "quintillion", 35: "sextillion",
            36: "septillion", 37: "octillion", 38: "nonillion", 39: "decillion",
            40: "<EOS>"
        }
        
        # Reverse mapping
        self.word2id = {v: k for k, v in self.id2word.items()}
    
    def _int_to_digits(self, n: int) -> list[int]:
        """Convert integer to list of digit indices."""
        if n == 0:
            return [0]
        digits = []
        while n > 0:
            digits.append(n % 10)
            n //= 10
        return digits[::-1]  # Reverse to get most significant digit first
    
    def _decode(self, token_ids: list[int]) -> str:
        """Decode token IDs to text, stopping at first EOS."""
        words = []
        eos_idx = self.model.config.eos_token_id  # Should be 40
        
        for idx in token_ids:
            if idx == eos_idx:  # Stop at EOS
                break
            if idx in self.id2word:
                word = self.id2word[idx]
                if word != "<EOS>":  # Skip EOS token itself
                    words.append(word)
        
        return " ".join(words) if words else "zero"
    
    def generate(self, text: Union[str, int], **kwargs) -> str:
        """Generate English name for a number.
        
        Args:
            text: Integer or string representation of integer
            
        Returns:
            English name of the number
        """
        # Parse input
        if isinstance(text, str):
            n = int(text.strip())
        else:
            n = int(text)
        
        # Convert to digits
        digits = self._int_to_digits(n)
        
        # Pad to max length (20)
        while len(digits) < 20:
            digits.append(10)  # padding token
        
        # Create tensor
        input_ids = torch.tensor([digits], dtype=torch.long).to(self.device)
        
        # Forward pass
        with torch.no_grad():
            outputs = self.model(input_ids)
            logits = outputs.logits
            predictions = logits.argmax(dim=-1)[0].cpu().tolist()
        
        # Decode
        return self._decode(predictions)
    
    def __call__(self, text: Union[str, int], **kwargs) -> dict:
        """Callable interface for pipeline.
        
        Returns dict with 'generated_text' key for HF pipeline compatibility.
        """
        result = self.generate(text, **kwargs)
        return {"generated_text": result}


def load_namer_pipeline(model_name_or_path: str = "edwinhere/namer", device: str = None, **kwargs):
    """Load a Namer pipeline with model.
    
    This is a convenience function that loads both the model and creates
    a pipeline for easy inference.
    
    Args:
        model_name_or_path: HuggingFace model ID or local path
        device: Device to run on ('cuda', 'cpu', or None for auto)
        **kwargs: Additional args passed to from_pretrained
        
    Returns:
        NamerPipeline instance ready for inference
        
    Example:
        >>> pipe = load_namer_pipeline("edwinhere/namer")
        >>> pipe.generate(42)
        'forty two'
        >>> pipe(123)
        {'generated_text': 'one hundred twenty three'}
    """
    from transformers import AutoModel
    
    model = AutoModel.from_pretrained(
        model_name_or_path,
        trust_remote_code=True,
        **kwargs
    )
    
    return NamerPipeline(model, device=device)