File size: 9,692 Bytes
6b7b403
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Compound (Octuple-style) GPT for bach-gpt.

Each input position carries N_AXES feature ids that are embedded in
parallel and summed before the transformer. Output is N_AXES separate
classification heads with their own softmaxes; the training loss is the
(optionally weighted) sum of per-axis cross-entropies.

Usage
-----
    from compound import AXIS_SIZES, AXIS_NAMES, STEP_PAD
    from compound_model import CompoundGPT, CompoundGPTConfig, compound_loss

    cfg = CompoundGPTConfig()              # axis_sizes default to compound.AXIS_SIZES
    model = CompoundGPT(cfg)
    # idx: (B, T, N_AXES) long; targets: same shape, shifted by one step
    logits = model(idx)                    # list of (B, T, axis_size_a) tensors
    loss = compound_loss(logits, targets)
"""

from __future__ import annotations

from dataclasses import dataclass, field
from typing import List, Optional, Tuple

import torch
import torch.nn as nn
import torch.nn.functional as F

from compound import AXIS_NAMES, AXIS_SIZES, N_AXES, STEP_PAD
from model import GPT, GPTConfig, TransformerBlock


@dataclass
class CompoundGPTConfig:
    axis_sizes: Tuple[int, ...] = field(default_factory=lambda: tuple(AXIS_SIZES))
    block_size: int = 1024
    d_model: int = 512
    n_layers: int = 6
    n_heads: int = 8
    d_ff: int = 2048
    dropout: float = 0.1
    # Optional per-axis loss weighting at training time. None = uniform.
    axis_loss_weights: Optional[Tuple[float, ...]] = None


def default_compound_config() -> CompoundGPTConfig:
    return CompoundGPTConfig()


class CompoundGPT(nn.Module):
    """Decoder-only transformer over compound (multi-axis) inputs."""

    def __init__(self, config: CompoundGPTConfig) -> None:
        super().__init__()
        self.config = config
        self.n_axes = len(config.axis_sizes)
        if self.n_axes != N_AXES:
            raise ValueError(
                f"axis_sizes has {self.n_axes} entries, expected {N_AXES}"
            )

        # Per-axis input embeddings, summed across axes.
        self.input_embeds = nn.ModuleList(
            nn.Embedding(s, config.d_model) for s in config.axis_sizes
        )
        self.wpe = nn.Embedding(config.block_size, config.d_model)
        self.drop = nn.Dropout(config.dropout)

        # Reuse the regular transformer blocks. vocab_size in this fake
        # GPTConfig is unused by TransformerBlock; we only need the
        # attention/MLP shape parameters.
        block_cfg = GPTConfig(
            vocab_size=1,
            block_size=config.block_size,
            d_model=config.d_model,
            n_layers=config.n_layers,
            n_heads=config.n_heads,
            d_ff=config.d_ff,
            dropout=config.dropout,
        )
        self.blocks = nn.ModuleList(
            TransformerBlock(block_cfg) for _ in range(config.n_layers)
        )
        self.ln_f = nn.LayerNorm(config.d_model)

        # Per-axis output heads. No weight tying — each axis has its own
        # classifier over a small vocabulary.
        self.heads = nn.ModuleList(
            nn.Linear(config.d_model, s, bias=False) for s in config.axis_sizes
        )

        self.apply(GPT._init_weights)

    def forward(
        self,
        idx: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.Tensor] = None,
        return_attn: bool = False,
        return_hidden: bool = False,
        use_cache: bool = False,
        past_key_values: Optional[Tuple[torch.Tensor, ...]] = None,
    ) -> List[torch.Tensor] | Tuple[List[torch.Tensor], List[torch.Tensor]] | torch.Tensor:
        """Forward pass for compound GPT.

        - return_hidden=False, return_attn=False: list of N_AXES logits.
        - return_hidden=True: post-LN hidden states (B, T, d_model). Used by
          the contrastive MIDI encoder for pooling. No classification heads run.
        - return_attn=True: (logits, attn_list).

        Input modes:
        - ``idx``: (B, T, N_AXES) of long feature ids.
        - ``inputs_embeds``: (B, T, d_model) precomputed embeddings.
        """
        if use_cache or past_key_values is not None:
            # Kept for API compatibility with GPT-style call sites.
            raise NotImplementedError(
                "CompoundGPT does not currently support KV caching."
            )
        if (idx is None) == (inputs_embeds is None):
            raise ValueError("Provide exactly one of idx or inputs_embeds.")

        if inputs_embeds is None:
            assert idx is not None
            if idx.dim() != 3 or idx.shape[-1] != self.n_axes:
                raise ValueError(
                    f"Expected idx of shape (B, T, {self.n_axes}); got {tuple(idx.shape)}"
                )
            B, T, _ = idx.shape
        else:
            if inputs_embeds.dim() != 3 or inputs_embeds.size(-1) != self.config.d_model:
                raise ValueError(
                    "Expected inputs_embeds shape "
                    f"(B, T, {self.config.d_model}); got {tuple(inputs_embeds.shape)}"
                )
            B, T, _ = inputs_embeds.shape

        if T > self.config.block_size:
            raise ValueError(
                f"Sequence length {T} exceeds block_size {self.config.block_size}"
            )

        if inputs_embeds is None:
            assert idx is not None
            x = self.input_embeds[0](idx[..., 0])
            for a in range(1, self.n_axes):
                x = x + self.input_embeds[a](idx[..., a])
            device = idx.device
        else:
            x = inputs_embeds
            device = inputs_embeds.device

        if position_ids is None:
            pos = torch.arange(T, device=device, dtype=torch.long).unsqueeze(0)
        else:
            if position_ids.dim() == 1:
                pos = position_ids.unsqueeze(0)
            elif position_ids.dim() == 2:
                pos = position_ids
            else:
                raise ValueError(
                    f"position_ids must be shape (T,) or (B, T); got {tuple(position_ids.shape)}"
                )
            if pos.size(1) != T:
                raise ValueError(
                    f"position_ids length {pos.size(1)} must equal sequence length {T}"
                )
            if pos.size(0) not in (1, B):
                raise ValueError(
                    f"position_ids batch dim {pos.size(0)} must be 1 or {B}"
                )
            pos = pos.to(device=device, dtype=torch.long)

        x = self.drop(x + self.wpe(pos))

        attn_list: List[torch.Tensor] = []
        for block in self.blocks:
            x, aw, _ = block(x, return_attn=return_attn)
            if aw is not None:
                attn_list.append(aw)
        x = self.ln_f(x)

        if return_hidden:
            return x

        logits_per_axis = [head(x) for head in self.heads]
        if return_attn:
            return logits_per_axis, attn_list
        return logits_per_axis

    @torch.no_grad()
    def count_parameters(self) -> int:
        return sum(p.numel() for p in self.parameters())


def compound_loss(
    logits_per_axis: List[torch.Tensor],
    targets: torch.Tensor,
    axis_weights: Optional[Tuple[float, ...]] = None,
    pad_step_value: int = STEP_PAD,
    ignore_pad_steps: bool = True,
) -> torch.Tensor:
    """Sum of per-axis cross-entropies. ``targets`` shape (B, T, N_AXES).

    If ``ignore_pad_steps`` is True, positions whose step-axis target is
    ``pad_step_value`` contribute zero loss (standard padding mask).
    """
    if targets.dim() != 3:
        raise ValueError(
            f"targets must be (B, T, N_AXES); got {tuple(targets.shape)}"
        )
    n_axes = len(logits_per_axis)
    if axis_weights is None:
        axis_weights = tuple([1.0] * n_axes)
    if len(axis_weights) != n_axes:
        raise ValueError(
            f"axis_weights length {len(axis_weights)} != {n_axes}"
        )

    step_targets = targets[..., 0]              # (B, T)
    valid = step_targets != pad_step_value if ignore_pad_steps else None

    total = torch.zeros((), device=targets.device, dtype=torch.float32)
    for a in range(n_axes):
        logits = logits_per_axis[a]             # (B, T, A_a)
        tgt = targets[..., a]                   # (B, T)
        flat_logits = logits.reshape(-1, logits.size(-1))
        flat_tgt = tgt.reshape(-1)
        if valid is not None:
            flat_mask = valid.reshape(-1)
            if flat_mask.sum() == 0:
                continue
            loss_a = F.cross_entropy(
                flat_logits[flat_mask], flat_tgt[flat_mask], reduction="mean"
            )
        else:
            loss_a = F.cross_entropy(flat_logits, flat_tgt, reduction="mean")
        total = total + axis_weights[a] * loss_a
    return total


# --- Smoke test --------------------------------------------------------------

if __name__ == "__main__":
    cfg = default_compound_config()
    cfg.block_size = 64
    cfg.n_layers = 2
    cfg.d_model = 128
    cfg.d_ff = 512
    model = CompoundGPT(cfg)
    print(f"Compound axes ({N_AXES}): {AXIS_NAMES}")
    print(f"Axis sizes: {AXIS_SIZES}")
    print(
        f"Parameters: {model.count_parameters():,} "
        f"(~{model.count_parameters()/1e6:.2f}M)"
    )

    B, T = 2, 32
    idx = torch.stack([
        torch.randint(0, AXIS_SIZES[a], (B, T)) for a in range(N_AXES)
    ], dim=-1).long()
    logits = model(idx)
    assert len(logits) == N_AXES
    for a, l in enumerate(logits):
        assert l.shape == (B, T, AXIS_SIZES[a]), (a, l.shape, AXIS_SIZES[a])

    loss = compound_loss(logits, idx)
    print(f"Forward + per-axis loss OK. loss={loss.item():.4f}")