Update modeling_esm_plusplus.py
Browse files- modeling_esm_plusplus.py +636 -635
modeling_esm_plusplus.py
CHANGED
@@ -1,635 +1,636 @@
|
|
1 |
-
### Modified from https://github.com/evolutionaryscale/esm
|
2 |
-
### License: https://www.evolutionaryscale.ai/policies/cambrian-non-commercial-license-agreement
|
3 |
-
import torch
|
4 |
-
import torch.nn as nn
|
5 |
-
import torch.nn.functional as F
|
6 |
-
import math
|
7 |
-
from dataclasses import dataclass
|
8 |
-
from transformers import PreTrainedModel, PretrainedConfig
|
9 |
-
from einops import rearrange, repeat
|
10 |
-
from functools import partial
|
11 |
-
from typing import Optional, Tuple
|
12 |
-
from transformers.modeling_outputs import ModelOutput
|
13 |
-
|
14 |
-
|
15 |
-
class ESMplusplusConfig(PretrainedConfig):
|
16 |
-
model_type = "ESMplusplus"
|
17 |
-
def __init__(
|
18 |
-
self,
|
19 |
-
vocab_size: int = 64,
|
20 |
-
hidden_size: int = 960,
|
21 |
-
num_attention_heads: int = 15,
|
22 |
-
num_hidden_layers: int = 30,
|
23 |
-
num_labels: int = 2,
|
24 |
-
problem_type: str | None = None,
|
25 |
-
**kwargs,
|
26 |
-
):
|
27 |
-
super().__init__(**kwargs)
|
28 |
-
self.vocab_size = vocab_size
|
29 |
-
self.hidden_size = hidden_size
|
30 |
-
self.num_attention_heads = num_attention_heads
|
31 |
-
self.num_hidden_layers = num_hidden_layers
|
32 |
-
self.num_labels = num_labels
|
33 |
-
self.problem_type = problem_type
|
34 |
-
|
35 |
-
|
36 |
-
### Rotary
|
37 |
-
# https://github.com/evolutionaryscale/esm/blob/main/esm/layers/rotary.py
|
38 |
-
# https://huggingface.co/togethercomputer/LLaMA-2-7B-32K/blob/08639a72e17836184096ae6a7e2766f2a34c3e36/modeling_flash_llama.py#L114
|
39 |
-
# Flash attention rotary implementation can be installed like so: `pip install git+https://github.com/HazyResearch/flash-attention.git#subdirectory=csrc/rotary`
|
40 |
-
def rotate_half(x, interleaved=False):
|
41 |
-
if not interleaved:
|
42 |
-
x1, x2 = x.chunk(2, dim=-1)
|
43 |
-
return torch.cat((-x2, x1), dim=-1)
|
44 |
-
else:
|
45 |
-
x1, x2 = x[..., ::2], x[..., 1::2]
|
46 |
-
return rearrange(
|
47 |
-
torch.stack((-x2, x1), dim=-1), "... d two -> ... (d two)", two=2
|
48 |
-
)
|
49 |
-
|
50 |
-
|
51 |
-
def apply_rotary_emb_torch(x, cos, sin, interleaved=False, _inplace=False):
|
52 |
-
"""
|
53 |
-
x: (batch_size, seqlen, nheads, headdim)
|
54 |
-
cos, sin: (seqlen, rotary_dim / 2)
|
55 |
-
"""
|
56 |
-
ro_dim = cos.shape[-1] * 2
|
57 |
-
assert ro_dim <= x.shape[-1]
|
58 |
-
seqlen = x.size(1)
|
59 |
-
cos = cos[:seqlen]
|
60 |
-
sin = sin[:seqlen]
|
61 |
-
cos = repeat(cos, "s d -> s 1 (2 d)")
|
62 |
-
sin = repeat(sin, "s d -> s 1 (2 d)")
|
63 |
-
return torch.cat(
|
64 |
-
[
|
65 |
-
x[..., :ro_dim] * cos + rotate_half(x[..., :ro_dim], interleaved) * sin,
|
66 |
-
x[..., ro_dim:],
|
67 |
-
],
|
68 |
-
dim=-1,
|
69 |
-
)
|
70 |
-
|
71 |
-
|
72 |
-
class RotaryEmbedding(torch.nn.Module):
|
73 |
-
def __init__(
|
74 |
-
self,
|
75 |
-
dim: int,
|
76 |
-
base=10000.0,
|
77 |
-
interleaved=False,
|
78 |
-
scale_base=None,
|
79 |
-
scaling_factor=1.0,
|
80 |
-
pos_idx_in_fp32=True,
|
81 |
-
device=None,
|
82 |
-
):
|
83 |
-
super().__init__()
|
84 |
-
self.dim = dim
|
85 |
-
self.base = float(base)
|
86 |
-
self.pos_idx_in_fp32 = pos_idx_in_fp32
|
87 |
-
# Generate and save the inverse frequency buffer (non trainable)
|
88 |
-
self.interleaved = interleaved
|
89 |
-
self.scale_base = scale_base
|
90 |
-
self.scaling_factor = scaling_factor
|
91 |
-
self.device = device
|
92 |
-
|
93 |
-
self._seq_len_cached = 0
|
94 |
-
self._cos_cached = None
|
95 |
-
self._sin_cached = None
|
96 |
-
self._cos_k_cached = None
|
97 |
-
self._sin_k_cached = None
|
98 |
-
self.reset_parameters()
|
99 |
-
|
100 |
-
def reset_parameters(self):
|
101 |
-
inv_freq = self._compute_inv_freq(self.device)
|
102 |
-
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
103 |
-
arange = torch.arange(0, self.dim, 2, device=self.device, dtype=torch.float32)
|
104 |
-
scale = (
|
105 |
-
(arange + 0.4 * self.dim) / (1.4 * self.dim)
|
106 |
-
if self.scale_base is not None
|
107 |
-
else None
|
108 |
-
)
|
109 |
-
self.register_buffer("scale", scale)
|
110 |
-
|
111 |
-
def _compute_inv_freq(self, device=None):
|
112 |
-
return 1 / (
|
113 |
-
self.base
|
114 |
-
** (
|
115 |
-
torch.arange(0, self.dim, 2, device=device, dtype=torch.float32)
|
116 |
-
/ self.dim
|
117 |
-
)
|
118 |
-
)
|
119 |
-
|
120 |
-
def _update_cos_sin_cache(self, seqlen, device=None, dtype=None):
|
121 |
-
if (
|
122 |
-
seqlen > self._seq_len_cached
|
123 |
-
or self._cos_cached is None
|
124 |
-
or self._cos_cached.device != device
|
125 |
-
or self._cos_cached.dtype != dtype
|
126 |
-
or (self.training and self._cos_cached.is_inference())
|
127 |
-
):
|
128 |
-
self._seq_len_cached = seqlen
|
129 |
-
if self.pos_idx_in_fp32:
|
130 |
-
t = torch.arange(seqlen, device=device, dtype=torch.float32)
|
131 |
-
t /= self.scaling_factor
|
132 |
-
if self.inv_freq.dtype != torch.float32:
|
133 |
-
inv_freq = self.inv_freq.to(torch.float32)
|
134 |
-
else:
|
135 |
-
inv_freq = self.inv_freq
|
136 |
-
else:
|
137 |
-
t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
|
138 |
-
t /= self.scaling_factor
|
139 |
-
inv_freq = self.inv_freq
|
140 |
-
freqs = torch.outer(t, inv_freq)
|
141 |
-
|
142 |
-
if self.scale is None:
|
143 |
-
self._cos_cached = torch.cos(freqs).to(dtype)
|
144 |
-
self._sin_cached = torch.sin(freqs).to(dtype)
|
145 |
-
else:
|
146 |
-
power = (
|
147 |
-
torch.arange(
|
148 |
-
seqlen, dtype=self.scale.dtype, device=self.scale.device
|
149 |
-
)
|
150 |
-
- seqlen // 2
|
151 |
-
) / self.scale_base
|
152 |
-
scale = self.scale.to(device=power.device) ** power.unsqueeze(-1)
|
153 |
-
self._cos_cached = (torch.cos(freqs) * scale).to(dtype)
|
154 |
-
self._sin_cached = (torch.sin(freqs) * scale).to(dtype)
|
155 |
-
self._cos_k_cached = (torch.cos(freqs) / scale).to(dtype)
|
156 |
-
self._sin_k_cached = (torch.sin(freqs) / scale).to(dtype)
|
157 |
-
|
158 |
-
def forward(self, q: torch.Tensor, k: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
|
159 |
-
"""
|
160 |
-
q: (batch, seqlen, nheads, headdim)
|
161 |
-
k: (batch, seqlen, nheads, headdim)
|
162 |
-
"""
|
163 |
-
self._update_cos_sin_cache(q.shape[1], device=q.device, dtype=q.dtype)
|
164 |
-
assert self._cos_cached is not None
|
165 |
-
assert self._sin_cached is not None
|
166 |
-
if self.scale is None:
|
167 |
-
return (
|
168 |
-
apply_rotary_emb_torch(
|
169 |
-
q,
|
170 |
-
self._cos_cached,
|
171 |
-
self._sin_cached,
|
172 |
-
self.interleaved,
|
173 |
-
True, # inplace=True
|
174 |
-
),
|
175 |
-
apply_rotary_emb_torch(
|
176 |
-
k,
|
177 |
-
self._cos_cached,
|
178 |
-
self._sin_cached,
|
179 |
-
self.interleaved,
|
180 |
-
True, # inplace=True
|
181 |
-
),
|
182 |
-
) # type: ignore
|
183 |
-
else:
|
184 |
-
assert False
|
185 |
-
|
186 |
-
|
187 |
-
### Feedforward
|
188 |
-
def swiglu_correction_fn(expansion_ratio: float, d_model: int) -> int:
|
189 |
-
return int(((expansion_ratio * d_model) + 255) // 256 * 256)
|
190 |
-
|
191 |
-
|
192 |
-
class SwiGLU(nn.Module):
|
193 |
-
def __init__(self):
|
194 |
-
super(SwiGLU, self).__init__()
|
195 |
-
|
196 |
-
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
197 |
-
x1, x2 = x.chunk(2, dim=-1)
|
198 |
-
return F.silu(x1) * x2
|
199 |
-
|
200 |
-
|
201 |
-
def swiglu_ln_ffn(d_model: int, expansion_ratio: float):
|
202 |
-
return nn.Sequential(
|
203 |
-
nn.LayerNorm(d_model),
|
204 |
-
nn.Linear(
|
205 |
-
d_model, swiglu_correction_fn(expansion_ratio, d_model) * 2, bias=False
|
206 |
-
),
|
207 |
-
SwiGLU(),
|
208 |
-
nn.Linear(swiglu_correction_fn(expansion_ratio, d_model), d_model, bias=False),
|
209 |
-
)
|
210 |
-
|
211 |
-
|
212 |
-
### Attention
|
213 |
-
class MultiHeadAttention(nn.Module):
|
214 |
-
def __init__(self, d_model: int, n_heads: int):
|
215 |
-
super().__init__()
|
216 |
-
self.d_model = d_model
|
217 |
-
self.n_heads = n_heads
|
218 |
-
self.d_head = self.d_model // self.n_heads
|
219 |
-
self.layernorm_qkv = nn.Sequential(
|
220 |
-
nn.LayerNorm(d_model), nn.Linear(d_model, d_model * 3, bias=False)
|
221 |
-
)
|
222 |
-
self.out_proj = nn.Linear(d_model, d_model, bias=False)
|
223 |
-
self.q_ln = nn.LayerNorm(d_model, bias=False)
|
224 |
-
self.k_ln = nn.LayerNorm(d_model, bias=False)
|
225 |
-
self.reshaper = partial(rearrange, pattern="b s (h d) -> b h s d", h=n_heads)
|
226 |
-
self.rotary = RotaryEmbedding(d_model // n_heads)
|
227 |
-
|
228 |
-
def _apply_rotary(self, q: torch.Tensor, k: torch.Tensor):
|
229 |
-
q = q.unflatten(-1, (self.n_heads, self.d_head))
|
230 |
-
k = k.unflatten(-1, (self.n_heads, self.d_head))
|
231 |
-
q, k = self.rotary(q, k)
|
232 |
-
q = q.flatten(-2, -1)
|
233 |
-
k = k.flatten(-2, -1)
|
234 |
-
return q, k
|
235 |
-
|
236 |
-
def forward(self, x, attention_mask=None):
|
237 |
-
qkv_BLD3 = self.layernorm_qkv(x)
|
238 |
-
query_BLD, key_BLD, value_BLD = torch.chunk(qkv_BLD3, 3, dim=-1)
|
239 |
-
query_BLD, key_BLD = (
|
240 |
-
self.q_ln(query_BLD).to(query_BLD.dtype),
|
241 |
-
self.k_ln(key_BLD).to(query_BLD.dtype),
|
242 |
-
)
|
243 |
-
query_BLD, key_BLD = self._apply_rotary(query_BLD, key_BLD)
|
244 |
-
query_BHLD, key_BHLD, value_BHLD = map(self.reshaper, (query_BLD, key_BLD, value_BLD))
|
245 |
-
context_BHLD = F.scaled_dot_product_attention(
|
246 |
-
query_BHLD, key_BHLD, value_BHLD, attention_mask
|
247 |
-
)
|
248 |
-
context_BLD = rearrange(context_BHLD, "b h s d -> b s (h d)")
|
249 |
-
return self.out_proj(context_BLD)
|
250 |
-
|
251 |
-
|
252 |
-
### LM Head
|
253 |
-
def RegressionHead(
|
254 |
-
d_model: int, output_dim: int, hidden_dim: int | None = None
|
255 |
-
) -> nn.Module:
|
256 |
-
hidden_dim = hidden_dim if hidden_dim is not None else d_model
|
257 |
-
return nn.Sequential(
|
258 |
-
nn.Linear(d_model, hidden_dim),
|
259 |
-
nn.GELU(),
|
260 |
-
nn.LayerNorm(hidden_dim),
|
261 |
-
nn.Linear(hidden_dim, output_dim),
|
262 |
-
)
|
263 |
-
|
264 |
-
|
265 |
-
### Transformer Block
|
266 |
-
class UnifiedTransformerBlock(nn.Module):
|
267 |
-
def __init__(
|
268 |
-
self,
|
269 |
-
d_model: int,
|
270 |
-
n_heads: int,
|
271 |
-
residue_scaling_factor: float = 1,
|
272 |
-
expansion_ratio: float = 8 / 3,
|
273 |
-
):
|
274 |
-
super().__init__()
|
275 |
-
self.attn = MultiHeadAttention(d_model, n_heads)
|
276 |
-
self.ffn = swiglu_ln_ffn(d_model, expansion_ratio)
|
277 |
-
self.scaling_factor = residue_scaling_factor
|
278 |
-
|
279 |
-
def forward(
|
280 |
-
self,
|
281 |
-
x: torch.Tensor,
|
282 |
-
attention_mask: Optional[torch.Tensor] = None,
|
283 |
-
) -> torch.Tensor:
|
284 |
-
r1 = self.attn(x, attention_mask)
|
285 |
-
x = x + r1 / self.scaling_factor
|
286 |
-
r3 = self.ffn(x) / self.scaling_factor
|
287 |
-
x = x + r3
|
288 |
-
return x
|
289 |
-
|
290 |
-
|
291 |
-
### Outputs
|
292 |
-
@dataclass
|
293 |
-
class TransformerOutput(ModelOutput):
|
294 |
-
last_hidden_state: torch.Tensor | None = None
|
295 |
-
hidden_states: tuple[torch.Tensor] | None = None
|
296 |
-
|
297 |
-
|
298 |
-
@dataclass
|
299 |
-
class ESMplusplusOutput(ModelOutput):
|
300 |
-
loss: torch.Tensor | None = None
|
301 |
-
logits: torch.Tensor | None = None
|
302 |
-
last_hidden_state: torch.Tensor | None = None
|
303 |
-
hidden_states: tuple[torch.Tensor] | None = None
|
304 |
-
|
305 |
-
|
306 |
-
### Transformer
|
307 |
-
class TransformerStack(nn.Module):
|
308 |
-
def __init__(
|
309 |
-
self,
|
310 |
-
d_model: int,
|
311 |
-
n_heads: int,
|
312 |
-
n_layers: int,
|
313 |
-
):
|
314 |
-
super().__init__()
|
315 |
-
self.blocks = nn.ModuleList(
|
316 |
-
[
|
317 |
-
UnifiedTransformerBlock(
|
318 |
-
d_model,
|
319 |
-
n_heads,
|
320 |
-
residue_scaling_factor=math.sqrt(n_layers / 36),
|
321 |
-
)
|
322 |
-
for i in range(n_layers)
|
323 |
-
]
|
324 |
-
)
|
325 |
-
self.norm = nn.LayerNorm(d_model, bias=False)
|
326 |
-
|
327 |
-
def forward(
|
328 |
-
self,
|
329 |
-
x: torch.Tensor,
|
330 |
-
attention_mask: Optional[torch.Tensor] = None,
|
331 |
-
output_hidden_states: bool = False,
|
332 |
-
) -> TransformerOutput:
|
333 |
-
batch_size, seq_len, _ = x.shape
|
334 |
-
hidden_states = ()
|
335 |
-
if attention_mask is not None:
|
336 |
-
attention_mask = attention_mask[:, None, None, :].expand(batch_size, 1, seq_len, seq_len).bool()
|
337 |
-
for block in self.blocks:
|
338 |
-
x = block(x, attention_mask)
|
339 |
-
if output_hidden_states:
|
340 |
-
hidden_states += (x,)
|
341 |
-
return TransformerOutput(last_hidden_state=self.norm(x), hidden_states=hidden_states)
|
342 |
-
|
343 |
-
|
344 |
-
### Full model
|
345 |
-
class ESMplusplusForMaskedLM(PreTrainedModel):
|
346 |
-
"""
|
347 |
-
ESM++ for masked language modeling.
|
348 |
-
"""
|
349 |
-
config_class = ESMplusplusConfig
|
350 |
-
def __init__(self, config: ESMplusplusConfig):
|
351 |
-
super().__init__(config)
|
352 |
-
self.config = config
|
353 |
-
self.vocab_size = config.vocab_size
|
354 |
-
self.embed = nn.Embedding(self.vocab_size, config.hidden_size)
|
355 |
-
self.transformer = TransformerStack(config.hidden_size, config.num_attention_heads, config.num_hidden_layers)
|
356 |
-
self.sequence_head = RegressionHead(config.hidden_size, self.vocab_size)
|
357 |
-
self.ce_loss = nn.CrossEntropyLoss()
|
358 |
-
self.tokenizer = EsmSequenceTokenizer()
|
359 |
-
|
360 |
-
@classmethod
|
361 |
-
def from_pretrained_esm(cls, model_name: str):
|
362 |
-
if '300' in model_name:
|
363 |
-
return ESMplusplus_300M()
|
364 |
-
elif '600' in model_name:
|
365 |
-
return ESMplusplus_600M()
|
366 |
-
else:
|
367 |
-
raise ValueError(f"Invalid model name: {model_name}")
|
368 |
-
|
369 |
-
@property
|
370 |
-
def device(self):
|
371 |
-
return next(self.parameters()).device
|
372 |
-
|
373 |
-
def forward(
|
374 |
-
self,
|
375 |
-
input_ids: torch.Tensor | None = None,
|
376 |
-
attention_mask: Optional[torch.Tensor] = None,
|
377 |
-
labels: Optional[torch.Tensor] = None,
|
378 |
-
output_hidden_states: bool = False,
|
379 |
-
) -> ESMplusplusOutput:
|
380 |
-
x = self.embed(input_ids)
|
381 |
-
output = self.transformer(x, attention_mask, output_hidden_states)
|
382 |
-
x = output.last_hidden_state
|
383 |
-
logits = self.sequence_head(x)
|
384 |
-
loss = None
|
385 |
-
if labels is not None:
|
386 |
-
loss = self.ce_loss(logits.view(-1, self.vocab_size), labels.view(-1))
|
387 |
-
return ESMplusplusOutput(
|
388 |
-
loss=loss,
|
389 |
-
logits=logits,
|
390 |
-
last_hidden_state=x,
|
391 |
-
hidden_states=output.hidden_states,
|
392 |
-
)
|
393 |
-
|
394 |
-
|
395 |
-
class ESMplusplusForSequenceClassification(ESMplusplusForMaskedLM):
|
396 |
-
"""
|
397 |
-
ESM++ for sequence classification.
|
398 |
-
"""
|
399 |
-
def __init__(self, config: ESMplusplusConfig):
|
400 |
-
super().__init__(config)
|
401 |
-
self.config = config
|
402 |
-
self.classifier = RegressionHead(config.hidden_size * 2, config.num_labels, config.hidden_size * 4)
|
403 |
-
# we find that large intermediate projections help with sequence classification tasks (*4)
|
404 |
-
self.mse = nn.MSELoss()
|
405 |
-
self.ce = nn.CrossEntropyLoss()
|
406 |
-
self.bce = nn.BCEWithLogitsLoss()
|
407 |
-
|
408 |
-
def mean_pooling(self, x: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
|
409 |
-
# x: (batch_size, seq_len, hidden_size)
|
410 |
-
# attention_mask: (batch_size, seq_len)
|
411 |
-
if attention_mask is None:
|
412 |
-
return x.mean(dim=1)
|
413 |
-
else:
|
414 |
-
|
415 |
-
|
416 |
-
|
417 |
-
|
418 |
-
|
419 |
-
|
420 |
-
|
421 |
-
|
422 |
-
|
423 |
-
|
424 |
-
|
425 |
-
|
426 |
-
|
427 |
-
|
428 |
-
features
|
429 |
-
|
430 |
-
|
431 |
-
|
432 |
-
|
433 |
-
|
434 |
-
|
435 |
-
|
436 |
-
|
437 |
-
|
438 |
-
|
439 |
-
|
440 |
-
|
441 |
-
|
442 |
-
|
443 |
-
|
444 |
-
|
445 |
-
|
446 |
-
|
447 |
-
|
448 |
-
|
449 |
-
|
450 |
-
|
451 |
-
|
452 |
-
|
453 |
-
|
454 |
-
|
455 |
-
|
456 |
-
|
457 |
-
|
458 |
-
|
459 |
-
|
460 |
-
|
461 |
-
|
462 |
-
|
463 |
-
|
464 |
-
|
465 |
-
self.
|
466 |
-
self.
|
467 |
-
|
468 |
-
|
469 |
-
|
470 |
-
|
471 |
-
|
472 |
-
|
473 |
-
|
474 |
-
|
475 |
-
|
476 |
-
|
477 |
-
|
478 |
-
|
479 |
-
|
480 |
-
|
481 |
-
|
482 |
-
|
483 |
-
|
484 |
-
|
485 |
-
|
486 |
-
|
487 |
-
|
488 |
-
|
489 |
-
|
490 |
-
|
491 |
-
|
492 |
-
|
493 |
-
|
494 |
-
from
|
495 |
-
from
|
496 |
-
|
497 |
-
|
498 |
-
|
499 |
-
@
|
500 |
-
|
501 |
-
|
502 |
-
|
503 |
-
|
504 |
-
if
|
505 |
-
|
506 |
-
|
507 |
-
|
508 |
-
|
509 |
-
|
510 |
-
|
511 |
-
|
512 |
-
|
513 |
-
|
514 |
-
|
515 |
-
|
516 |
-
|
517 |
-
|
518 |
-
|
519 |
-
|
520 |
-
|
521 |
-
|
522 |
-
|
523 |
-
|
524 |
-
|
525 |
-
|
526 |
-
|
527 |
-
|
528 |
-
|
529 |
-
|
530 |
-
|
531 |
-
|
532 |
-
|
533 |
-
|
534 |
-
|
535 |
-
|
536 |
-
|
537 |
-
|
538 |
-
|
539 |
-
|
540 |
-
|
541 |
-
|
542 |
-
|
543 |
-
|
544 |
-
|
545 |
-
|
546 |
-
|
547 |
-
from tokenizers
|
548 |
-
from tokenizers.
|
549 |
-
from
|
550 |
-
|
551 |
-
|
552 |
-
|
553 |
-
|
554 |
-
"
|
555 |
-
"
|
556 |
-
"
|
557 |
-
"
|
558 |
-
|
559 |
-
|
560 |
-
|
561 |
-
|
562 |
-
|
563 |
-
|
564 |
-
|
565 |
-
|
566 |
-
|
567 |
-
|
568 |
-
|
569 |
-
|
570 |
-
|
571 |
-
|
572 |
-
|
573 |
-
|
574 |
-
|
575 |
-
|
576 |
-
|
577 |
-
|
578 |
-
|
579 |
-
|
580 |
-
|
581 |
-
|
582 |
-
|
583 |
-
|
584 |
-
|
585 |
-
|
586 |
-
|
587 |
-
|
588 |
-
|
589 |
-
|
590 |
-
|
591 |
-
|
592 |
-
#
|
593 |
-
#
|
594 |
-
|
595 |
-
|
596 |
-
|
597 |
-
|
598 |
-
("<
|
599 |
-
|
600 |
-
|
601 |
-
|
602 |
-
|
603 |
-
|
604 |
-
|
605 |
-
|
606 |
-
|
607 |
-
|
608 |
-
|
609 |
-
|
610 |
-
|
611 |
-
|
612 |
-
|
613 |
-
|
614 |
-
|
615 |
-
|
616 |
-
|
617 |
-
|
618 |
-
|
619 |
-
|
620 |
-
|
621 |
-
|
622 |
-
|
623 |
-
|
624 |
-
|
625 |
-
|
626 |
-
|
627 |
-
|
628 |
-
|
629 |
-
|
630 |
-
|
631 |
-
|
632 |
-
|
633 |
-
|
634 |
-
|
635 |
-
|
|
|
|
1 |
+
### Modified from https://github.com/evolutionaryscale/esm
|
2 |
+
### License: https://www.evolutionaryscale.ai/policies/cambrian-non-commercial-license-agreement
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
import torch.nn.functional as F
|
6 |
+
import math
|
7 |
+
from dataclasses import dataclass
|
8 |
+
from transformers import PreTrainedModel, PretrainedConfig
|
9 |
+
from einops import rearrange, repeat
|
10 |
+
from functools import partial
|
11 |
+
from typing import Optional, Tuple
|
12 |
+
from transformers.modeling_outputs import ModelOutput
|
13 |
+
|
14 |
+
|
15 |
+
class ESMplusplusConfig(PretrainedConfig):
|
16 |
+
model_type = "ESMplusplus"
|
17 |
+
def __init__(
|
18 |
+
self,
|
19 |
+
vocab_size: int = 64,
|
20 |
+
hidden_size: int = 960,
|
21 |
+
num_attention_heads: int = 15,
|
22 |
+
num_hidden_layers: int = 30,
|
23 |
+
num_labels: int = 2,
|
24 |
+
problem_type: str | None = None,
|
25 |
+
**kwargs,
|
26 |
+
):
|
27 |
+
super().__init__(**kwargs)
|
28 |
+
self.vocab_size = vocab_size
|
29 |
+
self.hidden_size = hidden_size
|
30 |
+
self.num_attention_heads = num_attention_heads
|
31 |
+
self.num_hidden_layers = num_hidden_layers
|
32 |
+
self.num_labels = num_labels
|
33 |
+
self.problem_type = problem_type
|
34 |
+
|
35 |
+
|
36 |
+
### Rotary
|
37 |
+
# https://github.com/evolutionaryscale/esm/blob/main/esm/layers/rotary.py
|
38 |
+
# https://huggingface.co/togethercomputer/LLaMA-2-7B-32K/blob/08639a72e17836184096ae6a7e2766f2a34c3e36/modeling_flash_llama.py#L114
|
39 |
+
# Flash attention rotary implementation can be installed like so: `pip install git+https://github.com/HazyResearch/flash-attention.git#subdirectory=csrc/rotary`
|
40 |
+
def rotate_half(x, interleaved=False):
|
41 |
+
if not interleaved:
|
42 |
+
x1, x2 = x.chunk(2, dim=-1)
|
43 |
+
return torch.cat((-x2, x1), dim=-1)
|
44 |
+
else:
|
45 |
+
x1, x2 = x[..., ::2], x[..., 1::2]
|
46 |
+
return rearrange(
|
47 |
+
torch.stack((-x2, x1), dim=-1), "... d two -> ... (d two)", two=2
|
48 |
+
)
|
49 |
+
|
50 |
+
|
51 |
+
def apply_rotary_emb_torch(x, cos, sin, interleaved=False, _inplace=False):
|
52 |
+
"""
|
53 |
+
x: (batch_size, seqlen, nheads, headdim)
|
54 |
+
cos, sin: (seqlen, rotary_dim / 2)
|
55 |
+
"""
|
56 |
+
ro_dim = cos.shape[-1] * 2
|
57 |
+
assert ro_dim <= x.shape[-1]
|
58 |
+
seqlen = x.size(1)
|
59 |
+
cos = cos[:seqlen]
|
60 |
+
sin = sin[:seqlen]
|
61 |
+
cos = repeat(cos, "s d -> s 1 (2 d)")
|
62 |
+
sin = repeat(sin, "s d -> s 1 (2 d)")
|
63 |
+
return torch.cat(
|
64 |
+
[
|
65 |
+
x[..., :ro_dim] * cos + rotate_half(x[..., :ro_dim], interleaved) * sin,
|
66 |
+
x[..., ro_dim:],
|
67 |
+
],
|
68 |
+
dim=-1,
|
69 |
+
)
|
70 |
+
|
71 |
+
|
72 |
+
class RotaryEmbedding(torch.nn.Module):
|
73 |
+
def __init__(
|
74 |
+
self,
|
75 |
+
dim: int,
|
76 |
+
base=10000.0,
|
77 |
+
interleaved=False,
|
78 |
+
scale_base=None,
|
79 |
+
scaling_factor=1.0,
|
80 |
+
pos_idx_in_fp32=True,
|
81 |
+
device=None,
|
82 |
+
):
|
83 |
+
super().__init__()
|
84 |
+
self.dim = dim
|
85 |
+
self.base = float(base)
|
86 |
+
self.pos_idx_in_fp32 = pos_idx_in_fp32
|
87 |
+
# Generate and save the inverse frequency buffer (non trainable)
|
88 |
+
self.interleaved = interleaved
|
89 |
+
self.scale_base = scale_base
|
90 |
+
self.scaling_factor = scaling_factor
|
91 |
+
self.device = device
|
92 |
+
|
93 |
+
self._seq_len_cached = 0
|
94 |
+
self._cos_cached = None
|
95 |
+
self._sin_cached = None
|
96 |
+
self._cos_k_cached = None
|
97 |
+
self._sin_k_cached = None
|
98 |
+
self.reset_parameters()
|
99 |
+
|
100 |
+
def reset_parameters(self):
|
101 |
+
inv_freq = self._compute_inv_freq(self.device)
|
102 |
+
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
103 |
+
arange = torch.arange(0, self.dim, 2, device=self.device, dtype=torch.float32)
|
104 |
+
scale = (
|
105 |
+
(arange + 0.4 * self.dim) / (1.4 * self.dim)
|
106 |
+
if self.scale_base is not None
|
107 |
+
else None
|
108 |
+
)
|
109 |
+
self.register_buffer("scale", scale)
|
110 |
+
|
111 |
+
def _compute_inv_freq(self, device=None):
|
112 |
+
return 1 / (
|
113 |
+
self.base
|
114 |
+
** (
|
115 |
+
torch.arange(0, self.dim, 2, device=device, dtype=torch.float32)
|
116 |
+
/ self.dim
|
117 |
+
)
|
118 |
+
)
|
119 |
+
|
120 |
+
def _update_cos_sin_cache(self, seqlen, device=None, dtype=None):
|
121 |
+
if (
|
122 |
+
seqlen > self._seq_len_cached
|
123 |
+
or self._cos_cached is None
|
124 |
+
or self._cos_cached.device != device
|
125 |
+
or self._cos_cached.dtype != dtype
|
126 |
+
or (self.training and self._cos_cached.is_inference())
|
127 |
+
):
|
128 |
+
self._seq_len_cached = seqlen
|
129 |
+
if self.pos_idx_in_fp32:
|
130 |
+
t = torch.arange(seqlen, device=device, dtype=torch.float32)
|
131 |
+
t /= self.scaling_factor
|
132 |
+
if self.inv_freq.dtype != torch.float32:
|
133 |
+
inv_freq = self.inv_freq.to(torch.float32)
|
134 |
+
else:
|
135 |
+
inv_freq = self.inv_freq
|
136 |
+
else:
|
137 |
+
t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
|
138 |
+
t /= self.scaling_factor
|
139 |
+
inv_freq = self.inv_freq
|
140 |
+
freqs = torch.outer(t, inv_freq)
|
141 |
+
|
142 |
+
if self.scale is None:
|
143 |
+
self._cos_cached = torch.cos(freqs).to(dtype)
|
144 |
+
self._sin_cached = torch.sin(freqs).to(dtype)
|
145 |
+
else:
|
146 |
+
power = (
|
147 |
+
torch.arange(
|
148 |
+
seqlen, dtype=self.scale.dtype, device=self.scale.device
|
149 |
+
)
|
150 |
+
- seqlen // 2
|
151 |
+
) / self.scale_base
|
152 |
+
scale = self.scale.to(device=power.device) ** power.unsqueeze(-1)
|
153 |
+
self._cos_cached = (torch.cos(freqs) * scale).to(dtype)
|
154 |
+
self._sin_cached = (torch.sin(freqs) * scale).to(dtype)
|
155 |
+
self._cos_k_cached = (torch.cos(freqs) / scale).to(dtype)
|
156 |
+
self._sin_k_cached = (torch.sin(freqs) / scale).to(dtype)
|
157 |
+
|
158 |
+
def forward(self, q: torch.Tensor, k: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
|
159 |
+
"""
|
160 |
+
q: (batch, seqlen, nheads, headdim)
|
161 |
+
k: (batch, seqlen, nheads, headdim)
|
162 |
+
"""
|
163 |
+
self._update_cos_sin_cache(q.shape[1], device=q.device, dtype=q.dtype)
|
164 |
+
assert self._cos_cached is not None
|
165 |
+
assert self._sin_cached is not None
|
166 |
+
if self.scale is None:
|
167 |
+
return (
|
168 |
+
apply_rotary_emb_torch(
|
169 |
+
q,
|
170 |
+
self._cos_cached,
|
171 |
+
self._sin_cached,
|
172 |
+
self.interleaved,
|
173 |
+
True, # inplace=True
|
174 |
+
),
|
175 |
+
apply_rotary_emb_torch(
|
176 |
+
k,
|
177 |
+
self._cos_cached,
|
178 |
+
self._sin_cached,
|
179 |
+
self.interleaved,
|
180 |
+
True, # inplace=True
|
181 |
+
),
|
182 |
+
) # type: ignore
|
183 |
+
else:
|
184 |
+
assert False
|
185 |
+
|
186 |
+
|
187 |
+
### Feedforward
|
188 |
+
def swiglu_correction_fn(expansion_ratio: float, d_model: int) -> int:
|
189 |
+
return int(((expansion_ratio * d_model) + 255) // 256 * 256)
|
190 |
+
|
191 |
+
|
192 |
+
class SwiGLU(nn.Module):
|
193 |
+
def __init__(self):
|
194 |
+
super(SwiGLU, self).__init__()
|
195 |
+
|
196 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
197 |
+
x1, x2 = x.chunk(2, dim=-1)
|
198 |
+
return F.silu(x1) * x2
|
199 |
+
|
200 |
+
|
201 |
+
def swiglu_ln_ffn(d_model: int, expansion_ratio: float):
|
202 |
+
return nn.Sequential(
|
203 |
+
nn.LayerNorm(d_model),
|
204 |
+
nn.Linear(
|
205 |
+
d_model, swiglu_correction_fn(expansion_ratio, d_model) * 2, bias=False
|
206 |
+
),
|
207 |
+
SwiGLU(),
|
208 |
+
nn.Linear(swiglu_correction_fn(expansion_ratio, d_model), d_model, bias=False),
|
209 |
+
)
|
210 |
+
|
211 |
+
|
212 |
+
### Attention
|
213 |
+
class MultiHeadAttention(nn.Module):
|
214 |
+
def __init__(self, d_model: int, n_heads: int):
|
215 |
+
super().__init__()
|
216 |
+
self.d_model = d_model
|
217 |
+
self.n_heads = n_heads
|
218 |
+
self.d_head = self.d_model // self.n_heads
|
219 |
+
self.layernorm_qkv = nn.Sequential(
|
220 |
+
nn.LayerNorm(d_model), nn.Linear(d_model, d_model * 3, bias=False)
|
221 |
+
)
|
222 |
+
self.out_proj = nn.Linear(d_model, d_model, bias=False)
|
223 |
+
self.q_ln = nn.LayerNorm(d_model, bias=False)
|
224 |
+
self.k_ln = nn.LayerNorm(d_model, bias=False)
|
225 |
+
self.reshaper = partial(rearrange, pattern="b s (h d) -> b h s d", h=n_heads)
|
226 |
+
self.rotary = RotaryEmbedding(d_model // n_heads)
|
227 |
+
|
228 |
+
def _apply_rotary(self, q: torch.Tensor, k: torch.Tensor):
|
229 |
+
q = q.unflatten(-1, (self.n_heads, self.d_head))
|
230 |
+
k = k.unflatten(-1, (self.n_heads, self.d_head))
|
231 |
+
q, k = self.rotary(q, k)
|
232 |
+
q = q.flatten(-2, -1)
|
233 |
+
k = k.flatten(-2, -1)
|
234 |
+
return q, k
|
235 |
+
|
236 |
+
def forward(self, x, attention_mask=None):
|
237 |
+
qkv_BLD3 = self.layernorm_qkv(x)
|
238 |
+
query_BLD, key_BLD, value_BLD = torch.chunk(qkv_BLD3, 3, dim=-1)
|
239 |
+
query_BLD, key_BLD = (
|
240 |
+
self.q_ln(query_BLD).to(query_BLD.dtype),
|
241 |
+
self.k_ln(key_BLD).to(query_BLD.dtype),
|
242 |
+
)
|
243 |
+
query_BLD, key_BLD = self._apply_rotary(query_BLD, key_BLD)
|
244 |
+
query_BHLD, key_BHLD, value_BHLD = map(self.reshaper, (query_BLD, key_BLD, value_BLD))
|
245 |
+
context_BHLD = F.scaled_dot_product_attention(
|
246 |
+
query_BHLD, key_BHLD, value_BHLD, attention_mask
|
247 |
+
)
|
248 |
+
context_BLD = rearrange(context_BHLD, "b h s d -> b s (h d)")
|
249 |
+
return self.out_proj(context_BLD)
|
250 |
+
|
251 |
+
|
252 |
+
### LM Head
|
253 |
+
def RegressionHead(
|
254 |
+
d_model: int, output_dim: int, hidden_dim: int | None = None
|
255 |
+
) -> nn.Module:
|
256 |
+
hidden_dim = hidden_dim if hidden_dim is not None else d_model
|
257 |
+
return nn.Sequential(
|
258 |
+
nn.Linear(d_model, hidden_dim),
|
259 |
+
nn.GELU(),
|
260 |
+
nn.LayerNorm(hidden_dim),
|
261 |
+
nn.Linear(hidden_dim, output_dim),
|
262 |
+
)
|
263 |
+
|
264 |
+
|
265 |
+
### Transformer Block
|
266 |
+
class UnifiedTransformerBlock(nn.Module):
|
267 |
+
def __init__(
|
268 |
+
self,
|
269 |
+
d_model: int,
|
270 |
+
n_heads: int,
|
271 |
+
residue_scaling_factor: float = 1,
|
272 |
+
expansion_ratio: float = 8 / 3,
|
273 |
+
):
|
274 |
+
super().__init__()
|
275 |
+
self.attn = MultiHeadAttention(d_model, n_heads)
|
276 |
+
self.ffn = swiglu_ln_ffn(d_model, expansion_ratio)
|
277 |
+
self.scaling_factor = residue_scaling_factor
|
278 |
+
|
279 |
+
def forward(
|
280 |
+
self,
|
281 |
+
x: torch.Tensor,
|
282 |
+
attention_mask: Optional[torch.Tensor] = None,
|
283 |
+
) -> torch.Tensor:
|
284 |
+
r1 = self.attn(x, attention_mask)
|
285 |
+
x = x + r1 / self.scaling_factor
|
286 |
+
r3 = self.ffn(x) / self.scaling_factor
|
287 |
+
x = x + r3
|
288 |
+
return x
|
289 |
+
|
290 |
+
|
291 |
+
### Outputs
|
292 |
+
@dataclass
|
293 |
+
class TransformerOutput(ModelOutput):
|
294 |
+
last_hidden_state: torch.Tensor | None = None
|
295 |
+
hidden_states: tuple[torch.Tensor] | None = None
|
296 |
+
|
297 |
+
|
298 |
+
@dataclass
|
299 |
+
class ESMplusplusOutput(ModelOutput):
|
300 |
+
loss: torch.Tensor | None = None
|
301 |
+
logits: torch.Tensor | None = None
|
302 |
+
last_hidden_state: torch.Tensor | None = None
|
303 |
+
hidden_states: tuple[torch.Tensor] | None = None
|
304 |
+
|
305 |
+
|
306 |
+
### Transformer
|
307 |
+
class TransformerStack(nn.Module):
|
308 |
+
def __init__(
|
309 |
+
self,
|
310 |
+
d_model: int,
|
311 |
+
n_heads: int,
|
312 |
+
n_layers: int,
|
313 |
+
):
|
314 |
+
super().__init__()
|
315 |
+
self.blocks = nn.ModuleList(
|
316 |
+
[
|
317 |
+
UnifiedTransformerBlock(
|
318 |
+
d_model,
|
319 |
+
n_heads,
|
320 |
+
residue_scaling_factor=math.sqrt(n_layers / 36),
|
321 |
+
)
|
322 |
+
for i in range(n_layers)
|
323 |
+
]
|
324 |
+
)
|
325 |
+
self.norm = nn.LayerNorm(d_model, bias=False)
|
326 |
+
|
327 |
+
def forward(
|
328 |
+
self,
|
329 |
+
x: torch.Tensor,
|
330 |
+
attention_mask: Optional[torch.Tensor] = None,
|
331 |
+
output_hidden_states: bool = False,
|
332 |
+
) -> TransformerOutput:
|
333 |
+
batch_size, seq_len, _ = x.shape
|
334 |
+
hidden_states = ()
|
335 |
+
if attention_mask is not None:
|
336 |
+
attention_mask = attention_mask[:, None, None, :].expand(batch_size, 1, seq_len, seq_len).bool()
|
337 |
+
for block in self.blocks:
|
338 |
+
x = block(x, attention_mask)
|
339 |
+
if output_hidden_states:
|
340 |
+
hidden_states += (x,)
|
341 |
+
return TransformerOutput(last_hidden_state=self.norm(x), hidden_states=hidden_states)
|
342 |
+
|
343 |
+
|
344 |
+
### Full model
|
345 |
+
class ESMplusplusForMaskedLM(PreTrainedModel):
|
346 |
+
"""
|
347 |
+
ESM++ for masked language modeling.
|
348 |
+
"""
|
349 |
+
config_class = ESMplusplusConfig
|
350 |
+
def __init__(self, config: ESMplusplusConfig):
|
351 |
+
super().__init__(config)
|
352 |
+
self.config = config
|
353 |
+
self.vocab_size = config.vocab_size
|
354 |
+
self.embed = nn.Embedding(self.vocab_size, config.hidden_size)
|
355 |
+
self.transformer = TransformerStack(config.hidden_size, config.num_attention_heads, config.num_hidden_layers)
|
356 |
+
self.sequence_head = RegressionHead(config.hidden_size, self.vocab_size)
|
357 |
+
self.ce_loss = nn.CrossEntropyLoss()
|
358 |
+
self.tokenizer = EsmSequenceTokenizer()
|
359 |
+
|
360 |
+
@classmethod
|
361 |
+
def from_pretrained_esm(cls, model_name: str):
|
362 |
+
if '300' in model_name:
|
363 |
+
return ESMplusplus_300M()
|
364 |
+
elif '600' in model_name:
|
365 |
+
return ESMplusplus_600M()
|
366 |
+
else:
|
367 |
+
raise ValueError(f"Invalid model name: {model_name}")
|
368 |
+
|
369 |
+
@property
|
370 |
+
def device(self):
|
371 |
+
return next(self.parameters()).device
|
372 |
+
|
373 |
+
def forward(
|
374 |
+
self,
|
375 |
+
input_ids: torch.Tensor | None = None,
|
376 |
+
attention_mask: Optional[torch.Tensor] = None,
|
377 |
+
labels: Optional[torch.Tensor] = None,
|
378 |
+
output_hidden_states: bool = False,
|
379 |
+
) -> ESMplusplusOutput:
|
380 |
+
x = self.embed(input_ids)
|
381 |
+
output = self.transformer(x, attention_mask, output_hidden_states)
|
382 |
+
x = output.last_hidden_state
|
383 |
+
logits = self.sequence_head(x)
|
384 |
+
loss = None
|
385 |
+
if labels is not None:
|
386 |
+
loss = self.ce_loss(logits.view(-1, self.vocab_size), labels.view(-1))
|
387 |
+
return ESMplusplusOutput(
|
388 |
+
loss=loss,
|
389 |
+
logits=logits,
|
390 |
+
last_hidden_state=x,
|
391 |
+
hidden_states=output.hidden_states,
|
392 |
+
)
|
393 |
+
|
394 |
+
|
395 |
+
class ESMplusplusForSequenceClassification(ESMplusplusForMaskedLM):
|
396 |
+
"""
|
397 |
+
ESM++ for sequence classification.
|
398 |
+
"""
|
399 |
+
def __init__(self, config: ESMplusplusConfig):
|
400 |
+
super().__init__(config)
|
401 |
+
self.config = config
|
402 |
+
self.classifier = RegressionHead(config.hidden_size * 2, config.num_labels, config.hidden_size * 4)
|
403 |
+
# we find that large intermediate projections help with sequence classification tasks (*4)
|
404 |
+
self.mse = nn.MSELoss()
|
405 |
+
self.ce = nn.CrossEntropyLoss()
|
406 |
+
self.bce = nn.BCEWithLogitsLoss()
|
407 |
+
|
408 |
+
def mean_pooling(self, x: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
|
409 |
+
# x: (batch_size, seq_len, hidden_size)
|
410 |
+
# attention_mask: (batch_size, seq_len)
|
411 |
+
if attention_mask is None:
|
412 |
+
return x.mean(dim=1)
|
413 |
+
else:
|
414 |
+
attention_mask = attention_mask.unsqueeze(-1)
|
415 |
+
return (x * attention_mask).sum(dim=1) / attention_mask.sum(dim=1)
|
416 |
+
|
417 |
+
def forward(
|
418 |
+
self,
|
419 |
+
input_ids: torch.Tensor | None = None,
|
420 |
+
attention_mask: Optional[torch.Tensor] = None,
|
421 |
+
labels: Optional[torch.Tensor] = None,
|
422 |
+
output_hidden_states: bool = False,
|
423 |
+
) -> ESMplusplusOutput:
|
424 |
+
output = super().forward(input_ids, attention_mask, labels, output_hidden_states)
|
425 |
+
x = output.last_hidden_state
|
426 |
+
cls_features = x[:, 0, :]
|
427 |
+
mean_features = self.mean_pooling(x, attention_mask)
|
428 |
+
# we include mean pooling features to help with early convergence, the cost of this is basically zero
|
429 |
+
features = torch.cat([cls_features, mean_features], dim=-1)
|
430 |
+
logits = self.classifier(features)
|
431 |
+
loss = None
|
432 |
+
if labels is not None:
|
433 |
+
labels = labels.to(logits.device)
|
434 |
+
if self.config.problem_type is None:
|
435 |
+
if self.num_labels == 1:
|
436 |
+
self.config.problem_type = "regression"
|
437 |
+
elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
|
438 |
+
self.config.problem_type = "single_label_classification"
|
439 |
+
else:
|
440 |
+
self.config.problem_type = "multi_label_classification"
|
441 |
+
|
442 |
+
if self.config.problem_type == "regression":
|
443 |
+
if self.num_labels == 1:
|
444 |
+
loss = self.mse(logits.flatten(), labels.flatten())
|
445 |
+
else:
|
446 |
+
loss = self.mse(logits, labels)
|
447 |
+
elif self.config.problem_type == "single_label_classification":
|
448 |
+
loss = self.ce(logits.view(-1, self.num_labels), labels.view(-1))
|
449 |
+
elif self.config.problem_type == "multi_label_classification":
|
450 |
+
loss = self.bce(logits, labels)
|
451 |
+
return ESMplusplusOutput(
|
452 |
+
loss=loss,
|
453 |
+
logits=logits,
|
454 |
+
last_hidden_state=x,
|
455 |
+
hidden_states=output.hidden_states,
|
456 |
+
)
|
457 |
+
|
458 |
+
|
459 |
+
class ESMplusplusForTokenClassification(ESMplusplusForMaskedLM):
|
460 |
+
"""
|
461 |
+
ESM++ for token classification.
|
462 |
+
"""
|
463 |
+
def __init__(self, config: ESMplusplusConfig):
|
464 |
+
super().__init__(config)
|
465 |
+
self.config = config
|
466 |
+
self.num_labels = config.num_labels
|
467 |
+
self.classifier = RegressionHead(config.hidden_size, config.num_labels, config.hidden_size * 4)
|
468 |
+
# we find that large intermediate projections help with sequence classification tasks (*4)
|
469 |
+
self.loss_fct = nn.CrossEntropyLoss()
|
470 |
+
|
471 |
+
def forward(
|
472 |
+
self,
|
473 |
+
input_ids: torch.Tensor | None = None,
|
474 |
+
attention_mask: Optional[torch.Tensor] = None,
|
475 |
+
labels: Optional[torch.Tensor] = None,
|
476 |
+
output_hidden_states: bool = False,
|
477 |
+
) -> ESMplusplusOutput:
|
478 |
+
output = super().forward(input_ids, attention_mask, labels, output_hidden_states)
|
479 |
+
x = output.last_hidden_state
|
480 |
+
logits = self.classifier(x)
|
481 |
+
loss = None
|
482 |
+
if labels is not None:
|
483 |
+
loss = self.loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
|
484 |
+
return ESMplusplusOutput(
|
485 |
+
loss=loss,
|
486 |
+
logits=logits,
|
487 |
+
last_hidden_state=x,
|
488 |
+
hidden_states=output.hidden_states,
|
489 |
+
)
|
490 |
+
|
491 |
+
|
492 |
+
### Loading
|
493 |
+
import os
|
494 |
+
from functools import cache
|
495 |
+
from pathlib import Path
|
496 |
+
from huggingface_hub import snapshot_download
|
497 |
+
|
498 |
+
|
499 |
+
@staticmethod
|
500 |
+
@cache
|
501 |
+
def data_root(model: str):
|
502 |
+
if "INFRA_PROVIDER" in os.environ:
|
503 |
+
return Path("")
|
504 |
+
# Try to download from hugginface if it doesn't exist
|
505 |
+
if model.startswith("esmc-300"):
|
506 |
+
path = Path(snapshot_download(repo_id="EvolutionaryScale/esmc-300m-2024-12"))
|
507 |
+
elif model.startswith("esmc-600"):
|
508 |
+
path = Path(snapshot_download(repo_id="EvolutionaryScale/esmc-600m-2024-12"))
|
509 |
+
else:
|
510 |
+
raise ValueError(f"{model=} is an invalid model name.")
|
511 |
+
return path
|
512 |
+
|
513 |
+
|
514 |
+
def ESMplusplus_300M(device: torch.device | str = "cpu"):
|
515 |
+
with torch.device(device):
|
516 |
+
config = ESMplusplusConfig(
|
517 |
+
hidden_size=960,
|
518 |
+
num_attention_heads=15,
|
519 |
+
num_hidden_layers=30,
|
520 |
+
)
|
521 |
+
model = ESMplusplusForMaskedLM(config)
|
522 |
+
state_dict = torch.load(
|
523 |
+
data_root("esmc-300") / "data/weights/esmc_300m_2024_12_v0.pth",
|
524 |
+
map_location=device,
|
525 |
+
)
|
526 |
+
model.load_state_dict(state_dict)
|
527 |
+
return model
|
528 |
+
|
529 |
+
|
530 |
+
def ESMplusplus_600M(device: torch.device | str = "cpu"):
|
531 |
+
with torch.device(device):
|
532 |
+
config = ESMplusplusConfig(
|
533 |
+
hidden_size=1152,
|
534 |
+
num_attention_heads=18,
|
535 |
+
num_hidden_layers=36,
|
536 |
+
)
|
537 |
+
model = ESMplusplusForMaskedLM(config)
|
538 |
+
state_dict = torch.load(
|
539 |
+
data_root("esmc-600") / "data/weights/esmc_600m_2024_12_v0.pth",
|
540 |
+
map_location=device,
|
541 |
+
)
|
542 |
+
model.load_state_dict(state_dict)
|
543 |
+
return model
|
544 |
+
|
545 |
+
|
546 |
+
### Tokenization
|
547 |
+
from tokenizers import Tokenizer
|
548 |
+
from tokenizers.models import BPE
|
549 |
+
from tokenizers.processors import TemplateProcessing
|
550 |
+
from transformers import PreTrainedTokenizerFast
|
551 |
+
|
552 |
+
|
553 |
+
SEQUENCE_VOCAB = [
|
554 |
+
"<cls>", "<pad>", "<eos>", "<unk>",
|
555 |
+
"L", "A", "G", "V", "S", "E", "R", "T", "I", "D", "P", "K",
|
556 |
+
"Q", "N", "F", "Y", "M", "H", "W", "C", "X", "B", "U", "Z",
|
557 |
+
"O", ".", "-", "|",
|
558 |
+
"<mask>",
|
559 |
+
]
|
560 |
+
|
561 |
+
class EsmSequenceTokenizer(PreTrainedTokenizerFast):
|
562 |
+
model_input_names = ["input_ids", "attention_mask"]
|
563 |
+
|
564 |
+
def __init__(
|
565 |
+
self,
|
566 |
+
unk_token="<unk>",
|
567 |
+
cls_token="<cls>",
|
568 |
+
pad_token="<pad>",
|
569 |
+
mask_token="<mask>",
|
570 |
+
eos_token="<eos>",
|
571 |
+
chain_break_token="|",
|
572 |
+
**kwargs,
|
573 |
+
):
|
574 |
+
all_tokens = SEQUENCE_VOCAB
|
575 |
+
token_to_id = {tok: ind for ind, tok in enumerate(all_tokens)}
|
576 |
+
|
577 |
+
# a character-level tokenizer is the same as BPE with no token merges
|
578 |
+
bpe = BPE(token_to_id, merges=[], unk_token=unk_token)
|
579 |
+
tokenizer = Tokenizer(bpe)
|
580 |
+
special_tokens = [
|
581 |
+
cls_token,
|
582 |
+
pad_token,
|
583 |
+
mask_token,
|
584 |
+
eos_token,
|
585 |
+
chain_break_token,
|
586 |
+
]
|
587 |
+
self.cb_token = chain_break_token
|
588 |
+
additional_special_tokens = [chain_break_token]
|
589 |
+
|
590 |
+
tokenizer.add_special_tokens(special_tokens)
|
591 |
+
|
592 |
+
# This is where we configure the automatic addition of special tokens when we call
|
593 |
+
# tokenizer(text, add_special_tokens=True). Note that you can also configure how two
|
594 |
+
# sequences are merged if you want.
|
595 |
+
tokenizer.post_processor = TemplateProcessing( # type: ignore
|
596 |
+
single="<cls> $A <eos>",
|
597 |
+
special_tokens=[
|
598 |
+
("<cls>", tokenizer.token_to_id("<cls>")),
|
599 |
+
("<eos>", tokenizer.token_to_id("<eos>")),
|
600 |
+
],
|
601 |
+
)
|
602 |
+
super().__init__(
|
603 |
+
tokenizer_object=tokenizer,
|
604 |
+
unk_token=unk_token,
|
605 |
+
cls_token=cls_token,
|
606 |
+
pad_token=pad_token,
|
607 |
+
mask_token=mask_token,
|
608 |
+
eos_token=eos_token,
|
609 |
+
additional_special_tokens=additional_special_tokens,
|
610 |
+
**kwargs,
|
611 |
+
)
|
612 |
+
|
613 |
+
# These are a footgun, we never use the `bos` token anywhere so we're just overriding it here.
|
614 |
+
@property
|
615 |
+
def bos_token(self):
|
616 |
+
return self.cls_token
|
617 |
+
|
618 |
+
@property
|
619 |
+
def bos_token_id(self):
|
620 |
+
return self.cls_token_id
|
621 |
+
|
622 |
+
@property
|
623 |
+
def chain_break_token(self):
|
624 |
+
return self.cb_token
|
625 |
+
|
626 |
+
@property
|
627 |
+
def chain_break_token_id(self):
|
628 |
+
return self.convert_tokens_to_ids(self.chain_break_token)
|
629 |
+
|
630 |
+
@property
|
631 |
+
def all_token_ids(self):
|
632 |
+
return list(range(self.vocab_size))
|
633 |
+
|
634 |
+
@property
|
635 |
+
def special_token_ids(self):
|
636 |
+
return self.all_special_ids
|