Upload 2 files
Browse files- scripts/diffusion.py +293 -0
- scripts/train.py +3 -1
scripts/diffusion.py
ADDED
@@ -0,0 +1,293 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import itertools
|
2 |
+
import math
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
import numpy as np
|
6 |
+
import pytorch_lightning as L
|
7 |
+
import torchmetrics
|
8 |
+
from dataclasses import dataclass
|
9 |
+
from esm_utils import load_esm2_model
|
10 |
+
from transformers import AutoModel, AutoTokenizer
|
11 |
+
import dit, ema
|
12 |
+
import sys
|
13 |
+
import config
|
14 |
+
import wandb
|
15 |
+
import noise_schedule # Assuming this is part of the MDLM repository
|
16 |
+
|
17 |
+
wandb_key = "2b76a2fa2c1cdfddc5f443602c17b011fefb0a8f"
|
18 |
+
wandb.login(key=wandb_key)
|
19 |
+
wandb.init(project=config.Wandb.PROJECT, group=config.Wandb.GROUP)
|
20 |
+
|
21 |
+
LOG2 = math.log(2)
|
22 |
+
|
23 |
+
# Goal is to build an MDLM head on the BERT-style ESM model
|
24 |
+
# Wrap the ESM model to obtain embeddings and ignore sigma to work with MDLM codebase
|
25 |
+
class WrapESM(nn.Module):
|
26 |
+
def __init__(self, esm_model_path):
|
27 |
+
super(WrapESM, self).__init__()
|
28 |
+
self.esm_tokenizer, self.esm_model, _ = load_esm2_model(esm_model_path)
|
29 |
+
|
30 |
+
### Only fine-tune the last 3 layers of ESM
|
31 |
+
# Count number of encoder layers
|
32 |
+
model_layers = len(self.esm_model.esm.encoder.layer)
|
33 |
+
|
34 |
+
# Disable parameter updates for all layers
|
35 |
+
for param in self.esm_model.parameters():
|
36 |
+
param.requires_grad = False
|
37 |
+
|
38 |
+
# Now that all parameters are disabled, only enable updates for the last 3 layers
|
39 |
+
for i, layer in enumerate(self.esm_model.esm.encoder.layer):
|
40 |
+
if i >= model_layers-config.ESM_LAYERS:
|
41 |
+
for module in layer.attention.self.key.modules():
|
42 |
+
for param in module.parameters():
|
43 |
+
param.requires_grad = True
|
44 |
+
for module in layer.attention.self.query.modules():
|
45 |
+
for param in module.parameters():
|
46 |
+
param.requires_grad = True
|
47 |
+
for module in layer.attention.self.value.modules():
|
48 |
+
for param in module.parameters():
|
49 |
+
param.requires_grad = True
|
50 |
+
|
51 |
+
def forward(self, latents, sigma):
|
52 |
+
return latents
|
53 |
+
|
54 |
+
@dataclass
|
55 |
+
class Loss:
|
56 |
+
loss: torch.FloatTensor
|
57 |
+
nlls: torch.FloatTensor
|
58 |
+
token_mask: torch.FloatTensor
|
59 |
+
|
60 |
+
class NLL(torchmetrics.MeanMetric):
|
61 |
+
pass
|
62 |
+
|
63 |
+
class BPD(NLL):
|
64 |
+
def compute(self) -> torch.Tensor:
|
65 |
+
"""Computes the bits per dimension.
|
66 |
+
Returns:
|
67 |
+
bpd
|
68 |
+
"""
|
69 |
+
return self.mean_value / self.weight / LOG2
|
70 |
+
|
71 |
+
class Perplexity(NLL):
|
72 |
+
def compute(self) -> torch.Tensor:
|
73 |
+
"""Computes the Perplexity.
|
74 |
+
Returns:
|
75 |
+
Perplexity
|
76 |
+
"""
|
77 |
+
return torch.exp(self.mean_value / self.weight)
|
78 |
+
|
79 |
+
|
80 |
+
# Based on MDLM repo
|
81 |
+
class Diffusion(L.LightningModule):
|
82 |
+
def __init__(self, config, latent_dim, tokenizer):
|
83 |
+
super().__init__()
|
84 |
+
self.config = config
|
85 |
+
self.latent_dim = latent_dim
|
86 |
+
self.tokenizer = tokenizer
|
87 |
+
|
88 |
+
self.softplus = torch.nn.Softplus()
|
89 |
+
metrics = torchmetrics.MetricCollection({
|
90 |
+
'nll': NLL(),
|
91 |
+
'bpd': BPD(),
|
92 |
+
'ppl': Perplexity(),
|
93 |
+
})
|
94 |
+
metrics.set_dtype(torch.float64)
|
95 |
+
self.train_metrics = metrics.clone(prefix='train/')
|
96 |
+
self.valid_metrics = metrics.clone(prefix='val/')
|
97 |
+
self.test_metrics = metrics.clone(prefix='test/')
|
98 |
+
|
99 |
+
self.T = self.config.T
|
100 |
+
self.lr = self.config.Optim.LR
|
101 |
+
self.backbone = WrapESM(self.config.MODEL_NAME)
|
102 |
+
self.noise = noise_schedule.get_noise(self.config, dtype=self.dtype)
|
103 |
+
self.time_conditioning = self.config.TIME_CONDITIONING
|
104 |
+
self.subs_masking = self.config.SUBS_MASKING
|
105 |
+
self.mask_index = self.tokenizer.mask_token_id
|
106 |
+
self.antithetic_sampling = self.config.Training.ANTITHETIC_SAMPLING
|
107 |
+
self.sampling_eps = self.config.Training.SAMPLING_EPS
|
108 |
+
self.neg_infinity = -1000000.0
|
109 |
+
|
110 |
+
|
111 |
+
############ FORWARD DIFFUSION #########
|
112 |
+
def subs_parameterization(self, logits, noised_latents):
|
113 |
+
logits = logits.float()
|
114 |
+
logits[:, :, self.mask_index] += self.neg_infinity
|
115 |
+
|
116 |
+
# Normalize the logits such that x.exp() is a probability distribution over vocab_size.
|
117 |
+
logits = logits - torch.logsumexp(logits, dim=-1, keepdim=True)
|
118 |
+
|
119 |
+
unmasked_indices = (noised_latents != self.mask_index)
|
120 |
+
logits[unmasked_indices] = self.neg_infinity
|
121 |
+
logits[~unmasked_indices] = 0
|
122 |
+
|
123 |
+
return logits
|
124 |
+
|
125 |
+
# # -inf probability of selecting a masked token
|
126 |
+
# unmasked_indices = (noised_latents != self.mask_index)
|
127 |
+
# logits[unmasked_indices] = self.neg_infinity
|
128 |
+
|
129 |
+
# # Carry over unmasked tokens
|
130 |
+
# bsz, seq_len, input_dim = logits.shape
|
131 |
+
# for batch_idx in range(bsz):
|
132 |
+
# for residue in range(seq_len):
|
133 |
+
# logits[batch_idx, residue, noised_latents[batch_idx, residue]] = 0
|
134 |
+
|
135 |
+
# return logits
|
136 |
+
|
137 |
+
def forward(self, latents, sigma):
|
138 |
+
latents = latents.long()
|
139 |
+
logits = self.backbone(latents, sigma)
|
140 |
+
optimized_logits = self.subs_parameterization(logits, latents)
|
141 |
+
return optimized_logits
|
142 |
+
|
143 |
+
def q_xt(self, latents, move_chance):
|
144 |
+
"""
|
145 |
+
Computes the noisy sample xt.
|
146 |
+
Args:
|
147 |
+
x: int torch.Tensor with shape (batch_size, diffusion_model_input_length), input.
|
148 |
+
move_chance: float torch.Tensor with shape (batch_size, 1).
|
149 |
+
"""
|
150 |
+
#latents = latents.mean(dim=1) # [bsz x seq_len x 1280] --> [bsz x 1280] as per markdown
|
151 |
+
move_indices = torch.rand(* latents.shape, device=latents.device) < move_chance
|
152 |
+
noised_latents = torch.where(move_indices, self.mask_index, latents)
|
153 |
+
return noised_latents
|
154 |
+
|
155 |
+
def sample_timestep(self, n, device):
|
156 |
+
_eps_t = torch.rand(n, device=device)
|
157 |
+
if self.antithetic_sampling:
|
158 |
+
offset = torch.arange(n, device=device) / n
|
159 |
+
_eps_t = (_eps_t / n + offset) % 1
|
160 |
+
t = (1 - self.sampling_eps) * _eps_t + self.sampling_eps
|
161 |
+
# if self.importance_sampling:
|
162 |
+
# return self.noise.importance_sampling_transformation(t)
|
163 |
+
return t
|
164 |
+
|
165 |
+
def forward_diffusion(self, x0):
|
166 |
+
"""Forward diffusion process, adds noise to the latents."""
|
167 |
+
|
168 |
+
t = self.sample_timestep(x0.shape[0], x0.device)
|
169 |
+
sigma, dsigma = self.noise(t)
|
170 |
+
unet_conditioning = sigma[:, None]
|
171 |
+
move_chance = 1 - torch.exp(-sigma[:, None, None])
|
172 |
+
|
173 |
+
xt = self.q_xt(x0, move_chance)
|
174 |
+
model_output = self.forward(xt, unet_conditioning)
|
175 |
+
|
176 |
+
# SUBS parameterization, continuous time.
|
177 |
+
idx = x0.long()
|
178 |
+
print(f'idx: {idx.size()}')
|
179 |
+
print(f'idx min: {idx.min()}')
|
180 |
+
print(f'idx max: {idx.max()}')
|
181 |
+
print(f'model out: {model_output.size()}')
|
182 |
+
log_p_theta = torch.gather(input=model_output, dim=-1, index=idx).squeeze(-1)
|
183 |
+
scale = (dsigma / torch.expm1(sigma))[:, None]
|
184 |
+
return - log_p_theta * scale
|
185 |
+
|
186 |
+
|
187 |
+
######### LOSS CALCULATIONS #########
|
188 |
+
def compute_loss(self, latents, attention_mask):
|
189 |
+
""""Average of MLM losses to stabilize training"""
|
190 |
+
loss = self.forward_diffusion(latents)
|
191 |
+
|
192 |
+
nlls = loss * attention_mask
|
193 |
+
count = attention_mask.sum()
|
194 |
+
batch_nll = nlls.sum()
|
195 |
+
token_nll = batch_nll / count
|
196 |
+
|
197 |
+
return Loss(loss=token_nll, nlls=nlls, token_mask=attention_mask)
|
198 |
+
|
199 |
+
|
200 |
+
######### TRAINING #########
|
201 |
+
def training_step(self, batch):
|
202 |
+
latents, attention_mask = batch
|
203 |
+
loss = self.compute_loss(latents, attention_mask)
|
204 |
+
wandb.log({"train_loss": loss.loss.item()})
|
205 |
+
return loss.loss
|
206 |
+
|
207 |
+
def configure_optimizers(self):
|
208 |
+
optimizer = torch.optim.Adam(self.parameters(), lr=self.lr)
|
209 |
+
return optimizer
|
210 |
+
|
211 |
+
def validation_step(self, batch):
|
212 |
+
latents, attention_mask = batch
|
213 |
+
loss = self.compute_loss(latents, attention_mask)
|
214 |
+
wandb.log({"val_loss": loss.loss.item()})
|
215 |
+
return loss.loss
|
216 |
+
|
217 |
+
|
218 |
+
######### GENERATION #########
|
219 |
+
def sample_prior(self, *batch_dims):
|
220 |
+
return self.mask_index * torch.ones(* batch_dims, dtype=torch.int64)
|
221 |
+
|
222 |
+
def sample_categorical(categorical_probs):
|
223 |
+
gumbel_norm = (1e-10 - (torch.rand_like(categorical_probs) + 1e-10).log())
|
224 |
+
return (categorical_probs / gumbel_norm).argmax(dim=-1)
|
225 |
+
|
226 |
+
def ddpm_caching_update(self, x, t, dt, p_x0=None):
|
227 |
+
assert self.config.noise.type == 'loglinear'
|
228 |
+
sigma_t, _ = self.noise(t)
|
229 |
+
if t.ndim > 1:
|
230 |
+
t = t.squeeze(-1)
|
231 |
+
assert t.ndim == 1
|
232 |
+
move_chance_t = t[:, None, None]
|
233 |
+
move_chance_s = (t - dt)[:, None, None]
|
234 |
+
assert move_chance_t.ndim == 3, move_chance_t.shape
|
235 |
+
if p_x0 is None:
|
236 |
+
p_x0 = self.forward(x, sigma_t).exp()
|
237 |
+
|
238 |
+
assert move_chance_t.ndim == p_x0.ndim
|
239 |
+
q_xs = p_x0 * (move_chance_t - move_chance_s)
|
240 |
+
q_xs[:, :, self.mask_index] = move_chance_s[:, :, 0]
|
241 |
+
_x = self.sample_categorical(q_xs)
|
242 |
+
|
243 |
+
copy_flag = (x != self.mask_index).to(x.dtype)
|
244 |
+
return p_x0, copy_flag * x + (1 - copy_flag) * _x
|
245 |
+
|
246 |
+
|
247 |
+
@torch.no_grad()
|
248 |
+
def sample_subs_guidance(self, n_samples, stride_length, num_strides, dt=0.001):
|
249 |
+
ones = torch.ones(n_samples, dtype=self.dtype,device=self.device)
|
250 |
+
num_steps = int(1 / dt)
|
251 |
+
sampling_steps = 0
|
252 |
+
intermediate_tokens = []
|
253 |
+
target = None
|
254 |
+
|
255 |
+
for _ in range(num_strides + 1):
|
256 |
+
p_x0_cache = None
|
257 |
+
x = self._sample_prior(n_samples,self.config.model.length).to(self.device)
|
258 |
+
|
259 |
+
if target is not None:
|
260 |
+
x[:, : -stride_length] = target
|
261 |
+
|
262 |
+
for i in range(num_steps + 1):
|
263 |
+
p_x0_cache, x_next = self.ddpm_caching_update(x=x, t=(1 - i * dt) * ones, dt=dt, p_x0=p_x0_cache)
|
264 |
+
if (not torch.allclose(x_next, x) or self.time_conditioning):
|
265 |
+
p_x0_cache = None
|
266 |
+
sampling_steps += 1
|
267 |
+
x = x_next
|
268 |
+
x = self.forward(x, 0 * ones).argmax(dim=-1)
|
269 |
+
intermediate_tokens.append(x[:, :stride_length].cpu().numpy())
|
270 |
+
target = x[:, stride_length:]
|
271 |
+
|
272 |
+
intermediate_tokens.append(target.cpu().numpy())
|
273 |
+
intermediate_text_samples = []
|
274 |
+
sequence_lengths = ((np.concatenate(intermediate_tokens, axis=1)[:, 1:]
|
275 |
+
== self.tokenizer.eos_token_id).cumsum(-1) == 0).sum(-1)
|
276 |
+
|
277 |
+
for i in range(2, len(intermediate_tokens) + 1):
|
278 |
+
intermediate_text_samples.append(self.tokenizer.decode(np.concatenate(intermediate_tokens[:i], axis=1)))
|
279 |
+
|
280 |
+
return (sampling_steps, intermediate_text_samples,
|
281 |
+
sequence_lengths)
|
282 |
+
|
283 |
+
def restore_model_and_semi_ar_sample(self, stride_length, num_strides, dt=0.001):
|
284 |
+
"""Generate samples from the model."""
|
285 |
+
# Lightning auto-casting is not working in this method for some reason
|
286 |
+
self.backbone.eval()
|
287 |
+
self.noise.eval()
|
288 |
+
|
289 |
+
(sampling_steps, samples, sequence_lengths) = self.sample_subs_guidance(n_samples=self.config.Loader.BATCH_SIZE,stride_length=stride_length,num_strides=num_strides,dt=dt)
|
290 |
+
|
291 |
+
self.backbone.train()
|
292 |
+
self.noise.train()
|
293 |
+
return sampling_steps, samples, sequence_lengths
|
scripts/train.py
CHANGED
@@ -5,13 +5,14 @@ import config
|
|
5 |
from data_loader import get_dataloaders
|
6 |
from esm_utils import load_esm2_model
|
7 |
from diffusion import Diffusion
|
|
|
8 |
import sys
|
9 |
|
10 |
# Get dataloaders
|
11 |
train_loader, val_loader, _ = get_dataloaders(config)
|
12 |
|
13 |
# Initialize ESM tokenizer and model
|
14 |
-
tokenizer,
|
15 |
|
16 |
# Initialize diffusion model
|
17 |
latent_diffusion_model = Diffusion(config, latent_dim=config.LATENT_DIM, tokenizer=tokenizer)
|
@@ -46,3 +47,4 @@ sys.stdout.flush()
|
|
46 |
# Train the model
|
47 |
trainer.fit(latent_diffusion_model, train_loader, val_loader)
|
48 |
|
|
|
|
5 |
from data_loader import get_dataloaders
|
6 |
from esm_utils import load_esm2_model
|
7 |
from diffusion import Diffusion
|
8 |
+
import wandb
|
9 |
import sys
|
10 |
|
11 |
# Get dataloaders
|
12 |
train_loader, val_loader, _ = get_dataloaders(config)
|
13 |
|
14 |
# Initialize ESM tokenizer and model
|
15 |
+
tokenizer, _, _ = load_esm2_model(config.MODEL_NAME)
|
16 |
|
17 |
# Initialize diffusion model
|
18 |
latent_diffusion_model = Diffusion(config, latent_dim=config.LATENT_DIM, tokenizer=tokenizer)
|
|
|
47 |
# Train the model
|
48 |
trainer.fit(latent_diffusion_model, train_loader, val_loader)
|
49 |
|
50 |
+
wandb.finish()
|