Upload 3 files
Browse files- models/diffusion.py +246 -22
- models/dit.py +3 -4
models/diffusion.py
CHANGED
@@ -1,11 +1,11 @@
|
|
1 |
import itertools
|
2 |
import math
|
3 |
import torch
|
4 |
-
import
|
5 |
import pytorch_lightning as L
|
6 |
import torchmetrics
|
7 |
from dataclasses import dataclass
|
8 |
-
|
9 |
import noise_schedule # Assuming this is part of the MDLM repository
|
10 |
|
11 |
LOG2 = math.log(2)
|
@@ -22,7 +22,6 @@ class NLL(torchmetrics.MeanMetric):
|
|
22 |
class BPD(NLL):
|
23 |
def compute(self) -> torch.Tensor:
|
24 |
"""Computes the bits per dimension.
|
25 |
-
|
26 |
Returns:
|
27 |
bpd
|
28 |
"""
|
@@ -31,21 +30,24 @@ class BPD(NLL):
|
|
31 |
class Perplexity(NLL):
|
32 |
def compute(self) -> torch.Tensor:
|
33 |
"""Computes the Perplexity.
|
34 |
-
|
35 |
Returns:
|
36 |
Perplexity
|
37 |
"""
|
38 |
return torch.exp(self.mean_value / self.weight)
|
39 |
|
|
|
40 |
class Diffusion(L.LightningModule):
|
41 |
-
def __init__(self, config, latent_dim):
|
42 |
super().__init__()
|
43 |
self.config = config
|
44 |
self.latent_dim = latent_dim
|
|
|
45 |
|
46 |
-
self.backbone = dit.DIT(config, vocab_size=self.latent_dim)
|
47 |
self.T = self.config.T
|
48 |
-
self.subs_masking = self.config.
|
|
|
|
|
49 |
|
50 |
self.softplus = torch.nn.Softplus()
|
51 |
metrics = torchmetrics.MetricCollection({
|
@@ -59,30 +61,252 @@ class Diffusion(L.LightningModule):
|
|
59 |
self.test_metrics = metrics.clone(prefix='test/')
|
60 |
|
61 |
self.noise = noise_schedule.get_noise(self.config, dtype=self.dtype)
|
62 |
-
self.lr = self.config.
|
63 |
-
self.sampling_eps = self.config.
|
64 |
-
self.time_conditioning = self.config.
|
65 |
self.neg_infinity = -1000000.0
|
66 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
67 |
def forward(self, latents, sigma):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
68 |
"""Forward diffusion process, adds noise to the latents."""
|
69 |
-
noise = sigma * torch.randn_like(latents)
|
70 |
-
noisy_latents = latents + noise
|
71 |
-
return noisy_latents
|
72 |
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
77 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
78 |
def training_step(self, batch, batch_idx):
|
79 |
-
|
80 |
-
|
81 |
-
denoised_latents = self.reverse_diffusion(noisy_latents, sigma)
|
82 |
-
loss = F.mse_loss(denoised_latents, batch)
|
83 |
-
self.log("train_loss", loss)
|
84 |
return loss
|
85 |
|
86 |
def configure_optimizers(self):
|
87 |
optimizer = torch.optim.Adam(self.parameters(), lr=self.lr)
|
88 |
return optimizer
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import itertools
|
2 |
import math
|
3 |
import torch
|
4 |
+
import numpy as np
|
5 |
import pytorch_lightning as L
|
6 |
import torchmetrics
|
7 |
from dataclasses import dataclass
|
8 |
+
import dit, ema
|
9 |
import noise_schedule # Assuming this is part of the MDLM repository
|
10 |
|
11 |
LOG2 = math.log(2)
|
|
|
22 |
class BPD(NLL):
|
23 |
def compute(self) -> torch.Tensor:
|
24 |
"""Computes the bits per dimension.
|
|
|
25 |
Returns:
|
26 |
bpd
|
27 |
"""
|
|
|
30 |
class Perplexity(NLL):
|
31 |
def compute(self) -> torch.Tensor:
|
32 |
"""Computes the Perplexity.
|
|
|
33 |
Returns:
|
34 |
Perplexity
|
35 |
"""
|
36 |
return torch.exp(self.mean_value / self.weight)
|
37 |
|
38 |
+
# Based on MDLM repo
|
39 |
class Diffusion(L.LightningModule):
|
40 |
+
def __init__(self, config, latent_dim, tokenizer):
|
41 |
super().__init__()
|
42 |
self.config = config
|
43 |
self.latent_dim = latent_dim
|
44 |
+
self.tokenizer = tokenizer
|
45 |
|
46 |
+
self.backbone = dit.DIT(self.config, vocab_size=self.latent_dim)
|
47 |
self.T = self.config.T
|
48 |
+
self.subs_masking = self.config.SUBS_MASKING
|
49 |
+
self.antithetic_sampling = self.config.Training.ANTITHETIC_SAMPLING
|
50 |
+
self.mask_index = self.tokenizer.mask_token_id
|
51 |
|
52 |
self.softplus = torch.nn.Softplus()
|
53 |
metrics = torchmetrics.MetricCollection({
|
|
|
61 |
self.test_metrics = metrics.clone(prefix='test/')
|
62 |
|
63 |
self.noise = noise_schedule.get_noise(self.config, dtype=self.dtype)
|
64 |
+
self.lr = self.config.Optim.LR
|
65 |
+
self.sampling_eps = self.config.Training.SAMPLING_EPS
|
66 |
+
self.time_conditioning = self.config.TIME_CONDITIONING
|
67 |
self.neg_infinity = -1000000.0
|
68 |
|
69 |
+
|
70 |
+
############ FORWARD DIFFUSION #########
|
71 |
+
def subs_parameterization(self, logits, noised_latents):
|
72 |
+
# log prob at the mask index = - infinity
|
73 |
+
logits[:, :, self.mask_index] += self.neg_infinity
|
74 |
+
|
75 |
+
# Normalize the logits such that x.exp() is
|
76 |
+
# a probability distribution over vocab_size.
|
77 |
+
logits = logits - torch.logsumexp(logits, dim=-1,
|
78 |
+
keepdim=True)
|
79 |
+
|
80 |
+
# Apply updates directly in the logits matrix.
|
81 |
+
# For the logits of the unmasked tokens, set all values
|
82 |
+
# to -infinity except for the indices corresponding to
|
83 |
+
# the unmasked tokens.
|
84 |
+
unmasked_indices = (noised_latents != self.mask_index)
|
85 |
+
logits[unmasked_indices] = self.neg_infinity
|
86 |
+
logits[unmasked_indices, noised_latents[unmasked_indices]] = 0
|
87 |
+
return logits
|
88 |
+
|
89 |
def forward(self, latents, sigma):
|
90 |
+
latents = latents.long()
|
91 |
+
with torch.cuda.amp.autocast(dtype=torch.float32):
|
92 |
+
logits = self.backbone(latents, sigma)
|
93 |
+
print(logits)
|
94 |
+
optimized_logits = self.subs_parameterization(logits, latents)
|
95 |
+
return optimized_logits
|
96 |
+
|
97 |
+
def q_xt(self, latents, move_chance):
|
98 |
+
"""
|
99 |
+
Computes the noisy sample xt.
|
100 |
+
Args:
|
101 |
+
x: int torch.Tensor with shape (batch_size, diffusion_model_input_length), input.
|
102 |
+
move_chance: float torch.Tensor with shape (batch_size, 1).
|
103 |
+
"""
|
104 |
+
latents = latents.mean(dim=1) # [bsz x seq_len x 1280] --> [bsz x 1280] as per args
|
105 |
+
move_indices = torch.rand(* latents.shape, device=latents.device) < move_chance
|
106 |
+
noised_latents = torch.where(move_indices, self.mask_index, latents)
|
107 |
+
return noised_latents
|
108 |
+
|
109 |
+
def sample_timestep(self, n, device):
|
110 |
+
_eps_t = torch.rand(n, device=device)
|
111 |
+
if self.antithetic_sampling:
|
112 |
+
offset = torch.arange(n, device=device) / n
|
113 |
+
_eps_t = (_eps_t / n + offset) % 1
|
114 |
+
t = (1 - self.sampling_eps) * _eps_t + self.sampling_eps
|
115 |
+
# if self.importance_sampling:
|
116 |
+
# return self.noise.importance_sampling_transformation(t)
|
117 |
+
return t
|
118 |
+
|
119 |
+
|
120 |
+
def d3pm_loss(self, model_output, xt, x0, t):
|
121 |
+
"""Computes the D3PM loss between noisy latents and the original input at a given time step."""
|
122 |
+
dt = 1 / self.T
|
123 |
+
|
124 |
+
if torch.is_tensor(t):
|
125 |
+
t = t[:, None]
|
126 |
+
assert t.ndim == 2
|
127 |
+
t = t.clamp(0., 1. - 1e-4)
|
128 |
+
alpha_t = 1 - t + torch.zeros_like(xt)
|
129 |
+
alpha_s = 1 - (t - dt) + torch.zeros_like(xt)
|
130 |
+
|
131 |
+
x0 = x0.to(torch.int64)
|
132 |
+
log_x_theta_at_x0 = torch.gather(model_output, -1, x0[:, :, None]).squeeze(-1)
|
133 |
+
log_x_theta_at_m = model_output[:, :, self.mask_index]
|
134 |
+
x_theta_at_m = log_x_theta_at_m.exp()
|
135 |
+
|
136 |
+
term_1_coef = dt / t
|
137 |
+
term_1_log_nr = torch.log(alpha_t * x_theta_at_m / t + 1)
|
138 |
+
term_1_log_dr = log_x_theta_at_x0
|
139 |
+
|
140 |
+
term_2_coef = 1 - dt / t
|
141 |
+
term_2_log_nr = term_1_log_nr
|
142 |
+
term_2_log_dr = torch.log(alpha_s * x_theta_at_m / (t - dt) + 1)
|
143 |
+
|
144 |
+
L_vb_masked = (
|
145 |
+
term_1_coef * (term_1_log_nr - term_1_log_dr)
|
146 |
+
+ term_2_coef * (term_2_log_nr - term_2_log_dr))
|
147 |
+
|
148 |
+
L_vb = L_vb_masked * (xt == self.mask_index)
|
149 |
+
|
150 |
+
return self.T * L_vb
|
151 |
+
|
152 |
+
def forward_diffusion(self, latents):
|
153 |
"""Forward diffusion process, adds noise to the latents."""
|
|
|
|
|
|
|
154 |
|
155 |
+
t = self.sample_timestep(latents.shape[0], latents.device)
|
156 |
+
if self.T > 0:
|
157 |
+
t = (t * self.T).to(torch.int)
|
158 |
+
t = t / self.T
|
159 |
+
# t \in {1/T, 2/T, ..., 1}
|
160 |
+
t += (1 / self.T)
|
161 |
+
|
162 |
+
sigma, dsigma = self.noise(t)
|
163 |
+
unet_conditioning = sigma[:, None]
|
164 |
+
move_chance = 1 - torch.exp(-sigma[:, None])
|
165 |
+
|
166 |
+
noised_latents = self.q_xt(latents, move_chance)
|
167 |
+
model_output = self.forward(noised_latents, unet_conditioning)
|
168 |
+
|
169 |
+
if self.T > 0:
|
170 |
+
diffusion_loss = self.d3pm_loss(model_output=model_output, xt=noised_latents, x0=latents, t=t)
|
171 |
+
return diffusion_loss
|
172 |
+
# SUBS parameterization, continuous time.
|
173 |
+
else:
|
174 |
+
log_p_theta = torch.gather(input=model_output, dim=-1, index=latents[:, :, None]).squeeze(-1)
|
175 |
+
return - log_p_theta * (dsigma / torch.expm1(sigma))[:, None]
|
176 |
+
|
177 |
|
178 |
+
######### LOSS CALCULATIONS #########
|
179 |
+
def maybe_sub_sample(self, x0, attention_mask):
|
180 |
+
# seqlen = x0.shape[1]
|
181 |
+
# print(seqlen)
|
182 |
+
# if seqlen > self.config.model.length:
|
183 |
+
# assert seqlen == 2 * self.config.model.length
|
184 |
+
# # cropping is needed for text8-crop dataset
|
185 |
+
# # try the same starting point for now
|
186 |
+
# start = np.random.choice(self.config.model.length)
|
187 |
+
# end = start + self.config.model.length
|
188 |
+
# input_tokens = x0[:, start: end]
|
189 |
+
# output_tokens = x0[:, start + 1: end + 1]
|
190 |
+
# new_attention_mask = attention_mask[:, start: end]
|
191 |
+
|
192 |
+
# # Helps with validation PPL, since the val
|
193 |
+
# # examples will all start and end with BOS/EOS
|
194 |
+
# input_tokens[:, 0] = self.tokenizer.bos_token_id
|
195 |
+
# output_tokens[:, -1] = self.tokenizer.eos_token_id
|
196 |
+
|
197 |
+
# elif self.parameterization == 'ar':
|
198 |
+
# input_tokens = x0[:, :-1]
|
199 |
+
# output_tokens = x0[:, 1:]
|
200 |
+
# new_attention_mask = attention_mask[:, 1:]
|
201 |
+
# else:
|
202 |
+
input_tokens = x0
|
203 |
+
output_tokens = None
|
204 |
+
new_attention_mask = attention_mask
|
205 |
+
|
206 |
+
return input_tokens, output_tokens, new_attention_mask
|
207 |
+
|
208 |
+
def compute_loss(self, latents, attention_mask):
|
209 |
+
""""Average of MLM losses to stabilize training"""
|
210 |
+
(input_tokens, output_tokens, attention_mask) = self.maybe_sub_sample(latents, attention_mask)
|
211 |
+
loss = self.forward_diffusion(input_tokens)
|
212 |
+
|
213 |
+
nlls = loss * attention_mask
|
214 |
+
count = attention_mask.sum()
|
215 |
+
batch_nll = nlls.sum()
|
216 |
+
token_nll = batch_nll / count
|
217 |
+
|
218 |
+
return Loss(loss=token_nll, nlls=nlls, token_mask=attention_mask)
|
219 |
+
|
220 |
+
|
221 |
+
######### TRAINING #########
|
222 |
def training_step(self, batch, batch_idx):
|
223 |
+
latents, attention_mask = batch
|
224 |
+
loss = self.compute_loss(latents, attention_mask)
|
|
|
|
|
|
|
225 |
return loss
|
226 |
|
227 |
def configure_optimizers(self):
|
228 |
optimizer = torch.optim.Adam(self.parameters(), lr=self.lr)
|
229 |
return optimizer
|
230 |
+
|
231 |
+
def validation_step(self, batch):
|
232 |
+
latents, attention_mask = batch
|
233 |
+
loss = self.compute_loss(latents, attention_mask)
|
234 |
+
return loss
|
235 |
+
|
236 |
+
|
237 |
+
######### GENERATION #########
|
238 |
+
def sample_prior(self, *batch_dims):
|
239 |
+
return self.mask_index * torch.ones(* batch_dims, dtype=torch.int64)
|
240 |
+
|
241 |
+
def sample_categorical(categorical_probs):
|
242 |
+
gumbel_norm = (1e-10 - (torch.rand_like(categorical_probs) + 1e-10).log())
|
243 |
+
return (categorical_probs / gumbel_norm).argmax(dim=-1)
|
244 |
+
|
245 |
+
def ddpm_caching_update(self, x, t, dt, p_x0=None):
|
246 |
+
assert self.config.noise.type == 'loglinear'
|
247 |
+
sigma_t, _ = self.noise(t)
|
248 |
+
if t.ndim > 1:
|
249 |
+
t = t.squeeze(-1)
|
250 |
+
assert t.ndim == 1
|
251 |
+
move_chance_t = t[:, None, None]
|
252 |
+
move_chance_s = (t - dt)[:, None, None]
|
253 |
+
assert move_chance_t.ndim == 3, move_chance_t.shape
|
254 |
+
if p_x0 is None:
|
255 |
+
p_x0 = self.forward(x, sigma_t).exp()
|
256 |
+
|
257 |
+
assert move_chance_t.ndim == p_x0.ndim
|
258 |
+
q_xs = p_x0 * (move_chance_t - move_chance_s)
|
259 |
+
q_xs[:, :, self.mask_index] = move_chance_s[:, :, 0]
|
260 |
+
_x = self.sample_categorical(q_xs)
|
261 |
+
|
262 |
+
copy_flag = (x != self.mask_index).to(x.dtype)
|
263 |
+
return p_x0, copy_flag * x + (1 - copy_flag) * _x
|
264 |
+
|
265 |
+
|
266 |
+
@torch.no_grad()
|
267 |
+
def sample_subs_guidance(self, n_samples, stride_length, num_strides, dt=0.001):
|
268 |
+
ones = torch.ones(n_samples, dtype=self.dtype,device=self.device)
|
269 |
+
num_steps = int(1 / dt)
|
270 |
+
sampling_steps = 0
|
271 |
+
intermediate_tokens = []
|
272 |
+
target = None
|
273 |
+
|
274 |
+
for _ in range(num_strides + 1):
|
275 |
+
p_x0_cache = None
|
276 |
+
x = self._sample_prior(n_samples,self.config.model.length).to(self.device)
|
277 |
+
|
278 |
+
if target is not None:
|
279 |
+
x[:, : -stride_length] = target
|
280 |
+
|
281 |
+
for i in range(num_steps + 1):
|
282 |
+
p_x0_cache, x_next = self.ddpm_caching_update(x=x, t=(1 - i * dt) * ones, dt=dt, p_x0=p_x0_cache)
|
283 |
+
if (not torch.allclose(x_next, x) or self.time_conditioning):
|
284 |
+
p_x0_cache = None
|
285 |
+
sampling_steps += 1
|
286 |
+
x = x_next
|
287 |
+
x = self.forward(x, 0 * ones).argmax(dim=-1)
|
288 |
+
intermediate_tokens.append(x[:, :stride_length].cpu().numpy())
|
289 |
+
target = x[:, stride_length:]
|
290 |
+
|
291 |
+
intermediate_tokens.append(target.cpu().numpy())
|
292 |
+
intermediate_text_samples = []
|
293 |
+
sequence_lengths = ((np.concatenate(intermediate_tokens, axis=1)[:, 1:]
|
294 |
+
== self.tokenizer.eos_token_id).cumsum(-1) == 0).sum(-1)
|
295 |
+
|
296 |
+
for i in range(2, len(intermediate_tokens) + 1):
|
297 |
+
intermediate_text_samples.append(self.tokenizer.decode(np.concatenate(intermediate_tokens[:i], axis=1)))
|
298 |
+
|
299 |
+
return (sampling_steps, intermediate_text_samples,
|
300 |
+
sequence_lengths)
|
301 |
+
|
302 |
+
def restore_model_and_semi_ar_sample(self, stride_length, num_strides, dt=0.001):
|
303 |
+
"""Generate samples from the model."""
|
304 |
+
# Lightning auto-casting is not working in this method for some reason
|
305 |
+
self.backbone.eval()
|
306 |
+
self.noise.eval()
|
307 |
+
|
308 |
+
(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)
|
309 |
+
|
310 |
+
self.backbone.train()
|
311 |
+
self.noise.train()
|
312 |
+
return sampling_steps, samples, sequence_lengths
|
models/dit.py
CHANGED
@@ -246,8 +246,7 @@ class DDiTBlock(nn.Module):
|
|
246 |
|
247 |
bias_dropout_scale_fn = self._get_bias_dropout_scale()
|
248 |
|
249 |
-
(shift_msa, scale_msa, gate_msa, shift_mlp,
|
250 |
-
scale_mlp, gate_mlp) = self.adaLN_modulation(c)[:, None].chunk(6, dim=2)
|
251 |
|
252 |
# attention operation
|
253 |
x_skip = x
|
@@ -315,7 +314,7 @@ class DDitFinalLayer(nn.Module):
|
|
315 |
|
316 |
|
317 |
def forward(self, x, c):
|
318 |
-
shift, scale = self.adaLN_modulation(c)[:, None].chunk(2, dim=2)
|
319 |
x = modulate_fused(self.norm_final(x), shift, scale)
|
320 |
x = self.linear(x)
|
321 |
return x
|
@@ -348,7 +347,7 @@ class DIT(nn.Module, huggingface_hub.PyTorchModelHubMixin):
|
|
348 |
config.model.hidden_size,
|
349 |
vocab_size,
|
350 |
config.model.cond_dim)
|
351 |
-
self.scale_by_sigma = config.model.scale_by_sigma
|
352 |
|
353 |
def _get_bias_dropout_scale(self):
|
354 |
if self.training:
|
|
|
246 |
|
247 |
bias_dropout_scale_fn = self._get_bias_dropout_scale()
|
248 |
|
249 |
+
(shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp) = self.adaLN_modulation(c)[:, None][0].chunk(6, dim=2)
|
|
|
250 |
|
251 |
# attention operation
|
252 |
x_skip = x
|
|
|
314 |
|
315 |
|
316 |
def forward(self, x, c):
|
317 |
+
shift, scale = self.adaLN_modulation(c)[:, None][0].chunk(2, dim=2)
|
318 |
x = modulate_fused(self.norm_final(x), shift, scale)
|
319 |
x = self.linear(x)
|
320 |
return x
|
|
|
347 |
config.model.hidden_size,
|
348 |
vocab_size,
|
349 |
config.model.cond_dim)
|
350 |
+
#self.scale_by_sigma = config.model.scale_by_sigma
|
351 |
|
352 |
def _get_bias_dropout_scale(self):
|
353 |
if self.training:
|