-
Notifications
You must be signed in to change notification settings - Fork 1
/
clue.py
294 lines (258 loc) · 10.4 KB
/
clue.py
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
# AUTOGENERATED! DO NOT EDIT! File to edit: ../../nbs/methods/08_clue.ipynb.
# %% ../../nbs/methods/08_clue.ipynb 3
from __future__ import annotations
from ..import_essentials import *
from .base import CFModule, ParametricCFModule
from ..base import BaseConfig
from ..utils import auto_reshaping, validate_configs, get_config, grad_update
from ..ml_model import MLP, MLPBlock
from ..data_module import DataModule
from jax.scipy.stats.norm import logpdf as gaussian_logpdf
from keras.random import SeedGenerator
# %% auto 0
__all__ = ['Encoder', 'Decoder', 'kl_divergence', 'VAEGaussCat', 'CLUEConfig', 'get_reconstruction_loss_fn', 'CLUE']
# %% ../../nbs/methods/08_clue.ipynb 5
class Encoder(keras.layers.Layer):
def __init__(self, sizes: List[int], dropout: float = 0.1):
super().__init__()
assert sizes[-1] % 2 == 0, f"sizes[-1] must be even, but got {sizes[-1]}"
self.encoder = keras.Sequential([
MLPBlock(size, dropout_rate=dropout) for size in sizes
])
def call(self, x: Array, training: bool):
params = self.encoder(x, training=training)
d = params.shape[-1] // 2
mu, sigma = params[:, :d], params[:, d:]
sigma = jax.nn.softplus(sigma)
sigma = jnp.clip(sigma, 1e-3)
return mu, sigma
class Decoder(keras.layers.Layer):
def __init__(self, sizes: List[int], output_size: int, dropout: float = 0.1):
super().__init__()
self.decoder = MLP(
sizes, output_size=output_size,
dropout_rate=dropout, last_activation='sigmoid'
)
def __call__(self, z: Array, training: bool):
mu_dec = self.decoder(z, training=training)
return mu_dec
# %% ../../nbs/methods/08_clue.ipynb 7
@jit
def kl_divergence(p: Array, q: Array, eps: float = 2 ** -17) -> Array:
loss_pointwise = p * (jnp.log(p + eps) - jnp.log(q + eps))
return loss_pointwise
# %% ../../nbs/methods/08_clue.ipynb 8
class VAEGaussCat(keras.Model):
def __init__(
self,
enc_sizes: List[int] = [20, 16, 14, 12],
dec_sizes: List[int] = [12, 14, 16, 20],
dropout_rate: float = 0.1,
):
super().__init__()
self.enc_sizes = enc_sizes
self.dec_sizes = dec_sizes
self.dropout_rate = dropout_rate
self.seed_generator = SeedGenerator(get_config().global_seed)
# default reconstruction loss to l2 loss
self.reconstruction_loss = lambda x, y: optax.l2_loss(x, y).mean(-1)
def set_reconstruction_loss(self, fn):
self.reconstruction_loss = fn
def build(self, input_shape):
self.encoder = Encoder(self.enc_sizes, self.dropout_rate)
self.decoder = Decoder(self.dec_sizes, input_shape[-1], self.dropout_rate)
def encode(self, x, training=True):
mu_z, var_z = self.encoder(x, training=training)
return mu_z, var_z
def sample_latent(self, rng_key, mean, var):
key, _ = jax.random.split(rng_key)
std = jnp.exp(0.5 * var)
eps = jrand.normal(key, var.shape)
return mean + eps * std
def decode(self, z, training=True):
reconstruct_x = self.decoder(z, training=training)
return reconstruct_x
def sample_step(self, rng_key, mean, var, training=True):
z = self.sample_latent(rng_key, mean, var)
mu_x = self.decode(z, training=training)
return mu_x
def sample(self, x, mc_samples, training=True): # Shape: (mc_samples, batch_size, input_size)
mean, var = self.encode(x, training=training)
rng_key = self.seed_generator.next()
keys = jax.random.split(rng_key, mc_samples)
partial_sample_step = partial(
self.sample_step, mean=mean, var=var, training=training
)
reconstruct_x = jax.vmap(partial_sample_step)(keys)
return (mean, var, reconstruct_x)
def compute_vae_loss(self, inputs, mu_z, logvar_z, reconstruct_x):
kl_loss = -0.5 * (1 + logvar_z - jnp.power(mu_z, 2) - jnp.exp(logvar_z)).sum(-1)
rec = self.reconstruction_loss(inputs, reconstruct_x.reshape(inputs.shape)).sum(-1)
batchwise_loss = (rec + kl_loss) / inputs.shape[0]
return batchwise_loss.mean()
def call(self, inputs, training=True):
mu_z, logvar_z, reconstruct_x = self.sample(inputs, mc_samples=1, training=training)
loss = self.compute_vae_loss(inputs, mu_z, logvar_z, reconstruct_x)
self.add_loss(loss)
return reconstruct_x
# %% ../../nbs/methods/08_clue.ipynb 10
@ft.partial(jit, static_argnums=(3, 4, 6, 9, 12, 13))
def _clue_generate(
x: Array,
rng_key: jrand.PRNGKey,
y_target: Array,
pred_fn: Callable,
max_steps: int,
step_size: float,
vae_module: VAEGaussCat,
uncertainty_weight: float,
aleatoric_weight: float,
prior_weight: float,
distance_weight: float,
validity_weight: float,
validity_fn: Callable,
apply_fn: Callable
) -> Array:
@jit
def sample_latent_from_x(x: Array, rng_key: jrand.PRNGKey):
key_1, key_2 = jrand.split(rng_key)
mean, var = vae_module.encode(x, training=False)
z = vae_module.sample_latent(key_2, mean, var)
return z
@ft.partial(jit, static_argnums=(1))
def generate_from_z(z: Array, hard: bool):
cf = vae_module.decode(z, training=False)
cf = apply_fn(x, cf, hard=hard)
return cf
@jit
def uncertainty_from_z(z: Array):
cfs = generate_from_z(z, hard=False)
pred_cfs = pred_fn(cfs)
prob = pred_cfs[:, 1]
total_uncertainty = -(prob * jnp.log(prob + 1e-10)).sum(-1)
return total_uncertainty, cfs, pred_cfs
@jit
def compute_loss(z: Array):
uncertainty, cfs, pred_cfs = uncertainty_from_z(z)
loglik = gaussian_logpdf(z).sum(-1)
dist = jnp.abs(cfs - x).mean()
validity = validity_fn(y_target, pred_cfs).mean()
loss = (
(uncertainty_weight + aleatoric_weight) * uncertainty
+ prior_weight * loglik
+ distance_weight * dist
+ validity_weight * validity
)
return loss.mean()
@loop_tqdm(max_steps)
def step(i, z_opt_state):
z, opt_state = z_opt_state
z_grad = jax.grad(compute_loss)(z)
z, opt_state = grad_update(z_grad, z, opt_state, opt)
return z, opt_state
key_1, _ = jax.random.split(rng_key)
z = sample_latent_from_x(x, key_1)
opt = optax.adam(step_size)
opt_state = opt.init(z)
# Write a loop to optimize z using lax.fori_loop
z, opt_state = lax.fori_loop(0, max_steps, step, (z, opt_state))
cf = generate_from_z(z, hard=True)
return cf
# %% ../../nbs/methods/08_clue.ipynb 12
class CLUEConfig(BaseConfig):
enc_sizes: List[int] = Field(
[20, 16, 14, 12], description="Sequence of Encoder layer sizes."
)
dec_sizes: List[int] = Field(
[12, 14, 16, 20], description="Sequence of Decoder layer sizes."
)
dropout_rate: float = Field(0.1, description="Dropout rate")
encoded_size: int = Field(5, description="Encoded size")
lr: float = Field(0.001, description="Learning rate")
max_steps: int = Field(500, description="Max steps")
step_size: float = Field(0.01, description="Step size")
vae_n_epochs: int = Field(10, description="Number of epochs for VAE")
vae_batch_size: int = Field(128, description="Batch size for VAE")
seed: int = Field(0, description="Seed for random number generator")
# %% ../../nbs/methods/08_clue.ipynb 13
def get_reconstruction_loss_fn(dm: DataModule):
def reconstruction_loss(xs, cfs):
losses = []
for feat, (start, end) in dm.features.features_and_indices:
if feat.is_categorical:
losses.append(
optax.softmax_cross_entropy(cfs[:, start:end], xs[:, start:end]).reshape(-1, 1)
)
else:
losses.append(optax.l2_loss(cfs[:, start:end], xs[:, start:end]))
return jnp.concatenate(losses, axis=-1)
return reconstruction_loss
# %% ../../nbs/methods/08_clue.ipynb 15
class CLUE(ParametricCFModule):
def __init__(self, config: Dict | CLUEConfig = None, vae=None, name: str = 'CLUE'):
if config is None:
config = CLUEConfig()
config = validate_configs(config, CLUEConfig)
self.vae = vae
super().__init__(config, name=name)
def _init_model(self, config: CLUEConfig, model: VAEGaussCat):
if model is None:
model = VAEGaussCat(
enc_sizes=config.enc_sizes, dec_sizes=config.dec_sizes,
dropout_rate=config.dropout_rate
)
model.compile(optimizer=keras.optimizers.Adam(config.lr), loss=None)
return model
def train(
self,
data: DataModule, # data module
pred_fn: Callable = None,
batch_size: int = 128,
epochs: int = 10,
**fit_kwargs
):
if not isinstance(data, DataModule):
raise ValueError(f"Expected `data` to be `DataModule`, "
f"got type=`{type(data).__name__}` instead.")
X_train, y_train = data['train']
self.vae = self._init_model(self.config, self.vae)
self.vae.fit(
X_train, X_train,
batch_size=batch_size,
epochs=epochs,
**fit_kwargs
)
self._is_trained = True
return self
@auto_reshaping('x')
def generate_cf(
self,
x: Array,
pred_fn: Callable,
y_target: Array = None,
rng_key: jrand.PRNGKey = None,
**kwargs
) -> Array:
# TODO: Currently assumes binary classification.
if y_target is None:
y_target = 1 - pred_fn(x)
else:
y_target = jnp.array(y_target, copy=True)
if rng_key is None:
raise ValueError("`rng_key` must be provided, but got `None`.")
return _clue_generate(
x,
rng_key=rng_key,
y_target=y_target,
pred_fn=pred_fn,
max_steps=self.config.max_steps,
step_size=self.config.step_size,
vae_module=self.vae,
uncertainty_weight=.0,
aleatoric_weight=0.0,
prior_weight=0.0,
distance_weight=.1,
validity_weight=1.0,
validity_fn=keras.losses.get({'class_name': 'KLDivergence', 'config': {'reduction': None}}),
apply_fn=self.apply_constraints,
)