def pred_step(self, x, net_states, initializing):
if initializing:
states = elegy.States(net_states=0)
else:
states = elegy.States(net_states=net_states + 1)
return elegy.PredStep.simple(x + 1.0, states)
def test_step(self, metrics_states, initializing):
return elegy.TestStep(
loss=0.1,
logs=dict(loss=1.0),
states=elegy.States(metrics_states=0) if initializing else
elegy.States(metrics_states=metrics_states + 1),
)
def train_step(self, states, initializing):
return elegy.TrainStep(
logs=dict(loss=2.0),
states=elegy.States(
optimizer_states=0) if initializing else elegy.States(
optimizer_states=states.optimizer_states + 1),
)
def pred_step(self, x, initializing, states):
if initializing:
states = elegy.States(net_states=0)
else:
states = elegy.States(net_states=states.net_states + 1)
return elegy.PredStep(x + 1.0, states)
def pred_step(self, x, states, initializing):
if initializing:
states = elegy.States(net_states=0)
else:
states = elegy.States(net_states=states.net_states + 1)
return elegy.PredStep(
y_pred=1,
states=states,
)
def train_step(self, x, optimizer_states, initializing):
if initializing:
states = elegy.States(optimizer_states=0)
else:
states = elegy.States(optimizer_states=optimizer_states +
1)
return elegy.TrainStep(
logs=dict(loss=jnp.sum(x)),
states=states,
)
def test_step(self, x, metrics_states, initializing):
if initializing:
states = elegy.States(metrics_states=0)
else:
states = elegy.States(metrics_states=metrics_states + 1)
return elegy.TestStep(
loss=0.1,
logs=dict(loss=jnp.sum(x)),
states=states,
)
def states_step(self):
return elegy.States(
net_params=None,
net_states=None,
metrics_states=None,
optimizer_states=None,
)
def pred_step(self):
nonlocal N
N = N + 1
return elegy.PredStep.simple(
y_pred=None,
states=elegy.States(net_params=1, net_states=2),
)
def train_step(self, x, y_true, net_params, optimizer_states, initializing,
rng):
def loss_fn(net_params, x, y_true):
# flatten + scale
x = jnp.reshape(x, (x.shape[0], -1)) / 255
# model
if initializing:
w = jax.random.uniform(rng.next(),
shape=[np.prod(x.shape[1:]), 10],
minval=-1,
maxval=1)
b = jax.random.uniform(rng.next(),
shape=[1],
minval=-1,
maxval=1)
net_params = (w, b)
w, b = net_params
logits = jnp.dot(x, w) + b
# crossentropy loss
labels = jax.nn.one_hot(y_true, 10)
sample_loss = -jnp.sum(labels * jax.nn.log_softmax(logits),
axis=-1)
loss = jnp.mean(sample_loss)
# metrics
logs = dict(
accuracy=jnp.mean(jnp.argmax(logits, axis=-1) == y_true),
loss=loss,
)
return loss, (logs, net_params)
# train
(loss,
(logs,
net_params)), grads = jax.value_and_grad(loss_fn, has_aux=True)(
net_params, # gradients target
x,
y_true,
)
if initializing:
optimizer_states = self.optim.init(net_params)
else:
grads, optimizer_states = self.optim.update(
grads, optimizer_states, net_params)
net_params = optax.apply_updates(net_params, grads)
return logs, elegy.States(
net_params=net_params,
optimizer_states=optimizer_states,
)
def init_step(self, x):
rng = elegy.KeySeq(0)
gx, g_params, g_states = self.generator.init(rng=rng)(x)
dx, d_params, d_states = self.discriminator.init(rng=rng)(gx)
g_optimizer_states = self.g_optimizer.init(g_params)
d_optimizer_states = self.d_optimizer.init(d_params)
return elegy.States(
g_states=g_states,
d_states=d_states,
g_params=g_params,
d_params=d_params,
g_opt_states=g_optimizer_states,
d_opt_states=d_optimizer_states,
rng=rng,
step=0,
)
def main(
steps_per_epoch: int = 200,
batch_size: int = 64,
epochs: int = 50,
debug: bool = False,
eager: bool = False,
logdir: str = "runs",
):
if debug:
import debugpy
print("Waiting for debugger...")
debugpy.listen(5678)
debugpy.wait_for_client()
current_time = datetime.now().strftime("%b%d_%H-%M-%S")
logdir = os.path.join(logdir, current_time)
X_train, _1, X_test, _2 = dataget.image.mnist(global_cache=True).get()
# Now binarize data
X_train = (X_train > 0).astype(jnp.float32)
X_test = (X_test > 0).astype(jnp.float32)
print("X_train:", X_train.shape, X_train.dtype)
print("X_test:", X_test.shape, X_test.dtype)
model = VariationalAutoEncoder(
latent_size=LATENT_SIZE,
optimizer=optax.adam(1e-3),
run_eagerly=eager,
)
# Fit with datasets in memory
history = model.fit(
x=X_train,
epochs=epochs,
batch_size=batch_size,
steps_per_epoch=steps_per_epoch,
validation_data=(X_test, ),
shuffle=True,
callbacks=[TensorBoard(logdir)],
)
print(
"\n\n\nMetrics and images can be explored using tensorboard using:",
f"\n \t\t\t tensorboard --logdir {logdir}",
)
elegy.utils.plot_history(history)
# get random samples
idxs = np.random.randint(0, len(X_test), size=(5, ))
x_sample = X_test[idxs]
# get predictions
y_pred = model.predict(x=x_sample)
y_pred = jax.nn.sigmoid(y_pred)
# plot and save results
with SummaryWriter(os.path.join(logdir, "val")) as tbwriter:
figure = plt.figure(figsize=(12, 12))
for i in range(5):
plt.subplot(2, 5, i + 1)
plt.imshow(x_sample[i], cmap="gray")
plt.subplot(2, 5, 5 + i + 1)
plt.imshow(y_pred[i], cmap="gray")
# # tbwriter.add_figure("VAE Example", figure, epochs)
plt.show()
# sample
model_decoder = elegy.Model(
model.decoder,
states=elegy.States(
net_params=model.states.net_params[1],
net_states=model.states.net_states[1],
),
initialized=True,
)
z_samples = np.random.normal(size=(12, LATENT_SIZE))
logits = model_decoder.predict(z_samples)
samples = jax.nn.sigmoid(logits)
# plot and save results
# with SummaryWriter(os.path.join(logdir, "val")) as tbwriter:
figure = plt.figure(figsize=(5, 12))
plt.title("Generative Samples")
for i in range(5):
plt.subplot(2, 5, 2 * i + 1)
plt.imshow(samples[i], cmap="gray")
plt.subplot(2, 5, 2 * i + 2)
plt.imshow(samples[i + 1], cmap="gray")
# # tbwriter.add_figure("VAE Generative Example", figure, epochs)
plt.show()