latent_size,
kl_scale=10.,
opt=opt)
vae.summary()
encoder.summary()
decoder.summary()
start_nodes([
#RandomNode(12399, shape=[2]),
#CameraNode(12400, resize=32),
#CameraNode(12400, fovea_url="12399", fovea_size=1./4, resize=input_shape[0]),
CameraNode(12400, resize=input_shape[0]),
#CameraNode(12400, fovea_url="12399", fovea_size=1./2, resize=input_shape[0]),
#ImshowNode(12401, input_url="12400", title="source"),
GaussianLerp(12402, shape=[latent_size], steps=10),
VAENode(12403,
vae,
encoder,
decoder,
input_url="12400",
generator_url="12402").set_niceness(0.5),
#PrintNode(12404, input_url="12403/train_loss", prefix="train loss"),
#ImshowNode(12405, input_url="12403/reconstruction", title="reconstruction"),
#ImshowNode(12406, input_url="12403/generator", title="generation"),
GraphNode(12407,
input_url="12403/train_loss",
size=1000,
title="train_loss"),
TemporalHistogram(12408, input_url="12403/encoder", title="histogram")
])
return " ".join(args)
if __name__ == '__main__':
import argparse
from aegis_core.utils import start_nodes
parser = argparse.ArgumentParser()
parser.add_argument("--port", "-p", type=int)
parser.add_argument("--device", "-d", type=int, default=0)
parser.add_argument("--fovea-url", "-f", type=str, default=None)
parser.add_argument("--fovea-size", "-F", type=float, default=1./4)
parser.add_argument("--resize", "-r", type=int, default=None) #TODO: support 2d resize
parser.add_argument("--color", "-c", type=str, default="rgb")
parser.add_argument("--color-scale", "-s", type=float, default=1./255)
parser.add_argument("--niceness", "-N", type=float, default=1.)
parser.add_argument("--delay", "-D", type=float, default=1./100)
args = parser.parse_args()
node = CameraNode(
port=args.port,
device=args.device,
fovea_url=args.fovea_url,
fovea_size=args.fovea_size,
resize=args.resize,
color=args.color,
).set_niceness(args.niceness).set_delay(args.delay)
start_nodes([node])
vae, encoder, decoder = build_vae(encoder, decoder, input_shape,
latent_size)
vae.summary()
encoder.summary()
decoder.summary()
start_nodes([
#RandomNode(12399, shape=[2]),
#CameraNode(12400, resize=32),
#CameraNode(12400, fovea_url="12399", fovea_size=1./4, resize=32),
KerasDataset(12400, dataset="mnist", steps=10),
ImshowNode(12401, input_url="12400", title="source"),
GaussianLerp(12402, shape=[latent_size], steps=10),
VAENode(12403,
vae,
encoder,
decoder,
input_url="12400",
generator_url="12402"),
PrintNode(12404, input_url="12403/train_loss", prefix="train loss"),
ImshowNode(12405,
input_url="12403/reconstruction",
title="reconstruction"),
ImshowNode(12406, input_url="12403/generator", title="generation"),
GraphNode(12407,
input_url="12403/train_loss",
size=1000,
title="train_loss")
])
from aegis_core.camera_node import CameraNode from aegis_core.imshow_node import ImshowNode from aegis_core.utils import start_node, start_nodes start_nodes([ CameraNode(12400, resize=64), ImshowNode(12401, input_url="12400") ])
from aegis_core.utils import start_nodes, build_vae
from aegis_core.vae import VAENode
input_shape = [28, 28, 1]
latent_size = 2
acti = "tanh"
encoder = Sequential([
Flatten(input_shape=input_shape),
Dense(64, activation=acti),
Dense(32, activation=acti),
])
decoder = Sequential([
Dense(32, input_shape=[latent_size], activation=acti),
Dense(64, activation=acti),
Dense(np.prod(input_shape), activation="sigmoid"),
Reshape(input_shape)
])
vae, encoder, decoder = build_vae(encoder, decoder, input_shape, latent_size)
vae.summary()
encoder.summary()
decoder.summary()
start_nodes([
KerasDataset(12400, dataset="mnist", steps=100),
VAENode(12401, vae, encoder, decoder, input_url="12400").set_niceness(0),
PrintNode(12402, input_url="12401/train_loss")
])