class AttentionLayer(Config):
# Softmax temperature
tau: float = 10.0
# Config for the attention net. Final layer will be added automatically
mlp_config: MLP = MLP(
layers=(100, 50),
activation=None
)
# Number of input views
n_views: int = 2
from config.eamc.defaults import EAMCExperiment, EAMC, AttentionLayer, Discriminator, Loss, Optimizer
BACKBONE_MLP_LAYERS = (200, 200, 500)
CNN_LAYERS = (("conv", 5, 5, 32, "relu"), ("pool", 2, 2), ("conv", 5, 5, 64,
"relu"),
("pool", 2, 2), ("fc", 500), ("bn", ), ("relu", ))
CNN_BACKBONES = (
CNN(layers=CNN_LAYERS, input_size=(1, 28, 28)),
CNN(layers=CNN_LAYERS, input_size=(1, 28, 28)),
)
eamc_blobs_overlap = EAMCExperiment(
dataset_config=Dataset(name="blobs_overlap"),
model_config=EAMC(
backbone_configs=(
MLP(layers=[32, 32, 32], input_size=(2, )),
MLP(layers=[32, 32, 32], input_size=(2, )),
),
discriminator_config=Discriminator(mlp_config=MLP(layers=(32, 32,
32))),
loss_config=Loss(),
cm_config=DDC(n_clusters=3),
optimizer_config=Optimizer(lr_backbones=2e-4, lr_disc=1e-5)),
)
eamc_blobs_overlap_5 = EAMCExperiment(
dataset_config=Dataset(name="blobs_overlap_5"),
model_config=EAMC(
backbone_configs=(
MLP(layers=[32, 32, 32], input_size=(2, )),
MLP(layers=[32, 32, 32], input_size=(2, )),
from config.defaults import Experiment, Dataset, SiMVC, MLP, DDC, Fusion, Loss, CoMVC
blobs_overlap = Experiment(
dataset_config=Dataset(name="blobs_overlap"),
model_config=SiMVC(
backbone_configs=(
MLP(layers=[32, 32, 32], input_size=(2, )),
MLP(layers=[32, 32, 32], input_size=(2, )),
),
fusion_config=Fusion(method="weighted_mean", n_views=2),
cm_config=DDC(n_clusters=3),
loss_config=Loss(funcs="ddc_1|ddc_2|ddc_3", ),
),
n_runs=1,
n_epochs=10,
)
blobs_overlap_contrast = Experiment(
dataset_config=Dataset(name="blobs_overlap"),
model_config=CoMVC(backbone_configs=(
MLP(layers=[32, 32, 32], input_size=(2, )),
MLP(layers=[32, 32, 32], input_size=(2, )),
),
fusion_config=Fusion(method="weighted_mean", n_views=2),
projector_config=None,
cm_config=DDC(n_clusters=3),
loss_config=Loss(funcs="ddc_1|ddc_2|ddc_3|contrast", )),
n_runs=1,
)
blobs_overlap_5 = Experiment(
from config.defaults import Experiment, Dataset, SiMVC, DDC, Fusion, MLP, Loss, CoMVC, Optimizer
voc = Experiment(
dataset_config=Dataset(name="voc"),
model_config=SiMVC(
backbone_configs=(
MLP(input_size=(512,)),
MLP(input_size=(399,)),
),
fusion_config=Fusion(method="weighted_mean", n_views=2),
cm_config=DDC(n_clusters=20),
loss_config=Loss(
funcs="ddc_1|ddc_2|ddc_3",
),
optimizer_config=Optimizer(learning_rate=1e-3, scheduler_step_size=50, scheduler_gamma=0.1)
),
)
voc_contrast = Experiment(
dataset_config=Dataset(name="voc"),
model_config=CoMVC(
backbone_configs=(
MLP(input_size=(512,)),
MLP(input_size=(399,)),
),
projector_config=None,
fusion_config=Fusion(method="weighted_mean", n_views=2),
cm_config=DDC(n_clusters=20),
loss_config=Loss(
funcs="ddc_1|ddc_2|ddc_3|contrast",
),
from config.defaults import Experiment, SiMVC, DDC, Fusion, MLP, Loss, Dataset, CoMVC, Optimizer
ccv = Experiment(
dataset_config=Dataset(name="ccv"),
model_config=SiMVC(backbone_configs=(
MLP(input_size=(5000, )),
MLP(input_size=(5000, )),
MLP(input_size=(4000, )),
),
fusion_config=Fusion(method="weighted_mean", n_views=3),
cm_config=DDC(n_clusters=20),
loss_config=Loss(funcs="ddc_1|ddc_2|ddc_3", ),
optimizer_config=Optimizer()),
)
ccv_contrast = Experiment(
dataset_config=Dataset(name="ccv"),
model_config=CoMVC(backbone_configs=(
MLP(input_size=(5000, )),
MLP(input_size=(5000, )),
MLP(input_size=(4000, )),
),
fusion_config=Fusion(method="weighted_mean", n_views=3),
projector_config=None,
cm_config=DDC(n_clusters=20),
loss_config=Loss(funcs="ddc_1|ddc_2|ddc_3|contrast",
delta=20.0),
optimizer_config=Optimizer(scheduler_step_size=50,
scheduler_gamma=0.1)),
n_epochs=100)
from config.defaults import Experiment, Dataset, SiMVC, DDC, Fusion, MLP, Loss, CoMVC, Optimizer
rgbd = Experiment(
dataset_config=Dataset(name="rgbd"),
model_config=SiMVC(
backbone_configs=(
MLP(input_size=(2048,)),
MLP(input_size=(300,)),
),
fusion_config=Fusion(method="weighted_mean", n_views=2),
cm_config=DDC(n_clusters=13),
loss_config=Loss(
funcs="ddc_1|ddc_2|ddc_3",
)
),
)
rgbd_contrast = Experiment(
dataset_config=Dataset(name="rgbd"),
model_config=CoMVC(
backbone_configs=(
MLP(input_size=(2048,)),
MLP(input_size=(300,)),
),
fusion_config=Fusion(method="weighted_mean", n_views=2),
projector_config=None,
cm_config=DDC(n_clusters=13),
loss_config=Loss(
funcs="ddc_1|ddc_2|ddc_3|contrast",
),
class Discriminator(Config):
# Config for the discriminator
mlp_config: MLP = MLP(
layers=(256, 256, 128),
activation="leaky_relu:0.2"
)