num_components = 7 # number of components (K)
num_iters = 5
learn_rate = 0.001
batch_size = 4
stop_after_steps = int(1e6)
# Details for the dataset, model and optimizer are left empty here.
# They can be found in the configurations for individual datasets,
# which are provided in configurations.py and added as named configs.
data = {} # Dataset details will go here
model = {} # Model details will go here
optimizer = {} # Optimizer details will go here
ex.named_config(configurations.clevr6)
ex.named_config(configurations.multi_dsprites)
ex.named_config(configurations.tetrominoes)
ex.named_config(configurations.dots)
@ex.capture
def build(identifier, _config):
config_copy = deepcopy(_config[identifier])
return utils.build(config_copy, identifier=identifier)
def get_train_step(model, dataset, optimizer):
loss, scalars, _ = model(dataset("train"))
global_step = tf.train.get_or_create_global_step()
grads = optimizer.compute_gradients(loss)
