def next_frame_svglp():
"""SVG with learned prior model hparams."""
hparams = sv2p_params.next_frame_sv2p()
hparams.video_num_input_frames = 2
hparams.video_num_target_frames = 10
hparams.learning_rate_constant = 1e-4
seq_length = hparams.video_num_input_frames + hparams.video_num_target_frames
# The latent_loss_multiplier is divided by the number of frames because
# the image sequence loss in t2t is averaged instead of added through
# time as they do in the SVG-LP paper
hparams.latent_loss_multiplier = 1e-4 / seq_length
hparams.reward_prediction = False
hparams.num_iterations_1st_stage = -1
hparams.num_iterations_2nd_stage = -1
hparams.optimizer_adam_beta1 = 0.9
hparams.optimizer_adam_beta2 = 0.999
hparams.optimizer_adam_epsilon = 1e-08
hparams.anneal_end = -1
hparams.clip_grad_norm = 5.0
hparams.add_hparam("learned_prior", True)
hparams.add_hparam("z_dim", 64)
hparams.add_hparam("g_dim", 128)
hparams.add_hparam("rnn_size", 256)
hparams.add_hparam("rnn_type", "lstm")
hparams.add_hparam("prior_rnn_layers", 1)
hparams.add_hparam("posterior_rnn_layers", 1)
hparams.add_hparam("predictor_rnn_layers", 2)
hparams.add_hparam("has_skips", True)
hparams.add_hparam("has_batchnorm", True)
return hparams
def next_frame_svglp():
"""SVG with learned prior model hparams."""
hparams = sv2p_params.next_frame_sv2p()
hparams.video_num_input_frames = 2
hparams.video_num_target_frames = 10
hparams.learning_rate_constant = 1e-4
seq_length = hparams.video_num_input_frames + hparams.video_num_target_frames
# The latent_loss_multiplier is divided by the number of frames because
# the image sequence loss in t2t is averaged instead of added through
# time as they do in the SVG-LP paper
hparams.latent_loss_multiplier = 1e-4 / seq_length
hparams.reward_prediction = False
hparams.num_iterations_1st_stage = -1
hparams.num_iterations_2nd_stage = -1
hparams.optimizer_adam_beta1 = 0.9
hparams.optimizer_adam_beta2 = 0.999
hparams.optimizer_adam_epsilon = 1e-08
hparams.anneal_end = -1
hparams.clip_grad_norm = 5.0
hparams.add_hparam("learned_prior", True)
hparams.add_hparam("z_dim", 64)
hparams.add_hparam("g_dim", 128)
hparams.add_hparam("rnn_size", 256)
hparams.add_hparam("rnn_type", "lstm")
hparams.add_hparam("prior_rnn_layers", 1)
hparams.add_hparam("posterior_rnn_layers", 1)
hparams.add_hparam("predictor_rnn_layers", 2)
hparams.add_hparam("has_skips", True)
hparams.add_hparam("has_batchnorm", True)
return hparams
def next_frame_savp():
"""SAVP model hparams."""
hparams = sv2p_params.next_frame_sv2p()
hparams.add_hparam("z_dim", 8)
hparams.add_hparam("num_discriminator_filters", 32)
hparams.add_hparam("use_vae", True)
hparams.add_hparam("use_gan", False)
hparams.add_hparam("use_spectral_norm", True)
hparams.add_hparam("gan_loss", "cross_entropy")
hparams.add_hparam("gan_loss_multiplier", 0.01)
hparams.add_hparam("gan_vae_loss_multiplier", 0.01)
hparams.add_hparam("gan_optimization", "joint")
hparams.bottom = {
"inputs": modalities.video_raw_bottom,
"targets": modalities.video_raw_targets_bottom,
}
hparams.loss = {
"targets": modalities.video_l1_raw_loss,
}
hparams.top = {
"targets": modalities.video_raw_top,
}
hparams.latent_loss_multiplier_schedule = "linear"
hparams.upsample_method = "bilinear_upsample_conv"
hparams.internal_loss = False
hparams.reward_prediction = False
hparams.anneal_end = 100000
hparams.num_iterations_1st_stage = 0
hparams.num_iterations_2nd_stage = 50000
return hparams
def testSv2pWithActionsAndRewardsExternalLoss(self):
hp = sv2p_params.next_frame_sv2p()
hp.internal_loss = False
self.TestWithActionAndRewards(
hp,
sv2p.NextFrameSv2p,
1,
False)
def next_frame_emily():
"""Emily's model hparams."""
hparams = sv2p_params.next_frame_sv2p()
hparams.latent_loss_multiplier = 1e-4
hparams.learning_rate_constant = 1e-4
hparams.add_hparam("z_dim", 10)
hparams.add_hparam("g_dim", 128)
hparams.add_hparam("rnn_size", 256)
hparams.add_hparam("posterior_rnn_layers", 1)
hparams.add_hparam("predictor_rnn_layers", 2)
return hparams
def next_frame_savp():
"""SAVP model hparams."""
hparams = sv2p_params.next_frame_sv2p()
hparams.add_hparam("z_dim", 8)
hparams.add_hparam("num_discriminator_filters", 32)
hparams.add_hparam("use_vae", True)
hparams.add_hparam("use_gan", False)
hparams.add_hparam("use_spectral_norm", True)
hparams.add_hparam("gan_loss", "cross_entropy")
hparams.add_hparam("gan_loss_multiplier", 0.01)
hparams.add_hparam("gan_vae_loss_multiplier", 0.01)
hparams.add_hparam("gan_optimization", "joint")
hparams.target_modality = "video:l1raw"
hparams.input_modalities = "inputs:video:l1raw"
hparams.latent_loss_multiplier_schedule = "linear"
hparams.upsample_method = "bilinear_upsample_conv"
return hparams
def next_frame_savp():
"""SAVP model hparams."""
hparams = sv2p_params.next_frame_sv2p()
hparams.add_hparam("z_dim", 8)
hparams.add_hparam("num_discriminator_filters", 32)
hparams.add_hparam("use_vae", True)
hparams.add_hparam("use_gan", False)
hparams.add_hparam("use_spectral_norm", True)
hparams.add_hparam("gan_loss", "cross_entropy")
hparams.add_hparam("gan_loss_multiplier", 0.01)
hparams.add_hparam("gan_vae_loss_multiplier", 0.01)
hparams.add_hparam("gan_optimization", "joint")
hparams.modality = {
"inputs": modalities.VideoModalityL1Raw,
"targets": modalities.VideoModalityL1Raw,
}
hparams.latent_loss_multiplier_schedule = "linear"
hparams.upsample_method = "bilinear_upsample_conv"
hparams.internal_loss = False
hparams.reward_prediction = False
hparams.anneal_end = 100000
hparams.num_iterations_1st_stage = 0
hparams.num_iterations_2nd_stage = 50000
return hparams
def testSv2pTwoFrames(self):
self.TestOnVariousInputOutputSizes(sv2p_params.next_frame_sv2p(),
sv2p.NextFrameSv2pTwoFrames, 1,
False)
def testSv2pWithActions(self):
self.TestWithActions(sv2p_params.next_frame_sv2p(), sv2p.NextFrameSv2p,
1, False)
def testSv2pWithActionsAndRewards(self):
self.TestWithActionAndRewards(
sv2p_params.next_frame_sv2p(),
sv2p.NextFrameSv2p,
1)
def testSv2pTwoFrames(self):
self.TestOnVariousInputOutputSizes(
sv2p_params.next_frame_sv2p(),
sv2p.NextFrameSv2pTwoFrames,
1,
False)
def testSv2pWithActions(self):
self.TestWithActions(
sv2p_params.next_frame_sv2p(),
sv2p.NextFrameSv2p,
1,
False)
