def next_frame_sv2p():
"""SV2P model hparams."""
hparams = basic_stochastic.next_frame_basic_stochastic()
hparams.optimizer = "true_adam"
hparams.learning_rate_schedule = "constant"
hparams.learning_rate_constant = 1e-3
hparams.video_num_input_frames = 1
hparams.video_num_target_frames = 3
hparams.batch_size = 16
hparams.modality = {
"inputs": modalities.ModalityType.VIDEO_L2_RAW,
"targets": modalities.ModalityType.VIDEO_L2_RAW,
}
hparams.video_modality_loss_cutoff = 0.0
hparams.scheduled_sampling_mode = "count"
hparams.scheduled_sampling_k = 900.0
hparams.add_hparam("reward_prediction", True)
hparams.add_hparam("reward_prediction_stop_gradient", False)
hparams.add_hparam("reward_prediction_buffer_size", 0)
hparams.add_hparam("model_options", "CDNA")
hparams.add_hparam("num_masks", 10)
hparams.add_hparam("multi_latent", False)
hparams.add_hparam("relu_shift", 1e-12)
hparams.add_hparam("dna_kernel_size", 5)
hparams.add_hparam("upsample_method", "conv2d_transpose")
hparams.add_hparam("reward_model", "basic")
hparams.add_hparam("visualize_logits_histogram", True)
return hparams
def next_frame_sv2p():
"""SV2P model hparams."""
hparams = basic_stochastic.next_frame_basic_stochastic()
hparams.optimizer = "TrueAdam"
hparams.learning_rate_schedule = "constant"
hparams.learning_rate_constant = 1e-3
hparams.video_num_input_frames = 1
hparams.video_num_target_frames = 3
hparams.batch_size = 16
hparams.modality = {
"inputs": modalities.VideoModalityL2Raw,
"targets": modalities.VideoModalityL2Raw,
}
hparams.video_modality_loss_cutoff = 0.0
hparams.scheduled_sampling_mode = "count"
hparams.scheduled_sampling_k = 900.0
hparams.add_hparam("reward_prediction", True)
hparams.add_hparam("reward_prediction_stop_gradient", False)
hparams.add_hparam("reward_prediction_buffer_size", 0)
hparams.add_hparam("model_options", "CDNA")
hparams.add_hparam("num_masks", 10)
hparams.add_hparam("multi_latent", False)
hparams.add_hparam("relu_shift", 1e-12)
hparams.add_hparam("dna_kernel_size", 5)
hparams.add_hparam("upsample_method", "conv2d_transpose")
hparams.add_hparam("reward_model", "basic")
hparams.add_hparam("visualize_logits_histogram", True)
return hparams
def next_frame_sv2p():
"""SV2P model hparams."""
hparams = basic_stochastic.next_frame_basic_stochastic()
hparams.optimizer = "TrueAdam"
hparams.learning_rate_schedule = "constant"
hparams.learning_rate_constant = 1e-3
hparams.video_num_input_frames = 1
hparams.video_num_target_frames = 3
hparams.batch_size = 16
hparams.target_modality = "video:l2raw"
hparams.input_modalities = "inputs:video:l2raw"
hparams.video_modality_loss_cutoff = 0.0
hparams.add_hparam("reward_prediction", True)
hparams.add_hparam("reward_prediction_stop_gradient", False)
hparams.add_hparam("reward_prediction_buffer_size", 0)
hparams.add_hparam("model_options", "CDNA")
hparams.add_hparam("num_masks", 10)
hparams.add_hparam("multi_latent", False)
hparams.add_hparam("relu_shift", 1e-12)
hparams.add_hparam("dna_kernel_size", 5)
# Scheduled sampling method. Choose between prob or count.
hparams.add_hparam("scheduled_sampling_mode", "count")
hparams.add_hparam("scheduled_sampling_decay_steps", 10000)
hparams.add_hparam("scheduled_sampling_k", 900.0)
hparams.add_hparam("upsample_method", "conv2d_transpose")
hparams.add_hparam("internal_loss", True)
hparams.add_hparam("reward_model", "basic")
return hparams
def testBasicStochastic(self):
self.TestOnVariousInputOutputSizes(
basic_stochastic.next_frame_basic_stochastic(),
basic_stochastic.NextFrameBasicStochastic, 256, False)
def testBasicStochastic(self):
self.TestOnVariousInputOutputSizes(
basic_stochastic.next_frame_basic_stochastic(),
basic_stochastic.NextFrameBasicStochastic,
256,
False)
