def Task(self):
p = super(StarNetPedCycModel0704, self).Task()
p.train.learning_rate = 7e-4
builder = starnet.Builder()
builder.linear_params_init = py_utils.WeightInit.KaimingUniformFanInRelu(
)
gin_layer_sizes = [32, 256, 512, 256, 256, 128]
num_laser_features = 1
gin_layers = [
# Each layer should expect as input - 2 * dims of the last layer's
# output. We assume a middle layer that's the size of 2 * dim_out.
[dim_in * 2, dim_out * 2, dim_out]
for (dim_in,
dim_out) in zip(gin_layer_sizes[:-1], gin_layer_sizes[1:])
]
p.cell_feature_dims = sum(gin_layer_sizes)
# Disable BN on first layer
p.cell_featurizer = builder.GINFeaturizerV2('feat',
gin_layer_sizes[0],
gin_layers,
num_laser_features,
fc_use_bn=False)
p.anchor_projected_feature_dims = 512
class_name_to_idx = kitti_input_generator.KITTILabelExtractor.KITTI_CLASS_NAMES
num_classes = len(class_name_to_idx)
p.per_class_loss_weight = [0.] * num_classes
p.per_class_loss_weight[class_name_to_idx.index('Pedestrian')] = 3.5
p.per_class_loss_weight[class_name_to_idx.index('Cyclist')] = 3.25
p.focal_loss_alpha = 0.9
p.focal_loss_gamma = 1.25
p.use_oriented_per_class_nms = True
p.max_nms_boxes = 1024
p.nms_iou_threshold = [0.0] * num_classes
p.nms_iou_threshold[class_name_to_idx.index('Cyclist')] = 0.49
p.nms_iou_threshold[class_name_to_idx.index('Pedestrian')] = 0.32
p.nms_score_threshold = [1.0] * num_classes
p.nms_score_threshold[class_name_to_idx.index('Cyclist')] = 0.11
p.nms_score_threshold[class_name_to_idx.index('Pedestrian')] = 0.23
p.output_decoder.filter_predictions_outside_frustum = True
p.output_decoder.truncation_threshold = 0.65
# Equally weight pedestrian and cyclist moderate classes.
p.output_decoder.ap_metric.metric_weights = {
'easy': np.array([0.0, 0.0, 0.0]),
'moderate': np.array([0.0, 1.0, 1.0]),
'hard': np.array([0.0, 0.0, 0.0])
}
return p
def Task(self):
p = super(StarNetCarModel0701, self).Task()
# Builder configuration.
builder = starnet.Builder()
builder.linear_params_init = py_utils.WeightInit.KaimingUniformFanInRelu(
)
gin_layer_sizes = [32, 256, 512, 256, 256, 128]
num_laser_features = 1
gin_layers = [
# Each layer should expect as input - 2 * dims of the last layer's
# output. We assume a middle layer that's the size of 2 * dim_out.
[dim_in * 2, dim_out * 2, dim_out]
for (dim_in,
dim_out) in zip(gin_layer_sizes[:-1], gin_layer_sizes[1:])
]
p.cell_feature_dims = sum(gin_layer_sizes)
p.cell_featurizer = builder.GINFeaturizerV2(
name='feat',
fc_dims=gin_layer_sizes[0],
mlp_dims=gin_layers,
num_laser_features=num_laser_features,
fc_use_bn=False)
p.anchor_projected_feature_dims = 512
# Loss and training params
p.train.learning_rate = 0.001 / 2. # Divide by batch size.
p.focal_loss_alpha = 0.2
p.focal_loss_gamma = 3.0
class_name_to_idx = kitti_input_generator.KITTILabelExtractor.KITTI_CLASS_NAMES
num_classes = len(class_name_to_idx)
p.per_class_loss_weight = [0.] * num_classes
p.per_class_loss_weight[class_name_to_idx.index('Car')] = 1.
# Decoding / NMS params.
p.use_oriented_per_class_nms = True
p.max_nms_boxes = 512
p.nms_iou_threshold = [0.0] * num_classes
p.nms_iou_threshold[class_name_to_idx.index('Car')] = 0.0831011
p.nms_score_threshold = [1.0] * num_classes
p.nms_score_threshold[class_name_to_idx.index('Car')] = 0.321310
p.output_decoder.truncation_threshold = 0.65
p.output_decoder.filter_predictions_outside_frustum = True
return p
def Task(self):
metadata = waymo_metadata.WaymoMetadata()
num_classes = len(metadata.ClassNames())
p = starnet.ModelV2.Params(
num_classes,
num_anchor_bboxes_offsets=self.NUM_ANCHOR_BBOX_OFFSETS,
num_anchor_bboxes_rotations=self.NUM_ANCHOR_BBOX_ROTATIONS,
num_anchor_bboxes_dimensions=self.NUM_ANCHOR_BBOX_DIMENSIONS,
num_laser_features=3)
# Update the Point Cloud Featurizer architecture
starnet_builder = starnet.Builder()
starnet_builder.linear_params_init = (
py_utils.WeightInit.KaimingUniformFanInRelu())
gin_layers = [[
self.GIN_HIDDEN_DIMS * 2, self.GIN_HIDDEN_DIMS * 4,
self.GIN_HIDDEN_DIMS
]] * self.NUM_GIN_LAYERS # pyformat: disable
p.cell_featurizer = starnet_builder.GINFeaturizerV2(
'feat',
num_laser_features=3,
fc_dims=self.GIN_HIDDEN_DIMS,
mlp_dims=gin_layers,
fc_use_bn=False)
p.cell_feature_dims = self.GIN_HIDDEN_DIMS * (self.NUM_GIN_LAYERS + 1)
p.output_decoder = waymo_decoder.WaymoOpenDatasetDecoder.Params()
p.max_nms_boxes = 512
p.use_oriented_per_class_nms = True
# Note: Sub-classes need to set nms_iou_threshold and nms_score_threshold
# appropriately.
p.nms_iou_threshold = [0.0] * num_classes
# TODO(jngiam): 1.1 for untrained classes is needed to avoid an issue
# with boxutils error.
p.nms_score_threshold = [1.1] * num_classes
p.name = 'starnet'
tp = p.train
tp.optimizer = optimizer.Adam.Params()
tp.clip_gradient_norm_to_value = 5
ep = p.eval
# Train set uses a smaller decoding set, so we can
# safely eval over the entire input.
ep.samples_per_summary = 0
# To be tuned.
p.train.l2_regularizer_weight = 1e-8
cluster = cluster_factory.Current()
train_cluster_p = cluster.params.Copy()
train_cluster_p.job = 'trainer_client'
train_cluster_p.mode = 'sync'
# When running a decoding only job, there are no trainer workers, so we set
# worker replicas to 1 as a dummy value.
if train_cluster_p.worker.replicas <= 0:
train_cluster_p.worker.replicas = 1
# Set learning rate and schedule.
with cluster_factory.Cluster(train_cluster_p):
train_input_p = self.Train()
# Adapted from V1 tuning.
tp.ema_decay = 0.99
# TODO(b/148537111): consider setting this to True.
tp.ema_decay_moving_vars = False
tp.learning_rate = 0.001
lr_util.SetExponentialLR(train_p=tp,
train_input_p=train_input_p,
exp_start_epoch=5,
total_epoch=75)
p.dimension_loss_weight = .3
p.location_loss_weight = 3.
p.loss_weight_classification = 1.
p.loss_weight_localization = 3.
p.rotation_loss_weight = 0.3
return p
