We have a custom shoe dataset with 3 classes: air_max, air_force_1, and huaraches. It is a small dataset with 152 images and labels are in the training folder, and 41 images (20%) in the testing folder. When we run TLT training on the sample YOLO notebook, all of our AP values result in 0.0.
This is our config for generating the TFRecords:
kitti_config {
root_directory_path: "/workspace/shoe_experiment/data/training"
image_dir_name: "image_2"
label_dir_name: "label_2"
image_extension: ".png"
partition_mode: "random"
num_partitions: 2
val_split: 20
num_shards: 10
}
image_directory_path: "/workspace/shoe_experiment/data/training"
And these are our training specs:
random_seed: 42
yolo_config {
big_anchor_shape: "[(114.94, 60.67), (159.06, 114.59), (297.59, 176.38)]"
mid_anchor_shape: "[(42.99, 31.91), (79.57, 31.75), (56.80, 56.93)]"
small_anchor_shape: "[(15.60, 13.88), (30.25, 20.25), (20.67, 49.63)]"
matching_neutral_box_iou: 0.5
arch: "darknet"
nlayers: 19
arch_conv_blocks: 2
loss_loc_weight: 0.75
loss_neg_obj_weights: 200.0
loss_class_weights: 1.0
freeze_blocks: 0
freeze_bn: false
}
training_config {
batch_size_per_gpu: 16
num_epochs: 80
enable_qat: false
learning_rate {
soft_start_annealing_schedule {
min_learning_rate: 1e-8
max_learning_rate: 1e-2
soft_start: 0.1
annealing: 0.8
}
}
regularizer {
type: L1
weight: 5e-5
}
}
eval_config {
validation_period_during_training: 10
average_precision_mode: SAMPLE
batch_size: 16
matching_iou_threshold: 0.5
}
nms_config {
confidence_threshold: 0.01
clustering_iou_threshold: 0.6
top_k: 200
}
augmentation_config {
preprocessing {
output_image_width: 1248
output_image_height: 384
output_image_channel: 3
crop_right: 1248
crop_bottom: 384
min_bbox_width: 1.0
min_bbox_height: 1.0
}
spatial_augmentation {
hflip_probability: 0.5
vflip_probability: 0.0
zoom_min: 0.7
zoom_max: 1.8
translate_max_x: 8.0
translate_max_y: 8.0
}
color_augmentation {
hue_rotation_max: 25.0
saturation_shift_max: 0.20000000298
contrast_scale_max: 0.10000000149
contrast_center: 0.5
}
}
dataset_config {
data_sources: {
tfrecords_path: "/workspace/shoe_experiment/data/tfrecords/kitti_trainval/kitti_trainval*"
image_directory_path: "/workspace/shoe_experiment/data/training"
}
image_extension: "png"
target_class_mapping {
key: "air_force_1"
value: "air_force_1"
}
target_class_mapping {
key: "air_max"
value: "air_max"
}
target_class_mapping {
key: "huaraches"
value: "huaraches"
}
validation_fold: 0
}
All of the class names in labels are lowercase and we tried various min/max_learning_rate values: (1e-6,1e-4), (1e-14,1e-11), and (1e-8,1e-2) to no avail.
Would you have any suggestions for how to fix the AP 0.0 issue? Do we need to make our dataset larger and if so, how many images would we need per class?
