Tensorflow 对象检测 API 分类损失增加
Tensorflow object detection API classification loss increases
我正在使用我自己的数据训练 tensorflow 对象检测 API。我使用的模型是 ssd_mobilenet_v1，带有预训练的 coco 检查点。
我的数据集由 12 个 class 组成，每个 class 有 110 张图像，所以总共有 1320 张图像。
这工作正常，但 class化损失在某些时候会增加。
我认为数据集不足在训练阶段并不重要，因为它们都很相似；实际上我是从视频中提取的。
那该怎么办呢？我应该停止大约 10k 次迭代的训练吗？或者是否有任何可能的参数调整或数据扩充？
这是我刚刚调整目录和 data_augmentation_option 和 hard_example_miner 的配置文件。
model {
ssd {
num_classes: 12
box_coder {
faster_rcnn_box_coder {
y_scale: 10.0
x_scale: 10.0
height_scale: 5.0
width_scale: 5.0
}
}
matcher {
argmax_matcher {
matched_threshold: 0.5
unmatched_threshold: 0.5
ignore_thresholds: false
negatives_lower_than_unmatched: true
force_match_for_each_row: true
}
}
similarity_calculator {
iou_similarity {
}
}
anchor_generator {
ssd_anchor_generator {
num_layers: 6
min_scale: 0.2
max_scale: 0.95
aspect_ratios: 1.0
aspect_ratios: 2.0
aspect_ratios: 0.5
aspect_ratios: 3.0
aspect_ratios: 0.3333
}
}
image_resizer {
fixed_shape_resizer {
height: 320
width: 180
}
}
box_predictor {
convolutional_box_predictor {
min_depth: 0
max_depth: 0
num_layers_before_predictor: 0
use_dropout: false
dropout_keep_probability: 0.8
kernel_size: 1
box_code_size: 4
apply_sigmoid_to_scores: false
conv_hyperparams {
activation: RELU_6,
regularizer {
l2_regularizer {
weight: 0.00004
}
}
initializer {
truncated_normal_initializer {
stddev: 0.03
mean: 0.0
}
}
batch_norm {
train: true,
scale: true,
center: true,
decay: 0.9997,
epsilon: 0.001,
}
}
}
}
feature_extractor {
type: 'ssd_mobilenet_v1'
min_depth: 16
depth_multiplier: 1.0
conv_hyperparams {
activation: RELU_6,
regularizer {
l2_regularizer {
weight: 0.00004
}
}
initializer {
truncated_normal_initializer {
stddev: 0.03
mean: 0.0
}
}
batch_norm {
train: true,
scale: true,
center: true,
decay: 0.9997,
epsilon: 0.001,
}
}
}
loss {
classification_loss {
weighted_sigmoid {
}
}
localization_loss {
weighted_smooth_l1 {
}
}
hard_example_miner {
num_hard_examples: 600
iou_threshold: 0.99
loss_type: CLASSIFICATION
max_negatives_per_positive: 3
min_negatives_per_image: 0
}
classification_weight: 1.0
localization_weight: 1.0
}
normalize_loss_by_num_matches: true
post_processing {
batch_non_max_suppression {
score_threshold: 1e-8
iou_threshold: 0.6
max_detections_per_class: 100
max_total_detections: 100
}
score_converter: SIGMOID
}
}
}
train_config: {
batch_size: 96
optimizer {
rms_prop_optimizer: {
learning_rate: {
exponential_decay_learning_rate {
initial_learning_rate: 0.004
decay_steps: 800720
decay_factor: 0.95
}
}
momentum_optimizer_value: 0.9
decay: 0.9
epsilon: 1.0
}
}
fine_tune_checkpoint: "/home/dev1/tensorflow/training/data/checkpoint/ssd_mobilenet_v1_coco_2018_01_28/model.ckpt"
fine_tune_checkpoint_type: "detection"
from_detection_checkpoint: true
num_steps: 100000
data_augmentation_options {
random_horizontal_flip {
}
}
data_augmentation_options {
ssd_random_crop {
}
}
data_augmentation_options {
random_adjust_brightness {
}
}
}
train_input_reader: {
tf_record_input_reader {
input_path: "/home/dev1/tensorflow/training/data/train.record"
}
label_map_path: "/home/dev1/tensorflow/training/data/config/label.pbtxt"
}
eval_config: {
num_examples: 132
max_evals: 20
}
eval_input_reader: {
tf_record_input_reader {
input_path: "/home/dev1/tensorflow/training/data/test.record"
}
label_map_path: "/home/dev1/tensorflow/training/data/config/label.pbtxt"
shuffle: false
num_readers: 1
}
为了社区的利益，尽管它出现在评论部分（感谢 Shayan Tabatabaee），但在此处（回答部分）提供解决方案。
此问题是由于高学习率引起的，已通过将 decay_steps 从 800720 降低到 5000 来解决。
请参考下面修改后的配置文件
model {
ssd {
num_classes: 12
box_coder {
faster_rcnn_box_coder {
y_scale: 10.0
x_scale: 10.0
height_scale: 5.0
width_scale: 5.0
}
}
matcher {
argmax_matcher {
matched_threshold: 0.5
unmatched_threshold: 0.5
ignore_thresholds: false
negatives_lower_than_unmatched: true
force_match_for_each_row: true
}
}
similarity_calculator {
iou_similarity {
}
}
anchor_generator {
ssd_anchor_generator {
num_layers: 6
min_scale: 0.2
max_scale: 0.95
aspect_ratios: 1.0
aspect_ratios: 2.0
aspect_ratios: 0.5
aspect_ratios: 3.0
aspect_ratios: 0.3333
}
}
image_resizer {
fixed_shape_resizer {
height: 320
width: 180
}
}
box_predictor {
convolutional_box_predictor {
min_depth: 0
max_depth: 0
num_layers_before_predictor: 0
use_dropout: false
dropout_keep_probability: 0.8
kernel_size: 1
box_code_size: 4
apply_sigmoid_to_scores: false
conv_hyperparams {
activation: RELU_6,
regularizer {
l2_regularizer {
weight: 0.00004
}
}
initializer {
truncated_normal_initializer {
stddev: 0.03
mean: 0.0
}
}
batch_norm {
train: true,
scale: true,
center: true,
decay: 0.9997,
epsilon: 0.001,
}
}
}
}
feature_extractor {
type: 'ssd_mobilenet_v1'
min_depth: 16
depth_multiplier: 1.0
conv_hyperparams {
activation: RELU_6,
regularizer {
l2_regularizer {
weight: 0.00004
}
}
initializer {
truncated_normal_initializer {
stddev: 0.03
mean: 0.0
}
}
batch_norm {
train: true,
scale: true,
center: true,
decay: 0.9997,
epsilon: 0.001,
}
}
}
loss {
classification_loss {
weighted_sigmoid {
}
}
localization_loss {
weighted_smooth_l1 {
}
}
hard_example_miner {
num_hard_examples: 600
iou_threshold: 0.99
loss_type: CLASSIFICATION
max_negatives_per_positive: 3
min_negatives_per_image: 0
}
classification_weight: 1.0
localization_weight: 1.0
}
normalize_loss_by_num_matches: true
post_processing {
batch_non_max_suppression {
score_threshold: 1e-8
iou_threshold: 0.6
max_detections_per_class: 100
max_total_detections: 100
}
score_converter: SIGMOID
}
}
}
train_config: {
batch_size: 96
optimizer {
rms_prop_optimizer: {
learning_rate: {
exponential_decay_learning_rate {
initial_learning_rate: 0.004
decay_steps: 5000
decay_factor: 0.95
}
}
momentum_optimizer_value: 0.9
decay: 0.9
epsilon: 1.0
}
}
fine_tune_checkpoint: "/home/dev1/tensorflow/training/data/checkpoint/ssd_mobilenet_v1_coco_2018_01_28/model.ckpt"
fine_tune_checkpoint_type: "detection"
from_detection_checkpoint: true
num_steps: 100000
data_augmentation_options {
random_horizontal_flip {
}
}
data_augmentation_options {
ssd_random_crop {
}
}
data_augmentation_options {
random_adjust_brightness {
}
}
}
train_input_reader: {
tf_record_input_reader {
input_path: "/home/dev1/tensorflow/training/data/train.record"
}
label_map_path: "/home/dev1/tensorflow/training/data/config/label.pbtxt"
}
eval_config: {
num_examples: 132
max_evals: 20
}
eval_input_reader: {
tf_record_input_reader {
input_path: "/home/dev1/tensorflow/training/data/test.record"
}
label_map_path: "/home/dev1/tensorflow/training/data/config/label.pbtxt"
shuffle: false
num_readers: 1
}
我正在使用我自己的数据训练 tensorflow 对象检测 API。我使用的模型是 ssd_mobilenet_v1，带有预训练的 coco 检查点。
我的数据集由 12 个 class 组成，每个 class 有 110 张图像，所以总共有 1320 张图像。
这工作正常，但 class化损失在某些时候会增加。
我认为数据集不足在训练阶段并不重要，因为它们都很相似；实际上我是从视频中提取的。
那该怎么办呢？我应该停止大约 10k 次迭代的训练吗？或者是否有任何可能的参数调整或数据扩充？
这是我刚刚调整目录和 data_augmentation_option 和 hard_example_miner 的配置文件。
model {
ssd {
num_classes: 12
box_coder {
faster_rcnn_box_coder {
y_scale: 10.0
x_scale: 10.0
height_scale: 5.0
width_scale: 5.0
}
}
matcher {
argmax_matcher {
matched_threshold: 0.5
unmatched_threshold: 0.5
ignore_thresholds: false
negatives_lower_than_unmatched: true
force_match_for_each_row: true
}
}
similarity_calculator {
iou_similarity {
}
}
anchor_generator {
ssd_anchor_generator {
num_layers: 6
min_scale: 0.2
max_scale: 0.95
aspect_ratios: 1.0
aspect_ratios: 2.0
aspect_ratios: 0.5
aspect_ratios: 3.0
aspect_ratios: 0.3333
}
}
image_resizer {
fixed_shape_resizer {
height: 320
width: 180
}
}
box_predictor {
convolutional_box_predictor {
min_depth: 0
max_depth: 0
num_layers_before_predictor: 0
use_dropout: false
dropout_keep_probability: 0.8
kernel_size: 1
box_code_size: 4
apply_sigmoid_to_scores: false
conv_hyperparams {
activation: RELU_6,
regularizer {
l2_regularizer {
weight: 0.00004
}
}
initializer {
truncated_normal_initializer {
stddev: 0.03
mean: 0.0
}
}
batch_norm {
train: true,
scale: true,
center: true,
decay: 0.9997,
epsilon: 0.001,
}
}
}
}
feature_extractor {
type: 'ssd_mobilenet_v1'
min_depth: 16
depth_multiplier: 1.0
conv_hyperparams {
activation: RELU_6,
regularizer {
l2_regularizer {
weight: 0.00004
}
}
initializer {
truncated_normal_initializer {
stddev: 0.03
mean: 0.0
}
}
batch_norm {
train: true,
scale: true,
center: true,
decay: 0.9997,
epsilon: 0.001,
}
}
}
loss {
classification_loss {
weighted_sigmoid {
}
}
localization_loss {
weighted_smooth_l1 {
}
}
hard_example_miner {
num_hard_examples: 600
iou_threshold: 0.99
loss_type: CLASSIFICATION
max_negatives_per_positive: 3
min_negatives_per_image: 0
}
classification_weight: 1.0
localization_weight: 1.0
}
normalize_loss_by_num_matches: true
post_processing {
batch_non_max_suppression {
score_threshold: 1e-8
iou_threshold: 0.6
max_detections_per_class: 100
max_total_detections: 100
}
score_converter: SIGMOID
}
}
}
train_config: {
batch_size: 96
optimizer {
rms_prop_optimizer: {
learning_rate: {
exponential_decay_learning_rate {
initial_learning_rate: 0.004
decay_steps: 800720
decay_factor: 0.95
}
}
momentum_optimizer_value: 0.9
decay: 0.9
epsilon: 1.0
}
}
fine_tune_checkpoint: "/home/dev1/tensorflow/training/data/checkpoint/ssd_mobilenet_v1_coco_2018_01_28/model.ckpt"
fine_tune_checkpoint_type: "detection"
from_detection_checkpoint: true
num_steps: 100000
data_augmentation_options {
random_horizontal_flip {
}
}
data_augmentation_options {
ssd_random_crop {
}
}
data_augmentation_options {
random_adjust_brightness {
}
}
}
train_input_reader: {
tf_record_input_reader {
input_path: "/home/dev1/tensorflow/training/data/train.record"
}
label_map_path: "/home/dev1/tensorflow/training/data/config/label.pbtxt"
}
eval_config: {
num_examples: 132
max_evals: 20
}
eval_input_reader: {
tf_record_input_reader {
input_path: "/home/dev1/tensorflow/training/data/test.record"
}
label_map_path: "/home/dev1/tensorflow/training/data/config/label.pbtxt"
shuffle: false
num_readers: 1
}
为了社区的利益，尽管它出现在评论部分（感谢 Shayan Tabatabaee），但在此处（回答部分）提供解决方案。
此问题是由于高学习率引起的，已通过将 decay_steps 从 800720 降低到 5000 来解决。
请参考下面修改后的配置文件
model {
ssd {
num_classes: 12
box_coder {
faster_rcnn_box_coder {
y_scale: 10.0
x_scale: 10.0
height_scale: 5.0
width_scale: 5.0
}
}
matcher {
argmax_matcher {
matched_threshold: 0.5
unmatched_threshold: 0.5
ignore_thresholds: false
negatives_lower_than_unmatched: true
force_match_for_each_row: true
}
}
similarity_calculator {
iou_similarity {
}
}
anchor_generator {
ssd_anchor_generator {
num_layers: 6
min_scale: 0.2
max_scale: 0.95
aspect_ratios: 1.0
aspect_ratios: 2.0
aspect_ratios: 0.5
aspect_ratios: 3.0
aspect_ratios: 0.3333
}
}
image_resizer {
fixed_shape_resizer {
height: 320
width: 180
}
}
box_predictor {
convolutional_box_predictor {
min_depth: 0
max_depth: 0
num_layers_before_predictor: 0
use_dropout: false
dropout_keep_probability: 0.8
kernel_size: 1
box_code_size: 4
apply_sigmoid_to_scores: false
conv_hyperparams {
activation: RELU_6,
regularizer {
l2_regularizer {
weight: 0.00004
}
}
initializer {
truncated_normal_initializer {
stddev: 0.03
mean: 0.0
}
}
batch_norm {
train: true,
scale: true,
center: true,
decay: 0.9997,
epsilon: 0.001,
}
}
}
}
feature_extractor {
type: 'ssd_mobilenet_v1'
min_depth: 16
depth_multiplier: 1.0
conv_hyperparams {
activation: RELU_6,
regularizer {
l2_regularizer {
weight: 0.00004
}
}
initializer {
truncated_normal_initializer {
stddev: 0.03
mean: 0.0
}
}
batch_norm {
train: true,
scale: true,
center: true,
decay: 0.9997,
epsilon: 0.001,
}
}
}
loss {
classification_loss {
weighted_sigmoid {
}
}
localization_loss {
weighted_smooth_l1 {
}
}
hard_example_miner {
num_hard_examples: 600
iou_threshold: 0.99
loss_type: CLASSIFICATION
max_negatives_per_positive: 3
min_negatives_per_image: 0
}
classification_weight: 1.0
localization_weight: 1.0
}
normalize_loss_by_num_matches: true
post_processing {
batch_non_max_suppression {
score_threshold: 1e-8
iou_threshold: 0.6
max_detections_per_class: 100
max_total_detections: 100
}
score_converter: SIGMOID
}
}
}
train_config: {
batch_size: 96
optimizer {
rms_prop_optimizer: {
learning_rate: {
exponential_decay_learning_rate {
initial_learning_rate: 0.004
decay_steps: 5000
decay_factor: 0.95
}
}
momentum_optimizer_value: 0.9
decay: 0.9
epsilon: 1.0
}
}
fine_tune_checkpoint: "/home/dev1/tensorflow/training/data/checkpoint/ssd_mobilenet_v1_coco_2018_01_28/model.ckpt"
fine_tune_checkpoint_type: "detection"
from_detection_checkpoint: true
num_steps: 100000
data_augmentation_options {
random_horizontal_flip {
}
}
data_augmentation_options {
ssd_random_crop {
}
}
data_augmentation_options {
random_adjust_brightness {
}
}
}
train_input_reader: {
tf_record_input_reader {
input_path: "/home/dev1/tensorflow/training/data/train.record"
}
label_map_path: "/home/dev1/tensorflow/training/data/config/label.pbtxt"
}
eval_config: {
num_examples: 132
max_evals: 20
}
eval_input_reader: {
tf_record_input_reader {
input_path: "/home/dev1/tensorflow/training/data/test.record"
}
label_map_path: "/home/dev1/tensorflow/training/data/config/label.pbtxt"
shuffle: false
num_readers: 1
}