问题
System information Laptop: Linux Ubuntu Tensorflow 1.15.0
Raspi: Raspberry Pi 4 tflite-runtime 2.5.0 tensorflow-estimator 1.14.0 Coral Edge TPU
Hello everybody, I am stuck at getting my trained model running on raspi. I trained ssd_mobilenet_v2_coco model from tensorflow 1 modelzoo with my own custom dataset on google cloud with this config file where I did few changes:
model {
ssd {
num_classes: 3
image_resizer {
fixed_shape_resizer {
height: 720
width: 1280
}
}
feature_extractor {
type: "ssd_mobilenet_v2"
depth_multiplier: 1.0
min_depth: 16
conv_hyperparams {
regularizer {
l2_regularizer {
weight: 3.99999989895e-05
}
}
initializer {
truncated_normal_initializer {
mean: 0.0
stddev: 0.0299999993294
}
}
activation: RELU_6
batch_norm {
decay: 0.999700009823
center: true
scale: true
epsilon: 0.0010000000475
train: true
}
}
#batch_norm_trainable: true
use_depthwise: true
}
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 {
}
}
box_predictor {
convolutional_box_predictor {
conv_hyperparams {
regularizer {
l2_regularizer {
weight: 3.99999989895e-05
}
}
initializer {
truncated_normal_initializer {
mean: 0.0
stddev: 0.0299999993294
}
}
activation: RELU_6
batch_norm {
decay: 0.999700009823
center: true
scale: true
epsilon: 0.0010000000475
train: true
}
}
min_depth: 0
max_depth: 0
num_layers_before_predictor: 0
use_dropout: false
dropout_keep_probability: 0.800000011921
kernel_size: 3
box_code_size: 4
apply_sigmoid_to_scores: false
}
}
anchor_generator {
ssd_anchor_generator {
num_layers: 6
min_scale: 0.20000000298
max_scale: 0.949999988079
aspect_ratios: 1.0
aspect_ratios: 2.0
aspect_ratios: 0.5
aspect_ratios: 3.0
aspect_ratios: 0.333299994469
}
}
post_processing {
batch_non_max_suppression {
score_threshold: 0.300000011921
iou_threshold: 0.600000023842
max_detections_per_class: 100
max_total_detections: 100
}
score_converter: SIGMOID
}
normalize_loss_by_num_matches: true
loss {
localization_loss {
weighted_smooth_l1 {
}
}
classification_loss {
weighted_sigmoid_focal {
alpha: 0.25
gamma: 2.0
}
}
classification_weight: 1.0
localization_weight: 1.0
}
}
}
train_config {
batch_size: 4
data_augmentation_options {
random_horizontal_flip {
}
}
data_augmentation_options {
ssd_random_crop {
}
}
optimizer {
rms_prop_optimizer {
learning_rate {
exponential_decay_learning_rate {
initial_learning_rate: 0.00400000018999
decay_steps: 800720
decay_factor: 0.949999988079
}
}
momentum_optimizer_value: 0.899999976158
decay: 0.899999976158
epsilon: 1.0
}
}
fine_tune_checkpoint: "gs://objects1119/model.ckpt"
num_steps: 200000
fine_tune_checkpoint_type: "detection"
}
train_input_reader {
label_map_path: "gs://objects1119/labelmap.pbtxt"
tf_record_input_reader {
input_path: "gs://objects1119/train.record"
}
shuffle: true
}
eval_config {
num_examples: 150
max_evals: 10
use_moving_averages: false
}
eval_input_reader {
label_map_path: "gs://objects1119/labelmap.pbtxt"
shuffle: true
num_readers: 1
tf_record_input_reader {
input_path: "gs://objects1119/test.record"
}
}
After training I exported the frozen_inference_graph which worked pretty well on my laptop and prediction was good, so I exported the tflite graph with the following command:
python export_tflite_ssd_graph.py \ --pipeline_config_path=gs://objects1119/ssd_mobilenet_v2_coco.config \ --trained_checkpoint_prefix=gs://objects1119/model.ckpt-189879 \ --output_directory=tflite \ --add_postprocessing_op=true
after that I converted it to detect.tflite file with:
tflite_convert --graph_def_file=tflite/tflite_graph.pb --output_file=tflite/detect.tflite --output_format=TFLITE --input_arrays=normalized_input_image_tensor --output_arrays='TFLite_Detection_PostProcess','TFLite_Detection_PostProcess:1','TFLite_Detection_PostProcess:2','TFLite_Detection_PostProcess:3' --inference_type=QUANTIZED_UINT8 --mean_values=128 --std_dev_values=127 --change_concat_input_ranges=false --allow_custom_ops
After that I converted it to edgetpu-file with this google colab: https://colab.research.google.com/drive/1o6cNNNgGhoT7_DR4jhpMKpq3mZZ6Of4N?usp=sharing while using the EdjeElectronics Tutorial: https://github.com/EdjeElectronics/TensorFlow-Lite-Object-Detection-on-Android-and-Raspberry-Pi/blob/master/Raspberry_Pi_Guide.md everything worked without errors.
But when I finally try to run it on the raspi I am always facing the same output, no matter what input. There is always Class Zero predicted on the same areas of the image with prediction of 50%. What am I doing wrong? Is it because of the changed input shapes in the config file for training? In my case it would be really complicated to train again so I try to avoid this if its possible. Here is also the link to the Isuue in github https://github.com/tensorflow/tensorflow/issues/45148
Thank you very much in advance.
来源：https://stackoverflow.com/questions/65108043/tensoflow-trained-ssd-model-not-working-after-converting-to-tensorflow-lite-for