Re-Training EfficientDet drops accuracy to 0 - conv-neural-network

Question

Recently, I am trying to fine-tune the EfficientDet-D4 model available in TensorFlow model Zoo repository
(http://download.tensorflow.org/models/object_detection/tf2/20200711/efficientdet_d4_coco17_tpu-32.tar.gz)
To Fine-Tuning this model, I have followed the steps established in the following guide (https://tensorflow-object-detection-api-tutorial.readthedocs.io/en/latest/training.html). The dataset I want to use is the Visdrone dataset (http://aiskyeye.com/). When the model goes through the 2000 step, the accuracy decays to 0 and does not detect anything.
Below, I attach the configuration file used for fine-tuning:
# SSD with EfficientNet-b4 + BiFPN feature extractor,
# shared box predictor and focal loss (a.k.a EfficientDet-d4).
# See EfficientDet, Tan et al, https://arxiv.org/abs/1911.09070
# See Lin et al, https://arxiv.org/abs/1708.02002
# Trained on COCO, initialized from an EfficientNet-b4 checkpoint.
#
# Train on TPU-32
model {
ssd {
inplace_batchnorm_update: true
freeze_batchnorm: false
num_classes: 90
add_background_class: false
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
use_matmul_gather: true
}
}
similarity_calculator {
iou_similarity {
}
}
encode_background_as_zeros: true
anchor_generator {
multiscale_anchor_generator {
min_level: 3
max_level: 7
anchor_scale: 4.0
aspect_ratios: [1.0, 2.0, 0.5]
scales_per_octave: 3
}
}
image_resizer {
keep_aspect_ratio_resizer {
min_dimension: 1024
max_dimension: 1024
pad_to_max_dimension: true
}
}
box_predictor {
weight_shared_convolutional_box_predictor {
depth: 224
class_prediction_bias_init: -4.6
conv_hyperparams {
force_use_bias: true
activation: SWISH
regularizer {
l2_regularizer {
weight: 0.00004
}
}
initializer {
random_normal_initializer {
stddev: 0.01
mean: 0.0
}
}
batch_norm {
scale: true
decay: 0.99
epsilon: 0.001
}
}
num_layers_before_predictor: 4
kernel_size: 3
use_depthwise: true
}
}
feature_extractor {
type: 'ssd_efficientnet-b4_bifpn_keras'
bifpn {
min_level: 3
max_level: 7
num_iterations: 7
num_filters: 224
}
conv_hyperparams {
force_use_bias: true
activation: SWISH
regularizer {
l2_regularizer {
weight: 0.00004
}
}
initializer {
truncated_normal_initializer {
stddev: 0.03
mean: 0.0
}
}
batch_norm {
scale: true,
decay: 0.99,
epsilon: 0.001,
}
}
}
loss {
classification_loss {
weighted_sigmoid_focal {
alpha: 0.25
gamma: 1.5
}
}
localization_loss {
weighted_smooth_l1 {
}
}
classification_weight: 1.0
localization_weight: 1.0
}
normalize_loss_by_num_matches: true
normalize_loc_loss_by_codesize: true
post_processing {
batch_non_max_suppression {
score_threshold: 1e-8
iou_threshold: 0.5
max_detections_per_class: 100
max_total_detections: 100
}
score_converter: SIGMOID
}
}
}
train_config: {
fine_tune_checkpoint: "/home/models/efd4/checkpoint/ckpt-0"
fine_tune_checkpoint_version: V2
fine_tune_checkpoint_type: "detection"
batch_size: 1
sync_replicas: true
startup_delay_steps: 0
replicas_to_aggregate: 8
use_bfloat16: true
num_steps: 150000
data_augmentation_options {
random_horizontal_flip {
}
}
data_augmentation_options {
random_scale_crop_and_pad_to_square {
output_size: 1024
scale_min: 0.1
scale_max: 2.0
}
}
optimizer {
momentum_optimizer: {
learning_rate: {
cosine_decay_learning_rate {
learning_rate_base: 0.001
total_steps: 150000
warmup_learning_rate: .0001
warmup_steps: 500
}
}
momentum_optimizer_value: 0.9
}
use_moving_average: false
}
max_number_of_boxes: 100
unpad_groundtruth_tensors: false
}
train_input_reader: {
label_map_path: "/home/labels/label_map.txt"
tf_record_input_reader {
input_path: "/home/records/train.tfrecord"
}
}
eval_config: {
metrics_set: "coco_detection_metrics"
use_moving_averages: false
batch_size: 1;
}
eval_input_reader: {
label_map_path: "/home/labels/label_map.txt"
shuffle: false
num_epochs: 1
tf_record_input_reader {
input_path: "/home/records/validation.tfrecord"
}
}
I try modify parameters like learning-rate, etc but doesnt work