Optimizing Faster R-CNN Inception Resnet v2 for my need - tensorflow

Question

I'm using the Faster R-CNN Inception Resnet v2 model pre-trained on COCO to train my own object detector with the purpose of detecting objects from 3 classes. The objects are small compared to the size (resolution) of the image. I'm relatively new to ML and OD.
I wonder what changes I should make to the model to make it better fit my purpose. Is it a good idea to decrease the complexity of some parts of the model since I only detect 3 classes? Are there any feature extractors better suited for small objects? Is it generally best to train on a pre-trained model or should I train from scratch?
I'm aware that tuning the network to a specific need is a trial-and-error process, however, since it takes about 3 days to train the network I'm looking for some educated guesses.
Model configuration:
model {
faster_rcnn {
num_classes: 3
image_resizer {
keep_aspect_ratio_resizer {
min_dimension: 600
max_dimension: 4048
}
}
feature_extractor {
type: 'faster_rcnn_inception_resnet_v2'
first_stage_features_stride: 8
}
first_stage_anchor_generator {
# grid_anchor_generator {
# scales: [0.25, 0.5, 1.0, 2.0, 3.0]
# aspect_ratios: [0.25,0.5, 1.0, 2.0]
# height_stride: 8
# width_stride: 8
# }
grid_anchor_generator {
scales: [0.25, 0.5, 1.0, 2.0, 3.0]
aspect_ratios: [1.0, 2.0, 3.0]
height: 64
width: 64
height_stride: 8
width_stride: 8
}
}
first_stage_atrous_rate: 2
first_stage_box_predictor_conv_hyperparams {
op: CONV
regularizer {
l2_regularizer {
weight: 0.01
}
}
initializer {
truncated_normal_initializer {
stddev: 0.01
}
}
}
first_stage_nms_score_threshold: 0.0
first_stage_nms_iou_threshold: 0.4
first_stage_max_proposals: 1000
first_stage_localization_loss_weight: 2.0
first_stage_objectness_loss_weight: 1.0
initial_crop_size: 17
maxpool_kernel_size: 1
maxpool_stride: 1
second_stage_box_predictor {
mask_rcnn_box_predictor {
use_dropout: True
dropout_keep_probability: 0.9
fc_hyperparams {
op: FC
regularizer {
l2_regularizer {
weight: 0.01
}
}
initializer {
variance_scaling_initializer {
factor: 1.0
uniform: true
mode: FAN_AVG
}
}
}
}
}
second_stage_post_processing {
batch_non_max_suppression {
score_threshold: 0.0
iou_threshold: 0.5
max_detections_per_class: 20
max_total_detections: 20
}
score_converter: SOFTMAX
}
second_stage_localization_loss_weight: 2.0
second_stage_classification_loss_weight: 1.0
}
}
train_config: {
batch_size: 1
optimizer {
momentum_optimizer: {
learning_rate: {
manual_step_learning_rate {
initial_learning_rate: 0.00001
schedule {
step: 100000
learning_rate: .000001
}
schedule {
step: 150000
learning_rate: .0000001
}
}
}
momentum_optimizer_value: 0.9
}
use_moving_average: false
}
gradient_clipping_by_norm: 10.0
# PATH_TO_BE_CONFIGURED: Below line needs to match location of model checkpoint: Either use checkpoint from rcnn model, or checkpoint from previously trained model on other dataset.
fine_tune_checkpoint: "/.../model.ckpt"
from_detection_checkpoint: true
# Note: The below line limits the training process to 200K steps, which we
# empirically found to be sufficient enough to train the pets dataset. This
# effectively bypasses the learning rate schedule (the learning rate will
# never decay). Remove the below line to train indefinitely.
# num_steps: 200000
data_augmentation_options {
random_horizontal_flip {}
}
data_augmentation_options {
random_crop_image {
min_object_covered : 1.0
min_aspect_ratio: 0.5
max_aspect_ratio: 2
min_area: 0.2
max_area: 1.
}
}
data_augmentation_options {
random_distort_color {}
}
}
# PATH_TO_BE_CONFIGURED: Need to make sure folder structure below is correct for both train-record and label_map.pbtxt
train_input_reader: {
tf_record_input_reader {
input_path: "/.../train.record"
}
label_map_path: "/..../label_map.pbtxt"
queue_capacity: 500
min_after_dequeue: 250
}
#PATH_TO_BE_CONFIGURED: Make sure folder structure for eval_export, validation.record and label_map.pbtxt below are correct.
eval_config: {
num_examples: 30
# Note: The below line limits the evaluation process to 10 evaluations.
# Remove the below line to evaluate indefinitely.
max_evals: 10
num_visualizations: 30
eval_interval_secs: 600
visualization_export_dir: "/.../eval_export"
}
eval_input_reader: {
tf_record_input_reader {
input_path: "/.../test.record"
}
label_map_path: "/.../label_map.pbtxt"
shuffle: True
num_readers: 1
}