I am training a convolutional neural network to detect objects (weeds amongst crops, in my case) using TensorFlow. The original dimensions of the raw training photos are 4000 x 3000 pixels, which must be resized to become workable. The idea here is to label objects in the training images (using Label-Img), train the model, and use it to detect weeds in certain situations.
According to TensorFlow 2 Detection Model Zoo, there are algorithms designed for different speeds, which involves initially resizing the images to a specified dimension. Although this is not a coding question, here is an example of SSD ResNet-50, which initially resizes the input images to 1024 x 1024 pixels:
model {
ssd {
num_classes: 1
image_resizer {
fixed_shape_resizer {
height: 1024
width: 1024
}
}
feature_extractor {
type: "ssd_resnet50_v1_fpn_keras"
depth_multiplier: 1.0
min_depth: 16
conv_hyperparams {
regularizer {
l2_regularizer {
weight: 0.00039999998989515007
}
}
initializer {
truncated_normal_initializer {
mean: 0.0
stddev: 0.029999999329447746
}
}
activation: RELU_6
batch_norm {
decay: 0.996999979019165
scale: true
epsilon: 0.0010000000474974513
}
}
override_base_feature_extractor_hyperparams: true
fpn {
min_level: 3
max_level: 7
}
}
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 {
}
}
box_predictor {
weight_shared_convolutional_box_predictor {
conv_hyperparams {
regularizer {
l2_regularizer {
weight: 0.00039999998989515007
}
}
initializer {
random_normal_initializer {
mean: 0.0
stddev: 0.009999999776482582
}
}
activation: RELU_6
batch_norm {
decay: 0.996999979019165
scale: true
epsilon: 0.0010000000474974513
}
}
depth: 256
num_layers_before_predictor: 4
kernel_size: 3
class_prediction_bias_init: -4.599999904632568
}
}
anchor_generator {
multiscale_anchor_generator {
min_level: 3
max_level: 7
anchor_scale: 4.0
aspect_ratios: 1.0
aspect_ratios: 2.0
aspect_ratios: 0.5
scales_per_octave: 2
}
}
post_processing {
batch_non_max_suppression {
score_threshold: 9.99999993922529e-09
iou_threshold: 0.6000000238418579
max_detections_per_class: 100
max_total_detections: 100
use_static_shapes: false
}
score_converter: SIGMOID
}
normalize_loss_by_num_matches: true
loss {
localization_loss {
weighted_smooth_l1 {
}
}
classification_loss {
weighted_sigmoid_focal {
gamma: 2.0
alpha: 0.25
}
}
classification_weight: 1.0
localization_weight: 1.0
}
encode_background_as_zeros: true
normalize_loc_loss_by_codesize: true
inplace_batchnorm_update: true
freeze_batchnorm: false
}
}
train_config {
batch_size: 64
data_augmentation_options {
random_horizontal_flip {
}
}
data_augmentation_options {
random_crop_image {
min_object_covered: 0.0
min_aspect_ratio: 0.75
max_aspect_ratio: 3.0
min_area: 0.75
max_area: 1.0
overlap_thresh: 0.0
}
}
sync_replicas: true
optimizer {
momentum_optimizer {
learning_rate {
cosine_decay_learning_rate {
learning_rate_base: 0.03999999910593033
total_steps: 100000
warmup_learning_rate: 0.013333000242710114
warmup_steps: 2000
}
}
momentum_optimizer_value: 0.8999999761581421
}
use_moving_average: false
}
fine_tune_checkpoint: "PATH_TO_BE_CONFIGURED"
num_steps: 100000
startup_delay_steps: 0.0
replicas_to_aggregate: 8
max_number_of_boxes: 100
unpad_groundtruth_tensors: false
fine_tune_checkpoint_type: "classification"
use_bfloat16: true
fine_tune_checkpoint_version: V2
}
train_input_reader {
label_map_path: "PATH_TO_BE_CONFIGURED"
tf_record_input_reader {
input_path: "PATH_TO_BE_CONFIGURED"
}
}
eval_config {
metrics_set: "coco_detection_metrics"
use_moving_averages: false
}
eval_input_reader {
label_map_path: "PATH_TO_BE_CONFIGURED"
shuffle: false
num_epochs: 1
tf_record_input_reader {
input_path: "PATH_TO_BE_CONFIGURED"
}
}
Because I will be labeling many pictures in the future, I need to decide on a dimension to resize my original ones to (literature review says 1100 x 1100 has been used in previous projects).
If I were to change the image resizer in the code above to 1100 x 1100, for example, would that have any effect on model accuracy/training loss? Would it even run? I'm fairly new to this, so any insights on this would be greatly appreciated!
Note: I am using a NVIDIA GPU, so that helps speed the process quite a bit. Google Colab also can be used.
