I am training an image recognition algorithm (TesorFlow 1.15, Python 3.7.7) with 200 training images and 40 test images. Each image has dimensions of 4000 x 3000 pixels, so they are each quite large. I am training the following algorithm (SSD Mobilenet V1):
model {
ssd {
num_classes: 1
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: 300
width: 300
}
}
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: 3000
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: 3
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: "ssd_mobilenet_v1_coco_2018_01_28/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 {
ssd_random_crop {
}
}
}
train_input_reader: {
tf_record_input_reader {
input_path: "data/train.record"
}
label_map_path: "data/object-detection.pbtxt"
}
eval_config: {
num_examples: 40
# Note: The below line limits the evaluation process to 10 evaluations.
# Remove the below line to evaluate indefinitely.
max_evals: 10
}
eval_input_reader: {
tf_record_input_reader {
input_path: "data/test.record"
}
label_map_path: "training/object-detection.pbtxt"
shuffle: false
num_readers: 1
}
Modified classes include:
I do not wish for my images to be downsized, because I have a lot of labeled data that would be outside the bounds of the:
fixed_shape_resizer {
height: 300
width: 300
dimensions. Do I need to eliminate this? I'm new to TensorFlow and do not have much experience with these things, so any info would be helpful.
First of all, SSD needs to be fixed shape (same width as height). This is because of its configuration as a neural network with fully connected layers.
So you have to resize anyway. But I can see why you don't want to resize too much because of loss of information. I think in theory it's possible to train on such large images, but in practice it's not really. Training takes images in batches and loads them into memory. With sizes like those, you would probably run out of memory. Besides that, it would take way longer to train too.
So that's why resizing is used a lot. So there is two possible ways to go. Resize to a smaller size (you could try larger than 300x300 of course, but larger than 640x640 is not very realistic). But you could also split each image up into for example 4 images, where you train the model on these splits. This way, you lose less information. But it will take more time to train because of a larger dataset, which is good in some ways too. This would probably be the best way to go.