05. Deep dive into SSD training: 3 tips to boost performance¶
In the previous tutorial 04. Train SSD on Pascal VOC dataset, we briefly went through the basic APIs that help building the training pipeline of SSD.
In this article, we will dive deep into the details and introduce tricks that important for reproducing state-of-the-art performance. These are the hidden pitfalls that are usually missing in papers and tech reports.
The training objective mentioned in paper is a weighted summation of localization loss(loc) and the confidence loss(conf).
But the question is, what is the proper way to calculate
N? Should we sum up
N across the entire batch, or use per-sample
To illustrate this, please generate some dummy data:
import mxnet as mx x = mx.random.uniform(shape=(2, 3, 300, 300)) # use batch-size 2 # suppose image 1 has single object id1 = mx.nd.array() bbox1 = mx.nd.array([[10, 20, 80, 90]]) # xmin, ymin, xmax, ymax # suppose image 2 has 4 objects id2 = mx.nd.array([1, 3, 5, 7]) bbox2 = mx.nd.array([[10, 10, 30, 30], [40, 40, 60, 60], [50, 50, 90, 90], [100, 110, 120, 140]])
Then, combine them into a batch by padding -1 as sentinal values:
class_ids: [[ 1. -1. -1. -1.] [ 1. 3. 5. 7.]] <NDArray 2x4 @cpu(0)>
bounding boxes: [[[ 10. 20. 80. 90.] [ -1. -1. -1. -1.] [ -1. -1. -1. -1.] [ -1. -1. -1. -1.]] [[ 10. 10. 30. 30.] [ 40. 40. 60. 60.] [ 50. 50. 90. 90.] [100. 110. 120. 140.]]] <NDArray 2x4x4 @cpu(0)>
We use a vgg16 atrous 300x300 SSD model in this example. For demo purpose, we don’t use any pretrained weights here
Some preparation before training
Simulate the training steps by manually compute losses:
You can always use
gluoncv.loss.SSDMultiBoxLoss which fulfills this function.
from mxnet import autograd from gluoncv.model_zoo.ssd.target import SSDTargetGenerator target_generator = SSDTargetGenerator() with autograd.record(): # 1. forward pass cls_preds, box_preds, anchors = net(x) # 2. generate training targets cls_targets, box_targets, box_masks = target_generator( anchors, cls_preds, gt_boxes, gt_ids) num_positive = (cls_targets > 0).sum().asscalar() cls_mask = (cls_targets >= 0).expand_dims(axis=-1) # negative targets should be ignored in loss # 3 losses, here we have two options, batch-wise or sample wise norm # 3.1 batch wise normalization: divide loss by the summation of num positive targets in batch batch_conf_loss = conf_loss(cls_preds, cls_targets, cls_mask) / num_positive batch_loc_loss = loc_loss(box_preds, box_targets, box_masks) / num_positive # 3.2 sample wise normalization: divide by num positive targets in this sample(image) sample_num_positive = (cls_targets > 0).sum(axis=0, exclude=True) sample_conf_loss = conf_loss(cls_preds, cls_targets, cls_mask) / sample_num_positive sample_loc_loss = loc_loss(box_preds, box_targets, box_masks) / sample_num_positive # Since ``conf_loss`` and ``loc_loss`` calculate the mean of such loss, we want # to rescale it back to loss per image. rescale_conf = cls_preds.size / cls_preds.shape rescale_loc = box_preds.size / box_preds.shape # then call backward and step, to update the weights, etc.. # L = conf_loss + loc_loss * alpha # L.backward()
The norms are different, but sample-wise norms sum up to be the same with batch-wise norm
batch-wise num_positive: 36.0 sample-wise num_positive: [13. 23.] <NDArray 2 @cpu(0)>
The per image
num_positive is no longer 1 and 4 because multiple anchor
boxes can be matched to a single object
Compare the losses
batch-wise norm conf loss: [442.7147 675.863 ] <NDArray 2 @cpu(0)> sample-wise norm conf loss: [1225.9791 1057.8724] <NDArray 2 @cpu(0)>
batch-wise norm loc loss: [2.656074 2.1453514] <NDArray 2 @cpu(0)> sample-wise norm loc loss: [7.3552823 3.3579414] <NDArray 2 @cpu(0)>
Which one is better? At first glance, it is hard to say which one is theoretically better because batch-wise norm ensures loss is well normalized by global statistics while sample-wise norm ensures gradients won’t explode in some extreme cases where there are hundreds of objects in a single image. In such case it would cause other samples in the same batch to be suppressed by this unusually large norm.
In our experiments, batch-wise norm is always better on Pascal VOC dataset, contributing 1~2% mAP gain. However, you should definitely try both of them when you use a new dataset or a new model.
While SSD networks are based on pre-trained feature extractors (called the
we also append uninitialized convolutional layers to the
to extend the cascades of feature maps.
There are also convolutional predictors appended to each output feature map, serving as class predictors and bounding box offsets predictors.
For these added layers, we must initialize them before training.
from gluoncv import model_zoo import mxnet as mx # don't load pretrained for this demo net = model_zoo.get_model('ssd_300_vgg16_atrous_voc', pretrained=False, pretrained_base=False) # random init net.initialize() # gluon only infer shape when real input data is used net(mx.nd.zeros(shape=(1, 3, 300, 300))) # now we have real shape for each parameter predictors = [(k, v) for k, v in net.collect_params().items() if 'predictor' in k] name, pred = predictors print(name, pred)
ssd3_convpredictor0_conv0_weight Parameter ssd3_convpredictor0_conv0_weight (shape=(84, 512, 3, 3), dtype=<class 'numpy.float32'>)
we can initialize it with different initializers, such as
param shape: (84, 512, 3, 3) peek first 20 elem: [-0.04006358 0.04752301 -0.04936712 0.02708755 -0.06145268 -0.0103094 0.04445995 0.02895925 -0.01508887 -0.04410328 -0.05917829 0.00261795 0.02758304 0.02611597 0.06757144 0.03305504 0.01971556 -0.05105315 -0.03926021 0.04332945] <NDArray 20 @cpu(0)>
Simply switching from
Xavier can produce ~1% mAP gain.
param shape: (84, 512, 3, 3) peek first 20 elem: [ 0.05409709 -0.02777563 -0.05862886 0.0120097 -0.05354748 0.03673649 -0.01118423 -0.00505917 -0.07389503 -0.05523501 -0.05710729 0.05084738 -0.04024388 -0.06320304 0.00896897 0.09223884 -0.05637952 -0.00855709 -0.11271537 -0.01174088] <NDArray 20 @cpu(0)>
If we revisit the per-class confidence predictions, its shape is (
B is the batch size,
A is the number of anchor boxes,
N is the number of foreground classes.
print('class prediction shape:', cls_preds.shape)
class prediction shape: (2, 8732, 21)
There are two ways we can handle the prediction:
1. take argmax of the prediction along the class axis. This way, only the the most probable class is considered.
N foreground classes separately. This way, the second most
probable class, for example, still has a
chance of surviving as the final prediction.
Consider this example:
bg 0.00027409225003793836 apple 0.00010083290544571355 orange 0.014964930713176727 person 0.040678903460502625 dog 0.49557045102119446 cat 0.4484107196331024
The probabilities of dog and cat are so close that if we use method 1, we are quite likely to lose the bet when cat is the correct decision.
It turns out that by switching from method 1 to method 2, we gain 0.5~0.8 mAP in evaluation.
One obvious drawback of method 2 is that it is significantly slower than method 1. For N classes, method 2 has O(N) complexity while method 1 is always O(1). This may or may not be a problem depending on the use case, but feel free to switch between them if you want.
Total running time of the script: ( 0 minutes 1.306 seconds)