Open dorianbrown opened 4 years ago
Trying transfer-learning on COCO with 3 classes:
We'll try an even split of these 3 classes with a fourth containing none of these classes, with 1500 for each. We'll split these 6000 images into 4500 train and 1500 test.
The COCO dataset has about 82,000
training, and 40,000
validation
First trying transfer learning on the yolov3.weights
with
python train.py \
--data data/transfer.data \
--cfg cfg/transfer.cfg \
--weights weights/yolov3.weights \
--epochs 2000 \
--batch-size 128 \
--transfer
After 100 epochs this didn't work. Although the train_loss was slowly dropping, the mAP stayed around 0.01 and the F1 and GIoU both didn't improve.
Trying non-transfer learning starting from imagenet "starter" weights
python train.py \
--data data/transfer.data \
--cfg cfg/transfer.cfg \
--weights weights/darknet53.conv.74 \
--epochs 2000
For some reason the run keeps getting stuck after 10-15 epochs, not sure why
Now trying different starting weights
python train.py \
--data data/transfer.data \
--cfg cfg/transfer.cfg \
--weights weights/yolov3.weights \
--epochs 2000
So far it seems to be converging nicely, but to an mAP@0.5
of about 0.5. Not sure if this is good or not, but seems comparable to YoloV3 on COCO dataset:
python train.py \
--data data/transfer.data \
--cfg cfg/transfer.cfg \
--weights weights/yolov3.weights \
--epochs 2000 \
--adam
Transfer learning with smaller batch-size (32)
python train.py \
--data data/transfer.data \
--cfg cfg/transfer.cfg \
--weights weights/yolov3.weights \
--epochs 2000 \
--transfer
Not much of an improvement (grey is current run):
Problem
An issue that is currently causing problems is people on bicycles. We have been using the YoloV3 pretrained weights which have been trained on Pascal VOC. The problem is, this dataset contains
bicycle
andperson
as seperate labels, but not the two combined.This is causing a few different outcomes:
Possible Solutions
person
,car
,cyclist
. This seems to be the most robust, but also time consuming solution.outcome 2
above.