Help! error occurred when transform Darknet model!!!
i download Darknet model form that:
the tiny yolo cfg & weights
i add "pad=0" on " [maxpool]" followed by model compiler's "read me"
compiler error is
my command is :
python . --model_loader model_loader/darknet --cfg_path pb_files/yolov2-tiny.cfg --weights_path pb_files/yolov2-tiny-voc.weights --image_w 416 --image_h 416 --dataset_loader dataset_loader/img_0_1.py --image_h 416 --image_w 416 --dataset_pic_path dataset/yolo_416_416
the test picture has been resize to 416*416
Thanks a lot!!!
@Yu-Jiageng For the first, I think so. Re-training is needed. For the second, you are right. With no high fps requirement, the large model can be loaded in stages actually. However, we haven't tested this flow, and model compiler cannot convert so large a model now.
Yu Jiageng last edited by
@nathan For the first of your answer, should we re-training a new yolov2-tiny model with 320x240 inputs? Alternatively, could we reuse the pre-trained yolov2-tiny model with some tricks?
For the second, I think it is not absolutely unable to use the model larger than 6M according to your k210 datasheet. Actually, k210 has 16MiB SPI NandFlash, and the model can be loaded into k210 in stages. Am I right?
Thanks for your attention!
@123qwa First, I download
yolov2-tinycfg and weights files, and add
maxpoolin the cfg file. But I cannot reproduce your error message, instead another error as follows:
This is because KPU can only support the maximum size of 320 by 256.
My command is
python3 __main__.py --model_loader model_loader/darknet --cfg_path pb_files/tiny-yolo/yolov2-tiny.cfg --weights_path pb_files/tiny-yolo/yolov2-tiny.weights --image_w 416 --image_h 416 --dataset_input_name input --dataset_pic_path dataset/yolo_240_320/dog.bmp
yolov2-tiny's weights are two large that K210 cannot hold that. So the origin
yolov2-tinycannot be supported by K210 unless you crop the model to less than 6MiB with user program in total.
but why nan appear？
problem on model compiler or my model
(i download form darnet) or change the layers?
pl last edited by
seems like when run
batch = sess.run(tensor, dataset)with tensor is
netwok.png in tensorboard
@latyas there are still some options ,not understand there meaning （--tensor_input_min --tensor_input_max --tensor_input_minmax_auto --eight_bit_mode --layer_start_idx
still need more details。
there are still some options ,not understand there meaning （--tensor_input_min --tensor_input_max --tensor_input_minmax_auto --eight_bit_mode --layer_start_idx
still need more details。
yes followed by this
followed by this
See README.md ?