Hi @Frank , can you give me the argument to convert with SSD-Mobilenetv2? I followed these arguments and replaced Mobilenetv1 with Mobilenetv2 but it didn’t convert.
$ ./convert
–model-name mobilenet_ssd
–platform tensorflow
–model ~/yan/Yan/models-zoo/tensorflow/mobilenet_ssd/ssd_mobilenet_v1_coco_2017_11_17.pb
–input-size-list ‘300,300,3’
–inputs FeatureExtractor/MobilenetV1/MobilenetV1/Conv2d_0/BatchNorm/batchnorm/mul_1
–outputs “‘concat concat_1’”
–mean-values ‘127.5 127.5 127.5 0.007843137’
–quantized-dtype asymmetric_affine
–source-files ./data/dataset/dataset0.txt
–kboard VIM3 --print-level 1
**my arguments:
$ ./convert
–model-name mobilenet_ssd
–platform tensorflow
–model home/trong/Dowloads/save_model.pb
–input-size-list ‘320,320,3’
–inputs FeatureExtractor/MobilenetV2/MobilenetV2/Conv2d_0/BatchNorm/batchnorm/mul_1
–outputs “‘concat concat_1’”
–mean-values ‘127.5,127.5, 127.5, 127.5’
–quantized-dtype asymmetric_affine
–kboard VIM3 --print-level 1
@Trong_Huu The last layer of the SSD model does not support conversion. If you want to convert it smoothly, you must manually delete the last layer and implement it manually in the code. You can participate in the SDK documentation, it is written that those operators support conversion
@Frank I have switched back to using ssd-mobilenetv1 but I am facing problems. when I use the original model of ssd_mobilenet_v1_coco_2018_01_28 it runs but when I pretrain the model with my dataset it gets the same error as above I mentioned. Can you tell me why?
@Trong_Huu Regarding this issue, I suggest that you directly consult the maintenance of Tensorflow, which is better. We are more supportive for conversion tools. Ps: I personally think it is the TF version of your model pre-training, which is inconsistent with your current TF version, and some functions do not match
@Frank ,Thanks for the reply, can you tell me what version of Tensorflow your model trained on? I used tf version 1.15 on colab
@Trong_Huu I have not trained SSD on tensorflow, I recommend using the same version number as the conversion tool.
@Frank how do I know which tf version of Python API convert tool to use? I have trained ssd mobilenetv1 in many ways but the convert tool is not working
@Trong_Huu SSD cannot be directly converted. There are operators that are not supported by the conversion tool. You cannot solve this problem by modifying the TF version. The TF version of the Python conversion tool is the same as the C++ conversion tool, both are 2.0
@Frank can you suggest me any way to convert and run ssd-mobilenet on VIM3, I have done the conversion with Python API convert tool ksnn v1.2 many times but it doesn’t work. My model file: frozen_inference_graph.pb - Google Drive
@Trong_Huu KSNN support SSD, but you should use v1.3, not v1.2. And the post-processing part is still optimized, so the effect is not very good
@Frank can you tell me what platform you use to train ssd mobilenet to deploy using ksnn? when i used tensorflow 1.15.0 platform, my model didn’t convert
@Trong_Huu The SSD model needs to remove the last 2 layers first, and the last 2 layers do not support conversion
@Frank I have switched to SSD mobilenetv1 and I want to ask you what platform do you train SSD mobilenetv1?
@Frank actually I have converted many times but all failed, can you help me to convert. Here is my model: frozen_inference_graph.pb - Google Drive
@Trong_Huu My SSD model is caffe, and then I modified the last 2 layers through the configuration file. There is a fAQ document in the SDK, which explains how to deal with the SSD model.
