kendryte / tensorflow-workspace

Tensorflow workspaces for training your custom models of K210

Tensorflow Workspace for K210

Classifier for ImageNet

1. Download ImageNet dataset, extract it as the instructions to ILSVRC2012_img_train, it has 1000 folders, each folder's name represents its class id.
2. The model definition file is mobilenetv1/models/mobilenet_v1.py, ATTENTION, K210 does not support the method of SAME PADDING in tensorflow, so we need to add padding around the image manually before a stride=2 conv (in this situation, the padding method of conv layer with stride=2 should be set to VALID
3. Modify mobilenetv1/run_mobilenet_v1.sh and start your training.
4. Using freeze_graph.py to freeze your model from ckpt to pb, just run python mobilenetv1/freeze_graph.py model.mobilenet_v1 ckpt_fold pb_file
5. Test on ImageNet, you need a val dataset of ImageNet, then run python mobilenetv1/validation_imagenet.py pb_file val_set_fold
6. Estimate one image, run python mobilenetv1/predict_one_pic.py pb_file pic

ImageNet 分类器示例

1. 下载ImageNet数据集，按照说明解压缩训练数据集到文件夹ILSVRC2012_img_train，内含1000个子文件夹，每个子文件夹的命名为其分类代号（类似n02484975），每个子文件夹内为该分类的训练数据
2. mobilenet v1定义文件：mobilenetv1/models/mobilenet_v1.py，需要注意由于K210不支持tensorflow的SAME padding，所以在stride=2时先固定padding一圈0，然后再进行stride=2的卷积（padding=VALID）
3. 训练脚本 mobilenetv1/run_mobilenet_v1.sh，根据需要修改其中的参数，然后运行
4. freeze_graph.py将训练ckpt转成pb文件，命令格式如下：
   python mobilenetv1/freeze_graph.py model.mobilenet_v1 ckpt_fold pb_file
5. 测试在ImageNet验证集上的性能，下载验证集，将文件按类别解压好（与训练集类似），运行 python mobilenetv1/validation_imagenet.py pb_file（or ckpt folder） val_set_fold
6. 预测单张图片，python mobilenetv1/predict_one_pic.py pb_file（or ckpt folder） pic

tensorflow-workspace's Issues

There is no "stride=2" in mobilenet.py

only a few strides=(2,2) so what is stride=2?

How to train my own dataset model(not ImageNet dataset)?

Hi kendryter:
I have some datasets I collected myself, and I want to know how to use this data to train my own model? Is there a detailed tutorial on training my own model?

Restoring pretrained model error

It use ckpt in train_softmax.py, but use pb in run_mobilenet_v1.sh.
It case run_mobilenet_v1.sh error

  if pretrained_model:
                print('Restoring pretrained model: %s' % pretrained_model)
                saver_restore.restore(sess, tf.train.latest_checkpoint(pretrained_model))

Problem on using nncase to transfer tflite to C code

Hi,

I was following the instructions from d80f0d2#diff-7d9fc065a89438c0ca7e01e4611d0f3aR30 to generate C code from .pb file and I was using the mobilenetv1.pb file from the example. I just tried the steps from the solutions and am stuck in the following step:

.\ncc.exe -i tflite -o k210code --dataset .\dataset\ .\bin\mobilenetv1.tflite .\bin\mobilenetv1.c
It shows: Fatal: output-format.

I checked the output-format options and 'k210code' has been modified to 'k210model' since v0.1.0 RC1. So I chose 'k210model' and ran it again:
.\ncc.exe -i tflite -o k210model --dataset .\dataset\ .\bin\mobilenetv1.tflite .\bin\mobilenetv1.c
and then I got a .kmodel file rather than the .c file.

I wonder if anyone can help me with that. Thanks a lot.

TMaysGGS

Model cannot be trainned well by the scripts.

Hi, I try to use these scripts on my own datasets. I run the run_'mobilenet_v1.sh' but it cannot load pretrained model. So I train it without pretrained model. But I tried several times and found the saved model (ckpt) always output same result for different inputs. (by runing 'predict_one_pic.py ckpt_folder img') Could anybody solve my problem?

Compile mobilenet to K210

Hi,

Could you give instructions on how to use kendryte-model-compiler to compile the mobilenet checkpoint? Thanks.