caffe-jacindo-models

先 follow 這一篇 README.md
clone 下來，跟 caffe-jacindo 在同一層：

 
 ./
 |
  -- caffe-jacindo
 |
 |
  -- caffe-jacindo-models

clone 後，checkout 跟 caffe-jacindo 一樣的版本(branch): caffe-0.17

設定環境變數：

export PYTHONPATH=~/caffe-jacinto/python
export CAFFE_ROOT=~/caffe-jacinto

這樣就裝完了。
接下來依照需要做的 example 來 run script..

---

ref:

• Sparse, Quantized CNN training for segmentation
• EBC Exercise 39 setting up tidl on X15

SSD

scripts/train_image_object_detection.sh 就是用 caffe-ssd 的training data 做 SSD training 的 script。
可以知道，他用 caffe-ssd 的 training data:

if [ $dataset = "voc0712" ]
then
  train_data="../../caffe-jacinto/examples/VOC0712/VOC0712_trainval_lmdb"
  test_data="../../caffe-jacinto/examples/VOC0712/VOC0712_test_lmdb"

  name_size_file="../../caffe-jacinto/data/VOC0712/test_name_size.txt"
  label_map_file="../../caffe-jacinto/data/VOC0712/labelmap_voc.prototxt"

所以要先 follow caffe-ssd 的步驟，
download training data 到 ~/data
在 caffe-jacinto 中run create_data.sh, create_list.sh 把 VOC0712 的 data set 準備好。

chmod a+x data/VOC0712/*.sh
./data/VOC0712/create_list.sh
./data/VOC0712/create_data.sh

然後到 caffe-jacinto-models 下 run train_image_object_detection.sh. 根據 train_image_object_detection.sh 的內容，要在 scripts 目錄下執行。
結果：

2019-07-26 11:39:44 (10.2 MB/s) - ‘training/imagenet_jacintonet11v2_iter_320000.caffemodel’ saved [11516054/11516054]

Logging output to training/voc0712/JDetNet/20190726_11-39_ds_PSP_dsFac_32_hdDS8_1/train-log_20190726_11-39.txt
training/voc0712/JDetNet/20190726_11-39_ds_PSP_dsFac_32_hdDS8_1/initial
Traceback (most recent call last):
  File "./models/image_object_detection.py", line 5, in 
    from models.model_libs import *
ImportError: No module named models.model_libs
training/voc0712/JDetNet/20190726_11-39_ds_PSP_dsFac_32_hdDS8_1/l1reg
Traceback (most recent call last):
  File "./models/image_object_detection.py", line 5, in 
    from models.model_libs import *
ImportError: No module named models.model_libs
training/voc0712/JDetNet/20190726_11-39_ds_PSP_dsFac_32_hdDS8_1/sparse
Traceback (most recent call last):
  File "./models/image_object_detection.py", line 5, in 
    from models.model_libs import *
ImportError: No module named models.model_libs
training/voc0712/JDetNet/20190726_11-39_ds_PSP_dsFac_32_hdDS8_1/test
Traceback (most recent call last):
  File "./models/image_object_detection.py", line 5, in 
    from models.model_libs import *
ImportError: No module named models.model_libs
training/voc0712/JDetNet/20190726_11-39_ds_PSP_dsFac_32_hdDS8_1/test_quantize
Traceback (most recent call last):
  File "./models/image_object_detection.py", line 5, in 
    from models.model_libs import *
ImportError: No module named models.model_libs
cat: training/voc0712/JDetNet/20190726_11-39_ds_PSP_dsFac_32_hdDS8_1/test_quantize/deploy.prototxt: No such file or directory
cat: training/voc0712/JDetNet/20190726_11-39_ds_PSP_dsFac_32_hdDS8_1/test_quantize/test.prototxt: No such file or directory
./train_image_object_detection.sh: line 382: training/voc0712/JDetNet/20190726_11-39_ds_PSP_dsFac_32_hdDS8_1/initial/run.sh: No such file or directory
./train_image_object_detection.sh: line 382: training/voc0712/JDetNet/20190726_11-39_ds_PSP_dsFac_32_hdDS8_1/l1reg/run.sh: No such file or directory
./train_image_object_detection.sh: line 382: training/voc0712/JDetNet/20190726_11-39_ds_PSP_dsFac_32_hdDS8_1/sparse/run.sh: No such file or directory
./train_image_object_detection.sh: line 382: training/voc0712/JDetNet/20190726_11-39_ds_PSP_dsFac_32_hdDS8_1/test/run.sh: No such file or directory
./train_image_object_detection.sh: line 382: training/voc0712/JDetNet/20190726_11-39_ds_PSP_dsFac_32_hdDS8_1/test_quantize/run.sh: No such file or directory

好像是 module path 的問題。
image_object_detection.py 在 import 同目錄的 py module時，都加上了models。
把models 刪掉就沒有這個 Error 了。

然後是 Check failed: error == cudaSuccess (10 vs. 0) invalid device ordinal
一樣，把 scripts 下所有*.sh, *.py 中的 gpus= 找出來，都改成 gpus="0" 就可以了。
例如：

diff --git a/scripts/train_image_object_detection.sh b/scripts/train_image_object_detection.sh
index 0e5f99d..f9771f4 100755
--- a/scripts/train_image_object_detection.sh
+++ b/scripts/train_image_object_detection.sh
@@ -5,7 +5,7 @@ DATE_TIME=`date +'%Y%m%d_%H-%M'`
 #-------------------------------------------------------
 
 #------------------------------------------------
-gpus="0,1" #"0,1,2"
+gpus="0" #"0,1,2"


diff --git a/scripts/models/image_object_detection.py b/scripts/models/image_object_detection.py
index 67823a1..f625615 100644
--- a/scripts/models/image_object_detection.py
+++ b/scripts/models/image_object_detection.py
@@ -360,7 +360,7 @@ def main():
     # Which layers to freeze (no backward) during training.
     config_param.freeze_layers = []
     # Defining which GPUs to use.
-    config_param.gpus = "0,1" #gpus = "0"  
+    config_param.gpus = "0" #gpus = "0"  
 
     config_param.batch_size = 32
     config_param.accum_batch_size = 32

我把修改過的放到 https://github.com/checko/caffe-jacinto-models/tree/zoey-wpc
還有要注意 GPU 的 memory。
train_image_object_detection.sh 中的 batch_size 是 16，給 8G 的 GPU card 剛好
如果 GPU memory 只有 4G，要改成 8