有关Intel 计算棒NCS2的坑~~~~~~~~~~~Failed to read output from FIFO: NC_ERROR

第1个坑是ckpt文件转pb文件：

网上的做法是先保存为ckpt，再import_meta_graph，再用saver.restore取会保存的内容，最后graph_util.convert_variables_to_constants。这样做可以产生pb文件，但文件比较大（130M）。而且用transform_graph工具压缩去除无用节点没有效果，文件大小基本不变。个人分析是这种保存方法有很多冗余变量也被保存进去了。我最后跳过ckpt保存，直接convert_variables_to_constants。保存的文件非常小（12M），小了10倍多！！！

隔了一个周末，再尝试直接graph_util.convert_variables_to_constants，发现pb文件还是130M，而且Model.summary后发现参数个数*4大概就是130M bytes。所以之前模型突然变成12M是我失忆了？？无心再去找原因了。。。大家别当真就好

第2个坑是NCS2的Failed to read output from FIFO: NC_ERROR

E: [xLink] [    697998] [EventRead00Thr] eventReader:218    eventReader thread stopped (err -4)
E: [global] [    697998] [Scheduler00Thr] dispatcherEventSend:1004    Write failed (header) (err -4) | event XLINK_READ_REL_REQ

E: [xLink] [    697998] [Scheduler00Thr] eventSchedulerRun:626    Event sending failed
E: [global] [    698000] [Scheduler00Thr] dispatcherEventSend:1004    Write failed (header) (err -4) | event XLINK_WRITE_REQ

E: [xLink] [    698000] [Scheduler00Thr] eventSchedulerRun:626    Event sending failed
E: [xLink] [    698000] [Scheduler00Thr] eventSchedulerRun:576    Dispatcher received NULL event!
E: [ncAPI] [    698000] [python2.7] ncFifoReadElem:3510    Packet reading is failed.
E: [watchdog] [    698664] [WatchdogThread] sendPingMessage:121    Failed send ping message: X_LINK_ERROR
E: [watchdog] [    699663] [WatchdogThread] sendPingMessage:121    Failed send ping message: X_LINK_ERROR
E: [watchdog] [    700662] [WatchdogThread] sendPingMessage:121    Failed send ping message: X_LINK_ERROR
E: [watchdog] [    701662] [WatchdogThread] sendPingMessage:121    Failed send ping message: X_LINK_ERROR
E: [watchdog] [    702661] [WatchdogThread] sendPingMessage:121    Failed send ping message: X_LINK_ERROR
E: [watchdog] [    703660] [WatchdogThread] sendPingMessage:121    Failed send ping message: X_LINK_ERROR
E: [watchdog] [    704659] [WatchdogThread] sendPingMessage:121    Failed send ping message: X_LINK_ERROR
E: [watchdog] [    705658] [WatchdogThread] sendPingMessage:121    Failed send ping message: X_LINK_ERROR
E: [watchdog] [    706657] [WatchdogThread] sendPingMessage:121    Failed send ping message: X_LINK_ERROR
E: [watchdog] [    707656] [WatchdogThread] sendPingMessage:121    Failed send ping message: X_LINK_ERROR
E: [watchdog] [    708655] [WatchdogThread] sendPingMessage:121    Failed send ping message: X_LINK_ERROR
E: [watchdog] [    709654] [WatchdogThread] sendPingMessage:121    Failed send ping message: X_LINK_ERROR
E: [watchdog] [    710653] [WatchdogThread] sendPingMessage:121    Failed send ping message: X_LINK_ERROR
E: [watchdog] [    710653] [WatchdogThread] watchdog_routine:315    [0x2c6ff50] device, not respond, removing from watchdog

Traceback (most recent call last):
  File "/home/jojo/clion-2019.2.5/plugins/python/helpers/pydev/pydevd.py", line 1415, in _exec
    pydev_imports.execfile(file, globals, locals)  # execute the script
  File "/home/jojo/CLionProjects/samples/python_samples/classification_sample/classification_sample.py", line 147, in <module>
    sys.exit(main() or 0)
  File "/home/jojo/CLionProjects/samples/python_samples/classification_sample/classification_sample.py", line 115, in main
    res = exec_net.infer(inputs={input_blob: images})
  File "ie_api.pyx", line 241, in openvino.inference_engine.ie_api.ExecutableNetwork.infer
  File "ie_api.pyx", line 308, in openvino.inference_engine.ie_api.InferRequest.infer
  File "ie_api.pyx", line 312, in openvino.inference_engine.ie_api.InferRequest.infer
RuntimeError: Failed to read output from FIFO: NC_ERROR

一开始有个帖子说NC_ERROR是没有管理员权限造成的，sudo后就没问题了。其实这个方法没用。后来怀疑是产生的IR太大超过计算棒容量了，故压缩了IR文件到FP16，大小减少了一半，但问题依旧。其实如果ie.load_network没问题就说明模型装载没问题的。最后找到帖子：https://github.com/PINTO0309/OpenVINO-YoloV3/issues/7

发现是模型input_shape比较大时，计算棒功率猛增，电流不够了！！！！麻蛋！！还有这骚操作！！！故减少网络输入大小，居然可以了。解决方法：1、缩小网络入口大小，2、计算棒插到有独立供电，且电流大于3A的USB上（这个没试过，理论可行）