[mindspore] [training warning] warning when executing training code

1、 Use Ascend Training efficientnet The following warning appears when , Here I suspect mindspore Caused by environmental updates （ Environment from mindspore1.0.1 Updated to mindspore1.1.1）, Of course, some of these warnings have been seen before , There are also some that have never been seen . Want to know why these warnings appear and what solutions are available ？

2、 Use the training model obtained from the previous environmental training （ckpt file ） Changed the updated environment （mindspore1.1.1） The previous model cannot be loaded to continue training or verify the results （ precision ）.

【 Operation steps & Problem phenomenon 】

1、 For the problems described above 1, Use the updated environment mindspore1.1.1 When running training code , The following warning will appear , I have ignored similar warnings ：

【1】WARNING: 'ControlDepend' is deprecated from version 1.1 and will be removed in a future version, use 'Depend' instead.

【2】[WARNING] ME(8272:140026733385536,MainProcess):2021-02-17-09:45:52.440.210 [mindspore/ops/operations/array_ops.py:2302] WARN_DEPRECATED: The usage of Pack is deprecated. Please use Stack.

【3】[WARNING] DEVICE(8272,python):2021-02-17-09:47:03.284.745 [mindspore/ccsrc/runtime/device/ascend/kernel_select_ascend.cc:282] TagRaiseReduce] node:[DropoutGenMask]reduce precision from int64 to int32

【4】[WARNING] SESSION(8272,python):2021-02-17-09:47:19.915.525 [mindspore/ccsrc/backend/session/ascend_session.cc:1412] SelectKernel] There are 1 node/nodes used reduce precision to selected the kernel!

【5】[WARNING] KERNEL(8272,python):2021-02-17-09:47:47.034.963 [mindspore/ccsrc/backend/kernel_compiler/tbe/tbe_utils.cc:58] SaveJsonInfo] Create info file failed(./kernel_meta/te_fusion_conv2d_backprop_input_d_tile_d_7453389728167276878_0.info).

【6】[WARNING] KERNEL(8272,python):2021-02-17-09:47:47.081.467 [mindspore/ccsrc/backend/kernel_compiler/tbe/tbe_utils.cc:58] SaveJsonInfo] Create info file failed(./kernel_meta/te_fusion_depthwise_conv2d_bn_training_reduce_7468152902432878638_0.info).

【7】[WARNING] DEVICE(8272,python):2021-02-17-09:50:23.020.968 [mindspore/ccsrc/runtime/device/kernel_runtime.cc:339] AssignStaticMemoryInput] It is not suggested to use a lonely weight parameter as the output of graph

2、 Answer the above question 2, Use the model file generated for the environment （ Don't worry about the wrong naming of the model file , It's just that the code name hasn't changed ） And completely unchanged code loading and running , Will report a mistake , The error message is as follows , There is no problem after trying to load and run the old environment , I wonder if mindspore There is incompatibility after the update , Some operators （ The problem I have here is the convolution operator mindspore.nn.conv2d()） The calculation method has been modified from NCHW Revised to CNHW？ Do you think my judgment is correct ？

WARNING: 'ControlDepend' is deprecated from version 1.1 and will be removed in a future version, use 'Depend' instead.

[WARNING] ME(5604:140668626671424,MainProcess):2021-02-17-09:44:27.437.987 [mindspore/ops/operations/array_ops.py:2302] WARN_DEPRECATED: The usage of Pack is deprecated. Please use Stack.

[ERROR] ME(5604:140668626671424,MainProcess):2021-02-17-09:44:32.796.212 [mindspore/train/serialization.py:84] Failed to combine the net and the parameters for param blocks.0.layers.0.depth_wise_conv.0.weight.

Traceback (most recent call last):

  File "train.py", line 138, in <module>

    load_param_into_net(net, ckpt)

  File "/root/miniconda3/envs/ci3.7/lib/python3.7/site-packages/mindspore/train/serialization.py", line 373, in load_param_into_net

    _update_param(param, new_param)

  File "/root/miniconda3/envs/ci3.7/lib/python3.7/site-packages/mindspore/train/serialization.py", line 87, in _update_param

    raise RuntimeError(msg)

RuntimeError: Net parameters blocks.0.layers.0.depth_wise_conv.0.weight shape((32, 1, 3, 3)) different from parameter_dict's((1, 32, 3, 3))

【 Screenshot information 】

For question 1 ：

For question 2 ：

The error message is ：net The weight in shape and ckpt There's a discrepancy in .

net in conv.0.weight Of shape by （32, 1, 3, 3）,checkpoint in conv.0.weight Of shape by （1, 32, 3, 3）

If the network definition is the same , Is it across the back end , such as GPU Well trained checkpoint Got it Ascend Reasoning .

The solution is ：

1、 Same as hardware training and reasoning

2、 conversion conv.0.weight Of shape, send ckpt and net Of shape matching