Verify the loss value of ResNet training between Go and Python version

[Yancey1989]

To get a ResNet50 training baseline of loss value, I running the resnet.py example and got the following logs:

batch: 10, loss: 10.711030, acc1: 0.000000, acc5: 0.000000
batch: 20, loss: 7.499339, acc1: 0.000000, acc5: 0.000000
batch: 30, loss: 7.281894, acc1: 0.000000, acc5: 0.000000
batch: 40, loss: 7.059255, acc1: 0.000000, acc5: 0.000000
batch: 50, loss: 7.000484, acc1: 0.000000, acc5: 0.000000
batch: 60, loss: 6.871602, acc1: 0.000000, acc5: 3.125000
batch: 70, loss: 6.962079, acc1: 0.000000, acc5: 0.000000
batch: 80, loss: 6.872428, acc1: 0.000000, acc5: 0.000000
batch: 90, loss: 6.922100, acc1: 0.000000, acc5: 0.000000
batch: 100, loss: 6.918412, acc1: 0.000000, acc5: 0.000000
batch: 110, loss: 6.880023, acc1: 0.000000, acc5: 0.000000
batch: 120, loss: 6.936709, acc1: 0.000000, acc5: 3.125000
batch: 130, loss: 6.936309, acc1: 0.000000, acc5: 0.000000
batch: 140, loss: 6.923660, acc1: 0.000000, acc5: 0.000000
batch: 150, loss: 6.924109, acc1: 0.000000, acc5: 0.000000
batch: 160, loss: 6.923644, acc1: 0.000000, acc5: 3.125000
...

[Yancey1989]

Go version experiment

Run the following command to shuffle and creating.tgz files from the ImageNet training data:

$ find ./train | grep "JPEG" |  sort -R >  shuffle_list.txt
$ tar czf train_shuffle.tgz -T shuffle_list.txt

#239 can randomly skip some samples at the begging of each epoch, we use this PR to run the experiment and get the following logs:

2020/08/28 11:33:39 No CUDA found; CPU only
2020/08/28 11:51:41 building label vocabulary done.
2020/08/28 11:52:15 Epoch: 0, Batch: 10, loss:25.210730, acc1: 0.000000, acc5:0.000000, throughput: 0.028925 samples/secs
2020/08/28 11:52:48 Epoch: 0, Batch: 20, loss:18.479069, acc1: 0.000000, acc5:0.000000, throughput: 0.060924 samples/secs
2020/08/28 11:53:21 Epoch: 0, Batch: 30, loss:23.318371, acc1: 0.000000, acc5:0.000000, throughput: 0.091268 samples/secs
2020/08/28 11:53:54 Epoch: 0, Batch: 40, loss:17.084089, acc1: 0.000000, acc5:0.000000, throughput: 0.119941 samples/secs
2020/08/28 11:54:28 Epoch: 0, Batch: 50, loss:27.470881, acc1: 0.000000, acc5:0.000000, throughput: 0.150361 samples/secs
2020/08/28 11:55:01 Epoch: 0, Batch: 60, loss:13.764173, acc1: 0.000000, acc5:0.000000, throughput: 0.179411 samples/secs
2020/08/28 11:55:35 Epoch: 0, Batch: 70, loss:19.928579, acc1: 0.000000, acc5:0.000000, throughput: 0.208202 samples/secs
2020/08/28 11:56:08 Epoch: 0, Batch: 80, loss:9.244127, acc1: 0.000000, acc5:0.000000, throughput: 0.240541 samples/secs
2020/08/28 11:56:42 Epoch: 0, Batch: 90, loss:15.051638, acc1: 0.000000, acc5:0.000000, throughput: 0.263969 samples/secs
2020/08/28 11:57:15 Epoch: 0, Batch: 100, loss:18.548998, acc1: 0.000000, acc5:0.000000, throughput: 0.304695 samples/secs
2020/08/28 11:57:47 Epoch: 0, Batch: 110, loss:12.095874, acc1: 0.000000, acc5:0.000000, throughput: 0.339955 samples/secs
2020/08/28 11:58:20 Epoch: 0, Batch: 120, loss:23.189226, acc1: 0.000000, acc5:0.000000, throughput: 0.362004 samples/secs
2020/08/28 11:58:54 Epoch: 0, Batch: 130, loss:12.052785, acc1: 0.000000, acc5:0.000000, throughput: 0.391076 samples/secs
2020/08/28 11:59:26 Epoch: 0, Batch: 140, loss:16.743521, acc1: 0.000000, acc5:3.125000, throughput: 0.431979 samples/secs
2020/08/28 11:59:59 Epoch: 0, Batch: 150, loss:14.558453, acc1: 0.000000, acc5:3.125000, throughput: 0.456955 samples/secs
2020/08/28 12:00:31 Epoch: 0, Batch: 160, loss:16.397123, acc1: 0.000000, acc5:0.000000, throughput: 0.493441 samples/secs
...

but the program panics after 850 minibatch

2020/08/28 12:37:23 Epoch: 0, Batch: 850, loss:76.659737, acc1: 0.000000, acc5:0.000000, throughput: 2.500808 samples/secs
fatal error: unexpected signal during runtime execution
[signal SIGSEGV: segmentation violation code=0x1 addr=0x48 pc=0x7f515b598e39]

runtime stack:
runtime.throw(0x7bdeb4, 0x2a)
	/usr/local/go/src/runtime/panic.go:1116 +0x72
runtime.sigpanic()
	/usr/local/go/src/runtime/signal_unix.go:679 +0x46a

goroutine 1 [syscall]:
runtime.cgocall(0x706e60, 0xc000e57b80, 0xc000202790)
	/usr/local/go/src/runtime/cgocall.go:133 +0x5b fp=0xc000e57b50 sp=0xc000e57b18 pc=0x407e3b
github.com/wangkuiyi/gotorch/nn/functional._Cfunc_BatchNorm(0x7f4f0bc28f20, 0x36c0680, 0x36be550, 0x36c2240, 0x36c2260, 0xc000e57b01, 0x3fb999999999999a, 0x3ee4f8b588e368f1, 0xc000202790, 0x0)
	_cgo_gotypes.go:91 +0x4e fp=0xc000e57b80 sp=0xc000e57b50 pc=0x4f6a6e
github.com/wangkuiyi/gotorch/nn/functional.BatchNorm.func1(0xc000fac028, 0x36c0680, 0x36be550, 0x36c2240, 0x36c2260, 0x6fd701, 0x3fb999999999999a, 0x3ee4f8b588e368f1, 0xc000202790, 0xc000642060)
	/work/gotorch/nn/functional/functional.go:45 +0x18b fp=0xc000e57be8 sp=0xc000e57b80 pc=0x4f7b2b
github.com/wangkuiyi/gotorch/nn/functional.BatchNorm(0xc000fac028, 0xc0000100b0, 0xc0000100b8, 0xc0000100a0, 0xc0000100a8, 0x1, 0x3fb999999999999a, 0x3ee4f8b588e368f1, 0xc000642060)
	/work/gotorch/nn/functional/functional.go:45 +0xd6 fp=0xc000e57c50 sp=0xc000e57be8 pc=0x4f73c6
github.com/wangkuiyi/gotorch/nn.(*BatchNorm2dModule).Forward(0xc0001dc000, 0xc000fac028, 0xc000fac028)
	/work/gotorch/nn/batchnorm.go:73 +0x89 fp=0xc000e57ca8 sp=0xc000e57c50 pc=0x6fcba9
github.com/wangkuiyi/gotorch/vision/models.(*ResnetModule).Forward(0xc0001de000, 0xc000fac018, 0x735d60)
	/work/gotorch/vision/models/resnet.go:182 +0x70 fp=0xc000e57d68 sp=0xc000e57ca8 pc=0x703c60
main.trainOneBatch(0xc000fac018, 0xc000fac020, 0xc0001de000, 0xc000202f88, 0x426681f7, 0xc000000000)
	/work/gotorch/example/resnet/resnet.go:83 +0x3c fp=0xc000e57de0 sp=0xc000e57d68 pc=0x70538c
main.main()
	/work/gotorch/example/resnet/resnet.go:149 +0x51f fp=0xc000e57f88 sp=0xc000e57de0 pc=0x7059ff
runtime.main()
	/usr/local/go/src/runtime/proc.go:203 +0x1fa fp=0xc000e57fe0 sp=0xc000e57f88 pc=0x439f6a
runtime.goexit()
	/usr/local/go/src/runtime/asm_amd64.s:1373 +0x1 fp=0xc000e57fe8 sp=0xc000e57fe0 pc=0x466771
exit status 2

TODOs

1. The loss value does not decrease as we expected, we should find the reason.
2. fix panics on batch_norm function.