调试方法
计算图 Profiler
计算图 Profiler，显示完成 infer shape 操作后的已序列化 ir_graph 信息，用于确认 infer shape 是否正确。
使用方法
程序执行前，添加环境变量
export TG_DEBUG_GRAPH=1，启用计算图 Profiler 功能；删除环境变量
unset TG_DEBUG_GRAPH， 关闭计算图 Profiler 功能。
logo 信息
$./tm_classification -m mobilenet.tmfile -i cat.jpg -r 10
tengine-lite library version: 1.4-dev
graph node_num 86 tensor_num: 86 subgraph_num: 1
graph layout: NCHW model layout: NCHW model_format: tengine
node: 0 op: Const name: conv1-conv1/bn-conv1/scale.bias.bn.fused.fused
output tensors: 1
0: [id: 0] conv1-conv1/bn-conv1/scale.bias.bn.fused.fused type: fp32/const shape: [32] from node: 0 (consumer: 1)
node: 1 op: Const name: conv1/weight.fused.fused
output tensors: 1
0: [id: 1] conv1/weight.fused.fused type: fp32/const shape: [32,3,3,3] from node: 1 (consumer: 1)
node: 2 op: InputOp name: input
output tensors: 1
0: [id: 2] data type: fp32/input shape: [1,3,224,224] from node: 2 (consumer: 1)
node: 3 op: Const name: conv2_1/dw-conv2_1/dw/bn-conv2_1/dw/scale.bias.bn.fused.fused
output tensors: 1
0: [id: 3] conv2_1/dw-conv2_1/dw/bn-conv2_1/dw/scale.bias.bn.fused.fused type: fp32/const shape: [32] from node: 3 (consumer: 1)
(#### 太多了，直接跳到末尾 ####)
node: 84 op: Pooling name: pool6
input tensors: 1
0: [id: 81] relu6/sep/0 type: fp32/var shape: [1,1024,7,7] from node: 81 (consumer: 1)
output tensors: 1
0: [id: 84] pool6 type: fp32/var shape: [1,1024,1,1] from node: 84 (consumer: 1)
node: 85 op: Convolution name: fc7
input tensors: 3
0: [id: 84] pool6 type: fp32/var shape: [1,1024,1,1] from node: 84 (consumer: 1)
1: [id: 83] fc7/weight type: fp32/const shape: [1000,1024,1,1] from node: 83 (consumer: 1)
2: [id: 82] fc7/bias type: fp32/const shape: [1000] from node: 82 (consumer: 1)
output tensors: 1
0: [id: 85] fc7 type: fp32/var shape: [1,1000,1,1] from node: 85
graph inputs: 1
input
graph outputs: 1
fc7
model file : mobilenet.tmfile
image file : cat.jpg
img_h, img_w, scale[3], mean[3] : 224 224 , 0.017 0.017 0.017, 104.0 116.7 122.7
Repeat 10 times, thread 1, avg time 44.12 ms, max_time 73.76 ms, min_time 37.14 ms
--------------------------------------
8.574144, 282
7.880117, 277
7.812573, 278
7.286458, 263
6.357486, 281
--------------------------------------
Tengine plugin device CPU is unregistered.
性能 Profiler
性能 Profiler，用于逐层耗时统计，在网络模型运行时统计 CPU 上 kernel 耗时信息，用于分析潜在的耗时问题。
使用方法
程序执行前，添加环境变量
export TG_DEBUG_TIME=1，启用性能 Profiler 功能；删除环境变量
unset TG_DEBUG_TIME， 关闭性能 Profiler 功能。
logo 信息
$./tm_classification -m mobilenet.tmfile -i cat.jpg -r 10
tengine-lite library version: 1.4-dev
model file : mobilenet.tmfile
image file : cat.jpg
img_h, img_w, scale[3], mean[3] : 224 224 , 0.017 0.017 0.017, 104.0 116.7 122.7
Repeat 10 times, thread 1, avg time 42.36 ms, max_time 65.59 ms, min_time 38.26 ms
--------------------------------------
8.574144, 282
7.880117, 277
7.812573, 278
7.286458, 263
6.357486, 281
--------------------------------------
0 [ 3.42% : 1.2 ms] Convolution idx: 5 shape: {1 3 224 224} -> {1 32 112 112} fp32 -> fp32 K: 3x3 | S: 2x2 | P: 1 1 1 1 MFLOPS: 21.68 Rate:17722
1 [ 4.60% : 1.6 ms] Convolution idx: 8 shape: {1 32 112 112} -> {1 32 112 112} fp32 -> fp32 K: 3x3 | S: 1x1 | P: 1 1 1 1 DW( 32) MFLOPS: 7.23 Rate:4392
2 [ 5.66% : 2.0 ms] Convolution idx: 11 shape: {1 32 112 112} -> {1 64 112 112} fp32 -> fp32 K: 1x1 | S: 1x1 | P: 0 0 0 0 MFLOPS: 51.38 Rate:25423
3 [ 3.28% : 1.2 ms] Convolution idx: 14 shape: {1 64 112 112} -> {1 64 56 56} fp32 -> fp32 K: 3x3 | S: 2x2 | P: 1 1 1 1 DW( 64) MFLOPS: 3.61 Rate:3085
4 [ 4.25% : 1.5 ms] Convolution idx: 17 shape: {1 64 56 56} -> {1 128 56 56} fp32 -> fp32 K: 1x1 | S: 1x1 | P: 0 0 0 0 MFLOPS: 51.38 Rate:33824
5 [ 3.13% : 1.1 ms] Convolution idx: 20 shape: {1 128 56 56} -> {1 128 56 56} fp32 -> fp32 K: 3x3 | S: 1x1 | P: 1 1 1 1 DW(128) MFLOPS: 7.23 Rate:6458
6 [ 7.85% : 2.8 ms] Convolution idx: 23 shape: {1 128 56 56} -> {1 128 56 56} fp32 -> fp32 K: 1x1 | S: 1x1 | P: 0 0 0 0 MFLOPS:102.76 Rate:36658
7 [ 1.72% : 0.6 ms] Convolution idx: 26 shape: {1 128 56 56} -> {1 128 28 28} fp32 -> fp32 K: 3x3 | S: 2x2 | P: 1 1 1 1 DW(128) MFLOPS: 1.81 Rate:2937
8 [ 3.37% : 1.2 ms] Convolution idx: 29 shape: {1 128 28 28} -> {1 256 28 28} fp32 -> fp32 K: 1x1 | S: 1x1 | P: 0 0 0 0 MFLOPS: 51.38 Rate:42671
9 [ 1.32% : 0.5 ms] Convolution idx: 32 shape: {1 256 28 28} -> {1 256 28 28} fp32 -> fp32 K: 3x3 | S: 1x1 | P: 1 1 1 1 DW(256) MFLOPS: 3.61 Rate:7655
10 [ 6.45% : 2.3 ms] Convolution idx: 35 shape: {1 256 28 28} -> {1 256 28 28} fp32 -> fp32 K: 1x1 | S: 1x1 | P: 0 0 0 0 MFLOPS:102.76 Rate:44564
11 [ 0.78% : 0.3 ms] Convolution idx: 38 shape: {1 256 28 28} -> {1 256 14 14} fp32 -> fp32 K: 3x3 | S: 2x2 | P: 1 1 1 1 DW(256) MFLOPS: 0.90 Rate:3259
12 [ 3.27% : 1.2 ms] Convolution idx: 41 shape: {1 256 14 14} -> {1 512 14 14} fp32 -> fp32 K: 1x1 | S: 1x1 | P: 0 0 0 0 MFLOPS: 51.38 Rate:43954
13 [ 1.01% : 0.4 ms] Convolution idx: 44 shape: {1 512 14 14} -> {1 512 14 14} fp32 -> fp32 K: 3x3 | S: 1x1 | P: 1 1 1 1 DW(512) MFLOPS: 1.81 Rate:4989
14 [ 6.57% : 2.3 ms] Convolution idx: 47 shape: {1 512 14 14} -> {1 512 14 14} fp32 -> fp32 K: 1x1 | S: 1x1 | P: 0 0 0 0 MFLOPS:102.76 Rate:43767
15 [ 0.96% : 0.3 ms] Convolution idx: 50 shape: {1 512 14 14} -> {1 512 14 14} fp32 -> fp32 K: 3x3 | S: 1x1 | P: 1 1 1 1 DW(512) MFLOPS: 1.81 Rate:5266
16 [ 6.44% : 2.3 ms] Convolution idx: 53 shape: {1 512 14 14} -> {1 512 14 14} fp32 -> fp32 K: 1x1 | S: 1x1 | P: 0 0 0 0 MFLOPS:102.76 Rate:44659
17 [ 1.01% : 0.4 ms] Convolution idx: 56 shape: {1 512 14 14} -> {1 512 14 14} fp32 -> fp32 K: 3x3 | S: 1x1 | P: 1 1 1 1 DW(512) MFLOPS: 1.81 Rate:5003
18 [ 6.44% : 2.3 ms] Convolution idx: 59 shape: {1 512 14 14} -> {1 512 14 14} fp32 -> fp32 K: 1x1 | S: 1x1 | P: 0 0 0 0 MFLOPS:102.76 Rate:44678
19 [ 0.98% : 0.3 ms] Convolution idx: 62 shape: {1 512 14 14} -> {1 512 14 14} fp32 -> fp32 K: 3x3 | S: 1x1 | P: 1 1 1 1 DW(512) MFLOPS: 1.81 Rate:5174
20 [ 6.36% : 2.3 ms] Convolution idx: 65 shape: {1 512 14 14} -> {1 512 14 14} fp32 -> fp32 K: 1x1 | S: 1x1 | P: 0 0 0 0 MFLOPS:102.76 Rate:45249
21 [ 1.02% : 0.4 ms] Convolution idx: 68 shape: {1 512 14 14} -> {1 512 14 14} fp32 -> fp32 K: 3x3 | S: 1x1 | P: 1 1 1 1 DW(512) MFLOPS: 1.81 Rate:4976
22 [ 6.61% : 2.4 ms] Convolution idx: 71 shape: {1 512 14 14} -> {1 512 14 14} fp32 -> fp32 K: 1x1 | S: 1x1 | P: 0 0 0 0 MFLOPS:102.76 Rate:43523
23 [ 0.61% : 0.2 ms] Convolution idx: 74 shape: {1 512 14 14} -> {1 512 7 7} fp32 -> fp32 K: 3x3 | S: 2x2 | P: 1 1 1 1 DW(512) MFLOPS: 0.45 Rate:2062
24 [ 3.36% : 1.2 ms] Convolution idx: 77 shape: {1 512 7 7} -> {1 1024 7 7} fp32 -> fp32 K: 1x1 | S: 1x1 | P: 0 0 0 0 MFLOPS: 51.38 Rate:42853
25 [ 0.74% : 0.3 ms] Convolution idx: 80 shape: {1 1024 7 7} -> {1 1024 7 7} fp32 -> fp32 K: 3x3 | S: 1x1 | P: 1 1 1 1 DW(1024) MFLOPS: 0.90 Rate:3397
26 [ 7.65% : 2.7 ms] Convolution idx: 81 shape: {1 1024 7 7} -> {1 1024 7 7} fp32 -> fp32 K: 1x1 | S: 1x1 | P: 0 0 0 0 MFLOPS:102.76 Rate:37588
27 [ 0.08% : 0.0 ms] Pooling idx: 84 shape: {1 1024 7 7} -> {1 1024 1 1} fp32 -> fp32 K: 7x7 | S: 1x1 | P: 0 0 0 0 Avg
28 [ 1.07% : 0.4 ms] Convolution idx: 85 shape: {1 1024 1 1} -> {1 1000 1 1} fp32 -> fp32 K: 1x1 | S: 1x1 | P: 0 0 0 0 MFLOPS: 2.05 Rate:5360
total time: 422.97 ms. avg time: 42.30 ms. min time: 35.73 ms.
精度 Profiler
精度 Profiler，用于 CPU 后端运行网络模型后，导出每一层的 input/ouput tensor data，用于分析输出结果异常的问题。
使用方法
程序执行前，添加环境变量
export TG_DEBUG_DATA=1，启用精度 Profiler 功能；删除环境变量
unset TG_DEBUG_DATA， 关闭精度 Profiler 功能。
logo 信息
数据导出后，在程序执行的当前路径下生成 ./output 文件夹。
$ ./tm_classification -m models/squeezenet.tmfile -i images/cat.jpg
model file : models/squeezenet.tmfile
image file : images/cat.jpg
label_file : (null)
img_h, img_w, scale[3], mean[3] : 227 227 , 1.000 1.000 1.000, 104.0 116.7 122.7
Repeat 1 times, thread 1, avg time 4402.85 ms, max_time 4402.85 ms, min_time 4402.85 ms
--------------------------------------
0.273199, 281
0.267552, 282
0.181004, 278
0.081799, 285
0.072407, 151
--------------------------------------
$ ls ./output
conv1-conv1-bn-conv1-scale-relu1_in_blob_data.txt
conv1-conv1-bn-conv1-scale-relu1_out_blob_data.txt
conv2_1-dw-conv2_1-dw-bn-conv2_1-dw-scale-relu2_1-dw_in_blob_data.txt
conv2_1-dw-conv2_1-dw-bn-conv2_1-dw-scale-relu2_1-dw_out_blob_data.txt
conv2_1-sep-conv2_1-sep-bn-conv2_1-sep-scale-relu2_1-sep_in_blob_data.txt
conv2_1-sep-conv2_1-sep-bn-conv2_1-sep-scale-relu2_1-sep_out_blob_data.txt
conv2_2-dw-conv2_2-dw-bn-conv2_2-dw-scale-relu2_2-dw_in_blob_data.txt
conv2_2-dw-conv2_2-dw-bn-conv2_2-dw-scale-relu2_2-dw_out_blob_data.txt
conv2_2-sep-conv2_2-sep-bn-conv2_2-sep-scale-relu2_2-sep_in_blob_data.txt
conv2_2-sep-conv2_2-sep-bn-conv2_2-sep-scale-relu2_2-sep_out_blob_data.txt
conv3_1-dw-conv3_1-dw-bn-conv3_1-dw-scale-relu3_1-dw_in_blob_data.txt
conv3_1-dw-conv3_1-dw-bn-conv3_1-dw-scale-relu3_1-dw_out_blob_data.txt
conv3_1-sep-conv3_1-sep-bn-conv3_1-sep-scale-relu3_1-sep_in_blob_data.txt
conv3_1-sep-conv3_1-sep-bn-conv3_1-sep-scale-relu3_1-sep_out_blob_data.txt
conv3_2-dw-conv3_2-dw-bn-conv3_2-dw-scale-relu3_2-dw_in_blob_data.txt
conv3_2-dw-conv3_2-dw-bn-conv3_2-dw-scale-relu3_2-dw_out_blob_data.txt
conv3_2-sep-conv3_2-sep-bn-conv3_2-sep-scale-relu3_2-sep_in_blob_data.txt
conv3_2-sep-conv3_2-sep-bn-conv3_2-sep-scale-relu3_2-sep_out_blob_data.txt
conv4_1-dw-conv4_1-dw-bn-conv4_1-dw-scale-relu4_1-dw_in_blob_data.txt
conv4_1-dw-conv4_1-dw-bn-conv4_1-dw-scale-relu4_1-dw_out_blob_data.txt
conv4_1-sep-conv4_1-sep-bn-conv4_1-sep-scale-relu4_1-sep_in_blob_data.txt
conv4_1-sep-conv4_1-sep-bn-conv4_1-sep-scale-relu4_1-sep_out_blob_data.txt
conv4_2-dw-conv4_2-dw-bn-conv4_2-dw-scale-relu4_2-dw_in_blob_data.txt
conv4_2-dw-conv4_2-dw-bn-conv4_2-dw-scale-relu4_2-dw_out_blob_data.txt
conv4_2-sep-conv4_2-sep-bn-conv4_2-sep-scale-relu4_2-sep_in_blob_data.txt
conv4_2-sep-conv4_2-sep-bn-conv4_2-sep-scale-relu4_2-sep_out_blob_data.txt
conv5_1-dw-conv5_1-dw-bn-conv5_1-dw-scale-relu5_1-dw_in_blob_data.txt
conv5_1-dw-conv5_1-dw-bn-conv5_1-dw-scale-relu5_1-dw_out_blob_data.txt
conv5_1-sep-conv5_1-sep-bn-conv5_1-sep-scale-relu5_1-sep_in_blob_data.txt
conv5_1-sep-conv5_1-sep-bn-conv5_1-sep-scale-relu5_1-sep_out_blob_data.txt
conv5_2-dw-conv5_2-dw-bn-conv5_2-dw-scale-relu5_2-dw_in_blob_data.txt
conv5_2-dw-conv5_2-dw-bn-conv5_2-dw-scale-relu5_2-dw_out_blob_data.txt
conv5_2-sep-conv5_2-sep-bn-conv5_2-sep-scale-relu5_2-sep_in_blob_data.txt
conv5_2-sep-conv5_2-sep-bn-conv5_2-sep-scale-relu5_2-sep_out_blob_data.txt
conv5_3-dw-conv5_3-dw-bn-conv5_3-dw-scale-relu5_3-dw_in_blob_data.txt
conv5_3-dw-conv5_3-dw-bn-conv5_3-dw-scale-relu5_3-dw_out_blob_data.txt
conv5_3-sep-conv5_3-sep-bn-conv5_3-sep-scale-relu5_3-sep_in_blob_data.txt
conv5_3-sep-conv5_3-sep-bn-conv5_3-sep-scale-relu5_3-sep_out_blob_data.txt
conv5_4-dw-conv5_4-dw-bn-conv5_4-dw-scale-relu5_4-dw_in_blob_data.txt
conv5_4-dw-conv5_4-dw-bn-conv5_4-dw-scale-relu5_4-dw_out_blob_data.txt
conv5_4-sep-conv5_4-sep-bn-conv5_4-sep-scale-relu5_4-sep_in_blob_data.txt
conv5_4-sep-conv5_4-sep-bn-conv5_4-sep-scale-relu5_4-sep_out_blob_data.txt
conv5_5-dw-conv5_5-dw-bn-conv5_5-dw-scale-relu5_5-dw_in_blob_data.txt
conv5_5-dw-conv5_5-dw-bn-conv5_5-dw-scale-relu5_5-dw_out_blob_data.txt
conv5_5-sep-conv5_5-sep-bn-conv5_5-sep-scale-relu5_5-sep_in_blob_data.txt
conv5_5-sep-conv5_5-sep-bn-conv5_5-sep-scale-relu5_5-sep_out_blob_data.txt
conv5_6-dw-conv5_6-dw-bn-conv5_6-dw-scale-relu5_6-dw_in_blob_data.txt
conv5_6-dw-conv5_6-dw-bn-conv5_6-dw-scale-relu5_6-dw_out_blob_data.txt
conv5_6-sep-conv5_6-sep-bn-conv5_6-sep-scale-relu5_6-sep_in_blob_data.txt
conv5_6-sep-conv5_6-sep-bn-conv5_6-sep-scale-relu5_6-sep_out_blob_data.txt
conv6-dw-conv6-dw-bn-conv6-dw-scale-relu6-dw_in_blob_data.txt
conv6-dw-conv6-dw-bn-conv6-dw-scale-relu6-dw_out_blob_data.txt
conv6-sep-conv6-sep-bn-conv6-sep-scale-relu6-sep_in_blob_data.txt
conv6-sep-conv6-sep-bn-conv6-sep-scale-relu6-sep_out_blob_data.txt
fc7_in_blob_data.txt
fc7_out_blob_data.txt
pool6_in_blob_data.txt
pool6_out_blob_data.txt
Naive Profiler
Naive Profiler，用于关闭 CPU 性能算子，后端计算只使用 Naive C 实现的 reference op，用于对比分析性能算子的计算结果。
使用方法
程序执行前，添加环境变量
export TG_DEBUG_REF=1，启用 Naive Profiler 功能；删除环境变量
unset TG_DEBUG_REF， 关闭精度 Naive Profiler 功能。