"--fine_grained" has some problems when runing vaitrace on KV260, vitis 3.5, about vitis-ai

[jimheaton]

To use fine grained profiling debug mode needs to be enabled when compiling the xmodel.
Recompile the model in the vai compiler and add the following option: --options "{'mode':'debug'}"
The default option for mode is 'normal'. Using debug mode can slightly reduce the model performance by a few percent and its recommended to set this back to normal when after you are done profiling.

[TiantianZhang]

To use fine grained profiling debug mode needs to be enabled when compiling the xmodel. Recompile the model in the vai compiler and add the following option: --options "{'mode':'debug'}" The default option for mode is 'normal'. Using debug mode can slightly reduce the model performance by a few percent and its recommended to set this back to normal when after you are done profiling.

---

@jimheaton

I use the following command to achieve the compiling
vai_q_tensorflow2 -m quantized_model.h5 -a /opt/vitis_ai/compiler/arch/DPUCZDX8G/KV260/arch.json -o dpu_deploy.xmodel

it should be revised as
vai_q_tensorflow2 -m quantized_model.h5 -a /opt/vitis_ai/compiler/arch/DPUCZDX8G/KV260/arch.json -o dpu_deploy.xmodel --options "{'mode':'debug'}" ???

[jimheaton]

Yes, use: vai_q_tensorflow2 -m quantized_model.h5 -a /opt/vitis_ai/compiler/arch/DPUCZDX8G/KV260/arch.json -o dpu_deploy.xmodel --options "{'mode':'debug'}"

[TiantianZhang]

thank you！I will try

[TiantianZhang]

I recompile the xmodel, and run it in --fine_grained, the results are not in expectations, I'm not sure if there are any other solutions. I have unified all versions to Vitis ai2.5

vai_c_tensorflow2 -m quantized_model.h5 -a /opt/vitis_ai/compiler/arch/DPUCZDX8G/KV260/arch.json -o dpu_deploy.xmodel --options "{'mode':'debug'}"

---

• VITIS_AI Compilation - Xilinx Inc.

---

[INFO] Namespace(batchsize=1, inputs_shape=None, layout='NHWC', model_files=['quantized_model.h5'], model_type='tensorflow2', named_inputs_shape=None, out_filename='/tmp/deploy_DPUCZDX8G_ISA1_B4096_org.xmodel', proto=None)
[INFO] tensorflow2 model: /workspace/workspace/rff_fpga_workspace/resnet_fpga_rff_5000/quantized_model.h5
[INFO] keras version: 2.6.0
[INFO] Tensorflow Keras model type: functional
[INFO] parse raw model :100%|██████████████████████████████████████████████████████████████████████████████████████████████████| 51/51 [00:00<00:00, 54154.30it/s]
[INFO] infer shape (NHWC) :100%|██████████████████████████████████████████████████████████████████████████████████████████████████| 83/83 [00:00<00:00, 44608.82it/s]
[INFO] perform level-0 opt :100%|██████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 294.54it/s]
[INFO] perform level-1 opt :100%|██████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 1286.00it/s]
[INFO] generate xmodel :100%|██████████████████████████████████████████████████████████████████████████████████████████████████| 83/83 [00:00<00:00, 5635.59it/s]
[INFO] dump xmodel: /tmp/deploy_DPUCZDX8G_ISA1_B4096_org.xmodel
[UNILOG][INFO] Compile mode: dpu
[UNILOG][INFO] Debug mode: null
[UNILOG][INFO] Target architecture: DPUCZDX8G_ISA1_B4096
[UNILOG][INFO] Graph name: model, with op num: 165
[UNILOG][INFO] Begin to compile...
[UNILOG][INFO] Total device subgraph number 3, DPU subgraph number 1
[UNILOG][INFO] Compile done.
[UNILOG][INFO] The meta json is saved to "/workspace/workspace/dpu_deploy.xmodel/meta.json"
[UNILOG][INFO] The compiled xmodel is saved to "/workspace/workspace/dpu_deploy.xmodel/deploy.xmodel"
[UNILOG][INFO] The compiled xmodel's md5sum is d4223604745d48da5f4de8069d5ec89b, and has been saved to "/workspace/workspace//dpu_deploy.xmodel/md5sum.txt"

python3 -m vaitrace_py --fine_grained test_accuracy_core1.py deploy.xmodel

WARNING:root:The fine grained profiling model is enabled. Models will be run subgraph-by-subgraph.

1. It will interduce significant scheduling and memory access overhead, especially for small models.
2. Some xcompiler's optimizations will be turned-off.

INFO:root:VART will run xmodel in [DEBUG] mode
Analyzing symbol tables...
61 / 61
81 / 81
6 / 6
45 / 45
3 / 3
INFO:root:vaitrace compile python code: test_accuracy_core1.py
INFO:root:vaitrace exec poython code: test_accuracy_core1.py
['test_accuracy_core1.py', 'deploy.xmodel']
WARNING: Logging before InitGoogleLogging() is written to STDERR
I0311 18:56:16.364677 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Block1_conv has no mc_code
I0311 18:56:16.364840 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Block2_1_conv1 has no mc_code
I0311 18:56:16.364879 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Block2_1_conv2 has no mc_code
I0311 18:56:16.364918 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Block2_2_conv1 has no mc_code
I0311 18:56:16.364955 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Block2_2_conv2 has no mc_code
I0311 18:56:16.364997 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Block3_1_conv1 has no mc_code
I0311 18:56:16.365039 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Block3_1_conv2 has no mc_code
I0311 18:56:16.365082 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Block3_1_shortcut has no mc_code
I0311 18:56:16.365120 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Block3_2_conv1 has no mc_code
I0311 18:56:16.365159 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Block3_2_conv2 has no mc_code
I0311 18:56:16.365197 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Block4_1_conv1 has no mc_code
I0311 18:56:16.365236 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Block4_1_conv2 has no mc_code
I0311 18:56:16.365274 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Block4_1_shortcut has no mc_code
I0311 18:56:16.365314 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Block4_2_conv1 has no mc_code
I0311 18:56:16.365355 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Block4_2_conv2 has no mc_code
I0311 18:56:16.365394 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Block5_1_conv1 has no mc_code
I0311 18:56:16.365435 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Block5_1_conv2 has no mc_code
I0311 18:56:16.365476 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Block5_1_shortcut has no mc_code
I0311 18:56:16.365520 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Block5_2_conv1 has no mc_code
I0311 18:56:16.365562 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Block5_2_conv2 has no mc_code
I0311 18:56:16.365604 2135 dpu_kernel.cpp:178] child_name subgraph_quant_flatten_fix_reshaped has no mc_code
I0311 18:56:16.365641 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Dense_dense(TransferMatMulToConv2d) has no mc_code
I0311 18:56:16.365684 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Dense_dense(TransferMatMulToConv2d)_inserted_fix_2_reshaped has no mc_code
Number of RF Samples Tested is 3499
Batch Size is 1
Top1 accuracy = 0.0286
Pred Performance = 100
identification_time_per_slice=0.0006, total time = 2.0432 , total slice = 3500.0000
INFO:root:vaitrace timeout, waiting for Python thread terminated
APM Stop Collecting
INFO:root:Generating VTF
INFO:root:Processing xmodel information
INFO:root:[vart_trace.csv]: Events number: 0

[jimheaton]

What results were unexpected, are you referring to the bad Top1 accuracy you are getting?

[TiantianZhang]

there are errors in the log. address that there is no mc_code. if i run it without —fine_grained，there only 1 layer information and the accuracy will be normal WARNING: Logging before InitGoogleLogging() is written to STDERR I0311 18:56:16.364677 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Block1_conv has no mc_code I0311 18:56:16.364840 2135 dpu_kernel.cpp:178] child_name subgraph_quant_Block2_1_conv1 has no mc_code