J721EXSOMXEVM: Unable to compile a custom neural network(EAST) using TIDL

Vaibhav Kashera

Part Number: J721EXSOMXEVM

num_bits = 8
accuracy = 1
output_dir = 'model-artifacts/onnx/east'
compile_options = {
    'tidl_tools_path': os.environ['TIDL_TOOLS_PATH'],
    'artifacts_folder': output_dir,
    'tensor_bits': num_bits,
    'accuracy_level': accuracy,
    'advanced_options:calibration_frames': len(calib_images),
    'advanced_options:calibration_iterations': 3,  # used if accuracy_level = 1
    'debug_level': 3,
    # Comma separated string of operator types as defined by ONNX runtime, ex "MaxPool, Concat"
    'deny_list': "Resize"
}

This is the model configuration we are using.

This function is called from __main__

def test_onnx(onnx_path, test_image_path):
    print("CHECK - 1 ")
    try:
        onnx.shape_inference.infer_shapes_path(onnx_path, onnx_path)
    except Exception as e:
        print(e)
        # return
    so = onnxruntime.SessionOptions()
    # so.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_DISABLE_ALL
    EP_list = ['TIDLCompilationProvider','CPUExecutionProvider']
    try:
        ort_session = onnxruntime.InferenceSession(
            onnx_path, providers=EP_list, provider_options=[compile_options,{}], sess_options=so)
    except Exception as e:
        print("FAILED to load ort_session")
        print(e)
        return
    print("CHECK - 2")
    img = Image.open(test_image_path)
    # You may need the resize_img function from your original code
    img, ratio_h, ratio_w = resize_img(img)
    img_tensor = load_pil(img)
    print("IMAGE SHAPE TENSOR",np.shape(img_tensor))
    # Perform inference
    ort_inputs = {ort_session.get_inputs()[0].name: img_tensor.numpy()}
    for i in range(len(calib_images)):
        ort_outputs = ort_session.run(None, ort_inputs)

    # Post-process the inference results
    score = ort_outputs[0]
    geo = ort_outputs[1]
    boxes = get_boxes(score.squeeze(0), geo.squeeze(0))
    adjusted_boxes = adjust_ratio(boxes, ratio_w, ratio_h)

    # Visualize and save the results
    result_img = plot_boxes(img, adjusted_boxes)
    result_img.show()
    result_img.save("result.jpg")

Output logs: (Output includes the model structure of the model we are trying to use.)

opset version of the model: 13

Link to the original repository: https://github.com/SakuraRiven/EAST

The compilation is performed in a docker container with edgeai image (https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/advanced_setup.md#docker-based-setup-for-x86_pc)

root@904c189d1f61:/home/root# python3 onnx-detect.py 
graph torch_jit (
  %input[FLOAT, 1x3x704x1280]
) initializers (
  %output.conv1.weight[FLOAT, 1x32x1x1]
  %output.conv1.bias[FLOAT, 1]
  %output.conv2.weight[FLOAT, 4x32x1x1]
  %output.conv2.bias[FLOAT, 4]
  %output.conv3.weight[FLOAT, 1x32x1x1]
  %output.conv3.bias[FLOAT, 1]
  %onnx::Conv_244[FLOAT, 64x3x3x3]
  %onnx::Conv_245[FLOAT, 64]
  %onnx::Conv_247[FLOAT, 64x64x3x3]
  %onnx::Conv_248[FLOAT, 64]
  %onnx::Conv_250[FLOAT, 128x64x3x3]
  %onnx::Conv_251[FLOAT, 128]
  %onnx::Conv_253[FLOAT, 128x128x3x3]
  %onnx::Conv_254[FLOAT, 128]
  %onnx::Conv_256[FLOAT, 256x128x3x3]
  %onnx::Conv_257[FLOAT, 256]
  %onnx::Conv_259[FLOAT, 256x256x3x3]
  %onnx::Conv_260[FLOAT, 256]
  %onnx::Conv_262[FLOAT, 256x256x3x3]
  %onnx::Conv_263[FLOAT, 256]
  %onnx::Conv_265[FLOAT, 512x256x3x3]
  %onnx::Conv_266[FLOAT, 512]
  %onnx::Conv_268[FLOAT, 512x512x3x3]
  %onnx::Conv_269[FLOAT, 512]
  %onnx::Conv_271[FLOAT, 512x512x3x3]
  %onnx::Conv_272[FLOAT, 512]
  %onnx::Conv_274[FLOAT, 512x512x3x3]
  %onnx::Conv_275[FLOAT, 512]
  %onnx::Conv_277[FLOAT, 512x512x3x3]
  %onnx::Conv_278[FLOAT, 512]
  %onnx::Conv_280[FLOAT, 512x512x3x3]
  %onnx::Conv_281[FLOAT, 512]
  %onnx::Conv_283[FLOAT, 128x1024x1x1]
  %onnx::Conv_284[FLOAT, 128]
  %onnx::Conv_286[FLOAT, 128x128x3x3]
  %onnx::Conv_287[FLOAT, 128]
  %onnx::Conv_289[FLOAT, 64x384x1x1]
  %onnx::Conv_290[FLOAT, 64]
  %onnx::Conv_292[FLOAT, 64x64x3x3]
  %onnx::Conv_293[FLOAT, 64]
  %onnx::Conv_295[FLOAT, 32x192x1x1]
  %onnx::Conv_296[FLOAT, 32]
  %onnx::Conv_298[FLOAT, 32x32x3x3]
  %onnx::Conv_299[FLOAT, 32]
  %onnx::Conv_301[FLOAT, 32x32x3x3]
  %onnx::Conv_302[FLOAT, 32]
  %onnx::Resize_303[FLOAT, 4]
) {
  %onnx::Resize_305 = Identity(%onnx::Resize_303)
  %onnx::Resize_304 = Identity(%onnx::Resize_303)
  %input.4 = Conv[dilations = [1, 1], group = 1, kernel_shape = [3, 3], pads = [1, 1, 1, 1], strides = [1, 1]](%input, %onnx::Conv_244, %onnx::Conv_245)
  %onnx::Conv_149 = Relu(%input.4)
  %input.12 = Conv[dilations = [1, 1], group = 1, kernel_shape = [3, 3], pads = [1, 1, 1, 1], strides = [1, 1]](%onnx::Conv_149, %onnx::Conv_247, %onnx::Conv_248)
  %onnx::MaxPool_152 = Relu(%input.12)
  %input.16 = MaxPool[ceil_mode = 0, kernel_shape = [2, 2], pads = [0, 0, 0, 0], strides = [2, 2]](%onnx::MaxPool_152)
  %input.24 = Conv[dilations = [1, 1], group = 1, kernel_shape = [3, 3], pads = [1, 1, 1, 1], strides = [1, 1]](%input.16, %onnx::Conv_250, %onnx::Conv_251)
  %onnx::Conv_156 = Relu(%input.24)
  %input.32 = Conv[dilations = [1, 1], group = 1, kernel_shape = [3, 3], pads = [1, 1, 1, 1], strides = [1, 1]](%onnx::Conv_156, %onnx::Conv_253, %onnx::Conv_254)
  %onnx::MaxPool_159 = Relu(%input.32)
  %input.36 = MaxPool[ceil_mode = 0, kernel_shape = [2, 2], pads = [0, 0, 0, 0], strides = [2, 2]](%onnx::MaxPool_159)
  %input.44 = Conv[dilations = [1, 1], group = 1, kernel_shape = [3, 3], pads = [1, 1, 1, 1], strides = [1, 1]](%input.36, %onnx::Conv_256, %onnx::Conv_257)
  %onnx::Conv_163 = Relu(%input.44)
  %input.52 = Conv[dilations = [1, 1], group = 1, kernel_shape = [3, 3], pads = [1, 1, 1, 1], strides = [1, 1]](%onnx::Conv_163, %onnx::Conv_259, %onnx::Conv_260)
  %onnx::Conv_166 = Relu(%input.52)
  %input.60 = Conv[dilations = [1, 1], group = 1, kernel_shape = [3, 3], pads = [1, 1, 1, 1], strides = [1, 1]](%onnx::Conv_166, %onnx::Conv_262, %onnx::Conv_263)
  %onnx::MaxPool_169 = Relu(%input.60)
  %input.64 = MaxPool[ceil_mode = 0, kernel_shape = [2, 2], pads = [0, 0, 0, 0], strides = [2, 2]](%onnx::MaxPool_169)
  %input.72 = Conv[dilations = [1, 1], group = 1, kernel_shape = [3, 3], pads = [1, 1, 1, 1], strides = [1, 1]](%input.64, %onnx::Conv_265, %onnx::Conv_266)
  %onnx::Conv_173 = Relu(%input.72)
  %input.80 = Conv[dilations = [1, 1], group = 1, kernel_shape = [3, 3], pads = [1, 1, 1, 1], strides = [1, 1]](%onnx::Conv_173, %onnx::Conv_268, %onnx::Conv_269)
  %onnx::Conv_176 = Relu(%input.80)
  %input.88 = Conv[dilations = [1, 1], group = 1, kernel_shape = [3, 3], pads = [1, 1, 1, 1], strides = [1, 1]](%onnx::Conv_176, %onnx::Conv_271, %onnx::Conv_272)
  %onnx::MaxPool_179 = Relu(%input.88)
  %input.92 = MaxPool[ceil_mode = 0, kernel_shape = [2, 2], pads = [0, 0, 0, 0], strides = [2, 2]](%onnx::MaxPool_179)
  %input.100 = Conv[dilations = [1, 1], group = 1, kernel_shape = [3, 3], pads = [1, 1, 1, 1], strides = [1, 1]](%input.92, %onnx::Conv_274, %onnx::Conv_275)
  %onnx::Conv_183 = Relu(%input.100)
  %input.108 = Conv[dilations = [1, 1], group = 1, kernel_shape = [3, 3], pads = [1, 1, 1, 1], strides = [1, 1]](%onnx::Conv_183, %onnx::Conv_277, %onnx::Conv_278)
  %onnx::Conv_186 = Relu(%input.108)
  %input.116 = Conv[dilations = [1, 1], group = 1, kernel_shape = [3, 3], pads = [1, 1, 1, 1], strides = [1, 1]](%onnx::Conv_186, %onnx::Conv_280, %onnx::Conv_281)
  %onnx::MaxPool_189 = Relu(%input.116)
  %onnx::Resize_190 = MaxPool[ceil_mode = 0, kernel_shape = [2, 2], pads = [0, 0, 0, 0], strides = [2, 2]](%onnx::MaxPool_189)
  %onnx::Concat_195 = Resize[coordinate_transformation_mode = 'align_corners', cubic_coeff_a = -0.75, mode = 'linear', nearest_mode = 'floor'](%onnx::Resize_190, %, %onnx::Resize_303)
  %input.120 = Concat[axis = 1](%onnx::Concat_195, %input.92)
  %input.128 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%input.120, %onnx::Conv_283, %onnx::Conv_284)
  %input.132 = Relu(%input.128)
  %input.140 = Conv[dilations = [1, 1], group = 1, kernel_shape = [3, 3], pads = [1, 1, 1, 1], strides = [1, 1]](%input.132, %onnx::Conv_286, %onnx::Conv_287)
  %y = Relu(%input.140)
  %onnx::Concat_207 = Resize[coordinate_transformation_mode = 'align_corners', cubic_coeff_a = -0.75, mode = 'linear', nearest_mode = 'floor'](%y, %, %onnx::Resize_304)
  %input.144 = Concat[axis = 1](%onnx::Concat_207, %input.64)
  %input.152 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%input.144, %onnx::Conv_289, %onnx::Conv_290)
  %input.156 = Relu(%input.152)
  %input.164 = Conv[dilations = [1, 1], group = 1, kernel_shape = [3, 3], pads = [1, 1, 1, 1], strides = [1, 1]](%input.156, %onnx::Conv_292, %onnx::Conv_293)
  %y.3 = Relu(%input.164)
  %onnx::Concat_219 = Resize[coordinate_transformation_mode = 'align_corners', cubic_coeff_a = -0.75, mode = 'linear', nearest_mode = 'floor'](%y.3, %, %onnx::Resize_305)
  %input.168 = Concat[axis = 1](%onnx::Concat_219, %input.36)
  %input.176 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%input.168, %onnx::Conv_295, %onnx::Conv_296)
  %input.180 = Relu(%input.176)
  %input.188 = Conv[dilations = [1, 1], group = 1, kernel_shape = [3, 3], pads = [1, 1, 1, 1], strides = [1, 1]](%input.180, %onnx::Conv_298, %onnx::Conv_299)
  %input.192 = Relu(%input.188)
  %input.200 = Conv[dilations = [1, 1], group = 1, kernel_shape = [3, 3], pads = [1, 1, 1, 1], strides = [1, 1]](%input.192, %onnx::Conv_301, %onnx::Conv_302)
  %input.204 = Relu(%input.200)
  %onnx::Sigmoid_230 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%input.204, %output.conv1.weight, %output.conv1.bias)
  %output = Sigmoid(%onnx::Sigmoid_230)
  %onnx::Sigmoid_232 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%input.204, %output.conv2.weight, %output.conv2.bias)
  %onnx::Mul_233 = Sigmoid(%onnx::Sigmoid_232)
  %onnx::Mul_234 = Constant[value = <Scalar Tensor []>]()
  %onnx::Concat_235 = Mul(%onnx::Mul_233, %onnx::Mul_234)
  %onnx::Sigmoid_236 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%input.204, %output.conv3.weight, %output.conv3.bias)
  %onnx::Sub_237 = Sigmoid(%onnx::Sigmoid_236)
  %onnx::Sub_238 = Constant[value = <Scalar Tensor []>]()
  %onnx::Mul_239 = Sub(%onnx::Sub_237, %onnx::Sub_238)
  %onnx::Mul_240 = Constant[value = <Scalar Tensor []>]()
  %onnx::Concat_241 = Mul(%onnx::Mul_239, %onnx::Mul_240)
  %242 = Concat[axis = 1](%onnx::Concat_235, %onnx::Concat_241)
  return %output, %242
}
img_1.jpg
CHECK - 1 
tidl_tools_path                                 = /home/root/tidl_tools 
artifacts_folder                                = model-artifacts/onnx/east 
tidl_tensor_bits                                = 8 
debug_level                                     = 3 
num_tidl_subgraphs                              = 16 
tidl_denylist                                   = Resize   
tidl_denylist_layer_name                        = 
tidl_denylist_layer_type                         = 
tidl_allowlist_layer_name                        = 
model_type                                      =  
tidl_calibration_accuracy_level                 = 7 
tidl_calibration_options:num_frames_calibration = 2 
tidl_calibration_options:bias_calibration_iterations = 3 
mixed_precision_factor = -1.000000 
model_group_id = 0 
power_of_2_quantization                         = 2 
enable_high_resolution_optimization             = 0 
pre_batchnorm_fold                              = 1 
add_data_convert_ops                          = 0 
output_feature_16bit_names_list                 =  
m_params_16bit_names_list                       =  
reserved_compile_constraints_flag               = 1601 
ti_internal_reserved_1                          = 

 ****** WARNING : Network not identified as Object Detection network : (1) Ignore if network is not Object Detection network (2) If network is Object Detection network, please specify "model_type":"OD" as part of OSRT compilation options******

Supported TIDL layer type ---            Conv -- Conv_2 
Supported TIDL layer type ---            Relu -- Relu_3 
Supported TIDL layer type ---            Conv -- Conv_4 
Supported TIDL layer type ---            Relu -- Relu_5 
Supported TIDL layer type ---         MaxPool -- MaxPool_6 
Supported TIDL layer type ---            Conv -- Conv_7 
Supported TIDL layer type ---            Relu -- Relu_8 
Supported TIDL layer type ---            Conv -- Conv_9 
Supported TIDL layer type ---            Relu -- Relu_10 
Supported TIDL layer type ---         MaxPool -- MaxPool_11 
Supported TIDL layer type ---            Conv -- Conv_12 
Supported TIDL layer type ---            Relu -- Relu_13 
Supported TIDL layer type ---            Conv -- Conv_14 
Supported TIDL layer type ---            Relu -- Relu_15 
Supported TIDL layer type ---            Conv -- Conv_16 
Supported TIDL layer type ---            Relu -- Relu_17 
Supported TIDL layer type ---         MaxPool -- MaxPool_18 
Supported TIDL layer type ---            Conv -- Conv_19 
Supported TIDL layer type ---            Relu -- Relu_20 
Supported TIDL layer type ---            Conv -- Conv_21 
Supported TIDL layer type ---            Relu -- Relu_22 
Supported TIDL layer type ---            Conv -- Conv_23 
Supported TIDL layer type ---            Relu -- Relu_24 
Supported TIDL layer type ---         MaxPool -- MaxPool_25 
Supported TIDL layer type ---            Conv -- Conv_26 
Supported TIDL layer type ---            Relu -- Relu_27 
Supported TIDL layer type ---            Conv -- Conv_28 
Supported TIDL layer type ---            Relu -- Relu_29 
Supported TIDL layer type ---            Conv -- Conv_30 
Supported TIDL layer type ---            Relu -- Relu_31 
Supported TIDL layer type ---         MaxPool -- MaxPool_32 
Op type 'Resize'  added to unsupported nodes as specified in deny list 
Supported TIDL layer type ---          Concat -- Concat_34 
Supported TIDL layer type ---            Conv -- Conv_35 
Supported TIDL layer type ---            Relu -- Relu_36 
Supported TIDL layer type ---            Conv -- Conv_37 
Supported TIDL layer type ---            Relu -- Relu_38 
Op type 'Resize'  added to unsupported nodes as specified in deny list 
Supported TIDL layer type ---          Concat -- Concat_40 
Supported TIDL layer type ---            Conv -- Conv_41 
Supported TIDL layer type ---            Relu -- Relu_42 
Supported TIDL layer type ---            Conv -- Conv_43 
Supported TIDL layer type ---            Relu -- Relu_44 
Op type 'Resize'  added to unsupported nodes as specified in deny list 
Supported TIDL layer type ---          Concat -- Concat_46 
Supported TIDL layer type ---            Conv -- Conv_47 
Supported TIDL layer type ---            Relu -- Relu_48 
Supported TIDL layer type ---            Conv -- Conv_49 
Supported TIDL layer type ---            Relu -- Relu_50 
Supported TIDL layer type ---            Conv -- Conv_51 
Supported TIDL layer type ---            Relu -- Relu_52 
Supported TIDL layer type ---            Conv -- Conv_59 
Supported TIDL layer type ---         Sigmoid -- Sigmoid_60 
Supported TIDL layer type ---             Sub -- Sub_62 
Supported TIDL layer type ---             Mul -- Mul_64 
Supported TIDL layer type ---            Conv -- Conv_55 
Supported TIDL layer type ---         Sigmoid -- Sigmoid_56 
Supported TIDL layer type ---             Mul -- Mul_58 
Supported TIDL layer type ---          Concat -- Concat_65 
Supported TIDL layer type ---            Conv -- Conv_53 
Supported TIDL layer type ---         Sigmoid -- Sigmoid_54 

Preliminary subgraphs created = 4 
Final number of subgraphs created are : 4, - Offloaded Nodes - 58, Total Nodes - 61 
Node in deny list...delegated to ARM --- layer type - Resize, Node name - Resize_33  
Node in deny list...delegated to ARM --- layer type - Resize, Node name - Resize_39  
Node in deny list...delegated to ARM --- layer type - Resize, Node name - Resize_45  
/home/root/tidl_tools/tidl_graphVisualiser_runtimes.out: error while loading shared libraries: libcgraph.so.6: cannot open shared object file: No such file or directory
Running runtimes graphviz - /home/root/tidl_tools/tidl_graphVisualiser_runtimes.out model-artifacts/onnx/east/allowedNode.txt model-artifacts/onnx/east/tempDir/graphvizInfo.txt model-artifacts/onnx/east/tempDir/runtimes_visualization.svg 
*** In TIDL_createStateImportFunc *** 
Compute on node : TIDLExecutionProvider_TIDL_0_0
  0,            Conv, 3, 1, input, input.4
  1,            Relu, 1, 1, input.4, onnx::Conv_149
  2,            Conv, 3, 1, onnx::Conv_149, input.12
  3,            Relu, 1, 1, input.12, onnx::MaxPool_152
  4,         MaxPool, 1, 1, onnx::MaxPool_152, input.16
  5,            Conv, 3, 1, input.16, input.24
  6,            Relu, 1, 1, input.24, onnx::Conv_156
  7,            Conv, 3, 1, onnx::Conv_156, input.32
  8,            Relu, 1, 1, input.32, onnx::MaxPool_159
  9,         MaxPool, 1, 1, onnx::MaxPool_159, input.36
 10,            Conv, 3, 1, input.36, input.44
 11,            Relu, 1, 1, input.44, onnx::Conv_163
 12,            Conv, 3, 1, onnx::Conv_163, input.52
 13,            Relu, 1, 1, input.52, onnx::Conv_166
 14,            Conv, 3, 1, onnx::Conv_166, input.60
 15,            Relu, 1, 1, input.60, onnx::MaxPool_169
 16,         MaxPool, 1, 1, onnx::MaxPool_169, input.64
 17,            Conv, 3, 1, input.64, input.72
 18,            Relu, 1, 1, input.72, onnx::Conv_173
 19,            Conv, 3, 1, onnx::Conv_173, input.80
 20,            Relu, 1, 1, input.80, onnx::Conv_176
 21,            Conv, 3, 1, onnx::Conv_176, input.88
 22,            Relu, 1, 1, input.88, onnx::MaxPool_179
 23,         MaxPool, 1, 1, onnx::MaxPool_179, input.92
 24,            Conv, 3, 1, input.92, input.100
 25,            Relu, 1, 1, input.100, onnx::Conv_183
 26,            Conv, 3, 1, onnx::Conv_183, input.108
 27,            Relu, 1, 1, input.108, onnx::Conv_186
 28,            Conv, 3, 1, onnx::Conv_186, input.116
 29,            Relu, 1, 1, input.116, onnx::MaxPool_189
 30,         MaxPool, 1, 1, onnx::MaxPool_189, onnx::Resize_190

Input tensor name -  input 
Output tensor name - input.36 
Output tensor name - input.64 
Output tensor name - input.92 
Output tensor name - onnx::Resize_190 
*** In TIDL_createStateImportFunc *** 
Compute on node : TIDLExecutionProvider_TIDL_1_1
  0,          Concat, 2, 1, onnx::Concat_195, input.120
  1,            Conv, 3, 1, input.120, input.128
  2,            Relu, 1, 1, input.128, input.132
  3,            Conv, 3, 1, input.132, input.140
  4,            Relu, 1, 1, input.140, y

Input tensor name -  onnx::Concat_195 

Input tensor name -  input.92 
Output tensor name - y 
*** In TIDL_createStateImportFunc *** 
Compute on node : TIDLExecutionProvider_TIDL_2_2
  0,          Concat, 2, 1, onnx::Concat_207, input.144
  1,            Conv, 3, 1, input.144, input.152
  2,            Relu, 1, 1, input.152, input.156
  3,            Conv, 3, 1, input.156, input.164
  4,            Relu, 1, 1, input.164, y.3

Input tensor name -  onnx::Concat_207 

Input tensor name -  input.64 
Output tensor name - y.3 
*** In TIDL_createStateImportFunc *** 
Compute on node : TIDLExecutionProvider_TIDL_3_3
  0,          Concat, 2, 1, onnx::Concat_219, input.168
  1,            Conv, 3, 1, input.168, input.176
  2,            Relu, 1, 1, input.176, input.180
  3,            Conv, 3, 1, input.180, input.188
  4,            Relu, 1, 1, input.188, input.192
  5,            Conv, 3, 1, input.192, input.200
  6,            Relu, 1, 1, input.200, input.204
  7,            Conv, 3, 1, input.204, onnx::Sigmoid_230
  8,         Sigmoid, 1, 1, onnx::Sigmoid_230, output
  9,            Conv, 3, 1, input.204, onnx::Sigmoid_232
 10,         Sigmoid, 1, 1, onnx::Sigmoid_232, onnx::Mul_233
 11,             Mul, 2, 1, onnx::Mul_233, onnx::Concat_235
 12,            Conv, 3, 1, input.204, onnx::Sigmoid_236
 13,         Sigmoid, 1, 1, onnx::Sigmoid_236, onnx::Sub_237
 14,             Sub, 2, 1, onnx::Sub_237, onnx::Mul_239
 15,             Mul, 2, 1, onnx::Mul_239, onnx::Concat_241
 16,          Concat, 2, 1, onnx::Concat_235, 242

Input tensor name -  onnx::Concat_219 

Input tensor name -  input.36 
Output tensor name - 242 
Output tensor name - output 
CHECK - 2
IMAGE SHAPE TENSOR torch.Size([1, 3, 704, 1280])
In TIDL_onnxRtImportInit subgraph_name=onnx::Resize_190
Layer 0, subgraph id onnx::Resize_190, name=input.36
Layer 1, subgraph id onnx::Resize_190, name=input.64
Layer 2, subgraph id onnx::Resize_190, name=input.92
Layer 3, subgraph id onnx::Resize_190, name=onnx::Resize_190
Layer 4, subgraph id onnx::Resize_190, name=input
In TIDL_runtimesOptimizeNet: LayerIndex = 36, dataIndex = 32 

 ************** Frame index 1 : Running float import ************* 
In TIDL_runtimesPostProcessNet 
/home/root/tidl_tools/tidl_graphVisualiser.out: error while loading shared libraries: libcgraph.so.6: cannot open shared object file: No such file or directory
****************************************************
**                ALL MODEL CHECK PASSED          **
****************************************************

************ in TIDL_subgraphRtCreate ************ 
 The soft limit is 2048
The hard limit is 2048
MEM: Init ... !!!
MEM: Init ... Done !!!
 0.0s:  VX_ZONE_INIT:Enabled
 0.8s:  VX_ZONE_ERROR:Enabled
 0.10s:  VX_ZONE_WARNING:Enabled
 0.1596s:  VX_ZONE_INIT:[tivxInit:184] Initialization Done !!!
************ TIDL_subgraphRtCreate done ************ 
 *******   In TIDL_subgraphRtInvoke  ******** 
   0         1.00000        -1.00000         1.00000 6
   1         1.00000         0.00000        10.97984 6
   2         1.00000         0.00000         8.12576 6
   3         1.00000         0.00000         8.12576 6
   4         1.00000         0.00000         7.14118 6
   5         1.00000         0.00000         3.65851 6
   6         1.00000         0.00000         3.65851 6
   7         1.00000         0.00000         3.80828 6
   8         1.00000         0.00000         2.07062 6
   9         1.00000         0.00000         2.23841 6
  10         1.00000         0.00000         2.23841 6
  11         1.00000         0.00000         3.03889 6
  12         1.00000         0.00000         2.26227 6
  13         1.00000         0.00000         3.44511 6
  14         1.00000         0.00000         3.44511 6
  15         1.00000         0.00000         2.78007 6
  16         1.00000         0.00000         1.93240 6
  17         1.00000         0.00000         3.68336 6
  18         1.00000         0.00000         3.68336 6
 Layer,   Layer Cycles,kernelOnlyCycles, coreLoopCycles,LayerSetupCycles,dmaPipeupCycles, dmaPipeDownCycles, PrefetchCycles,copyKerCoeffCycles,LayerDeinitCycles,LastBlockCycles, paddingTrigger,    paddingWait,LayerWithoutPad,LayerHandleCopy,   BackupCycles,  RestoreCycles,

     1,              0,              0,              0,              0,              0,                 0,              0,                 0,              0,              0,              0,              0,              0,              0,              0,              0,
     2,              0,              0,              0,              0,              0,                 0,              0,                 0,              0,              0,              0,              0,              0,              0,              0,              0,
     3,              0,              0,              0,              0,              0,                 0,              0,                 0,              0,              0,              0,              0,              0,              0,              0,              0,
     4,              0,              0,              0,              0,              0,                 0,              0,                 0,              0,              0,              0,              0,              0,              0,              0,              0,
     5,              0,              0,              0,              0,              0,                 0,              0,                 0,              0,              0,              0,              0,              0,              0,              0,              0,
     6,              0,              0,              0,              0,              0,                 0,              0,                 0,              0,              0,              0,              0,              0,              0,              0,              0,
     7,              0,              0,              0,              0,              0,                 0,              0,                 0,              0,              0,              0,              0,              0,              0,              0,              0,
     8,              0,              0,              0,              0,              0,                 0,              0,                 0,              0,              0,              0,              0,              0,              0,              0,              0,
     9,              0,              0,              0,              0,              0,                 0,              0,                 0,              0,              0,              0,              0,              0,              0,              0,              0,
    10,              0,              0,              0,              0,              0,                 0,              0,                 0,              0,              0,              0,              0,              0,              0,              0,              0,
    11,              0,              0,              0,              0,              0,                 0,              0,                 0,              0,              0,              0,              0,              0,              0,              0,              0,
    12,              0,              0,              0,              0,              0,                 0,              0,                 0,              0,              0,              0,              0,              0,              0,              0,              0,
    13,              0,              0,              0,              0,              0,                 0,              0,                 0,              0,              0,              0,              0,              0,              0,              0,              0,
    14,              0,              0,              0,              0,              0,                 0,              0,                 0,              0,              0,              0,              0,              0,              0,              0,              0,
    15,              0,              0,              0,              0,              0,                 0,              0,                 0,              0,              0,              0,              0,              0,              0,              0,              0,
    16,              0,              0,              0,              0,              0,                 0,              0,                 0,              0,              0,              0,              0,              0,              0,              0,              0,
    17,              0,              0,              0,              0,              0,                 0,              0,                 0,              0,              0,              0,              0,              0,              0,              0,              0,
    18,              0,              0,              0,              0,              0,                 0,              0,                 0,              0,              0,              0,              0,              0,              0,              0,              0,




 Sum of Layer Cycles 0 
Segmentation fault (core dumped)

over 1 year ago

0 Pratik Kedar over 1 year ago

TI__Mastermind 24041 points

Hi,

We currently support TIDL till opset version 11, opset version 13 which you are using above are not supported.

I would suggest to migrate to opset 11 if its possible to do so and redo the above experiment.

Please note that opset version 18 support will be available in next SDK release.

0 Vaibhav Kashera over 1 year ago in reply to Pratik Kedar

Prodigy 60 points

Hi Pratik,

We tried our model with opset 11. We are getting the same error while compiling the model. Request you to suggest how to move forward.

We also tried with making accuracy = 0 instead of accuracy=1

Thanks,

Vaibhav

0 Pratik Kedar over 1 year ago in reply to Vaibhav Kashera

TI__Mastermind 24041 points

Docker logs shared above are for opset 11 based model ?

If not could you shared the same ?

Also what SDK version you are using ?

0 Vaibhav Kashera over 1 year ago in reply to Pratik Kedar

Prodigy 60 points

Yes, the docker logs shared above are for the opset 11 based model. To add, both opset 11 and 13 models go through the same stages before throwing a segmentation fault error.

Please find the model here: https://drive.google.com/file/d/1z88wxG1niZmBGok-RiwcfbpMT0q3zto5/view?usp=sharing

We ran this on the Docker setup based on ubuntu 18.04.6 LTS, edge ai tidl tools tag 08_06_00_03 (https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/version_compatibility_table.md)

Thanks,

Vaibhav

0 Vaibhav Kashera over 1 year ago in reply to Vaibhav Kashera

Prodigy 60 points

Hi,

We tried compiling the model using version 9 of psdk tools. We used PROCESSOR-SDK-LINUX-SK-TDA4VM ( Version: 09.00.00.08 Release date: 11 Aug) (https://www.ti.com/tool/download/PROCESSOR-SDK-LINUX-SK-TDA4VM/09.00.00.08)

Along with this we used edgeai-tidl tools based on ubuntu 22, github tag 09_00_00_06 to compile the same model shared above.

Here are the compile options:

calib_images = [
    'img_2.jpg',
    'img_3.jpg',
    'img_1.jpg'
]

num_bits = 8
accuracy = 1
output_dir = 'model-artifacts/onnx/east-artifacts-deny-resize'
compile_options = {
    'tidl_tools_path': os.environ['TIDL_TOOLS_PATH'],
    'artifacts_folder': output_dir,
    'tensor_bits': num_bits,
    'accuracy_level': accuracy,
    'advanced_options:calibration_frames': len(calib_images),
    'advanced_options:calibration_iterations': 3,  # used if accuracy_level = 1
    'debug_level': 3,
    # Comma separated string of operator types as defined by ONNX runtime, ex "MaxPool, Concat"
    'deny_list': "Resize"
}

While the compilation and heterogenous inference on PC is successful without errors, executing the same model on the evm throws errors:

root@j721e-evm:/opt/edgeai-tidl-tools# python3 onnx-detect.py 
img_1.jpg
CHECK - 1 
libtidl_onnxrt_EP loaded 0x2742ff20 
artifacts_folder                                = model-artifacts/onnx/east-artifacts-deny-resize 
debug_level                                     = 3 
target_priority                                 = 0 
max_pre_empt_delay                              = 340282346638528859811704183484516925440.000000 
Final number of subgraphs created are : 4, - Offloaded Nodes - 58, Total Nodes - 61 
In TIDL_createStateInfer 
Compute on node : TIDLExecutionProvider_TIDL_0_0
************ in TIDL_subgraphRtCreate ************ 
 APP: Init ... !!!
MEM: Init ... !!!
MEM: Initialized DMA HEAP (fd=5) !!!
MEM: Init ... Done !!!
IPC: Init ... !!!
IPC: Init ... Done !!!
REMOTE_SERVICE: Init ... !!!
REMOTE_SERVICE: Init ... Done !!!
   263.987894 s: GTC Frequency = 200 MHz
APP: Init ... Done !!!
   263.991144 s:  VX_ZONE_INIT:Enabled
   263.991155 s:  VX_ZONE_ERROR:Enabled
   263.991161 s:  VX_ZONE_WARNING:Enabled
   263.994832 s:  VX_ZONE_INIT:[tivxInitLocal:130] Initialization Done !!!
   263.997454 s:  VX_ZONE_INIT:[tivxHostInitLocal:96] Initialization Done for HOST !!!
   264.114180 s:  VX_ZONE_ERROR:[ownContextSendCmd:822] Command ack message returned failure cmd_status: -1
   264.114233 s:  VX_ZONE_ERROR:[ownContextSendCmd:862] tivxEventWait() failed.
   264.114241 s:  VX_ZONE_ERROR:[ownNodeKernelInit:527] Target kernel, TIVX_CMD_NODE_CREATE failed for node TIDLNode
   264.114248 s:  VX_ZONE_ERROR:[ownNodeKernelInit:528] Please be sure the target callbacks have been registered for this core
   264.114254 s:  VX_ZONE_ERROR:[ownNodeKernelInit:529] If the target callbacks have been registered, please ensure no errors are occurring within the create callback of this kernel
   264.114262 s:  VX_ZONE_ERROR:[ownGraphNodeKernelInit:583] kernel init for node 0, kernel com.ti.tidl:1:4 ... failed !!!
   264.114273 s:  VX_ZONE_ERROR:[vxVerifyGraph:2055] Node kernel init failed
   264.114280 s:  VX_ZONE_ERROR:[vxVerifyGraph:2109] Graph verify failed
TIDL_RT_OVX: ERROR: Verifying TIDL graph ... Failed !!!
TIDL_RT_OVX: ERROR: Verify OpenVX graph failed
************ TIDL_subgraphRtCreate done ************ 
 In TIDL_createStateInfer 
Compute on node : TIDLExecutionProvider_TIDL_1_1
************ in TIDL_subgraphRtCreate ************ 
 ************ TIDL_subgraphRtCreate done ************ 
 In TIDL_createStateInfer 
Compute on node : TIDLExecutionProvider_TIDL_2_2
************ in TIDL_subgraphRtCreate ************ 
 ************ TIDL_subgraphRtCreate done ************ 
 In TIDL_createStateInfer 
Compute on node : TIDLExecutionProvider_TIDL_3_3
************ in TIDL_subgraphRtCreate ************ 
 ************ TIDL_subgraphRtCreate done ************ 
CHECK - 2
IMAGE SHAPE TENSOR torch.Size([1, 3, 704, 1280])
 *******   In TIDL_subgraphRtInvoke  ******** 
   264.699962 s:  VX_ZONE_ERROR:[ownContextSendCmd:822] Command ack message returned failure cmd_status: -1
   264.699984 s:  VX_ZONE_ERROR:[ownContextSendCmd:862] tivxEventWait() failed.
   264.699992 s:  VX_ZONE_ERROR:[ownNodeKernelInit:527] Target kernel, TIVX_CMD_NODE_CREATE failed for node TIDLNode
   264.699999 s:  VX_ZONE_ERROR:[ownNodeKernelInit:528] Please be sure the target callbacks have been registered for this core
   264.700005 s:  VX_ZONE_ERROR:[ownNodeKernelInit:529] If the target callbacks have been registered, please ensure no errors are occurring within the create callback of this kernel
   264.700012 s:  VX_ZONE_ERROR:[ownGraphNodeKernelInit:583] kernel init for node 0, kernel com.ti.tidl:1:4 ... failed !!!
   264.700020 s:  VX_ZONE_ERROR:[vxVerifyGraph:2055] Node kernel init failed
   264.700026 s:  VX_ZONE_ERROR:[vxVerifyGraph:2109] Graph verify failed
   264.700147 s:  VX_ZONE_ERROR:[ownGraphScheduleGraphWrapper:799] graph is not in a state required to be scheduled
   264.700153 s:  VX_ZONE_ERROR:[vxProcessGraph:734] schedule graph failed
   264.700159 s:  VX_ZONE_ERROR:[vxProcessGraph:739] wait graph failed
ERROR: Running TIDL graph ... Failed !!!
Sub Graph Stats 4185.000000 880.000000 9725265032722382.000000 
*******  TIDL_subgraphRtInvoke done  ******** 
*******   In TIDL_subgraphRtInvoke  ********

Please respond on how we can go about fixing this error to execute on the evaluation board (J721EXSOMXEVM). We have used the tisdk-edgeai-image.wic.xz (https://dr-download.ti.com/software-development/software-development-kit-sdk/MD-4K6R4tqhZI/09.00.00.08/tisdk-edgeai-image-j721e-evm.wic.xz) to flash the SD card for the v9 of psdk.

Regards,

Vaibhav

0 Pratik Kedar over 1 year ago in reply to Vaibhav Kashera

TI__Mastermind 24041 points

Thanks for sharing the logs.

Will get back to you on this.

0 Vaibhav Kashera over 1 year ago in reply to Pratik Kedar

Prodigy 60 points

Hi Pratik,

I am unable to reproduce the successful compilation for "Resize" in the deny list. The compilation and PC inference are still successful when no operators are denied to the accelerator manually:

num_bits = 8
accuracy = 1
output_dir = f'model-artifacts/onnx/east-original'
print(output_dir)
compile_options = {
    'tidl_tools_path': os.environ['TIDL_TOOLS_PATH'],
    'artifacts_folder': output_dir,
    'tensor_bits': num_bits,
    'accuracy_level': accuracy,
    'advanced_options:calibration_frames': len(calib_images),
    'advanced_options:calibration_iterations': 3,  # used if accuracy_level = 1
    'debug_level': 3,
    # Comma separated string of operator types as defined by ONNX runtime, ex "MaxPool, Concat"
    'deny_list': ""
    }

I am requesting you to look into this configuration instead. The output logs contain the same graph verification errors as uploaded for the previous configuration.

One more issue I am facing while using the docker container with edgeai v9 is that I need to restart the container otherwise I keep getting Bus error(core dumped) for the same scripts which executed successfully. Is this normal or something that can be fixed?

Regards,

Vaibhav

0 Pratik Kedar over 1 year ago in reply to Vaibhav Kashera

TI__Mastermind 24041 points

Hi,

Please expect slight delay in responses as our india team is on festive break.

will get back to you

0 Vaibhav Kashera over 1 year ago in reply to Pratik Kedar

Prodigy 60 points

Hi Pratik,

Thank you for your response. There is an update noticed on our end, which I have elaborated in this post: https://e2e.ti.com/support/processors-group/processors/f/processors-forum/1285195/j721exsomxevm-facing-compile-issues-with-edgeai-tidl-tools-on-edgeai-zoo-models

Could you please look at this at reach back?

Thanks & Regards,

Vaibhav

0 Pratik Kedar over 1 year ago in reply to Vaibhav Kashera

TI__Mastermind 24041 points

Hi,

Could you attach the zip file of model here ? External site access is restricted at our end.

Can you confirm this are you final compilation flags ?

compile_options = {
'tidl_tools_path': os.environ['TIDL_TOOLS_PATH'],
'artifacts_folder': output_dir,
'tensor_bits': num_bits,
'accuracy_level': accuracy,
'advanced_options:calibration_frames': len(calib_images),
'advanced_options:calibration_iterations': 3, # used if accuracy_level = 1
'debug_level': 3,
# Comma separated string of operator types as defined by ONNX runtime, ex "MaxPool, Concat"
'deny_list': ""
}

0 Vaibhav Kashera over 1 year ago in reply to Pratik Kedar

Prodigy 60 points

east-opset-11.zip

Hi Pratik,

On the previous reply, we figured out that the issue lies in the https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/advanced_setup.md docker setup instructions file. The instructions/version compatibility documentation have not been updated after the latest release of bug fixes for illegal instruction(core dumped). The documentation(https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/version_compatibility_table.md) points to use 09_00_00_06 branch of edgeai-tidl-tools as 09_00_00_07 does not work.

To fix the issue, one must stay on the master branch while compiling the setup for the TIDL tools when using v9 of tidl-tools. Also request you to please update this on the github repository as the current version table/documentation is very misleading.

The EAST model compiles on v9 of tidl tools, but gives segmentation fault on v8. Attached is the model for your reference. The issue can be marked as non urgent for now, as we are able to use it with v9 (master branch edgeai-tidl-tools) to compile and run inference.

Regards,

Vaibhav Kashera

0 Pratik Kedar over 1 year ago in reply to Vaibhav Kashera

TI__Mastermind 24041 points

Hi Vaibhav,

We look forward for such positive feedback and personally want to thanks you for pointing this out to us.

As a action item, I will file JIRA for above mentioned changes, please be assured.

Thanks again.

Closing this thread.

Processors

Processors forum

J721EXSOMXEVM: Unable to compile a custom neural network(EAST) using TIDL