• State TI Thinks Resolved
  • Locked Locked
  • Replies 4 replies
  • Answers 2 answers
  • Subscribers 98 subscribers
  • Views 442 views
  • Users 0 members are here
Support feedback
Options
Options
Related

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

Original question:

TDA2EVM5777: [TIDL] Queries about caffe-jacinto layers

TDA2SG: jacinto-ai-devkit

reset_zhang
Intellectual 850 points
Part Number: TDA2SG
Other Parts Discussed in Thread: TDA2

I  Read  caffe-jacinto-model CNN train；as below

Stage-1: Initial stage with L2 regularization training

Stage-2: L1 regularization training

Stage-3: Sparsification training

The main training script is located ../scripts/train_image_object_detection.sh.

My question  Is：

1，Is  stage-2  train   need depend Stage-1，Stage-3  train   need depend Stage-2.

or  Stage-1，Stage-2，Stage-3 Is independent，I  can train this  stage 1 2 3 at the  same time。,

2，I see Stage 1,2,3 mAp is almost  equal，so  if I can  train one of them，not need to train  every Stage。

3，CAN   TDA4 and TDA2S  ues the same weight  file（caffemodel）   to run the  TIDL  on the usecase。

 

  • 0
    TI__Genius 11466 points

    Hi,

    If your purpose is to simply run on TDA2, the Stage1 is sufficient. The additional stages after this are required to generate a sparse model, which will run faster on TDA2. They have to be run sequentially, not in parallel.

    Yes, the same model is expected to run on TDA4 as well. The only difference is that for TDA4 Stage1 itself offer very high speed (other stages are not required as TDA4 is already much faster than TDA2 interms of CNN performance and sparse model doesn't offer additional speedup)

    Best regards,

  • 0 in reply to Manu Mathew
    Intellectual 850 points

    thanks。

    1,when use tidl_model_import.out.exe,  how can I   get   layersGroupId  and  conv2dKernelType  value,when I change the model (link ssd 512*512),

    2 ,how I test the  result (stats_tool_out.bin) in the pc.  

  • 0 in reply to reset_zhang
    Intellectual 850 points

    hi,what`s the value of layersGroupId  and  conv2dKernelType  in the follow  deploy?

    name: "jsegnet21v2_deploy"
    input: "data"
    input_shape {
      dim: 1
      dim: 3
      dim: 512
      dim: 1024
    }
    layer {
      name: "data/bias"
      type: "Bias"
      bottom: "data"
      top: "data/bias"
      param {
        lr_mult: 0
        decay_mult: 0
      }
      bias_param {
        filler {
          type: "constant"
          value: -128
        }
      }
    }
    layer {
      name: "conv1a"
      type: "Convolution"
      bottom: "data/bias"
      top: "conv1a"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 32
        bias_term: true
        pad: 2
        kernel_size: 5
        group: 1
        stride: 2
        weight_filler {
          type: "msra"
        }
        bias_filler {
          type: "constant"
          value: 0
        }
        dilation: 1
      }
    }
    layer {
      name: "conv1a/bn"
      type: "BatchNorm"
      bottom: "conv1a"
      top: "conv1a"
      batch_norm_param {
        moving_average_fraction: 0.99
        eps: 0.0001
        scale_bias: true
      }
    }
    layer {
      name: "conv1a/relu"
      type: "ReLU"
      bottom: "conv1a"
      top: "conv1a"
    }
    layer {
      name: "conv1b"
      type: "Convolution"
      bottom: "conv1a"
      top: "conv1b"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 32
        bias_term: true
        pad: 1
        kernel_size: 3
        group: 4
        stride: 1
        weight_filler {
          type: "msra"
        }
        bias_filler {
          type: "constant"
          value: 0
        }
        dilation: 1
      }
    }
    layer {
      name: "conv1b/bn"
      type: "BatchNorm"
      bottom: "conv1b"
      top: "conv1b"
      batch_norm_param {
        moving_average_fraction: 0.99
        eps: 0.0001
        scale_bias: true
      }
    }
    layer {
      name: "conv1b/relu"
      type: "ReLU"
      bottom: "conv1b"
      top: "conv1b"
    }
    layer {
      name: "pool1"
      type: "Pooling"
      bottom: "conv1b"
      top: "pool1"
      pooling_param {
        pool: MAX
        kernel_size: 2
        stride: 2
      }
    }
    layer {
      name: "res2a_branch2a"
      type: "Convolution"
      bottom: "pool1"
      top: "res2a_branch2a"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 64
        bias_term: true
        pad: 1
        kernel_size: 3
        group: 1
        stride: 1
        weight_filler {
          type: "msra"
        }
        bias_filler {
          type: "constant"
          value: 0
        }
        dilation: 1
      }
    }
    layer {
      name: "res2a_branch2a/bn"
      type: "BatchNorm"
      bottom: "res2a_branch2a"
      top: "res2a_branch2a"
      batch_norm_param {
        moving_average_fraction: 0.99
        eps: 0.0001
        scale_bias: true
      }
    }
    layer {
      name: "res2a_branch2a/relu"
      type: "ReLU"
      bottom: "res2a_branch2a"
      top: "res2a_branch2a"
    }
    layer {
      name: "res2a_branch2b"
      type: "Convolution"
      bottom: "res2a_branch2a"
      top: "res2a_branch2b"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 64
        bias_term: true
        pad: 1
        kernel_size: 3
        group: 4
        stride: 1
        weight_filler {
          type: "msra"
        }
        bias_filler {
          type: "constant"
          value: 0
        }
        dilation: 1
      }
    }
    layer {
      name: "res2a_branch2b/bn"
      type: "BatchNorm"
      bottom: "res2a_branch2b"
      top: "res2a_branch2b"
      batch_norm_param {
        moving_average_fraction: 0.99
        eps: 0.0001
        scale_bias: true
      }
    }
    layer {
      name: "res2a_branch2b/relu"
      type: "ReLU"
      bottom: "res2a_branch2b"
      top: "res2a_branch2b"
    }
    layer {
      name: "pool2"
      type: "Pooling"
      bottom: "res2a_branch2b"
      top: "pool2"
      pooling_param {
        pool: MAX
        kernel_size: 2
        stride: 2
      }
    }
    layer {
      name: "res3a_branch2a"
      type: "Convolution"
      bottom: "pool2"
      top: "res3a_branch2a"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 128
        bias_term: true
        pad: 1
        kernel_size: 3
        group: 1
        stride: 1
        weight_filler {
          type: "msra"
        }
        bias_filler {
          type: "constant"
          value: 0
        }
        dilation: 1
      }
    }
    layer {
      name: "res3a_branch2a/bn"
      type: "BatchNorm"
      bottom: "res3a_branch2a"
      top: "res3a_branch2a"
      batch_norm_param {
        moving_average_fraction: 0.99
        eps: 0.0001
        scale_bias: true
      }
    }
    layer {
      name: "res3a_branch2a/relu"
      type: "ReLU"
      bottom: "res3a_branch2a"
      top: "res3a_branch2a"
    }
    layer {
      name: "res3a_branch2b"
      type: "Convolution"
      bottom: "res3a_branch2a"
      top: "res3a_branch2b"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 128
        bias_term: true
        pad: 1
        kernel_size: 3
        group: 4
        stride: 1
        weight_filler {
          type: "msra"
        }
        bias_filler {
          type: "constant"
          value: 0
        }
        dilation: 1
      }
    }
    layer {
      name: "res3a_branch2b/bn"
      type: "BatchNorm"
      bottom: "res3a_branch2b"
      top: "res3a_branch2b"
      batch_norm_param {
        moving_average_fraction: 0.99
        eps: 0.0001
        scale_bias: true
      }
    }
    layer {
      name: "res3a_branch2b/relu"
      type: "ReLU"
      bottom: "res3a_branch2b"
      top: "res3a_branch2b"
    }
    layer {
      name: "pool3"
      type: "Pooling"
      bottom: "res3a_branch2b"
      top: "pool3"
      pooling_param {
        pool: MAX
        kernel_size: 2
        stride: 2
      }
    }
    layer {
      name: "res4a_branch2a"
      type: "Convolution"
      bottom: "pool3"
      top: "res4a_branch2a"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 256
        bias_term: true
        pad: 1
        kernel_size: 3
        group: 1
        stride: 1
        weight_filler {
          type: "msra"
        }
        bias_filler {
          type: "constant"
          value: 0
        }
        dilation: 1
      }
    }
    layer {
      name: "res4a_branch2a/bn"
      type: "BatchNorm"
      bottom: "res4a_branch2a"
      top: "res4a_branch2a"
      batch_norm_param {
        moving_average_fraction: 0.99
        eps: 0.0001
        scale_bias: true
      }
    }
    layer {
      name: "res4a_branch2a/relu"
      type: "ReLU"
      bottom: "res4a_branch2a"
      top: "res4a_branch2a"
    }
    layer {
      name: "res4a_branch2b"
      type: "Convolution"
      bottom: "res4a_branch2a"
      top: "res4a_branch2b"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 256
        bias_term: true
        pad: 1
        kernel_size: 3
        group: 4
        stride: 1
        weight_filler {
          type: "msra"
        }
        bias_filler {
          type: "constant"
          value: 0
        }
        dilation: 1
      }
    }
    layer {
      name: "res4a_branch2b/bn"
      type: "BatchNorm"
      bottom: "res4a_branch2b"
      top: "res4a_branch2b"
      batch_norm_param {
        moving_average_fraction: 0.99
        eps: 0.0001
        scale_bias: true
      }
    }
    layer {
      name: "res4a_branch2b/relu"
      type: "ReLU"
      bottom: "res4a_branch2b"
      top: "res4a_branch2b"
    }
    layer {
      name: "pool4"
      type: "Pooling"
      bottom: "res4a_branch2b"
      top: "pool4"
      pooling_param {
        pool: MAX
        kernel_size: 1
        stride: 1
      }
    }
    layer {
      name: "res5a_branch2a"
      type: "Convolution"
      bottom: "pool4"
      top: "res5a_branch2a"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 512
        bias_term: true
        pad: 2
        kernel_size: 3
        group: 1
        stride: 1
        weight_filler {
          type: "msra"
        }
        bias_filler {
          type: "constant"
          value: 0
        }
        dilation: 2
      }
    }
    layer {
      name: "res5a_branch2a/bn"
      type: "BatchNorm"
      bottom: "res5a_branch2a"
      top: "res5a_branch2a"
      batch_norm_param {
        moving_average_fraction: 0.99
        eps: 0.0001
        scale_bias: true
      }
    }
    layer {
      name: "res5a_branch2a/relu"
      type: "ReLU"
      bottom: "res5a_branch2a"
      top: "res5a_branch2a"
    }
    layer {
      name: "res5a_branch2b"
      type: "Convolution"
      bottom: "res5a_branch2a"
      top: "res5a_branch2b"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 512
        bias_term: true
        pad: 2
        kernel_size: 3
        group: 4
        stride: 1
        weight_filler {
          type: "msra"
        }
        bias_filler {
          type: "constant"
          value: 0
        }
        dilation: 2
      }
    }
    layer {
      name: "res5a_branch2b/bn"
      type: "BatchNorm"
      bottom: "res5a_branch2b"
      top: "res5a_branch2b"
      batch_norm_param {
        moving_average_fraction: 0.99
        eps: 0.0001
        scale_bias: true
      }
    }
    layer {
      name: "res5a_branch2b/relu"
      type: "ReLU"
      bottom: "res5a_branch2b"
      top: "res5a_branch2b"
    }
    layer {
      name: "out5a"
      type: "Convolution"
      bottom: "res5a_branch2b"
      top: "out5a"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 64
        bias_term: true
        pad: 4
        kernel_size: 3
        group: 2
        stride: 1
        weight_filler {
          type: "msra"
        }
        bias_filler {
          type: "constant"
          value: 0
        }
        dilation: 4
      }
    }
    layer {
      name: "out5a/bn"
      type: "BatchNorm"
      bottom: "out5a"
      top: "out5a"
      batch_norm_param {
        moving_average_fraction: 0.99
        eps: 0.0001
        scale_bias: true
      }
    }
    layer {
      name: "out5a/relu"
      type: "ReLU"
      bottom: "out5a"
      top: "out5a"
    }
    layer {
      name: "out5a_up2"
      type: "Deconvolution"
      bottom: "out5a"
      top: "out5a_up2"
      param {
        lr_mult: 0
        decay_mult: 0
      }
      convolution_param {
        num_output: 64
        bias_term: false
        pad: 1
        kernel_size: 4
        group: 64
        stride: 2
        weight_filler {
          type: "bilinear"
        }
      }
    }
    layer {
      name: "out3a"
      type: "Convolution"
      bottom: "res3a_branch2b"
      top: "out3a"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 64
        bias_term: true
        pad: 1
        kernel_size: 3
        group: 2
        stride: 1
        weight_filler {
          type: "msra"
        }
        bias_filler {
          type: "constant"
          value: 0
        }
        dilation: 1
      }
    }
    layer {
      name: "out3a/bn"
      type: "BatchNorm"
      bottom: "out3a"
      top: "out3a"
      batch_norm_param {
        moving_average_fraction: 0.99
        eps: 0.0001
        scale_bias: true
      }
    }
    layer {
      name: "out3a/relu"
      type: "ReLU"
      bottom: "out3a"
      top: "out3a"
    }
    layer {
      name: "out3_out5_combined"
      type: "Eltwise"
      bottom: "out5a_up2"
      bottom: "out3a"
      top: "out3_out5_combined"
    }
    layer {
      name: "ctx_conv1"
      type: "Convolution"
      bottom: "out3_out5_combined"
      top: "ctx_conv1"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 64
        bias_term: true
        pad: 1
        kernel_size: 3
        group: 1
        stride: 1
        weight_filler {
          type: "msra"
        }
        bias_filler {
          type: "constant"
          value: 0
        }
        dilation: 1
      }
    }
    layer {
      name: "ctx_conv1/bn"
      type: "BatchNorm"
      bottom: "ctx_conv1"
      top: "ctx_conv1"
      batch_norm_param {
        moving_average_fraction: 0.99
        eps: 0.0001
        scale_bias: true
      }
    }
    layer {
      name: "ctx_conv1/relu"
      type: "ReLU"
      bottom: "ctx_conv1"
      top: "ctx_conv1"
    }
    layer {
      name: "ctx_conv2"
      type: "Convolution"
      bottom: "ctx_conv1"
      top: "ctx_conv2"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 64
        bias_term: true
        pad: 4
        kernel_size: 3
        group: 1
        stride: 1
        weight_filler {
          type: "msra"
        }
        bias_filler {
          type: "constant"
          value: 0
        }
        dilation: 4
      }
    }
    layer {
      name: "ctx_conv2/bn"
      type: "BatchNorm"
      bottom: "ctx_conv2"
      top: "ctx_conv2"
      batch_norm_param {
        moving_average_fraction: 0.99
        eps: 0.0001
        scale_bias: true
      }
    }
    layer {
      name: "ctx_conv2/relu"
      type: "ReLU"
      bottom: "ctx_conv2"
      top: "ctx_conv2"
    }
    layer {
      name: "ctx_conv3"
      type: "Convolution"
      bottom: "ctx_conv2"
      top: "ctx_conv3"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 64
        bias_term: true
        pad: 4
        kernel_size: 3
        group: 1
        stride: 1
        weight_filler {
          type: "msra"
        }
        bias_filler {
          type: "constant"
          value: 0
        }
        dilation: 4
      }
    }
    layer {
      name: "ctx_conv3/bn"
      type: "BatchNorm"
      bottom: "ctx_conv3"
      top: "ctx_conv3"
      batch_norm_param {
        moving_average_fraction: 0.99
        eps: 0.0001
        scale_bias: true
      }
    }
    layer {
      name: "ctx_conv3/relu"
      type: "ReLU"
      bottom: "ctx_conv3"
      top: "ctx_conv3"
    }
    layer {
      name: "ctx_conv4"
      type: "Convolution"
      bottom: "ctx_conv3"
      top: "ctx_conv4"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 64
        bias_term: true
        pad: 4
        kernel_size: 3
        group: 1
        stride: 1
        weight_filler {
          type: "msra"
        }
        bias_filler {
          type: "constant"
          value: 0
        }
        dilation: 4
      }
    }
    layer {
      name: "ctx_conv4/bn"
      type: "BatchNorm"
      bottom: "ctx_conv4"
      top: "ctx_conv4"
      batch_norm_param {
        moving_average_fraction: 0.99
        eps: 0.0001
        scale_bias: true
      }
    }
    layer {
      name: "ctx_conv4/relu"
      type: "ReLU"
      bottom: "ctx_conv4"
      top: "ctx_conv4"
    }
    layer {
      name: "ctx_final"
      type: "Convolution"
      bottom: "ctx_conv4"
      top: "ctx_final"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 8
        bias_term: true
        pad: 1
        kernel_size: 3
        kernel_size: 3
        group: 1
        stride: 1
        weight_filler {
          type: "msra"
        }
        bias_filler {
          type: "constant"
          value: 0
        }
        dilation: 1
      }
    }
    layer {
      name: "ctx_final/relu"
      type: "ReLU"
      bottom: "ctx_final"
      top: "ctx_final"
    }
    layer {
      name: "out_deconv_final_up2"
      type: "Deconvolution"
      bottom: "ctx_final"
      top: "out_deconv_final_up2"
      param {
        lr_mult: 0
        decay_mult: 0
      }
      convolution_param {
        num_output: 8
        bias_term: false
        pad: 1
        kernel_size: 4
        group: 8
        stride: 2
        weight_filler {
          type: "bilinear"
        }
      }
    }
    layer {
      name: "out_deconv_final_up4"
      type: "Deconvolution"
      bottom: "out_deconv_final_up2"
      top: "out_deconv_final_up4"
      param {
        lr_mult: 0
        decay_mult: 0
      }
      convolution_param {
        num_output: 8
        bias_term: false
        pad: 1
        kernel_size: 4
        group: 8
        stride: 2
        weight_filler {
          type: "bilinear"
        }
      }
    }
    layer {
      name: "out_deconv_final_up8"
      type: "Deconvolution"
      bottom: "out_deconv_final_up4"
      top: "out_deconv_final_up8"
      param {
        lr_mult: 0
        decay_mult: 0
      }
      convolution_param {
        num_output: 8
        bias_term: false
        pad: 1
        kernel_size: 4
        group: 8
        stride: 2
        weight_filler {
          type: "bilinear"
        }
      }
    }
    layer {
      name: "argMaxOut"
      type: "ArgMax"
      bottom: "out_deconv_final_up8"
      top: "argMaxOut"
      argmax_param {
        axis: 1
      }
    }



  • 0 in reply to reset_zhang
    TI__Genius 11466 points

    The parameters that you mentioned are not part of Caffe / Caffe-jacinto - but I am guessing it may be for TIDL. You would have to consult the TIDL user guide or ask questions to TIDL experts (use TIDL as tag).