Getting 'Check failed' error when loading a toy LSTM prototxt file

[aurotripathy]

Also posted on Caffe Users. However, looks like an issue internal to Caffe.

Below is the error. I don't understand the error message (2 vs. 2).

I0801 09:50:09.308823 21586 layer_factory.hpp:77] Creating layer lstm1_x_transform
I0801 09:50:09.308851 21586 net.cpp:91] Creating Layer lstm1_x_transform
I0801 09:50:09.308863 21586 net.cpp:425] lstm1_x_transform <- x
I0801 09:50:09.308879 21586 net.cpp:399] lstm1_x_transform -> W_xc_x
F0801 09:50:09.308907 21586 blob.hpp:122] Check failed: axis_index < num_axes() (2 vs. 2) axis 2 out of range for 2-D Blob with shape 320 1 (320)
*** Check failure stack trace: ***
Aborted (core dumped)

Here's how I load the prototxt file (derived from https://github.com/junhyukoh/caffe-lstm/blob/master/examples/lstm_sequence/lstm_short.prototxt).

solver = caffe.get_solver('solver.prototxt')

Here's the prototxt file in its entirety.

name: "LSTM"
input: "data"
input_shape { dim: 320 dim: 1 }
input: "clip"
input_shape { dim: 320 dim: 1 }
input: "label"
input_shape { dim: 320 dim: 1 }
layer {
  name: "Silence"
  type: "Silence"
  bottom: "label"
  include: { phase: TEST }
}
layer {
  name: "lstm1"
  type: "LSTM"
  bottom: "data"
  bottom: "clip"
  top: "lstm1"

  param {
    lr_mult: 1
  }
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 2
  }

  recurrent_param {
    num_output: 7
    weight_filler {
      type: "gaussian"
      std: 0.1
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "lstm2"
  type: "Lstm"
  bottom: "lstm1"
  bottom: "clip"
  top: "lstm2"

  recurrent_param {
    num_output: 7
    weight_filler {
      type: "gaussian"
      std: 0.1
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "lstm3"
  type: "Lstm"
  bottom: "lstm2"
  bottom: "clip"
  top: "lstm3"

  recurrent_param {
    num_output: 7
    weight_filler {
      type: "gaussian"
      std: 0.1
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "ip1"
  type: "InnerProduct"
  bottom: "lstm3"
  top: "ip1"

  inner_product_param {
    num_output: 1
    weight_filler {
      type: "gaussian"
      std: 0.1
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "loss"
  type: "EuclideanLoss"
  bottom: "ip1"
  bottom: "label"
  top: "loss"
  include: { phase: TRAIN }
}

[aurotripathy]

As was pointed out in #2033, the format for the input data is T x N ...
So the following toy lstm prototxt meets the validation checks.

 name: "TOY_LSTM"
input: "data"
# T = 320 time_steps, N = 1 streams, 1-D data
input_shape { dim: 320 dim: 1 dim: 1}
input: "clip"
input_shape { dim: 320 dim: 1 }
input: "label"
input_shape { dim: 320 dim: 1 }

layer {
  name: "Silence"
  type: "Silence"
  bottom: "label"
  include: { phase: TEST }
}

layer {
  name: "lstm"
  type: "LSTM"
  bottom: "data"
  bottom: "clip"
  top: "lstm"
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 2
  }
  recurrent_param {
    num_output: 15
    weight_filler {
      type: "gaussian"
      std: 0.1
    }
    bias_filler {
      type: "constant"
    }
  }
}

layer {
  name: "ip"
  type: "InnerProduct"
  bottom: "lstm"
  top: "ip"
  inner_product_param {
    num_output: 1
    weight_filler {
      type: "gaussian"
      std: 0.1
    }
    bias_filler {
      type: "constant"
    }
  }
}

layer {
  name: "loss"
  type: "EuclideanLoss"
  bottom: "ip"
  bottom: "label"
  top: "loss"
  include: { phase: TRAIN }
}

[ubenz55555]

HI @aurotripathy

it shows up error with

Iteration 4800, lr = 0.0001
F0306 20:33:55.743507 10364 recurrent_layer.cpp:187] Check failed: T_ == bottom[0]->shape(0) (320 vs. 2) input number of timesteps changed

do you know how to solve it ?

[elinx]

@ubenz55555 I met the same problem, how do you solve it?