tests/nnapi/specs/skip/V1_2/lstm_state2_float16.mod.py

   1 #
   2 # Copyright (C) 2017 The Android Open Source Project
   3 #
   4 # Licensed under the Apache License, Version 2.0 (the "License");
   5 # you may not use this file except in compliance with the License.
   6 # You may obtain a copy of the License at
   7 #
   8 #      http://www.apache.org/licenses/LICENSE-2.0
   9 #
  10 # Unless required by applicable law or agreed to in writing, software
  11 # distributed under the License is distributed on an "AS IS" BASIS,
  12 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13 # See the License for the specific language governing permissions and
  14 # limitations under the License.
  15 #
  16
  17 # LSTM Test: No Cifg, No Peephole, No Projection, and No Clipping.
  18
  19 model = Model()
  20
  21 n_batch = 1
  22 n_input = 2
  23 # n_cell and n_output have the same size when there is no projection.
  24 n_cell = 4
  25 n_output = 4
  26
  27 input = Input("input", "TENSOR_FLOAT16", "{%d, %d}" % (n_batch, n_input))
  28
  29 input_to_input_weights = Input("input_to_input_weights", "TENSOR_FLOAT16", "{%d, %d}" % (n_cell, n_input))
  30 input_to_forget_weights = Input("input_to_forget_weights", "TENSOR_FLOAT16", "{%d, %d}" % (n_cell, n_input))
  31 input_to_cell_weights = Input("input_to_cell_weights", "TENSOR_FLOAT16", "{%d, %d}" % (n_cell, n_input))
  32 input_to_output_weights = Input("input_to_output_weights", "TENSOR_FLOAT16", "{%d, %d}" % (n_cell, n_input))
  33
  34 recurrent_to_input_weights = Input("recurrent_to_intput_weights", "TENSOR_FLOAT16", "{%d, %d}" % (n_cell, n_output))
  35 recurrent_to_forget_weights = Input("recurrent_to_forget_weights", "TENSOR_FLOAT16", "{%d, %d}" % (n_cell, n_output))
  36 recurrent_to_cell_weights = Input("recurrent_to_cell_weights", "TENSOR_FLOAT16", "{%d, %d}" % (n_cell, n_output))
  37 recurrent_to_output_weights = Input("recurrent_to_output_weights", "TENSOR_FLOAT16", "{%d, %d}" % (n_cell, n_output))
  38
  39 cell_to_input_weights = Input("cell_to_input_weights", "TENSOR_FLOAT16", "{0}")
  40 cell_to_forget_weights = Input("cell_to_forget_weights", "TENSOR_FLOAT16", "{0}")
  41 cell_to_output_weights = Input("cell_to_output_weights", "TENSOR_FLOAT16", "{0}")
  42
  43 input_gate_bias = Input("input_gate_bias", "TENSOR_FLOAT16", "{%d}"%(n_cell))
  44 forget_gate_bias = Input("forget_gate_bias", "TENSOR_FLOAT16", "{%d}"%(n_cell))
  45 cell_gate_bias = Input("cell_gate_bias", "TENSOR_FLOAT16", "{%d}"%(n_cell))
  46 output_gate_bias = Input("output_gate_bias", "TENSOR_FLOAT16", "{%d}"%(n_cell))
  47
  48 projection_weights = Input("projection_weights", "TENSOR_FLOAT16", "{0,0}")
  49 projection_bias = Input("projection_bias", "TENSOR_FLOAT16", "{0}")
  50
  51 output_state_in = Input("output_state_in", "TENSOR_FLOAT16", "{%d, %d}" % (n_batch, n_output))
  52 cell_state_in = Input("cell_state_in", "TENSOR_FLOAT16", "{%d, %d}" % (n_batch, n_cell))
  53
  54 activation_param = Int32Scalar("activation_param", 4)  # Tanh
  55 cell_clip_param = Float16Scalar("cell_clip_param", 0.)
  56 proj_clip_param = Float16Scalar("proj_clip_param", 0.)
  57
  58 scratch_buffer = IgnoredOutput("scratch_buffer", "TENSOR_FLOAT16", "{%d, %d}" % (n_batch, (n_cell * 4)))
  59 output_state_out = IgnoredOutput("output_state_out", "TENSOR_FLOAT16", "{%d, %d}" % (n_batch, n_output))
  60 cell_state_out = IgnoredOutput("cell_state_out", "TENSOR_FLOAT16", "{%d, %d}" % (n_batch, n_cell))
  61 output = Output("output", "TENSOR_FLOAT16", "{%d, %d}" % (n_batch, n_output))
  62
  63 model = model.Operation("LSTM",
  64                         input,
  65
  66                         input_to_input_weights,
  67                         input_to_forget_weights,
  68                         input_to_cell_weights,
  69                         input_to_output_weights,
  70
  71                         recurrent_to_input_weights,
  72                         recurrent_to_forget_weights,
  73                         recurrent_to_cell_weights,
  74                         recurrent_to_output_weights,
  75
  76                         cell_to_input_weights,
  77                         cell_to_forget_weights,
  78                         cell_to_output_weights,
  79
  80                         input_gate_bias,
  81                         forget_gate_bias,
  82                         cell_gate_bias,
  83                         output_gate_bias,
  84
  85                         projection_weights,
  86                         projection_bias,
  87
  88                         output_state_in,
  89                         cell_state_in,
  90
  91                         activation_param,
  92                         cell_clip_param,
  93                         proj_clip_param
  94 ).To([scratch_buffer, output_state_out, cell_state_out, output])
  95
  96 # Example 1. Input in operand 0,
  97 input0 = {input_to_input_weights:  [-0.45018822, -0.02338299, -0.0870589, -0.34550029, 0.04266912, -0.15680569, -0.34856534, 0.43890524],
  98           input_to_forget_weights: [0.09701663, 0.20334584, -0.50592935, -0.31343272, -0.40032279, 0.44781327, 0.01387155, -0.35593212],
  99           input_to_cell_weights:   [-0.50013041, 0.1370284, 0.11810488, 0.2013163, -0.20583314, 0.44344562, 0.22077113, -0.29909778],
 100           input_to_output_weights: [-0.25065863, -0.28290087, 0.04613829, 0.40525138, 0.44272184, 0.03897077, -0.1556896, 0.19487578],
 101
 102           input_gate_bias:  [0.,0.,0.,0.],
 103           forget_gate_bias: [1.,1.,1.,1.],
 104           cell_gate_bias:   [0.,0.,0.,0.],
 105           output_gate_bias: [0.,0.,0.,0.],
 106
 107           recurrent_to_input_weights: [
 108               -0.0063535, -0.2042388, 0.31454784, -0.35746509, 0.28902304, 0.08183324,
 109             -0.16555229, 0.02286911, -0.13566875, 0.03034258, 0.48091322,
 110             -0.12528998, 0.24077177, -0.51332325, -0.33502164, 0.10629296],
 111
 112           recurrent_to_cell_weights: [
 113               -0.3407414, 0.24443203, -0.2078532, 0.26320225, 0.05695659, -0.00123841,
 114             -0.4744786, -0.35869038, -0.06418842, -0.13502428, -0.501764, 0.22830659,
 115             -0.46367589, 0.26016325, -0.03894562, -0.16368064],
 116
 117           recurrent_to_forget_weights: [
 118               -0.48684245, -0.06655136, 0.42224967, 0.2112639, 0.27654213, 0.20864892,
 119             -0.07646349, 0.45877004, 0.00141793, -0.14609534, 0.36447752, 0.09196436,
 120             0.28053468, 0.01560611, -0.20127171, -0.01140004],
 121
 122           recurrent_to_output_weights: [
 123               0.43385774, -0.17194885, 0.2718237, 0.09215671, 0.24107647, -0.39835793,
 124               0.18212086, 0.01301402, 0.48572797, -0.50656658, 0.20047462, -0.20607421,
 125               -0.51818722, -0.15390486, 0.0468148, 0.39922136],
 126
 127           cell_to_input_weights: [],
 128           cell_to_forget_weights: [],
 129           cell_to_output_weights: [],
 130
 131           projection_weights: [],
 132           projection_bias: [],
 133 }
 134
 135 test_input = [1., 1.]
 136 output_state = [-0.0371611, 0.125073, 0.411934, -0.208605]
 137 cell_state = [-0.287121, 0.148115, 0.556837, -0.388276]
 138 golden_output = [-0.15053082, 0.09120187,  0.24278517,  -0.12222792]
 139 output0 = {
 140     scratch_buffer: [ 0 for x in range(n_batch * n_cell * 4) ],
 141     cell_state_out: [ 0 for x in range(n_batch * n_cell) ],
 142     output_state_out: [ 0 for x in range(n_batch * n_output) ],
 143     output: golden_output
 144 }
 145 input0[input] = test_input
 146 input0[output_state_in] = output_state
 147 input0[cell_state_in] = cell_state
 148 Example((input0, output0))