webnn/validation_tests/lstm.https.any.js

// META: title=validation tests for WebNN API lstm operation
// META: global=window,dedicatedworker
// META: script=../resources/utils_validation.js
// META: timeout=long

'use strict';

const steps = 10, batchSize = 5, inputSize = 3, hiddenSize = 8,
      numDirections = 1;

const tests = [
  {
    name: '[lstm] Test with default options',
    input: {dataType: 'float16', dimensions: [steps, batchSize, inputSize]},
    weight: {
      dataType: 'float16',
      dimensions: [numDirections, 4 * hiddenSize, inputSize]
    },
    recurrentWeight: {
      dataType: 'float16',
      dimensions: [numDirections, 4 * hiddenSize, hiddenSize]
    },
    steps: steps,
    hiddenSize: hiddenSize,
    outputs: [
      {dataType: 'float16', dimensions: [numDirections, batchSize, hiddenSize]},
      {dataType: 'float16', dimensions: [numDirections, batchSize, hiddenSize]}
    ]
  },
  {
    name: '[lstm] Test with given options',
    input: {dataType: 'float32', dimensions: [steps, batchSize, inputSize]},
    weight: {
      dataType: 'float32',
      dimensions: [/*numDirections=*/ 2, 4 * hiddenSize, inputSize]
    },
    recurrentWeight: {
      dataType: 'float32',
      dimensions: [/*numDirections=*/ 2, 4 * hiddenSize, hiddenSize]
    },
    steps: steps,
    hiddenSize: hiddenSize,
    options: {
      bias: {
        dataType: 'float32',
        dimensions: [/*numDirections=*/ 2, 4 * hiddenSize]
      },
      recurrentBias: {
        dataType: 'float32',
        dimensions: [/*numDirections=*/ 2, 4 * hiddenSize]
      },
      peepholeWeight: {
        dataType: 'float32',
        dimensions: [/*numDirections=*/ 2, 3 * hiddenSize]
      },
      initialHiddenState: {
        dataType: 'float32',
        dimensions: [/*numDirections=*/ 2, batchSize, hiddenSize]
      },
      initialCellState: {
        dataType: 'float32',
        dimensions: [/*numDirections=*/ 2, batchSize, hiddenSize]
      },
      returnSequence: true,
      direction: 'both',
      layout: 'ifgo',
      activations: ['sigmoid', 'relu', 'tanh']
    },
    outputs: [
      {
        dataType: 'float32',
        dimensions: [/*numDirections=*/ 2, batchSize, hiddenSize]
      },
      {
        dataType: 'float32',
        dimensions: [/*numDirections=*/ 2, batchSize, hiddenSize]
      },
      {
        dataType: 'float32',
        dimensions: [steps, /*numDirections=*/ 2, batchSize, hiddenSize]
      }
    ]
  },
  {
    name: '[lstm] DataError is expected if hiddenSize equals to zero',
    input: {dataType: 'float32', dimensions: [steps, batchSize, inputSize]},
    weight: {
      dataType: 'float32',
      dimensions: [numDirections, 4 * hiddenSize, inputSize]
    },
    recurrentWeight: {
      dataType: 'float32',
      dimensions: [numDirections, 4 * hiddenSize, hiddenSize]
    },
    steps: steps,
    hiddenSize: 0
  },
  {
    name: '[lstm] DataError is expected if hiddenSize is too large',
    input: {dataType: 'float32', dimensions: [steps, batchSize, inputSize]},
    weight: {
      dataType: 'float32',
      dimensions: [numDirections, 4 * hiddenSize, inputSize]
    },
    recurrentWeight: {
      dataType: 'float32',
      dimensions: [numDirections, 4 * hiddenSize, hiddenSize]
    },
    steps: steps,
    hiddenSize: 4294967295,
  },
  {
    name: '[lstm] DataError is expected if steps equals to zero',
    input: {dataType: 'float32', dimensions: [steps, batchSize, inputSize]},
    weight: {
      dataType: 'float32',
      dimensions: [numDirections, 4 * hiddenSize, inputSize]
    },
    recurrentWeight: {
      dataType: 'float32',
      dimensions: [numDirections, 4 * hiddenSize, hiddenSize]
    },
    steps: 0,
    hiddenSize: hiddenSize,
  },
  {
    name:
        '[lstm] DataError is expected if the data type is not one of the floating point types',
    input: {dataType: 'uint32', dimensions: [steps, batchSize, inputSize]},
    weight: {
      dataType: 'uint32',
      dimensions: [numDirections, 4 * hiddenSize, inputSize]
    },
    recurrentWeight: {
      dataType: 'uint32',
      dimensions: [numDirections, 4 * hiddenSize, hiddenSize]
    },
    steps: steps,
    hiddenSize: hiddenSize
  },
  {
    name:
        '[lstm] DataError is expected if the rank of input is not 3',
    input: {dataType: 'float32', dimensions: [steps, batchSize]},
    weight: {
      dataType: 'float32',
      dimensions: [numDirections, 4 * hiddenSize, inputSize]
    },
    recurrentWeight: {
      dataType: 'float32',
      dimensions: [numDirections, 4 * hiddenSize, hiddenSize]
    },
    steps: steps,
    hiddenSize: hiddenSize
  },
  {
    name:
        '[lstm] DataError is expected if input.dimensions[0] is not equal to steps',
    input: {dataType: 'float32', dimensions: [1000, batchSize, inputSize]},
    weight: {
      dataType: 'float32',
      dimensions: [numDirections, 4 * hiddenSize, inputSize]
    },
    recurrentWeight: {
      dataType: 'float32',
      dimensions: [numDirections, 4 * hiddenSize, hiddenSize]
    },
    steps: steps,
    hiddenSize: hiddenSize
  },
  {
    name: '[lstm] DataError is expected if the shape of weight is incorrect',
    input: {dataType: 'float32', dimensions: [steps, batchSize, inputSize]},
    weight: {
      dataType: 'float32',
      dimensions: [numDirections, 4 * hiddenSize, 1000]
    },
    recurrentWeight: {
      dataType: 'float32',
      dimensions: [numDirections, 4 * hiddenSize, hiddenSize]
    },
    steps: steps,
    hiddenSize: hiddenSize
  },
  {
    name:
        '[lstm] DataError is expected if the rank of recurrentWeight is not 3',
    input: {dataType: 'float32', dimensions: [steps, batchSize, inputSize]},
    weight: {
      dataType: 'float32',
      dimensions: [numDirections, 4 * hiddenSize, inputSize]
    },
    recurrentWeight:
        {dataType: 'float32', dimensions: [numDirections, 4 * hiddenSize]},
    steps: steps,
    hiddenSize: hiddenSize
  },
  {
    name:
        '[lstm] DataError is expected if the size of options.activations is not 3',
    input: {dataType: 'float32', dimensions: [steps, batchSize, inputSize]},
    weight: {
      dataType: 'float32',
      dimensions: [numDirections, 4 * hiddenSize, inputSize]
    },
    recurrentWeight: {
      dataType: 'float32',
      dimensions: [numDirections, 4 * hiddenSize, hiddenSize]
    },
    steps: steps,
    hiddenSize: hiddenSize,
    options: {activations: ['sigmoid', 'tanh']}
  },
  {
    name:
        '[lstm] DataError is expected if the rank of options.bias is not 2',
    input: {dataType: 'float16', dimensions: [steps, batchSize, inputSize]},
    weight: {
      dataType: 'float16',
      dimensions: [numDirections, 4 * hiddenSize, inputSize]
    },
    recurrentWeight: {
      dataType: 'float16',
      dimensions: [numDirections, 4 * hiddenSize, hiddenSize]
    },
    steps: steps,
    hiddenSize: hiddenSize,
    options: {bias: {dataType: 'float16', dimensions: [numDirections]}}
  },
  {
    name:
        '[lstm] DataError is expected if the shape of options.recurrentBias.dimensions is incorrect',
    input: {dataType: 'float16', dimensions: [steps, batchSize, inputSize]},
    weight: {
      dataType: 'float16',
      dimensions: [numDirections, 4 * hiddenSize, inputSize]
    },
    recurrentWeight: {
      dataType: 'float16',
      dimensions: [numDirections, 4 * hiddenSize, hiddenSize]
    },
    steps: steps,
    hiddenSize: hiddenSize,
    options: {
      recurrentBias: {dataType: 'float16', dimensions: [numDirections, 1000]}
    }
  },
  {
    name:
        '[lstm] DataError is expected if the dataType of options.peepholeWeight is incorrect',
    input: {dataType: 'float16', dimensions: [steps, batchSize, inputSize]},
    weight: {
      dataType: 'float16',
      dimensions: [numDirections, 4 * hiddenSize, inputSize]
    },
    recurrentWeight: {
      dataType: 'float16',
      dimensions: [numDirections, 4 * hiddenSize, hiddenSize]
    },
    steps: steps,
    hiddenSize: hiddenSize,
    options: {
      peepholeWeight:
          {dataType: 'float32', dimensions: [numDirections, 3 * hiddenSize]}
    }
  },
  {
    name:
        '[lstm] DataError is expected if the dataType of options.initialHiddenState is incorrect',
    input: {dataType: 'float16', dimensions: [steps, batchSize, inputSize]},
    weight: {
      dataType: 'float16',
      dimensions: [numDirections, 4 * hiddenSize, inputSize]
    },
    recurrentWeight: {
      dataType: 'float16',
      dimensions: [numDirections, 4 * hiddenSize, hiddenSize]
    },
    steps: steps,
    hiddenSize: hiddenSize,
    options: {
      initialHiddenState: {
        dataType: 'uint64',
        dimensions: [numDirections, batchSize, hiddenSize]
      }
    }
  },
  {
    name:
        '[lstm] DataError is expected if the shape of options.initialCellState is incorrect',
    input: {dataType: 'float32', dimensions: [steps, batchSize, inputSize]},
    weight: {
      dataType: 'float32',
      dimensions: [numDirections, 4 * hiddenSize, inputSize]
    },
    recurrentWeight: {
      dataType: 'float32',
      dimensions: [numDirections, 4 * hiddenSize, hiddenSize]
    },
    steps: steps,
    hiddenSize: hiddenSize,
    options: {
      initialCellState:
          {dataType: 'float32', dimensions: [numDirections, batchSize, 1000]}
    }
  }
];

tests.forEach(
    test => promise_test(async t => {
      const input = builder.input(
          'input',
          {dataType: test.input.dataType, dimensions: test.input.dimensions});
      const weight = builder.input(
          'weight',
          {dataType: test.weight.dataType, dimensions: test.weight.dimensions});
      const recurrentWeight = builder.input('recurrentWeight', {
        dataType: test.recurrentWeight.dataType,
        dimensions: test.recurrentWeight.dimensions
      });

      const options = {};
      if (test.options) {
        if (test.options.bias) {
          options.bias = builder.input('bias', {
            dataType: test.options.bias.dataType,
            dimensions: test.options.bias.dimensions
          });
        }
        if (test.options.recurrentBias) {
          options.bias = builder.input('recurrentBias', {
            dataType: test.options.recurrentBias.dataType,
            dimensions: test.options.recurrentBias.dimensions
          });
        }
        if (test.options.peepholeWeight) {
          options.peepholeWeight = builder.input('peepholeWeight', {
            dataType: test.options.peepholeWeight.dataType,
            dimensions: test.options.peepholeWeight.dimensions
          });
        }
        if (test.options.initialHiddenState) {
          options.initialHiddenState = builder.input('initialHiddenState', {
            dataType: test.options.initialHiddenState.dataType,
            dimensions: test.options.initialHiddenState.dimensions
          });
        }
        if (test.options.initialCellState) {
          options.initialCellState = builder.input('initialCellState', {
            dataType: test.options.initialCellState.dataType,
            dimensions: test.options.initialCellState.dimensions
          });
        }
        if (test.options.returnSequence) {
          options.returnSequence = test.options.returnSequence;
        }
        if (test.options.direction) {
          options.direction = test.options.direction;
        }
        if (test.options.layout) {
          options.layout = test.options.layout;
        }
        if (test.options.activations) {
          options.activations = [];
          test.options.activations.forEach(
              activation => options.activations.push(builder[activation]()));
        }
      }

      if (test.outputs) {
        const outputs = builder.lstm(
            input, weight, recurrentWeight, test.steps, test.hiddenSize,
            options);
        assert_equals(outputs.length, test.outputs.length);
        for (let i = 0; i < outputs.length; ++i) {
          assert_equals(outputs[i].dataType(), test.outputs[i].dataType);
          assert_array_equals(outputs[i].shape(), test.outputs[i].dimensions);
        }
      } else {
        assert_throws_dom(
            'DataError',
            () => builder.lstm(
                input, weight, recurrentWeight, test.steps, test.hiddenSize,
                options));
      }
    }, test.name));