Python gradient_checker.compute_gradient_error函数代码示例

  def testGradient(self):
    if np.__version__ == "1.13.0":
      self.skipTest("numpy 1.13.0 bug")

    np.random.seed(8161)
    test_dims = [(11, 1, 5, 7, 1), (2, 2)]
    with self.test_session(use_gpu=False):
      for dims in test_dims:
        sp_t, nnz = _sparsify(np.random.randn(*dims))
        # reduce random axes from 1D to N-D
        for d in range(1, len(dims) + 1):
          axes = np.random.choice(len(dims), size=d, replace=False).tolist()
          reduced = sparse_ops.sparse_reduce_sum(sp_t, axes)

          err = gradient_checker.compute_gradient_error(sp_t.values, (nnz,),
                                                        reduced,
                                                        reduced.eval().shape)
          self.assertLess(err, 1e-3)

        # Tests for negative axes.
        reduced = sparse_ops.sparse_reduce_sum(sp_t, -1)
        err = gradient_checker.compute_gradient_error(sp_t.values, (nnz,),
                                                      reduced,
                                                      reduced.eval().shape)
        self.assertLess(err, 1e-3)

  def testFillFloat(self):
    with self.test_session(use_gpu=False) as sess:
      values = constant_op.constant(
          [0.0, 10.0, 13.0, 14.0, 32.0, 33.0], dtype=dtypes.float64)
      default_value = constant_op.constant(-1.0, dtype=dtypes.float64)
      sp_input = sparse_tensor.SparseTensorValue(
          indices=np.array([[0, 0], [1, 0], [1, 3], [1, 4], [3, 2], [3, 3]]),
          values=values,
          dense_shape=np.array([5, 6]))
      sp_output, empty_row_indicator = (sparse_ops.sparse_fill_empty_rows(
          sp_input, default_value))
      output, empty_row_indicator_out = sess.run(
          [sp_output, empty_row_indicator])

      self.assertAllEqual(output.indices, [[0, 0], [1, 0], [1, 3], [1, 4],
                                           [2, 0], [3, 2], [3, 3], [4, 0]])
      self.assertAllClose(output.values, [0, 10, 13, 14, -1, 32, 33, -1])
      self.assertAllEqual(output.dense_shape, [5, 6])
      self.assertAllEqual(empty_row_indicator_out,
                          np.array([0, 0, 1, 0, 1]).astype(np.bool))

      values_grad_err = gradient_checker.compute_gradient_error(
          values, values.shape.as_list(), sp_output.values, [8], delta=1e-8)
      self.assertGreater(values_grad_err, 0)
      self.assertLess(values_grad_err, 1e-8)

      default_value_grad_err = gradient_checker.compute_gradient_error(
          default_value,
          default_value.shape.as_list(),
          sp_output.values, [8],
          delta=1e-8)
      self.assertGreater(default_value_grad_err, 0)
      self.assertLess(default_value_grad_err, 1e-8)

  def doOutputTest(self, input_shape, moments_axes, tol=1e-4,
                   check_gradients=False):
    for mu in [0.0, 1.0, 1e3]:
      for sigma in [1.0, 0.1]:
        for keep_dims in [True, False]:
          input_values = np.random.rand(*input_shape) * sigma + mu
          expected_mean = np.mean(
              input_values, axis=moments_axes, keepdims=keep_dims)
          expected_var = np.var(
              input_values, axis=moments_axes, keepdims=keep_dims)
          with ops.Graph().as_default() as g:
            with self.test_session(graph=g) as sess:
              inputs = constant_op.constant(
                  input_values, shape=input_shape, dtype=dtypes.float32)
              mean, variance = nn_impl.moments(
                  inputs, moments_axes, keep_dims=keep_dims)

              if check_gradients:
                err = gradient_checker.compute_gradient_error(
                    inputs, input_shape, mean, mean.shape.as_list())
                self.assertLess(err, 1e-3)
                err = gradient_checker.compute_gradient_error(
                    inputs, input_shape, variance, variance.shape.as_list())
                self.assertLess(err, 1e-3)

              # Evaluate.
              [mean, variance] = sess.run([mean, variance])
              # Make sure that there are no NaNs
              self.assertFalse(np.isnan(mean).any())
              self.assertFalse(np.isnan(variance).any())
              self.assertAllClose(mean, expected_mean, rtol=tol, atol=tol)
              self.assertAllClose(variance, expected_var, rtol=tol, atol=tol)

  def testGradients(self):
    np.random.seed(1618)
    sp_shapes = [(10, 10, 10), (5, 5), (1618,), (3, 3, 7)]
    dense_shapes = [(10, 10, 1), (5, 5), (1,), (1, 7)]

    with self.test_session(use_gpu=False):
      for dtype in [np.float32, np.float64]:
        for sp_shape, dense_shape in zip(sp_shapes, dense_shapes):
          sp_vals_np = np.random.rand(*sp_shape).astype(dtype) + 1
          dense_vals_np = np.random.rand(*dense_shape).astype(dtype) + 1
          sp_t, nnz = _sparsify(sp_vals_np, thresh=1.5)
          dense_t = constant_op.constant(dense_vals_np)

          cmul = sp_t * dense_t
          err = gradient_checker.compute_gradient_error([sp_t.values, dense_t],
                                                        [(nnz,), dense_shape],
                                                        cmul.values, (nnz,))
          self.assertLess(err, 1e-4)

          cdiv = sp_t / dense_t
          err = gradient_checker.compute_gradient_error(sp_t.values, (nnz,),
                                                        cdiv.values, (nnz,))
          self.assertLess(err, 1e-4)
          err = gradient_checker.compute_gradient_error(
              dense_t,
              dense_shape,
              cdiv.values, (nnz,),
              x_init_value=dense_vals_np)
          self.assertLess(err, 2e-4)

 def testGradient(self):
   np.random.seed(1)  # Make it reproducible.
   x_shape = [5, 10]
   x_np = np.random.randn(*x_shape).astype(np.float32)
   alpha_np = np.float32(np.random.rand(1, x_shape[1]) + 0.01)
   clip_np = np.float32(np.random.rand(x_shape[0], 1) * 5.)
   with self.test_session(use_gpu=True):
     x_tf = constant_op.constant(x_np)
     alpha_tf = constant_op.constant(alpha_np)
     clip_tf = constant_op.constant(clip_np)
     y_tf = scaled_softplus(x_tf, alpha_tf)
     z_tf = scaled_softplus(x_tf, alpha_tf, clip_tf * 0.1)
     err = gradient_checker.compute_gradient_error([x_tf, alpha_tf],
                                                   [x_shape, alpha_np.shape],
                                                   y_tf, x_shape,
                                                   [x_np, alpha_np],
                                                   delta=0.002)
     err_clip = gradient_checker.compute_gradient_error(
         [x_tf, alpha_tf, clip_tf],
         [x_shape, alpha_np.shape, clip_np.shape],
         z_tf, x_shape,
         [x_np, alpha_np, clip_np],
         delta=0.002)
   eps = 2e-4
   self.assertLess(err, eps)
   self.assertLess(err_clip, eps)

  def _testGradients(self, tr_a, tr_b, sp_a, sp_b, a_dtype, b_dtype, delta,
                     name):
    with self.test_session():
      a = constant_op.constant(
          RandMatrix(
              3, 2, tr_a, round_bfloat=True), dtype=dtypes.float32)
      b = constant_op.constant(
          RandMatrix(
              2, 4, tr_b, round_bfloat=True), dtype=dtypes.float32)
      tf_a = math_ops.cast(a, a_dtype) if a_dtype != dtypes.float32 else a
      tf_b = math_ops.cast(b, b_dtype) if b_dtype != dtypes.float32 else b

      m = math_ops.matmul(
          tf_a,
          tf_b,
          name=name,
          transpose_a=tr_a,
          transpose_b=tr_b,
          a_is_sparse=sp_a,
          b_is_sparse=sp_b)
      err = (gradient_checker.compute_gradient_error(
          a, [2, 3] if tr_a else [3, 2],
          m, [3, 4],
          x_init_value=a.eval(),
          delta=delta) + gradient_checker.compute_gradient_error(
              b, [4, 2] if tr_b else [2, 4],
              m, [3, 4],
              x_init_value=b.eval(),
              delta=delta))
    self.assertLessEqual(err, delta / 2.)

  def _test_gradient(self,
                     x_shape,
                     scale_shape,
                     use_gpu=True,
                     data_format='NHWC'):
    np.random.seed(1)
    x_val = np.random.random_sample(x_shape).astype(np.float32)
    scale_val = np.random.random_sample(scale_shape).astype(np.float32)
    offset_val = np.random.random_sample(scale_shape).astype(np.float32)

    with self.test_session(use_gpu=use_gpu):
      x = constant_op.constant(x_val, name='x')
      scale = constant_op.constant(scale_val, name='scale')
      offset = constant_op.constant(offset_val, name='offset')
      y, _, _ = nn_impl.fused_batch_norm(
          x, scale, offset, data_format=data_format)
      err_x = gradient_checker.compute_gradient_error(x, x_shape, y, x_shape)
      err_scale = gradient_checker.compute_gradient_error(scale, scale_shape, y,
                                                          x_shape)
      err_offset = gradient_checker.compute_gradient_error(offset, scale_shape,
                                                           y, x_shape)
    err_tolerance = 1e-3
    self.assertLess(err_x, err_tolerance)
    self.assertLess(err_scale, err_tolerance)
    self.assertLess(err_offset, err_tolerance)

  def _test_gradient(self,
                     x_shape,
                     x_dtype,
                     scale_shape,
                     scale_dtype,
                     use_gpu=True,
                     data_format='NHWC',
                     is_training=True):
    np.random.seed(1)
    x_val = np.random.random_sample(x_shape).astype(x_dtype)
    scale_val = np.random.random_sample(scale_shape).astype(scale_dtype)
    offset_val = np.random.random_sample(scale_shape).astype(scale_dtype)

    with self.test_session(use_gpu=use_gpu):
      x = constant_op.constant(x_val, name='x')
      scale = constant_op.constant(scale_val, name='scale')
      offset = constant_op.constant(offset_val, name='offset')
      if is_training:
        pop_mean = None
        pop_var = None
      else:
        pop_mean = np.random.random_sample(scale_shape).astype(scale_dtype)
        pop_var = np.random.random_sample(scale_shape).astype(scale_dtype)
      y, _, _ = nn_impl.fused_batch_norm(
          x,
          scale,
          offset,
          mean=pop_mean,
          variance=pop_var,
          data_format=data_format,
          is_training=is_training)
      if x_dtype != np.float16:
        err_x = gradient_checker.compute_gradient_error(x, x_shape, y, x_shape)
        err_scale = gradient_checker.compute_gradient_error(
            scale, scale_shape, y, x_shape)
        err_offset = gradient_checker.compute_gradient_error(
            offset, scale_shape, y, x_shape)
      else:
        x32 = constant_op.constant(x_val, name='x32', dtype=dtypes.float32)
        y32, _, _ = nn_impl.fused_batch_norm(
            x32,
            scale,
            offset,
            mean=pop_mean,
            variance=pop_var,
            data_format=data_format,
            is_training=is_training)
        err_x = self._compute_gradient_error_float16(x, x32, x_shape, y, y32,
                                                     x_shape)
        err_scale = self._compute_gradient_error_float16(
            scale, scale, scale_shape, y, y32, x_shape)
        err_offset = self._compute_gradient_error_float16(
            offset, offset, scale_shape, y, y32, x_shape)

    x_err_tolerance = 2e-3 if x_dtype == np.float16 else 1e-3
    scale_err_tolerance = 1e-3
    self.assertLess(err_x, x_err_tolerance)
    self.assertLess(err_scale, scale_err_tolerance)
    self.assertLess(err_offset, scale_err_tolerance)

 def testEmptyFails(self):
   with ops.Graph().as_default() as g:
     with self.session(graph=g):
       x = array_ops.placeholder(dtypes.float32)
       with g.gradient_override_map({"Identity": "BadGrad"}):
         y = array_ops.identity(x)
       bad = r"Empty gradient has wrong shape: expected \(0, 3\), got \(3, 0\)"
       with self.assertRaisesRegexp(ValueError, bad):
         gradient_checker.compute_gradient(x, (0, 3), y, (0, 3))
       with self.assertRaisesRegexp(ValueError, bad):
         gradient_checker.compute_gradient_error(x, (0, 3), y, (0, 3))

  def testClipByValueGradient(self):
    inputs = constant_op.constant([1.0, 2.0, 3.0, 4.0], dtype=dtypes.float32)
    outputs_1 = clip_ops.clip_by_value(inputs, 0.5, 3.5)
    min_val = constant_op.constant([0.5, 0.5, 0.5, 0.5], dtype=dtypes.float32)
    max_val = constant_op.constant([3.5, 3.5, 3.5, 3.5], dtype=dtypes.float32)
    outputs_2 = clip_ops.clip_by_value(inputs, min_val, max_val)
    with self.test_session():
      error_1 = gradient_checker.compute_gradient_error(inputs, [4],
                                                        outputs_1, [4])
      self.assertLess(error_1, 1e-4)

      error_2 = gradient_checker.compute_gradient_error(inputs, [4],
                                                        outputs_2, [4])
      self.assertLess(error_2, 1e-4)

  def testGradientsAxis0(self):
    np.random.seed(7)
    for shape in (2,), (3,), (2, 3), (3, 2), (4, 3, 2):
      data = np.random.randn(*shape)
      shapes = [shape[1:]] * shape[0]
      with self.test_session(use_gpu=True):
        # TODO(irving): Remove list() once we handle maps correctly
        xs = list(map(constant_op.constant, data))
        c = array_ops.pack(xs)
        err = gradient_checker.compute_gradient_error(xs, shapes, c, shape)
        self.assertLess(err, 1e-6)

        c = array_ops.stack(xs)
        err = gradient_checker.compute_gradient_error(xs, shapes, c, shape)
        self.assertLess(err, 1e-6)

 def testDifferentTensorShapesThroughGradientError(self):
   pseudo_random = True
   overlapping = True
   pooling_ratio = [1, math.sqrt(3), math.sqrt(2), 1]
   for num_batches in [1, 2]:
     for num_rows in [5, 13]:
       for num_cols in [5, 11]:
         for num_channels in [1, 3]:
           input_shape = (num_batches, num_rows, num_cols, num_channels)
           input_data = self._GenerateUniqueRandomInputTensor(input_shape)
           # Add some randomness to make input_data not so 'integer'
           input_data += self._PRNG.random_sample(input_shape)
           with self.cached_session() as _:
             input_tensor = constant_op.constant(input_data, shape=input_shape)
             output_tensor, unused_a, unused_b = nn_ops.fractional_max_pool_v2(
                 input_tensor,
                 pooling_ratio,
                 pseudo_random=pseudo_random,
                 overlapping=overlapping,
                 seed=self._SEED)
             output_data = self.evaluate(output_tensor)
             output_shape = output_data.shape
             # error_margin and delta setting is similar to max_pool_grad.
             error_margin = 1e-3
             gradient_error = gradient_checker.compute_gradient_error(
                 input_tensor,
                 input_shape,
                 output_tensor,
                 output_shape,
                 x_init_value=input_data.reshape(input_shape),
                 delta=1e-2)
             self.assertLess(gradient_error, error_margin)

  def testAllInputOptionsThroughGradientError(self):
    input_shape = (1, 7, 13, 1)
    input_data = self._GenerateUniqueRandomInputTensor(input_shape)
    # Add some randomness to make input_data not so 'integer'
    input_data += self._PRNG.random_sample(input_shape)
    pooling_ratio = [1, math.sqrt(2), math.sqrt(3), 1]

    for pseudo_random in True, False:
      for overlapping in True, False:
        with self.cached_session() as _:
          input_tensor = constant_op.constant(input_data, shape=input_shape)
          output_tensor, unused_a, unused_b = nn_ops.fractional_max_pool_v2(
              input_tensor,
              pooling_ratio,
              pseudo_random=pseudo_random,
              overlapping=overlapping,
              seed=self._SEED)
          output_data = self.evaluate(output_tensor)
          output_shape = output_data.shape
          # error_margin and delta setting is similar to max_pool_grad.
          error_margin = 1e-3
          gradient_error = gradient_checker.compute_gradient_error(
              input_tensor,
              input_shape,
              output_tensor,
              output_shape,
              x_init_value=input_data.reshape(input_shape),
              delta=1e-2)
          self.assertLess(gradient_error, error_margin)

  def testSecondGradient(self):
    with self.test_session() as sess:
      l = constant_op.constant(
          [
              0.0, 0.0, 1.0 / 3, 0.0, 1.0 / 3, 0.0, 0.0, 0.0, 0.0, 0.5 / 3, 0.0,
              0.5 / 3
          ],
          shape=[12],
          dtype=dtypes.float64,
          name="l")
      f = constant_op.constant(
          [0.1, 0.2, 0.3, 0.4, 0.1, 0.4, 0.9, 1.6, 0.1, 0.8, 2.7, 6.4],
          shape=[12],
          dtype=dtypes.float64,
          name="f")
      x = nn_ops.softmax_cross_entropy_with_logits(
          labels=l, logits=f, name="xent")
      loss = math_ops.reduce_sum(x)

      gradients = gradients_impl.gradients(loss, [f])[0]

      err = gradient_checker.compute_gradient_error(f, [12], gradients, [12])

      # Check that second derivative is calculated.
      # (it is equivalent to being `BatchMatMul` op in the graph because of implementation of xentropy grad)
      op_names = [
          op.op_def.name for op in sess.graph.get_operations() if op.op_def
      ]
      self.assertIn("BatchMatMul", op_names)

    print("cross entropy hessian err = ", err)
    self.assertLess(err, 5e-8)

 def testDifferentTensorShapesThroughGradientError(self):
   pseudo_random = True
   overlapping = True
   pooling_ratio = [1, math.sqrt(3), math.sqrt(2), 1]
   for num_batches in [1, 2]:
     for num_rows in [5, 13]:
       for num_cols in [5, 11]:
         for num_channels in [1, 3]:
           input_shape = (num_batches, num_rows, num_cols, num_channels)
           input_data = self._GenerateRandomInputTensor(input_shape)
           with self.cached_session() as _:
             input_tensor = constant_op.constant(input_data, shape=input_shape)
             output_tensor, unused_a, unused_b = nn_ops.fractional_avg_pool(
                 input_tensor,
                 pooling_ratio,
                 pseudo_random=pseudo_random,
                 overlapping=overlapping,
                 deterministic=True,
                 seed=self._SEED,
                 seed2=self._SEED2)
             output_data = output_tensor.eval()
             output_shape = output_data.shape
             # error_margin and delta setting is similar to avg_pool_grad.
             error_margin = 1e-4
             gradient_error = gradient_checker.compute_gradient_error(
                 input_tensor,
                 input_shape,
                 output_tensor,
                 output_shape,
                 x_init_value=input_data.reshape(input_shape),
                 delta=1e-2)
             self.assertLess(gradient_error, error_margin)