def probs_to_logits(probs, is_binary=False):
"""
converts probabilities into logits.
Args:
probs (Tensor)
is_binary (bool)
"""
ps_clamped = clamp_probs(probs)
if is_binary:
return P.Log()(ps_clamped) - P.Log()(1-ps_clamped)
return P.Log()(ps_clamped)
def __init__(self,
probs=None,
seed=0,
dtype=mstype.int32,
name="Bernoulli"):
"""
Constructor of Bernoulli distribution.
"""
param = dict(locals())
valid_dtype = mstype.int_type + mstype.uint_type
check_type(dtype, valid_dtype, "Bernoulli")
super(Bernoulli, self).__init__(seed, dtype, name, param)
self.parameter_type = mstype.float32
if probs is not None:
self._probs = cast_to_tensor(probs, mstype.float32)
check_prob(self.probs)
else:
self._probs = probs
# ops needed for the class
self.squeeze = P.Squeeze(0)
self.cast = P.Cast()
self.const = P.ScalarToArray()
self.dtypeop = P.DType()
self.erf = P.Erf()
self.exp = P.Exp()
self.floor = P.Floor()
self.fill = P.Fill()
self.log = P.Log()
self.less = P.Less()
self.shape = P.Shape()
self.select = P.Select()
self.sq = P.Square()
self.sqrt = P.Sqrt()
self.uniform = C.uniform
def __init__(self,
mean=None,
sd=None,
seed=0,
dtype=mstype.float32,
name="Normal"):
"""
Constructor of normal distribution.
"""
param = dict(locals())
super(Normal, self).__init__(dtype, name, param)
if mean is not None and sd is not None:
self._mean_value = convert_to_batch(mean, self._broadcast_shape,
dtype)
self._sd_value = convert_to_batch(sd, self._broadcast_shape, dtype)
check_greater_equal_zero(self._sd_value, "Standard deviation")
else:
self._mean_value = mean
self._sd_value = sd
#ops needed for the class
self.exp = P.Exp()
self.add = P.TensorAdd()
self.mul = P.Mul()
self.sq = P.Square()
self.log = P.Log()
self.sqrt = P.Sqrt()
self.realdiv = P.RealDiv()
self.expm1 = P.Expm1() if get_context(
'device_target') == 'Ascend' else self._expm1_by_step
self.normal = P.Normal(seed=seed)
self.shape = P.Shape()
self.zeroslike = P.ZerosLike()
self.const = P.ScalarToArray()
def __init__(self, num_classes, num_boxes, neg_pre_positive, batch_size):
super(MultiBoxLoss, self).__init__()
self.num_classes = num_classes
self.num_boxes = num_boxes
self.neg_pre_positive = neg_pre_positive
self.notequal = P.NotEqual()
self.less = P.Less()
self.tile = P.Tile()
self.reduce_sum = P.ReduceSum()
self.reduce_mean = P.ReduceMean()
self.expand_dims = P.ExpandDims()
self.smooth_l1_loss = P.SmoothL1Loss()
self.cross_entropy = SoftmaxCrossEntropyWithLogits()
self.maximum = P.Maximum()
self.minimum = P.Minimum()
self.sort_descend = P.TopK(True)
self.sort = P.TopK(True)
self.gather = P.GatherNd()
self.max = P.ReduceMax()
self.log = P.Log()
self.exp = P.Exp()
self.concat = P.Concat(axis=1)
self.reduce_sum2 = P.ReduceSum(keep_dims=True)
self.idx = Tensor(
np.reshape(np.arange(batch_size * num_boxes), (-1, 1)), ms.int32)
def __init__(self,
mean=None,
sd=None,
seed=0,
dtype=mstype.float32,
name="Normal"):
"""
Constructor of normal distribution.
"""
param = dict(locals())
super(Normal, self).__init__(dtype, name, param)
if mean is not None and sd is not None:
self._mean_value = convert_to_batch(mean, self._broadcast_shape, dtype)
self._sd_value = convert_to_batch(sd, self._broadcast_shape, dtype)
check_greater_equal_zero(self._sd_value, "Standard deviation")
else:
self._mean_value = mean
self._sd_value = sd
self.seed = seed
#ops needed for the class
self.const = P.ScalarToArray()
self.erf = P.Erf()
self.exp = P.Exp()
self.expm1 = self._expm1_by_step
self.fill = P.Fill()
self.log = P.Log()
self.shape = P.Shape()
self.sq = P.Square()
self.sqrt = P.Sqrt()
self.zeroslike = P.ZerosLike()
def __init__(self,
probs=None,
seed=0,
dtype=mstype.int32,
name="Geometric"):
"""
Constructor of Geometric distribution.
"""
param = dict(locals())
super(Geometric, self).__init__(dtype, name, param)
if probs is not None:
self._probs = cast_to_tensor(probs, dtype=mstype.float32)
check_prob(self._probs)
else:
self._probs = probs
self.minval = np.finfo(np.float).tiny
# ops needed for the class
self.const = P.ScalarToArray()
self.dtypeop = P.DType()
self.fill = P.Fill()
self.floor = P.Floor()
self.issubclass = P.IsSubClass()
self.less = P.Less()
self.log = P.Log()
self.pow = P.Pow()
self.select = P.Select()
self.shape = P.Shape()
self.sq = P.Square()
self.sqrt = P.Sqrt()
self.uniform = P.UniformReal(seed=seed)
def __init__(self, num_sampled, num_classes, num_true=1,
sampled_values=None, remove_accidental_hits=True, seed=0,
reduction='none'):
super(SampledSoftmaxLoss, self).__init__()
self.num_sampled = num_sampled
self.num_classes = num_classes
self.num_true = num_true
self.sampled_values = sampled_values
self.remove_accidental_hits = remove_accidental_hits
self.seed = seed
self.sampler = P.UniformSampler(
num_true,
num_sampled,
True,
num_classes,
seed,
remove_accidental_hits)
self.cast = P.Cast()
self.reshape = P.Reshape()
self.shape = P.Shape()
self.exp = P.Exp()
self.log = P.Log()
self.slice_op = P.Slice()
self.matmul = P.MatMul(False, True)
self.gather_v2 = P.GatherV2()
self.reduce_max_true = P.ReduceMax(True)
self.reduce_sum = P.ReduceSum()
self.reduce_sum_true = P.ReduceSum(True)
self.concat_dim0 = P.Concat(0)
self.concat_dim1 = P.Concat(1)
self.ones_like = P.OnesLike()
self.zeros_like = P.ZerosLike()
self.mul = P.Mul()
self.expand_dims = P.ExpandDims()
def __init__(self,
rate=None,
seed=0,
dtype=mstype.float32,
name="Exponential"):
"""
Constructor of Exponential distribution.
"""
param = dict(locals())
valid_dtype = mstype.float_type
check_type(dtype, valid_dtype, "Exponential")
super(Exponential, self).__init__(seed, dtype, name, param)
if rate is not None:
self._rate = cast_to_tensor(rate, dtype)
check_greater_zero(self._rate, "rate")
else:
self._rate = rate
self.minval = np.finfo(np.float).tiny
# ops needed for the class
self.cast = P.Cast()
self.const = P.ScalarToArray()
self.dtypeop = P.DType()
self.exp = P.Exp()
self.fill = P.Fill()
self.less = P.Less()
self.log = P.Log()
self.select = P.Select()
self.shape = P.Shape()
self.sqrt = P.Sqrt()
self.sq = P.Square()
self.uniform = C.uniform
def __init__(self, sparse=False, stra_list=None):
super(SoftmaxCrossEntropyExpand, self).__init__()
if stra_list is None:
stra_list = []
if len(stra_list) < 11:
stra_list = [None] * 11
self.exp = P.Exp()
self.reduce_sum = P.ReduceSum(keep_dims=True).set_strategy(
strategy=stra_list[1])
self.onehot = P.OneHot().set_strategy(strategy=stra_list[2])
self.on_value = Tensor(1.0, mstype.float32)
self.off_value = Tensor(0.0, mstype.float32)
self.div = P.Div().set_strategy(strategy=stra_list[3])
self.log = P.Log().set_strategy(strategy=stra_list[4])
self.sum_cross_entropy = P.ReduceSum(keep_dims=False).set_strategy(
strategy=stra_list[5])
self.mul = P.Mul().set_strategy(strategy=stra_list[6])
self.mul2 = P.Mul().set_strategy(strategy=stra_list[7])
self.cast = P.Cast()
self.reduce_mean = P.ReduceMean(keep_dims=False).set_strategy(
strategy=stra_list[8])
self.sparse = sparse
self.reduce_max = P.ReduceMax(keep_dims=True).set_strategy(
strategy=stra_list[9])
self.sub = P.Sub().set_strategy(strategy=stra_list[10])
def __init__(self, tag_to_index, batch_size=1, seq_length=128, is_training=True):
super(CRF, self).__init__()
self.target_size = len(tag_to_index)
self.is_training = is_training
self.tag_to_index = tag_to_index
self.batch_size = batch_size
self.seq_length = seq_length
self.START_TAG = "<START>"
self.STOP_TAG = "<STOP>"
self.START_VALUE = Tensor(self.target_size-2, dtype=mstype.int32)
self.STOP_VALUE = Tensor(self.target_size-1, dtype=mstype.int32)
transitions = np.random.normal(size=(self.target_size, self.target_size)).astype(np.float32)
transitions[tag_to_index[self.START_TAG], :] = -10000
transitions[:, tag_to_index[self.STOP_TAG]] = -10000
self.transitions = Parameter(Tensor(transitions), name="transition_matrix")
self.cat = P.Concat(axis=-1)
self.argmax = P.ArgMaxWithValue(axis=-1)
self.log = P.Log()
self.exp = P.Exp()
self.sum = P.ReduceSum()
self.tile = P.Tile()
self.reduce_sum = P.ReduceSum(keep_dims=True)
self.reshape = P.Reshape()
self.expand = P.ExpandDims()
self.mean = P.ReduceMean()
init_alphas = np.ones(shape=(self.batch_size, self.target_size)) * -10000.0
init_alphas[:, self.tag_to_index[self.START_TAG]] = 0.
self.init_alphas = Tensor(init_alphas, dtype=mstype.float32)
self.cast = P.Cast()
self.reduce_max = P.ReduceMax(keep_dims=True)
self.on_value = Tensor(1.0, dtype=mstype.float32)
self.off_value = Tensor(0.0, dtype=mstype.float32)
self.onehot = P.OneHot()
def __init__(self,
config,
roi_layer,
out_channels,
featmap_strides,
batch_size=1,
finest_scale=56,
mask=False):
super(SingleRoIExtractor, self).__init__()
cfg = config
self.train_batch_size = batch_size
self.out_channels = out_channels
self.featmap_strides = featmap_strides
self.num_levels = len(self.featmap_strides)
self.out_size = roi_layer['mask_out_size'] if mask else roi_layer['out_size']
self.mask = mask
self.sample_num = roi_layer['sample_num']
self.roi_layers = self.build_roi_layers(self.featmap_strides)
self.roi_layers = L.CellList(self.roi_layers)
self.sqrt = P.Sqrt()
self.log = P.Log()
self.finest_scale_ = finest_scale
self.clamp = C.clip_by_value
self.cast = P.Cast()
self.equal = P.Equal()
self.select = P.Select()
_mode_16 = False
self.dtype = np.float16 if _mode_16 else np.float32
self.ms_dtype = mstype.float16 if _mode_16 else mstype.float32
self.set_train_local(cfg, training=True)
def log_generic(input_x):
"""
Log op on Ascend is calculated as log(abs(x)).
Fix this with putting negative values as nan.
And log op on Ascend doesn't supprot int types.
Fix this with casting the type.
"""
log = P.Log()
less = P.Less()
lessequal = P.LessEqual()
fill = P.Fill()
cast = P.Cast()
dtype = P.DType()
shape = P.Shape()
select = P.Select()
checktype = P.IsSubClass()
if not checktype(dtype(input_x), mstype.float_):
input_x = cast(input_x, mstype.float32)
nan = fill(dtype(input_x), shape(input_x), np.nan)
inf = fill(dtype(input_x), shape(input_x), np.inf)
neg_x = less(input_x, 0.0)
nonpos_x = lessequal(input_x, 0.0)
log_x = log(input_x)
result = select(nonpos_x, -inf, log_x)
return select(neg_x, nan, result)
def __init__(self, config):
super(CrossEntropyLoss, self).__init__()
self.mean = P.ReduceMean()
self.sum = P.ReduceSum().shard(((config.dp, config.mp),))
self.onehot = P.OneHot().shard(((config.dp, config.mp), (), ()))
self.on_value = Tensor(1.0, mstype.float32)
self.off_value = Tensor(0.0, mstype.float32)
self.vocab_size = config.vocab_size
self.max = P.ArgMaxWithValue(axis=-1, keep_dims=True).shard(
((config.dp, config.mp),))
self.eps_const = Tensor(1e-24, mstype.float32)
self.sub = P.Sub().shard(((config.dp, config.mp), (config.dp, 1)))
self.exp = P.Exp().shard(((config.dp, config.mp),))
self.div = P.RealDiv().shard(((config.dp, config.mp), (config.dp, 1)))
self.log = P.Log().shard(((config.dp, config.mp),))
self.add = P.TensorAdd().shard(((config.dp, config.mp), ()))
self.mul = P.Mul().shard(
((config.dp, config.mp), (config.dp, config.mp)))
self.neg = P.Neg().shard(((config.dp, config.mp),))
self.sum2 = P.ReduceSum().shard(((1,),))
self.mul2 = P.Mul().shard(((1,), (1,)))
self.add2 = P.TensorAdd()
self.div2 = P.RealDiv()
if config.stage_num > 1:
self.div2.add_prim_attr("end", 0)
def __init__(self, params, learning_rate, momentum, matrix_A, matrix_G, A_inv_max, G_inv_max, weight_decay=0.0,
loss_scale=1.0,
decay_filter=lambda x: x.name not in []):
super(THOR, self).__init__(learning_rate, params, weight_decay, loss_scale)
if isinstance(momentum, float) and momentum < 0.0:
raise ValueError("momentum should be at least 0.0, but got momentum {}".format(momentum))
self.momentum = Parameter(Tensor(momentum, mstype.float32))
self.params = self.parameters
self.moments = self.params.clone(prefix="moments", init='zeros')
self.hyper_map = C.HyperMap()
self.opt = P.ApplyMomentum()
self.matrix_A = ParameterTuple(matrix_A)
self.matrix_G = ParameterTuple(matrix_G)
self.A_inv_max = ParameterTuple(A_inv_max)
self.G_inv_max = ParameterTuple(G_inv_max)
self.cube_matmul_left = P.CusMatMulCubeFraczLeftCast()
self.cube_matmul_left_fc = P.CusMatMulCubeDenseLeft()
self.cube_matmul_right_fc = P.CusMatMulCubeDenseRight()
self.cube_matmul_right_mul = P.CusMatMulCubeFraczRightMul()
self.transpose = P.Transpose()
self.shape = P.Shape()
self.reshape = P.Reshape()
self.mul = P.Mul()
self.weight_idx = []
for i in range(len(self.params)):
if "conv" in self.params[i].name or "end_point" in self.params[i].name:
self.weight_idx.append(i)
self.weight_idx.append(len(self.params))
self.feature_map = [1.0 / 12544, 1.0 / 3136, 1.0 / 3136, 1.0 / 3136, 1.0 / 3136, 1.0 / 3136, 1.0 / 3136,
1.0 / 3136, 1.0 / 3136, 1.0 / 3136, 1.0 / 3136, 1.0 / 3136,
1.0 / 784, 1.0 / 784, 1.0 / 784, 1.0 / 784, 1.0 / 784, 1.0 / 784, 1.0 / 784, 1.0 / 784,
1.0 / 784, 1.0 / 784, 1.0 / 784, 1.0 / 784, 1.0 / 784,
1.0 / 196, 1.0 / 196, 1.0 / 196, 1.0 / 196, 1.0 / 196, 1.0 / 196, 1.0 / 196, 1.0 / 196,
1.0 / 196, 1.0 / 196, 1.0 / 196, 1.0 / 196, 1.0 / 196, 1.0 / 196, 1.0 / 196, 1.0 / 196,
1.0 / 196, 1.0 / 196, 1.0 / 196,
1.0 / 49, 1.0 / 49, 1.0 / 49, 1.0 / 49, 1.0 / 49, 1.0 / 49, 1.0 / 49, 1.0 / 49, 1.0 / 49,
1.0]
mean = _get_gradients_mean()
degree = _get_device_num()
parameter_length = len(self.feature_map)
self.grad_reducer_Amax = DistributedGradReducerThor(parameter_length, ((27,), 2), mean, degree)
self.grad_reducer_Gmax = DistributedGradReducerThor(parameter_length, ((27,), 4), mean, degree)
self.grad_reducer_A = DistributedGradReducerThor(parameter_length, ((27,), 6), mean, degree)
self.grad_reducer_G = DistributedGradReducerThor(parameter_length, ((27,), 8), mean, degree)
self.matrix_A_inv = ()
self.matrix_G_inv = ()
self.matrix_max_inv = ()
for i in range(54):
self.matrix_max_inv = self.matrix_max_inv + (
Parameter(initializer(1, [1], mstype.float32), name="matrix_max" + str(i), requires_grad=False),)
self.log = P.Log()
self.exp = P.Exp()
self.sqrt = P.Sqrt()
self.matrix_max_inv = ParameterTuple(self.matrix_max_inv)
self.assign = P.Assign()
self.cast = P.Cast()
self.thor = True
self.weight_decay = weight_decay * loss_scale
self.decay_flags = tuple(decay_filter(x) for x in self.parameters)
def __init__(self,
probs=None,
seed=0,
dtype=mstype.int32,
name="Bernoulli"):
"""
Constructor of Bernoulli distribution.
"""
param = dict(locals())
super(Bernoulli, self).__init__(dtype, name, param)
if probs is not None:
self._probs = cast_to_tensor(probs, dtype=mstype.float32)
check_prob(self.probs)
else:
self._probs = probs
self.seed = seed
# ops needed for the class
self.cast = P.Cast()
self.const = P.ScalarToArray()
self.dtypeop = P.DType()
self.erf = P.Erf()
self.fill = P.Fill()
self.log = P.Log()
self.less = P.Less()
self.shape = P.Shape()
self.select = P.Select()
self.sq = P.Square()
self.sqrt = P.Sqrt()
self.uniform = P.UniformReal(seed=seed)
def __init__(self):
super(IGamma, self).__init__()
# const numbers
# If more data types are supported, this float max value need to be selected.
self.log_maxfloat32 = Tensor(np.log(np.finfo(np.float32).max),
mstype.float32)
# operations
self.logicaland = P.LogicalAnd()
self.logicalor = P.LogicalOr()
self.logicalnot = P.LogicalNot()
self.equal = P.Equal()
self.greater = P.Greater()
self.less = P.Less()
self.neg = P.Neg()
self.log = P.Log()
self.exp = P.Exp()
self.select = P.Select()
self.zeroslike = P.ZerosLike()
self.fill = P.Fill()
self.shape = P.Shape()
self.dtype = P.DType()
self.lgamma = LGamma()
self.const = P.ScalarToArray()
self.cast = P.Cast()
def __init__(self):
super(LogSigmoid, self).__init__()
self.mul = P.Mul()
self.exp = P.Exp()
self.add = P.TensorAdd()
self.rec = P.Reciprocal()
self.log = P.Log()
def __init__(self,
probs=None,
seed=0,
dtype=mstype.int32,
name="Bernoulli"):
"""
Constructor of Bernoulli distribution.
"""
param = dict(locals())
super(Bernoulli, self).__init__(dtype, name, param)
if probs is not None:
self._probs = cast_to_tensor(probs)
check_prob(self._probs)
else:
self._probs = probs
# ops needed for the class
self.log = P.Log()
self.add = P.TensorAdd()
self.mul = P.Mul()
self.sqrt = P.Sqrt()
self.realdiv = P.RealDiv()
self.shape = P.Shape()
self.const = P.ScalarToArray()
self.less = P.Less()
self.cast = P.Cast()
self.normal = P.Normal(seed=seed)
self.erf = P.Erf()
self.sqrt = P.Sqrt()
def __init__(self):
super(BoundingBoxEncode, self).__init__()
self.split = P.Split(axis=1, output_num=4)
self.ones = 1.0
self.half = 0.5
self.log = P.Log()
self.concat = P.Concat(axis=1)
def __init__(self,
low=None,
high=None,
seed=0,
dtype=mstype.float32,
name="Uniform"):
"""
Constructor of Uniform distribution.
"""
param = dict(locals())
super(Uniform, self).__init__(dtype, name, param)
if low is not None and high is not None:
self._low = convert_to_batch(low, self._broadcast_shape, dtype)
self._high = convert_to_batch(high, self._broadcast_shape, dtype)
check_greater(self.low, self.high, "low value", "high value")
else:
self._low = low
self._high = high
# ops needed for the class
self.const = P.ScalarToArray()
self.dtypeop = P.DType()
self.exp = P.Exp()
self.fill = P.Fill()
self.less = P.Less()
self.lessequal = P.LessEqual()
self.log = P.Log()
self.logicaland = P.LogicalAnd()
self.select = P.Select()
self.shape = P.Shape()
self.sq = P.Square()
self.sqrt = P.Sqrt()
self.uniform = P.UniformReal(seed=seed)
self.zeroslike = P.ZerosLike()
def __init__(self, axis, keep_dims=False):
super(ReduceLogSumExp, self).__init__()
validator.check_value_type('axis', axis, [int, list, tuple], self.cls_name)
validator.check_value_type('keep_dims', keep_dims, [bool], self.cls_name)
self.axis = axis
self.exp = P.Exp()
self.sum = P.ReduceSum(keep_dims)
self.log = P.Log()
def construct(self, prediction, pred_xy, pred_wh, y_true, gt_box, input_shape):
"""
prediction : origin output from yolo
pred_xy: (sigmoid(xy)+grid)/grid_size
pred_wh: (exp(wh)*anchors)/input_shape
y_true : after normalize
gt_box: [batch, maxboxes, xyhw] after normalize
"""
object_mask = y_true[:, :, :, :, 4:5]
class_probs = y_true[:, :, :, :, 5:]
true_boxes = y_true[:, :, :, :, :4]
grid_shape = P.Shape()(prediction)[1:3]
grid_shape = P.Cast()(F.tuple_to_array(grid_shape[::-1]), ms.float32)
pred_boxes = self.concat((pred_xy, pred_wh))
true_wh = y_true[:, :, :, :, 2:4]
true_wh = P.Select()(P.Equal()(true_wh, 0.0),
P.Fill()(P.DType()(true_wh),
P.Shape()(true_wh), 1.0),
true_wh)
true_wh = P.Log()(true_wh / self.anchors * input_shape)
# 2-w*h for large picture, use small scale, since small obj need more precise
box_loss_scale = 2 - y_true[:, :, :, :, 2:3] * y_true[:, :, :, :, 3:4]
gt_shape = P.Shape()(gt_box)
gt_box = P.Reshape()(gt_box, (gt_shape[0], 1, 1, 1, gt_shape[1], gt_shape[2]))
# add one more dimension for broadcast
iou = self.iou(P.ExpandDims()(pred_boxes, -2), gt_box)
# gt_box is x,y,h,w after normalize
# [batch, grid[0], grid[1], num_anchor, num_gt]
best_iou = self.reduce_max(iou, -1)
# [batch, grid[0], grid[1], num_anchor]
# ignore_mask IOU too small
ignore_mask = best_iou < self.ignore_threshold
ignore_mask = P.Cast()(ignore_mask, ms.float32)
ignore_mask = P.ExpandDims()(ignore_mask, -1)
# ignore_mask backpro will cause a lot maximunGrad and minimumGrad time consume.
# so we turn off its gradient
ignore_mask = F.stop_gradient(ignore_mask)
confidence_loss = self.confidence_loss(object_mask, prediction[:, :, :, :, 4:5], ignore_mask)
class_loss = self.class_loss(object_mask, prediction[:, :, :, :, 5:], class_probs)
object_mask_me = P.Reshape()(object_mask, (-1, 1)) # [8, 72, 72, 3, 1]
box_loss_scale_me = P.Reshape()(box_loss_scale, (-1, 1))
pred_boxes_me = xywh2x1y1x2y2(pred_boxes)
pred_boxes_me = P.Reshape()(pred_boxes_me, (-1, 4))
true_boxes_me = xywh2x1y1x2y2(true_boxes)
true_boxes_me = P.Reshape()(true_boxes_me, (-1, 4))
ciou = self.giou(pred_boxes_me, true_boxes_me)
ciou_loss = object_mask_me * box_loss_scale_me * (1 - ciou)
ciou_loss_me = self.reduce_sum(ciou_loss, ())
loss = ciou_loss_me * 10 + confidence_loss + class_loss
batch_size = P.Shape()(prediction)[0]
return loss / batch_size
def __init__(self,
num_sampled,
num_classes,
num_true=1,
sampled_values=None,
remove_accidental_hits=True,
seed=0,
reduction='none'):
super(SampledSoftmaxLoss, self).__init__(reduction)
if num_true < 1:
raise ValueError(f"num_true {num_true} is less than 1.")
if seed < 0:
raise ValueError(f"seed {seed} is less than 0.")
if num_sampled > num_classes:
raise ValueError(
f"num_sampled {num_sampled} is great than num_classes {num_classes}."
)
if num_true > num_classes:
raise ValueError(
f"num_true {num_true} is great than num_classes {num_classes}."
)
if sampled_values is not None:
if not isinstance(sampled_values, (list, tuple)):
raise TypeError(
f"sampled_values {sampled_values} is not a list.")
if len(sampled_values) != 3:
raise ValueError(
f"sampled_values size {len(sampled_values)} is not 3.")
self.num_sampled = num_sampled
self.num_classes = num_classes
self.num_true = num_true
self.sampled_values = sampled_values
self.remove_accidental_hits = remove_accidental_hits
self.seed = seed
self.sampler = P.LogUniformCandidateSampler(num_true, num_sampled,
True, num_classes, seed)
self.cast = P.Cast()
self.reshape = P.Reshape()
self.shape = P.Shape()
self.exp = P.Exp()
self.log = P.Log()
self.slice_op = P.Slice()
self.matmul = P.MatMul(False, True)
self.gather_v2 = P.Gather()
self.reduce_max_true = P.ReduceMax(True)
self.reduce_sum = P.ReduceSum()
self.reduce_sum_true = P.ReduceSum(True)
self.concat_dim0 = P.Concat(0)
self.concat_dim1 = P.Concat(1)
self.ones_like = P.OnesLike()
self.zeros_like = P.ZerosLike()
self.mul = P.Mul()
self.expand_dims = P.ExpandDims()
self.dtype = P.DType()
self.compute_accidental_hits = P.ComputeAccidentalHits(num_true)
self.scatter_nd = P.ScatterNd()
def __init__(self, power=0, name='PowerTransform', param=None):
param = dict(locals()) if param is None else param
super(PowerTransform, self).__init__(name=name, param=param)
validator.check_value_type('power', power, [int, float], self.name)
self._power = power
self.pow = P.Pow()
self.exp = P.Exp()
self.log = P.Log()
self.log1p = self._log1p_by_step
self.expm1 = self._expm1_by_step
def construct(self, img1, img2):
max_val = _convert_img_dtype_to_float32(self.max_val, self.max_val)
img1 = _convert_img_dtype_to_float32(img1, self.max_val)
img2 = _convert_img_dtype_to_float32(img2, self.max_val)
mse = P.ReduceMean()(F.square(img1 - img2), (-3, -2, -1))
# 10*log_10(max_val^2/MSE)
psnr = 10 * P.Log()(F.square(max_val) / mse) / F.scalar_log(10.0)
return psnr
def __init__(self, weight_angle=10):
super(LossFunc, self).__init__()
self.split = P.Split(1, 5)
self.min = P.Minimum()
self.log = P.Log()
self.cos = P.Cos()
self.mean = P.ReduceMean()
#self.flatten = P.Flatten()
self.sum = P.ReduceSum()
self.weight_angle = weight_angle
self.max = P.Maximum()
self.print = P.Print()
def __init__(self):
super(CrossEntropy, self).__init__()
self.exp = P.Exp()
self.sum = P.ReduceSum()
self.reshape = P.Reshape()
self.log = P.Log()
self.cast = P.Cast()
self.eps_const = Tensor(eps, dtype=mstype.float32)
self.onehot = P.OneHot()
self.on_value = Tensor(1.0, mstype.float32)
self.off_value = Tensor(0.0, mstype.float32)
def construct(self, img1, img2):
_check_input_4d(F.shape(img1), "img1", self.cls_name)
_check_input_4d(F.shape(img2), "img2", self.cls_name)
P.SameTypeShape()(img1, img2)
max_val = _convert_img_dtype_to_float32(self.max_val, self.max_val)
img1 = _convert_img_dtype_to_float32(img1, self.max_val)
img2 = _convert_img_dtype_to_float32(img2, self.max_val)
mse = P.ReduceMean()(F.square(img1 - img2), (-3, -2, -1))
psnr = 10 * P.Log()(F.square(max_val) / mse) / F.scalar_log(10.0)
return psnr
def __init__(self, temp=0.1):
super(LossNet, self).__init__()
self.concat = P.Concat()
self.exp = P.Exp()
self.t = P.Transpose()
self.diag_part = P.DiagPart()
self.matmul = P.MatMul()
self.sum = P.ReduceSum()
self.sum_keep_dim = P.ReduceSum(keep_dims=True)
self.log = P.Log()
self.mean = P.ReduceMean()
self.shape = P.Shape()
self.eye = P.Eye()
self.temp = temp
def construct(self, grid, prediction, pred_xy, pred_wh, y_true, gt_box):
object_mask = y_true[:, :, :, :, 4:5]
class_probs = y_true[:, :, :, :, 5:]
grid_shape = P.Shape()(prediction)[1:3]
grid_shape = P.Cast()(F.tuple_to_array(grid_shape[::-1]), ms.float32)
pred_boxes = self.concat((pred_xy, pred_wh))
true_xy = y_true[:, :, :, :, :2] * grid_shape - grid
true_wh = y_true[:, :, :, :, 2:4]
true_wh = P.Select()(P.Equal()(true_wh,
0.0), P.Fill()(P.DType()(true_wh),
P.Shape()(true_wh), 1.0),
true_wh)
true_wh = P.Log()(true_wh / self.anchors * self.input_shape)
box_loss_scale = 2 - y_true[:, :, :, :, 2:3] * y_true[:, :, :, :, 3:4]
gt_shape = P.Shape()(gt_box)
gt_box = P.Reshape()(gt_box,
(gt_shape[0], 1, 1, 1, gt_shape[1], gt_shape[2]))
iou = self.iou(P.ExpandDims()(pred_boxes, -2),
gt_box) # [batch, grid[0], grid[1], num_anchor, num_gt]
best_iou = self.reduce_max(iou,
-1) # [batch, grid[0], grid[1], num_anchor]
ignore_mask = best_iou < self.ignore_threshold
ignore_mask = P.Cast()(ignore_mask, ms.float32)
ignore_mask = P.ExpandDims()(ignore_mask, -1)
ignore_mask = F.stop_gradient(ignore_mask)
xy_loss = object_mask * box_loss_scale * self.cross_entropy(
prediction[:, :, :, :, :2], true_xy)
wh_loss = object_mask * box_loss_scale * 0.5 * P.Square()(
true_wh - prediction[:, :, :, :, 2:4])
confidence_loss = self.cross_entropy(prediction[:, :, :, :, 4:5],
object_mask)
confidence_loss = object_mask * confidence_loss + (
1 - object_mask) * confidence_loss * ignore_mask
class_loss = object_mask * self.cross_entropy(
prediction[:, :, :, :, 5:], class_probs)
# Get smooth loss
xy_loss = self.reduce_sum(xy_loss, ())
wh_loss = self.reduce_sum(wh_loss, ())
confidence_loss = self.reduce_sum(confidence_loss, ())
class_loss = self.reduce_sum(class_loss, ())
loss = xy_loss + wh_loss + confidence_loss + class_loss
return loss / P.Shape()(prediction)[0]
