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,
config,
batch_size,
num_classes,
use_sigmoid_cls,
target_means=(.0, .0, .0, .0),
target_stds=(1.0, 1.0, 1.0, 1.0)
):
super(Proposal, self).__init__()
cfg = config
self.batch_size = batch_size
self.num_classes = num_classes
self.target_means = target_means
self.target_stds = target_stds
self.use_sigmoid_cls = config.use_sigmoid_cls
if self.use_sigmoid_cls:
self.cls_out_channels = 1
self.activation = P.Sigmoid()
self.reshape_shape = (-1, 1)
else:
self.cls_out_channels = num_classes
self.activation = P.Softmax(axis=1)
self.reshape_shape = (-1, 2)
if self.cls_out_channels <= 0:
raise ValueError('num_classes={} is too small'.format(num_classes))
self.num_pre = cfg.rpn_proposal_nms_pre
self.min_box_size = cfg.rpn_proposal_min_bbox_size
self.nms_thr = cfg.rpn_proposal_nms_thr
self.nms_post = cfg.rpn_proposal_nms_post
self.nms_across_levels = cfg.rpn_proposal_nms_across_levels
self.max_num = cfg.rpn_proposal_max_num
# Op Define
self.squeeze = P.Squeeze()
self.reshape = P.Reshape()
self.cast = P.Cast()
self.feature_shapes = cfg.feature_shapes
self.transpose_shape = (1, 2, 0)
self.decode = BoundingBoxDecode()
self.nms = P.NMSWithMask(self.nms_thr)
self.concat_axis0 = P.Concat(axis=0)
self.concat_axis1 = P.Concat(axis=1)
self.split = P.Split(axis=1, output_num=5)
self.min = P.Minimum()
self.gatherND = P.GatherNd()
self.slice = P.Slice()
self.select = P.Select()
self.greater = P.Greater()
self.transpose = P.Transpose()
self.tile = P.Tile()
self.set_train_local(config, training=True)
self.multi_10 = Tensor(10.0, mstype.float16)
def __init__(self,
probs=None,
seed=None,
dtype=mstype.int32,
name="Categorical"):
param = dict(locals())
param['param_dict'] = {'probs': probs}
valid_dtype = mstype.uint_type + mstype.int_type + mstype.float_type
Validator.check_type_name("dtype", dtype, valid_dtype,
type(self).__name__)
super(Categorical, self).__init__(seed, dtype, name, param)
self._probs = self._add_parameter(probs, 'probs')
if self.probs is not None:
check_rank(self.probs)
check_prob(self.probs)
check_sum_equal_one(probs)
# update is_scalar_batch and broadcast_shape
# drop one dimension
if self.probs.shape[:-1] == ():
self._is_scalar_batch = True
self._broadcast_shape = self._broadcast_shape[:-1]
self.argmax = P.ArgMaxWithValue(axis=-1)
self.broadcast = broadcast_to
self.cast = P.Cast()
self.clip_by_value = C.clip_by_value
self.concat = P.Concat(-1)
self.cumsum = P.CumSum()
self.dtypeop = P.DType()
self.exp = exp_generic
self.expand_dim = P.ExpandDims()
self.fill = P.Fill()
self.gather = P.GatherNd()
self.greater = P.Greater()
self.issubclass = P.IsSubClass()
self.less = P.Less()
self.log = log_generic
self.log_softmax = P.LogSoftmax()
self.logicor = P.LogicalOr()
self.logicand = P.LogicalAnd()
self.multinomial = P.Multinomial(seed=self.seed)
self.reshape = P.Reshape()
self.reduce_sum = P.ReduceSum(keep_dims=True)
self.select = P.Select()
self.shape = P.Shape()
self.softmax = P.Softmax()
self.squeeze = P.Squeeze()
self.squeeze_first_axis = P.Squeeze(0)
self.squeeze_last_axis = P.Squeeze(-1)
self.square = P.Square()
self.transpose = P.Transpose()
self.is_nan = P.IsNan()
self.index_type = mstype.int32
self.nan = np.nan
def test_case2_indices_to_matrix():
op = P.GatherNd()
op_wrapper = OpNetWrapper(op)
indices = Tensor(np.array([[1], [0]]), mindspore.int32)
params = Tensor(np.array([[0, 1], [2, 3]]), mindspore.float32)
outputs = op_wrapper(params, indices)
print(outputs)
expected = [[2, 3], [0, 1]]
assert np.allclose(outputs.asnumpy(), np.array(expected))
def test_case1_basic_func():
op = P.GatherNd()
op_wrapper = OpNetWrapper(op)
indices = Tensor(np.array([[0, 0], [1, 1]]), mindspore.int32)
params = Tensor(np.array([[0, 1], [2, 3]]), mindspore.float32)
outputs = op_wrapper(params, indices)
print(outputs)
expected = [0, 3]
assert np.allclose(outputs.asnumpy(), np.array(expected))
def __init__(self, config, batch_size, num_bboxes, add_gt_as_proposals):
super(BboxAssignSample, self).__init__()
cfg = config
self.batch_size = batch_size
self.neg_iou_thr = Tensor(cfg.neg_iou_thr, mstype.float16)
self.pos_iou_thr = Tensor(cfg.pos_iou_thr, mstype.float16)
self.min_pos_iou = Tensor(cfg.min_pos_iou, mstype.float16)
self.zero_thr = Tensor(0.0, mstype.float16)
self.num_bboxes = num_bboxes
self.num_gts = cfg.num_gts
self.num_expected_pos = cfg.num_expected_pos
self.num_expected_neg = cfg.num_expected_neg
self.add_gt_as_proposals = add_gt_as_proposals
if self.add_gt_as_proposals:
self.label_inds = Tensor(np.arange(1, self.num_gts + 1))
self.concat = P.Concat(axis=0)
self.max_gt = P.ArgMaxWithValue(axis=0)
self.max_anchor = P.ArgMaxWithValue(axis=1)
self.sum_inds = P.ReduceSum()
self.iou = P.IOU()
self.greaterequal = P.GreaterEqual()
self.greater = P.Greater()
self.select = P.Select()
self.gatherND = P.GatherNd()
self.squeeze = P.Squeeze()
self.cast = P.Cast()
self.logicaland = P.LogicalAnd()
self.less = P.Less()
self.random_choice_with_mask_pos = P.RandomChoiceWithMask(self.num_expected_pos)
self.random_choice_with_mask_neg = P.RandomChoiceWithMask(self.num_expected_neg)
self.reshape = P.Reshape()
self.equal = P.Equal()
self.bounding_box_encode = BoundingBoxEncode()
self.scatterNdUpdate = P.ScatterNdUpdate()
self.scatterNd = P.ScatterNd()
self.logicalnot = P.LogicalNot()
self.tile = P.Tile()
self.zeros_like = P.ZerosLike()
self.assigned_gt_inds = Tensor(np.array(-1 * np.ones(num_bboxes), dtype=np.int32))
self.assigned_gt_zeros = Tensor(np.array(np.zeros(num_bboxes), dtype=np.int32))
self.assigned_gt_ones = Tensor(np.array(np.ones(num_bboxes), dtype=np.int32))
self.assigned_gt_ignores = Tensor(np.array(-1 * np.ones(num_bboxes), dtype=np.int32))
self.assigned_pos_ones = Tensor(np.array(np.ones(self.num_expected_pos), dtype=np.int32))
self.check_neg_mask = Tensor(np.array(np.ones(self.num_expected_neg - self.num_expected_pos), dtype=np.bool))
self.range_pos_size = Tensor(np.arange(self.num_expected_pos).astype(np.float16))
self.check_gt_one = Tensor(np.array(-1 * np.ones((self.num_gts, 4)), dtype=np.float16))
self.check_anchor_two = Tensor(np.array(-2 * np.ones((self.num_bboxes, 4)), dtype=np.float16))
self.print = P.Print()
def test_case4():
op = P.GatherNd()
op_wrapper = OpNetWrapper(op)
indices = Tensor(np.array([[0, 1], [1, 0]]), mindspore.int32) # (2, 2)
params = Tensor(np.array([[[0, 1], [2, 3]], [[4, 5], [6, 7]]]),
mindspore.float32) # (2, 2, 2)
outputs = op_wrapper(params, indices)
print(outputs)
expected = [[2, 3], [4, 5]] # (2, 2)
assert np.allclose(outputs.asnumpy(), np.array(expected))
def __init__(self,
config,
batch_size,
num_classes,
target_means=(0., 0., 0., 0.),
target_stds=(0.1, 0.1, 0.2, 0.2)):
super(RcnnCls, self).__init__()
cfg = config
self.rcnn_loss_cls_weight = Tensor(
np.array(cfg.rcnn_loss_cls_weight).astype(np.float16))
self.rcnn_loss_reg_weight = Tensor(
np.array(cfg.rcnn_loss_reg_weight).astype(np.float16))
self.rcnn_fc_out_channels = cfg.rcnn_fc_out_channels
self.target_means = target_means
self.target_stds = target_stds
self.num_classes = num_classes
self.in_channels = cfg.rcnn_in_channels
self.train_batch_size = batch_size
self.test_batch_size = cfg.test_batch_size
self.fpn_cls = FpnCls(self.in_channels, self.rcnn_fc_out_channels,
self.num_classes, cfg.roi_layer["out_size"])
self.relu = P.ReLU()
self.logicaland = P.LogicalAnd()
self.loss_cls = P.SoftmaxCrossEntropyWithLogits()
self.loss_bbox = P.SmoothL1Loss(beta=1.0)
self.loss_mask = P.SigmoidCrossEntropyWithLogits()
self.reshape = P.Reshape()
self.onehot = P.OneHot()
self.greater = P.Greater()
self.cast = P.Cast()
self.sum_loss = P.ReduceSum()
self.tile = P.Tile()
self.expandims = P.ExpandDims()
self.gather = P.GatherNd()
self.argmax = P.ArgMaxWithValue(axis=1)
self.on_value = Tensor(1.0, mstype.float32)
self.off_value = Tensor(0.0, mstype.float32)
self.value = Tensor(1.0, mstype.float16)
self.num_bboxes = (cfg.num_expected_pos_stage2 +
cfg.num_expected_neg_stage2) * batch_size
rmv_first = np.ones((self.num_bboxes, self.num_classes))
rmv_first[:, 0] = np.zeros((self.num_bboxes, ))
self.rmv_first_tensor = Tensor(rmv_first.astype(np.float16))
self.num_bboxes_test = cfg.rpn_max_num * cfg.test_batch_size
def __init__(self, config):
super(LabelSmoothedCrossEntropyCriterion, self).__init__()
self.vocab_size = config.vocab_size
self.batch_size = config.batch_size
self.smoothing = 0.1
self.confidence = 0.9
self.last_idx = (-1,)
self.reduce_sum = P.ReduceSum()
self.reduce_mean = P.ReduceMean()
self.reshape = P.Reshape()
self.neg = P.Neg()
self.cast = P.Cast()
self.index_ids = Tensor(np.arange(config.batch_size * config.max_decode_length).reshape((-1, 1)), mstype.int32)
self.gather_nd = P.GatherNd()
self.expand = P.ExpandDims()
self.concat = P.Concat(axis=-1)
def _get_real_word(self, select_word, real_word_index):
"""
Get word index in UNSORTED(logits generated from GPT2Model without passing through a TopKTopPFilter) ids and logits.
"""
# mindspore.ops.Gather is supported in MindSpore v.1.0 on Ascend
if self.device_target == "Ascend" or self.device_target == "GPU":
select_word_np = select_word.asnumpy()
range_index = np.arange(0, self.batch_size)
select_word_merge = np.array([[index, word] for index, word in zip(range_index, select_word_np)], dtype=int)
word_index_2D = Tensor(select_word_merge, dtype=mstype.int32)
real_selected_word_ids = P.GatherNd()( real_word_index,word_index_2D)
#Tensor shape: (batch_size,)
else:
raise NotImplementedError('CPU and other backend types have not been supported yet')
return real_selected_word_ids
def __init__(self,
probs=None,
logits=None,
seed=None,
dtype=mstype.int32,
name="Categorical"):
param = dict(locals())
valid_dtype = mstype.int_type
check_type(dtype, valid_dtype, "Categorical")
super(Categorical, self).__init__(seed, dtype, name, param)
if (probs is None) == (logits is None):
raise_probs_logits_error()
self.reduce_sum = P.ReduceSum(keep_dims=True)
self.reduce_sum1 = P.ReduceSum(keep_dims=False)
self.log = P.Log()
self.exp = P.Exp()
self.shape = P.Shape()
self.reshape = P.Reshape()
self.div = P.RealDiv()
self.size = P.Size()
self.mutinomial = P.Multinomial(seed=self.seed)
self.cast = P.Cast()
self.expandim = P.ExpandDims()
self.gather = P.GatherNd()
self.concat = P.Concat(-1)
self.transpose = P.Transpose()
if probs is not None:
self._probs = cast_to_tensor(probs, mstype.float32)
input_sum = self.reduce_sum(self._probs, -1)
self._probs = self.div(self._probs, input_sum)
self._logits = probs_to_logits(self._probs)
self._param = self._probs
else:
self._logits = cast_to_tensor(logits, mstype.float32)
input_sum = self.reduce_sum(self.exp(self._logits), -1)
self._logits = self._logits - self.log(input_sum)
self._probs = logits_to_probs(self._logits)
self._param = self._logits
self._num_events = self.shape(self._param)[-1]
self._param2d = self.reshape(self._param, (-1, self._num_events))
self._batch_shape = self.shape(self._param)[:-1]
self._batch_shape_n = (1, ) * len(self._batch_shape)
def test_mode_init(self, config):
self.test_batch_size = config.test_batch_size
self.split = P.Split(axis=0, output_num=self.test_batch_size)
self.split_shape = P.Split(axis=0, output_num=4)
self.split_scores = P.Split(axis=1, output_num=self.num_classes)
self.split_cls = P.Split(axis=0, output_num=self.num_classes - 1)
self.tile = P.Tile()
self.gather = P.GatherNd()
self.rpn_max_num = config.rpn_max_num
self.zeros_for_nms = Tensor(
np.zeros((self.rpn_max_num, 3)).astype(self.dtype))
self.ones_mask = np.ones((self.rpn_max_num, 1)).astype(np.bool)
self.zeros_mask = np.zeros((self.rpn_max_num, 1)).astype(np.bool)
self.bbox_mask = Tensor(
np.concatenate((self.ones_mask, self.zeros_mask, self.ones_mask,
self.zeros_mask),
axis=1))
self.nms_pad_mask = Tensor(
np.concatenate((self.ones_mask, self.ones_mask, self.ones_mask,
self.ones_mask, self.zeros_mask),
axis=1))
self.test_score_thresh = Tensor(
np.ones((self.rpn_max_num, 1)).astype(self.dtype) *
config.test_score_thr)
self.test_score_zeros = Tensor(
np.ones((self.rpn_max_num, 1)).astype(self.dtype) * 0)
self.test_box_zeros = Tensor(
np.ones((self.rpn_max_num, 4)).astype(self.dtype) * -1)
self.test_iou_thr = Tensor(
np.ones((self.rpn_max_num, 1)).astype(self.dtype) *
config.test_iou_thr)
self.test_max_per_img = config.test_max_per_img
self.nms_test = P.NMSWithMask(config.test_iou_thr)
self.softmax = P.Softmax(axis=1)
self.logicand = P.LogicalAnd()
self.oneslike = P.OnesLike()
self.test_topk = P.TopK(sorted=True)
self.test_num_proposal = self.test_batch_size * self.rpn_max_num
def get_pixel_value(self, img, x, y):
"""
Utility function to get pixel value for coordinate
vectors x and y from a 4D tensor image.
Input
-----
- img: tensor of shape (B, H, W, C)
- x: flattened tensor of shape (B*H*W,)
- y: flattened tensor of shape (B*H*W,)
Returns
-------
- output: tensor of shape (B, H, W, C)
"""
shape = P.Shape()
img_shape = shape(x)
batch_size = img_shape[0]
height = img_shape[1]
width = img_shape[2]
img[:, 0, :, :] = self.zero
img[:, height - 1, :, :] = self.zero
img[:, :, 0, :] = self.zero
img[:, :, width - 1, :] = self.zero
tile = P.Tile()
batch_idx = P.Slice()(self.batch_idx, (0, 0, 0), (batch_size, 1, 1))
b = tile(batch_idx, (1, height, width))
expand_dims = P.ExpandDims()
b = expand_dims(b, 3)
x = expand_dims(x, 3)
y = expand_dims(y, 3)
concat = P.Concat(3)
indices = concat((b, y, x))
cast = P.Cast()
indices = cast(indices, mindspore.int32)
gather_nd = P.GatherNd()
return cast(gather_nd(img, indices), mindspore.float32)
def __init__(self,
probs=None,
logits=None,
seed=0,
dtype=mstype.int32,
name="Categorical"):
param = dict(locals())
super(Categorical, self).__init__(seed, dtype, name, param)
if (probs is None) == (logits is None):
raise ValueError(
"Either 'prob' or 'logits' must be specified, but not both.")
self.reduce_sum = P.ReduceSum(keep_dims=True)
self.log = P.Log()
self.exp = P.Exp()
self.shape = P.Shape()
self.reshape = P.Reshape()
self.div = P.RealDiv()
self.size = P.Size()
self.mutinomial = P.Multinomial(seed=seed)
self.cast = P.Cast()
self.expandim = P.ExpandDims()
self.gather = P.GatherNd()
self.concat = P.Concat(-1)
if probs is not None:
self._probs = cast_to_tensor(probs, mstype.float32)
input_sum = self.reduce_sum(self._probs, -1)
self._probs = self.div(self._probs, input_sum)
self._logits = probs_to_logits(self._probs)
self._param = self._probs
else:
self._logits = cast_to_tensor(logits, mstype.float32)
input_sum = self.reduce_sum(self.exp(self._logits), -1)
self._logits = self._logits - self.log(input_sum)
self._probs = logits_to_probs(self._logits)
self._param = self._logits
self._num_events = self.shape(self._param)[-1]
self._param2d = self.reshape(self._param, (-1, self._num_events))
self._batch_shape = self.shape(self._param2d)[0]
'desc_const': [0.0],
'desc_inputs': [],
'desc_bprop': [1],
'skip': ['backward']}),
('BoundingBoxEncode', {
'block': P.BoundingBoxEncode(means=(0.0, 0.0, 0.0, 0.0), stds=(1.0, 1.0, 1.0, 1.0)),
'desc_inputs': [[256, 4], [256, 4]],
'desc_bprop': [[256, 4]],
'skip': ['backward']}),
('BoundingBoxDecode', {
'block': P.BoundingBoxDecode(means=(0.0, 0.0, 0.0, 0.0), stds=(1.0, 1.0, 1.0, 1.0), max_shape=(768, 1280)),
'desc_inputs': [[256, 4], [256, 4]],
'desc_bprop': [[256, 4]],
'skip': ['backward']}),
('GatherNd', {
'block': P.GatherNd(),
'desc_inputs': (Tensor(np.ones((1, 3, 6, 6), np.float32)),
Tensor(np.ones((2, 4), np.int32))),
'desc_bprop': [[2]]}),
('ScatterNdUpdate', {
'block': P.ScatterNdUpdate(),
'desc_inputs': (Tensor(np.ones((2, 3), np.float32)),
Tensor(np.ones((2, 2), np.int32)),
Tensor(np.ones((2,), np.float32))),
'desc_bprop': [[2, 3]]}),
('ScatterNd', {
'block': P.ScatterNd(),
'desc_const': [(3, 3)],
'desc_inputs': (Tensor(np.ones((2, 2), np.int32)),
Tensor(np.ones((2,), np.int32))),
'desc_bprop': [[3, 3]]}),
def __init__(self,
batch_size,
seq_length,
vocab_size,
decoder,
beam_width=4,
decoder_layers_nums=4,
length_penalty_weight=0.6,
cov_penalty_factor=0.1,
hidden_size=1024,
max_decode_length=64,
sos_id=2,
eos_id=3,
compute_type=mstype.float32):
super(BeamSearchDecoder, self).__init__()
self.encoder_length = seq_length
self.hidden_size = hidden_size
self.batch_size = batch_size
self.vocab_size = vocab_size
self.beam_width = beam_width
self.decoder_layers_nums = decoder_layers_nums
self.length_penalty_weight = length_penalty_weight
self.cov_penalty_factor = cov_penalty_factor
self.max_decode_length = max_decode_length
self.decoder = decoder
self.add = P.TensorAdd()
self.expand = P.ExpandDims()
self.reshape = P.Reshape()
self.shape_flat = (-1,)
self.shape = P.Shape()
self.zero_tensor = Tensor(np.zeros([batch_size, beam_width]), mstype.float32)
self.ninf_tensor = Tensor(np.full([batch_size, beam_width], -INF), mstype.float32)
self.select = P.Select()
self.flat_shape = (batch_size, beam_width * vocab_size)
self.topk = P.TopK(sorted=True)
self.floor_div = P.FloorDiv()
self.vocab_size_tensor = Tensor(self.vocab_size, mstype.int32)
self.real_div = P.RealDiv()
self.mod = Mod()
self.equal = P.Equal()
self.eos_ids = Tensor(np.full([batch_size, beam_width], eos_id), mstype.int32)
beam_ids = np.tile(np.arange(beam_width).reshape((1, beam_width)), [batch_size, 1])
self.beam_ids = Tensor(beam_ids, mstype.int32)
batch_ids = np.arange(batch_size * beam_width).reshape((batch_size, beam_width)) // beam_width
self.batch_ids = Tensor(batch_ids, mstype.int32)
self.concat = P.Concat(axis=-1)
self.gather_nd = P.GatherNd()
self.start = Tensor(0, dtype=mstype.int32)
self.start_ids = Tensor(np.full([batch_size * beam_width, 1], sos_id), mstype.int32)
self.init_seq = Tensor(np.full([batch_size, beam_width, self.max_decode_length], sos_id), mstype.int32)
init_scores = np.tile(np.array([[0.] + [-INF] * (beam_width - 1)]), [batch_size, 1])
self.init_scores = Tensor(init_scores, mstype.float32)
self.init_finished = Tensor(np.zeros([batch_size, beam_width], dtype=np.bool))
self.init_length = Tensor(np.zeros([batch_size, beam_width], dtype=np.int32))
self.length_penalty = LengthPenalty(weight=length_penalty_weight)
self.one = Tensor(1, mstype.int32)
self.prob_concat = P.Concat(axis=1)
self.cast = P.Cast()
self.decoder_hidden_state = Tensor(np.zeros([self.decoder_layers_nums, 2,
self.batch_size * self.beam_width,
hidden_size]), mstype.float32)
self.zeros_scores = Tensor(np.zeros([batch_size, beam_width], dtype=np.float))
self.active_index = Tensor(np.ones([batch_size, beam_width], dtype=np.int32))
self.init_zeros = Tensor(np.zeros([batch_size, beam_width], dtype=np.int32))
self.init_ones = Tensor(np.ones([batch_size, beam_width], dtype=np.float32))
self.accu_attn_scores = Tensor(np.zeros([batch_size, beam_width, self.encoder_length], dtype=np.float32))
self.zeros = Tensor([0], mstype.int32)
self.eos_tensor = Tensor(np.full([batch_size, beam_width, beam_width], eos_id), mstype.int32)
self.ones_3d = Tensor(np.full([batch_size, beam_width, self.encoder_length], 1), mstype.float32)
self.neg_inf_3d = Tensor(np.full([batch_size, beam_width, self.encoder_length], -INF), mstype.float32)
self.zeros_3d = Tensor(np.full([batch_size, beam_width, self.encoder_length], 0), mstype.float32)
self.zeros_2d = Tensor(np.full([batch_size * beam_width, self.encoder_length], 0), mstype.int32)
self.argmin = P.ArgMinWithValue(axis=1)
self.reducesum = P.ReduceSum()
self.div = P.Div()
self.shape_op = P.Shape()
self.mul = P.Mul()
self.log = P.Log()
self.less = P.Less()
self.tile = P.Tile()
self.noteq = P.Neg()
self.zeroslike = P.ZerosLike()
self.greater_equal = P.GreaterEqual()
self.sub = P.Sub()
def __init__(self, config, batch_size, num_bboxes, add_gt_as_proposals):
super(BboxAssignSampleForRcnn, self).__init__()
cfg = config
self.dtype = np.float32
self.ms_type = mstype.float32
self.batch_size = batch_size
self.neg_iou_thr = cfg.neg_iou_thr_stage2
self.pos_iou_thr = cfg.pos_iou_thr_stage2
self.min_pos_iou = cfg.min_pos_iou_stage2
self.num_gts = cfg.num_gts
self.num_bboxes = num_bboxes
self.num_expected_pos = cfg.num_expected_pos_stage2
self.num_expected_neg = cfg.num_expected_neg_stage2
self.num_expected_total = cfg.num_expected_total_stage2
self.add_gt_as_proposals = add_gt_as_proposals
self.label_inds = Tensor(
np.arange(1, self.num_gts + 1).astype(np.int32))
self.add_gt_as_proposals_valid = Tensor(
np.full(self.num_gts, self.add_gt_as_proposals, dtype=np.int32))
self.concat = P.Concat(axis=0)
self.max_gt = P.ArgMaxWithValue(axis=0)
self.max_anchor = P.ArgMaxWithValue(axis=1)
self.sum_inds = P.ReduceSum()
self.iou = P.IOU()
self.greaterequal = P.GreaterEqual()
self.greater = P.Greater()
self.select = P.Select()
self.gatherND = P.GatherNd()
self.squeeze = P.Squeeze()
self.cast = P.Cast()
self.logicaland = P.LogicalAnd()
self.less = P.Less()
self.random_choice_with_mask_pos = P.RandomChoiceWithMask(
self.num_expected_pos)
self.random_choice_with_mask_neg = P.RandomChoiceWithMask(
self.num_expected_neg)
self.reshape = P.Reshape()
self.equal = P.Equal()
self.bounding_box_encode = P.BoundingBoxEncode(means=(0.0, 0.0, 0.0,
0.0),
stds=(0.1, 0.1, 0.2,
0.2))
self.concat_axis1 = P.Concat(axis=1)
self.logicalnot = P.LogicalNot()
self.tile = P.Tile()
# Check
self.check_gt_one = Tensor(
np.full((self.num_gts, 4), -1, dtype=self.dtype))
self.check_anchor_two = Tensor(
np.full((self.num_bboxes, 4), -2, dtype=self.dtype))
# Init tensor
self.assigned_gt_inds = Tensor(np.full(num_bboxes, -1, dtype=np.int32))
self.assigned_gt_zeros = Tensor(
np.array(np.zeros(num_bboxes), dtype=np.int32))
self.assigned_gt_ones = Tensor(
np.array(np.ones(num_bboxes), dtype=np.int32))
self.assigned_gt_ignores = Tensor(
np.full(num_bboxes, -1, dtype=np.int32))
self.assigned_pos_ones = Tensor(
np.array(np.ones(self.num_expected_pos), dtype=np.int32))
self.gt_ignores = Tensor(np.full(self.num_gts, -1, dtype=np.int32))
self.range_pos_size = Tensor(
np.arange(self.num_expected_pos).astype(self.dtype))
self.check_neg_mask = Tensor(
np.array(np.ones(self.num_expected_neg - self.num_expected_pos),
dtype=np.bool))
self.bboxs_neg_mask = Tensor(
np.zeros((self.num_expected_neg, 4), dtype=self.dtype))
self.labels_neg_mask = Tensor(
np.array(np.zeros(self.num_expected_neg), dtype=np.uint8))
self.reshape_shape_pos = (self.num_expected_pos, 1)
self.reshape_shape_neg = (self.num_expected_neg, 1)
self.scalar_zero = Tensor(0.0, dtype=self.ms_type)
self.scalar_neg_iou_thr = Tensor(self.neg_iou_thr, dtype=self.ms_type)
self.scalar_pos_iou_thr = Tensor(self.pos_iou_thr, dtype=self.ms_type)
self.scalar_min_pos_iou = Tensor(self.min_pos_iou, dtype=self.ms_type)
def __init__(self):
super(GatherNdNet, self).__init__()
self.gathernd = P.GatherNd()
def __init__(self, config):
super(Deeptext_VGG16, self).__init__()
self.train_batch_size = config.batch_size
self.num_classes = config.num_classes
self.anchor_scales = config.anchor_scales
self.anchor_ratios = config.anchor_ratios
self.anchor_strides = config.anchor_strides
self.target_means = tuple(config.rcnn_target_means)
self.target_stds = tuple(config.rcnn_target_stds)
# Anchor generator
anchor_base_sizes = None
self.anchor_base_sizes = list(
self.anchor_strides
) if anchor_base_sizes is None else anchor_base_sizes
self.anchor_generators = []
for anchor_base in self.anchor_base_sizes:
self.anchor_generators.append(
AnchorGenerator(anchor_base, self.anchor_scales,
self.anchor_ratios))
self.num_anchors = len(self.anchor_ratios) * len(self.anchor_scales)
featmap_sizes = config.feature_shapes
assert len(featmap_sizes) == len(self.anchor_generators)
self.anchor_list = self.get_anchors(featmap_sizes)
# Rpn and rpn loss
self.gt_labels_stage1 = Tensor(
np.ones((self.train_batch_size, config.num_gts)).astype(np.uint8))
self.rpn_with_loss = RPN(config, self.train_batch_size,
config.rpn_in_channels,
config.rpn_feat_channels, config.num_anchors,
config.rpn_cls_out_channels)
# Proposal
self.proposal_generator = Proposal(config, self.train_batch_size,
config.activate_num_classes,
config.use_sigmoid_cls)
self.proposal_generator.set_train_local(config, True)
self.proposal_generator_test = Proposal(config, config.test_batch_size,
config.activate_num_classes,
config.use_sigmoid_cls)
self.proposal_generator_test.set_train_local(config, False)
# Assign and sampler stage two
self.bbox_assigner_sampler_for_rcnn = BboxAssignSampleForRcnn(
config, self.train_batch_size, config.num_bboxes_stage2, True)
self.decode = P.BoundingBoxDecode(max_shape=(576, 960), means=self.target_means, \
stds=self.target_stds)
# Rcnn
self.rcnn = Rcnn(
config, config.rcnn_in_channels * config.roi_layer['out_size'] *
config.roi_layer['out_size'], self.train_batch_size,
self.num_classes)
# Op declare
self.squeeze = P.Squeeze()
self.cast = P.Cast()
self.concat = P.Concat(axis=0)
self.concat_1 = P.Concat(axis=1)
self.concat_2 = P.Concat(axis=2)
self.reshape = P.Reshape()
self.select = P.Select()
self.greater = P.Greater()
self.transpose = P.Transpose()
# Test mode
self.test_batch_size = config.test_batch_size
self.split = P.Split(axis=0, output_num=self.test_batch_size)
self.split_shape = P.Split(axis=0, output_num=4)
self.split_scores = P.Split(axis=1, output_num=self.num_classes)
self.split_cls = P.Split(axis=0, output_num=self.num_classes - 1)
self.tile = P.Tile()
self.gather = P.GatherNd()
self.rpn_max_num = config.rpn_max_num
self.zeros_for_nms = Tensor(
np.zeros((self.rpn_max_num, 3)).astype(np.float32))
self.ones_mask = np.ones((self.rpn_max_num, 1)).astype(np.bool)
self.zeros_mask = np.zeros((self.rpn_max_num, 1)).astype(np.bool)
self.bbox_mask = Tensor(
np.concatenate((self.ones_mask, self.zeros_mask, self.ones_mask,
self.zeros_mask),
axis=1))
self.nms_pad_mask = Tensor(
np.concatenate((self.ones_mask, self.ones_mask, self.ones_mask,
self.ones_mask, self.zeros_mask),
axis=1))
self.test_score_thresh = Tensor(
np.ones((self.rpn_max_num, 1)).astype(np.float32) *
config.test_score_thr)
self.test_score_zeros = Tensor(
np.ones((self.rpn_max_num, 1)).astype(np.float32) * 0)
self.test_box_zeros = Tensor(
np.ones((self.rpn_max_num, 4)).astype(np.float32) * -1)
self.test_iou_thr = Tensor(
np.ones((self.rpn_max_num, 1)).astype(np.float32) *
config.test_iou_thr)
self.test_max_per_img = config.test_max_per_img
self.nms_test = P.NMSWithMask(config.test_iou_thr)
self.softmax = P.Softmax(axis=1)
self.logicand = P.LogicalAnd()
self.oneslike = P.OnesLike()
self.test_topk = P.TopK(sorted=True)
self.test_num_proposal = self.test_batch_size * self.rpn_max_num
# Improve speed
self.concat_start = (self.num_classes - 2)
self.concat_end = (self.num_classes - 1)
# Init tensor
self.use_ambigous_sample = config.use_ambigous_sample
roi_align_index = [
np.array(np.ones((config.num_expected_pos_stage2 +
config.num_expected_neg_stage2, 1)) * i,
dtype=np.float32) for i in range(self.train_batch_size)
]
if self.use_ambigous_sample:
roi_align_index = [
np.array(np.ones((config.num_expected_pos_stage2 +
config.num_expected_amb_stage2 +
config.num_expected_neg_stage2, 1)) * i,
dtype=np.float32)
for i in range(self.train_batch_size)
]
roi_align_index_test = [np.array(np.ones((config.rpn_max_num, 1)) * i, dtype=np.float32) \
for i in range(self.test_batch_size)]
self.roi_align_index_tensor = Tensor(np.concatenate(roi_align_index))
self.roi_align_index_test_tensor = Tensor(
np.concatenate(roi_align_index_test))
self.roi_align4 = P.ROIAlign(pooled_width=7,
pooled_height=7,
spatial_scale=0.125)
self.roi_align5 = P.ROIAlign(pooled_width=7,
pooled_height=7,
spatial_scale=0.0625)
self.concat1 = P.Concat(axis=1)
self.roi_align_fuse = _conv(in_channels=1024,
out_channels=512,
kernel_size=1,
padding=0,
stride=1)
self.vgg16_feature_extractor = VGG16FeatureExtraction()
def __init__(self, config):
super(Mask_Rcnn_Resnet50, self).__init__()
self.train_batch_size = config.batch_size
self.num_classes = config.num_classes
self.anchor_scales = config.anchor_scales
self.anchor_ratios = config.anchor_ratios
self.anchor_strides = config.anchor_strides
self.target_means = tuple(config.rcnn_target_means)
self.target_stds = tuple(config.rcnn_target_stds)
# Anchor generator
anchor_base_sizes = None
self.anchor_base_sizes = list(
self.anchor_strides) if anchor_base_sizes is None else anchor_base_sizes
self.anchor_generators = []
for anchor_base in self.anchor_base_sizes:
self.anchor_generators.append(
AnchorGenerator(anchor_base, self.anchor_scales, self.anchor_ratios))
self.num_anchors = len(self.anchor_ratios) * len(self.anchor_scales)
featmap_sizes = config.feature_shapes
assert len(featmap_sizes) == len(self.anchor_generators)
self.anchor_list = self.get_anchors(featmap_sizes)
# Backbone resnet50
self.backbone = ResNetFea(ResidualBlockUsing,
config.resnet_block,
config.resnet_in_channels,
config.resnet_out_channels,
False)
# Fpn
self.fpn_ncek = FeatPyramidNeck(config.fpn_in_channels,
config.fpn_out_channels,
config.fpn_num_outs)
# Rpn and rpn loss
self.gt_labels_stage1 = Tensor(np.ones((self.train_batch_size, config.num_gts)).astype(np.uint8))
self.rpn_with_loss = RPN(config,
self.train_batch_size,
config.rpn_in_channels,
config.rpn_feat_channels,
config.num_anchors,
config.rpn_cls_out_channels)
# Proposal
self.proposal_generator = Proposal(config,
self.train_batch_size,
config.activate_num_classes,
config.use_sigmoid_cls)
self.proposal_generator.set_train_local(config, True)
self.proposal_generator_test = Proposal(config,
config.test_batch_size,
config.activate_num_classes,
config.use_sigmoid_cls)
self.proposal_generator_test.set_train_local(config, False)
# Assign and sampler stage two
self.bbox_assigner_sampler_for_rcnn = BboxAssignSampleForRcnn(config, self.train_batch_size,
config.num_bboxes_stage2, True)
self.decode = P.BoundingBoxDecode(max_shape=(768, 1280), means=self.target_means, \
stds=self.target_stds)
# Roi
self.roi_align = SingleRoIExtractor(config,
config.roi_layer,
config.roi_align_out_channels,
config.roi_align_featmap_strides,
self.train_batch_size,
config.roi_align_finest_scale,
mask=False)
self.roi_align.set_train_local(config, True)
self.roi_align_mask = SingleRoIExtractor(config,
config.roi_layer,
config.roi_align_out_channels,
config.roi_align_featmap_strides,
self.train_batch_size,
config.roi_align_finest_scale,
mask=True)
self.roi_align_mask.set_train_local(config, True)
self.roi_align_test = SingleRoIExtractor(config,
config.roi_layer,
config.roi_align_out_channels,
config.roi_align_featmap_strides,
1,
config.roi_align_finest_scale,
mask=False)
self.roi_align_test.set_train_local(config, False)
self.roi_align_mask_test = SingleRoIExtractor(config,
config.roi_layer,
config.roi_align_out_channels,
config.roi_align_featmap_strides,
1,
config.roi_align_finest_scale,
mask=True)
self.roi_align_mask_test.set_train_local(config, False)
# Rcnn
self.rcnn_cls = RcnnCls(config, self.train_batch_size, self.num_classes)
self.rcnn_mask = RcnnMask(config, self.train_batch_size, self.num_classes)
# Op declare
self.squeeze = P.Squeeze()
self.cast = P.Cast()
self.concat = P.Concat(axis=0)
self.concat_1 = P.Concat(axis=1)
self.concat_2 = P.Concat(axis=2)
self.reshape = P.Reshape()
self.select = P.Select()
self.greater = P.Greater()
self.transpose = P.Transpose()
# Test mode
self.test_batch_size = config.test_batch_size
self.split = P.Split(axis=0, output_num=self.test_batch_size)
self.split_shape = P.Split(axis=0, output_num=4)
self.split_scores = P.Split(axis=1, output_num=self.num_classes)
self.split_fb_mask = P.Split(axis=1, output_num=self.num_classes)
self.split_cls = P.Split(axis=0, output_num=self.num_classes-1)
self.tile = P.Tile()
self.gather = P.GatherNd()
self.rpn_max_num = config.rpn_max_num
self.zeros_for_nms = Tensor(np.zeros((self.rpn_max_num, 3)).astype(np.float16))
self.ones_mask = np.ones((self.rpn_max_num, 1)).astype(np.bool)
self.zeros_mask = np.zeros((self.rpn_max_num, 1)).astype(np.bool)
self.bbox_mask = Tensor(np.concatenate((self.ones_mask, self.zeros_mask,
self.ones_mask, self.zeros_mask), axis=1))
self.nms_pad_mask = Tensor(np.concatenate((self.ones_mask, self.ones_mask,
self.ones_mask, self.ones_mask, self.zeros_mask), axis=1))
self.test_score_thresh = Tensor(np.ones((self.rpn_max_num, 1)).astype(np.float16) * config.test_score_thr)
self.test_score_zeros = Tensor(np.ones((self.rpn_max_num, 1)).astype(np.float16) * 0)
self.test_box_zeros = Tensor(np.ones((self.rpn_max_num, 4)).astype(np.float16) * -1)
self.test_iou_thr = Tensor(np.ones((self.rpn_max_num, 1)).astype(np.float16) * config.test_iou_thr)
self.test_max_per_img = config.test_max_per_img
self.nms_test = P.NMSWithMask(config.test_iou_thr)
self.softmax = P.Softmax(axis=1)
self.logicand = P.LogicalAnd()
self.oneslike = P.OnesLike()
self.test_topk = P.TopK(sorted=True)
self.test_num_proposal = self.test_batch_size * self.rpn_max_num
# Improve speed
self.concat_start = min(self.num_classes - 2, 55)
self.concat_end = (self.num_classes - 1)
# Init tensor
roi_align_index = [np.array(np.ones((config.num_expected_pos_stage2 + config.num_expected_neg_stage2, 1)) * i,
dtype=np.float16) for i in range(self.train_batch_size)]
roi_align_index_test = [np.array(np.ones((config.rpn_max_num, 1)) * i, dtype=np.float16) \
for i in range(self.test_batch_size)]
self.roi_align_index_tensor = Tensor(np.concatenate(roi_align_index))
self.roi_align_index_test_tensor = Tensor(np.concatenate(roi_align_index_test))
roi_align_index_pos = [np.array(np.ones((config.num_expected_pos_stage2, 1)) * i,
dtype=np.float16) for i in range(self.train_batch_size)]
self.roi_align_index_tensor_pos = Tensor(np.concatenate(roi_align_index_pos))
self.rcnn_loss_cls_weight = Tensor(np.array(config.rcnn_loss_cls_weight).astype(np.float16))
self.rcnn_loss_reg_weight = Tensor(np.array(config.rcnn_loss_reg_weight).astype(np.float16))
self.rcnn_loss_mask_fb_weight = Tensor(np.array(config.rcnn_loss_mask_fb_weight).astype(np.float16))
self.argmax_with_value = P.ArgMaxWithValue(axis=1)
self.on_value = Tensor(1.0, mstype.float32)
self.off_value = Tensor(0.0, mstype.float32)
self.onehot = P.OneHot()
self.reducesum = P.ReduceSum()
self.sigmoid = P.Sigmoid()
self.expand_dims = P.ExpandDims()
self.test_mask_fb_zeros = Tensor(np.zeros((self.rpn_max_num, 28, 28)).astype(np.float16))
self.value = Tensor(1.0, mstype.float16)
def __init__(self, w1, strategy1=None, strategy2=None):
super().__init__()
self.mul = P.Mul().shard(strategy1)
self.w1 = Parameter(w1, "w1")
self.indices = Tensor(np.ones([16, 2]), dtype=ms.int32)
self.gathernd = P.GatherNd().shard(strategy2)
def __init__(self,
config,
representation_size,
batch_size,
num_classes,
target_means=(0., 0., 0., 0.),
target_stds=(0.1, 0.1, 0.2, 0.2)):
super(Rcnn, self).__init__()
cfg = config
self.rcnn_loss_cls_weight = Tensor(
np.array(cfg.rcnn_loss_cls_weight).astype(np.float16))
self.rcnn_loss_reg_weight = Tensor(
np.array(cfg.rcnn_loss_reg_weight).astype(np.float16))
self.rcnn_fc_out_channels = cfg.rcnn_fc_out_channels
self.target_means = target_means
self.target_stds = target_stds
self.num_classes = num_classes
self.in_channels = cfg.rcnn_in_channels
self.train_batch_size = batch_size
self.test_batch_size = cfg.test_batch_size
self.use_ambigous_sample = cfg.use_ambigous_sample
shape_0 = (self.rcnn_fc_out_channels, representation_size)
weights_0 = initializer("XavierUniform",
shape=shape_0[::-1],
dtype=mstype.float16).to_tensor()
shape_1 = (self.rcnn_fc_out_channels, self.rcnn_fc_out_channels)
weights_1 = initializer("XavierUniform",
shape=shape_1[::-1],
dtype=mstype.float16).to_tensor()
self.shared_fc_0 = DenseNoTranpose(representation_size,
self.rcnn_fc_out_channels,
weights_0)
self.shared_fc_1 = DenseNoTranpose(self.rcnn_fc_out_channels,
self.rcnn_fc_out_channels,
weights_1)
cls_weight = initializer(
'Normal',
shape=[num_classes, self.rcnn_fc_out_channels][::-1],
dtype=mstype.float16).to_tensor()
reg_weight = initializer(
'Normal',
shape=[num_classes * 4, self.rcnn_fc_out_channels][::-1],
dtype=mstype.float16).to_tensor()
self.cls_scores = DenseNoTranpose(self.rcnn_fc_out_channels,
num_classes, cls_weight)
self.reg_scores = DenseNoTranpose(self.rcnn_fc_out_channels,
num_classes * 4, reg_weight)
self.flatten = P.Flatten()
self.relu = P.ReLU()
self.logicaland = P.LogicalAnd()
self.loss_cls = P.SoftmaxCrossEntropyWithLogits()
self.loss_bbox = P.SmoothL1Loss(beta=1.0)
self.reshape = P.Reshape()
self.onehot = P.OneHot()
self.greater = P.Greater()
self.equal = P.Equal()
self.cast = P.Cast()
self.sum_loss = P.ReduceSum()
self.tile = P.Tile()
self.expandims = P.ExpandDims()
self.gather = P.GatherNd()
self.argmax = P.ArgMaxWithValue(axis=1)
self.on_value = Tensor(1.0, mstype.float32)
self.off_value = Tensor(0.0, mstype.float32)
self.value = Tensor(1.0, mstype.float16)
self.num_bboxes = (cfg.num_expected_pos_stage2 +
cfg.num_expected_neg_stage2) * batch_size
if self.use_ambigous_sample:
self.num_bboxes = (cfg.num_expected_pos_stage2 +
cfg.num_expected_amb_stage2 +
cfg.num_expected_neg_stage2) * batch_size
rmv_first = np.ones((self.num_bboxes, self.num_classes))
rmv_first[:, 0] = np.zeros((self.num_bboxes, ))
self.rmv_first_tensor = Tensor(rmv_first.astype(np.float16))
self.num_bboxes_test = cfg.rpn_max_num * cfg.test_batch_size
range_max = np.arange(self.num_bboxes_test).astype(np.int32)
self.range_max = Tensor(range_max)
def __init__(self, config, batch_size, num_bboxes, add_gt_as_proposals):
super(BboxAssignSampleForRcnn, self).__init__()
cfg = config
self.batch_size = batch_size
self.neg_iou_thr = cfg.neg_iou_thr_stage2
self.pos_iou_thr = cfg.pos_iou_thr_stage2
self.min_pos_iou = cfg.min_pos_iou_stage2
self.num_gts = cfg.num_gts
self.num_bboxes = num_bboxes
self.num_expected_pos = cfg.num_expected_pos_stage2
self.num_expected_neg = cfg.num_expected_neg_stage2
self.num_expected_total = cfg.num_expected_total_stage2
self.add_gt_as_proposals = add_gt_as_proposals
self.label_inds = Tensor(
np.arange(1, self.num_gts + 1).astype(np.int32))
self.add_gt_as_proposals_valid = Tensor(
np.array(self.add_gt_as_proposals * np.ones(self.num_gts),
dtype=np.int32))
self.concat = P.Concat(axis=0)
self.max_gt = P.ArgMaxWithValue(axis=0)
self.max_anchor = P.ArgMaxWithValue(axis=1)
self.sum_inds = P.ReduceSum()
self.iou = P.IOU()
self.greaterequal = P.GreaterEqual()
self.greater = P.Greater()
self.select = P.Select()
self.gatherND = P.GatherNd()
self.squeeze = P.Squeeze()
self.cast = P.Cast()
self.logicaland = P.LogicalAnd()
self.less = P.Less()
self.random_choice_with_mask_pos = P.RandomChoiceWithMask(
self.num_expected_pos)
self.random_choice_with_mask_neg = P.RandomChoiceWithMask(
self.num_expected_neg)
self.reshape = P.Reshape()
self.equal = P.Equal()
self.bounding_box_encode = P.BoundingBoxEncode(means=(0.0, 0.0, 0.0,
0.0),
stds=(0.1, 0.1, 0.2,
0.2))
self.concat_axis1 = P.Concat(axis=1)
self.logicalnot = P.LogicalNot()
self.tile = P.Tile()
# Check
self.check_gt_one = Tensor(
np.array(-1 * np.ones((self.num_gts, 4)), dtype=np.float16))
self.check_anchor_two = Tensor(
np.array(-2 * np.ones((self.num_bboxes, 4)), dtype=np.float16))
# Init tensor
self.assigned_gt_inds = Tensor(
np.array(-1 * np.ones(num_bboxes), dtype=np.int32))
self.assigned_gt_zeros = Tensor(
np.array(np.zeros(num_bboxes), dtype=np.int32))
self.assigned_gt_ones = Tensor(
np.array(np.ones(num_bboxes), dtype=np.int32))
self.assigned_gt_ignores = Tensor(
np.array(-1 * np.ones(num_bboxes), dtype=np.int32))
self.assigned_pos_ones = Tensor(
np.array(np.ones(self.num_expected_pos), dtype=np.int32))
self.gt_ignores = Tensor(
np.array(-1 * np.ones(self.num_gts), dtype=np.int32))
self.range_pos_size = Tensor(
np.arange(self.num_expected_pos).astype(np.float16))
self.check_neg_mask = Tensor(
np.array(np.ones(self.num_expected_neg - self.num_expected_pos),
dtype=np.bool))
self.bboxs_neg_mask = Tensor(
np.zeros((self.num_expected_neg, 4), dtype=np.float16))
self.labels_neg_mask = Tensor(
np.array(np.zeros(self.num_expected_neg), dtype=np.uint8))
self.reshape_shape_pos = (self.num_expected_pos, 1)
self.reshape_shape_neg = (self.num_expected_neg, 1)
self.scalar_zero = Tensor(0.0, dtype=mstype.float16)
self.scalar_neg_iou_thr = Tensor(self.neg_iou_thr,
dtype=mstype.float16)
self.scalar_pos_iou_thr = Tensor(self.pos_iou_thr,
dtype=mstype.float16)
self.scalar_min_pos_iou = Tensor(self.min_pos_iou,
dtype=mstype.float16)
self.expand_dims = P.ExpandDims()
self.split = P.Split(axis=1, output_num=4)
self.concat_last_axis = P.Concat(axis=-1)
self.round = P.Round()
self.image_h_w = Tensor(
[cfg.img_height, cfg.img_width, cfg.img_height, cfg.img_width],
dtype=mstype.float16)
self.range = nn.Range(start=0, limit=cfg.num_expected_pos_stage2)
self.crop_and_resize = P.CropAndResize(method="bilinear_v2")
self.mask_shape = (cfg.mask_shape[0], cfg.mask_shape[1])
self.squeeze_mask_last = P.Squeeze(axis=-1)
def __init__(self,
batch_size,
seq_length,
vocab_size,
decoder,
beam_width=4,
length_penalty_weight=1.0,
max_decode_length=128,
sos_id=1,
eos_id=2,
compute_type=mstype.float32):
super(BeamSearchDecoder, self).__init__(auto_prefix=False)
self.seq_length = seq_length
self.batch_size = batch_size
self.vocab_size = vocab_size
self.beam_width = beam_width
self.length_penalty_weight = length_penalty_weight
self.max_decode_length = max_decode_length
self.decoder = decoder
self.add = P.TensorAdd()
self.expand = P.ExpandDims()
self.reshape = P.Reshape()
self.shape_flat = (-1, )
self.shape = P.Shape()
self.zero_tensor = Tensor(np.zeros([batch_size, beam_width]),
mstype.float32)
self.ninf_tensor = Tensor(np.full([batch_size, beam_width], -INF),
mstype.float32)
self.select = P.Select()
self.flat_shape = (batch_size, beam_width * vocab_size)
self.topk = P.TopK(sorted=True)
self.floor_div = P.FloorDiv()
self.vocab_size_tensor = Tensor(self.vocab_size, mstype.int32)
self.real_div = P.RealDiv()
self.mod = Mod()
self.equal = P.Equal()
self.eos_ids = Tensor(np.full([batch_size, beam_width], eos_id),
mstype.int32)
beam_ids = np.tile(
np.arange(beam_width).reshape((1, beam_width)), [batch_size, 1])
self.beam_ids = Tensor(beam_ids, mstype.int32)
batch_ids = np.arange(batch_size * beam_width).reshape(
(batch_size, beam_width)) // beam_width
self.batch_ids = Tensor(batch_ids, mstype.int32)
self.concat = P.Concat(axis=-1)
self.gather_nd = P.GatherNd()
self.greater_equal = P.GreaterEqual()
self.sub = P.Sub()
self.cast = P.Cast()
self.zeroslike = P.ZerosLike()
# init inputs and states
self.start_ids = Tensor(np.full([batch_size * beam_width, 1], sos_id),
mstype.int32)
self.init_seq = Tensor(np.full([batch_size, beam_width, 1], sos_id),
mstype.int32)
init_scores = np.tile(np.array([[0.] + [-INF] * (beam_width - 1)]),
[batch_size, 1])
self.init_scores = Tensor(init_scores, mstype.float32)
self.init_finished = Tensor(
np.zeros([batch_size, beam_width], dtype=np.bool))
self.init_length = Tensor(
np.zeros([batch_size, beam_width], dtype=np.int32))
self.length_penalty = LengthPenalty(weight=length_penalty_weight)
self.one = Tensor(1, mstype.int32)
'desc_inputs': [[256, 4], [256, 4]],
'desc_bprop': [[256, 4]],
'skip': ['backward']
}),
('BoundingBoxDecode', {
'block':
P.BoundingBoxDecode(means=(0.0, 0.0, 0.0, 0.0),
stds=(1.0, 1.0, 1.0, 1.0),
max_shape=(768, 1280)),
'desc_inputs': [[256, 4], [256, 4]],
'desc_bprop': [[256, 4]],
'skip': ['backward']
}),
('GatherNd', {
'block':
P.GatherNd(),
'desc_inputs': (Tensor(np.ones(
(1, 3, 6, 6), np.float32)), Tensor(np.ones((2, 4), np.int32))),
'desc_bprop': [[2]]
}),
('ScatterNdUpdate', {
'block':
P.ScatterNdUpdate(),
'desc_inputs': (Tensor(np.ones(
(2, 3), np.float32)), Tensor(np.ones(
(2, 2), np.int32)), Tensor(np.ones((2, ), np.float32))),
'desc_bprop': [[2, 3]]
}),
('ScatterNd', {
'block':
P.ScatterNd(),