def __init__(self, in_channels=None, return_2d=False):
super(LayerNorm, self).__init__()
self.return_2d = return_2d
self.layer_norm = nn.LayerNorm((in_channels, ))
self.cast = P.Cast()
self.get_dtype = P.DType()
self.reshape = P.Reshape()
self.get_shape = P.Shape()
def __init__(self, config: Callable[..., None]) -> None:
super().__init__()
self.dense = nn.Dense(
config.hidden_size,
config.hidden_size,
weight_init=TruncatedNormal(config.initializer_range),
)
self.transform_act_fn = get_activation(config.hidden_act)
self.layer_norm = nn.LayerNorm((config.hidden_size, ), epsilon=1e-12)
def __init__(self,
in_channels,
out_channels,
initializer_range=0.02,
dropout_prob=0.1,
compute_type=mstype.float32):
super(BertOutput, self).__init__()
self.dense = nn.Dense(in_channels, out_channels,
weight_init=TruncatedNormal(initializer_range)).to_float(compute_type)
self.dropout = nn.Dropout(1 - dropout_prob)
self.add = P.TensorAdd()
self.is_gpu = context.get_context('device_target') == "GPU"
if self.is_gpu:
self.layernorm = nn.LayerNorm((out_channels,)).to_float(mstype.float32)
self.compute_type = compute_type
else:
self.layernorm = nn.LayerNorm((out_channels,)).to_float(compute_type)
self.cast = P.Cast()
def __init__(self,
embedding_size,
embedding_shape,
use_relative_positions=False,
use_token_type=False,
token_type_vocab_size=16,
use_one_hot_embeddings=False,
initializer_range=0.02,
max_position_embeddings=512,
dropout_prob=0.1):
super(EmbeddingPostprocessor, self).__init__()
self.use_token_type = use_token_type
self.token_type_vocab_size = token_type_vocab_size
self.use_one_hot_embeddings = use_one_hot_embeddings
self.max_position_embeddings = max_position_embeddings
self.token_type_embedding = Embedding_Thor(
vocab_size=token_type_vocab_size,
embedding_size=embedding_size,
embedding_shape=embedding_shape,
use_one_hot_embeddings=use_one_hot_embeddings,
initializer_range=initializer_range,
name='embedding_table',
is_expand=False,
batch_size=batch_size,
damping=damping,
loss_scale=loss_scale,
frequency=1)
self.shape_flat = (-1, )
self.one_hot = P.OneHot()
self.on_value = Tensor(1.0, mstype.float32)
self.off_value = Tensor(0.1, mstype.float32)
self.array_mul = P.MatMul()
self.reshape = P.Reshape()
self.shape = tuple(embedding_shape)
self.dropout = nn.Dropout(1 - dropout_prob)
self.gather = P.GatherV2()
self.use_relative_positions = use_relative_positions
self.slice = P.StridedSlice()
_, seq, width = self.shape
position_embedding_shape = [1, seq, width]
self.full_position_embedding = Embedding_Thor(
vocab_size=max_position_embeddings,
embedding_size=embedding_size,
embedding_shape=position_embedding_shape,
use_one_hot_embeddings=use_one_hot_embeddings,
initializer_range=initializer_range,
name='full_position_embeddings',
is_expand=False,
batch_size=batch_size,
damping=damping,
loss_scale=loss_scale,
frequency=1)
self.position_ids = Tensor(
np.arange(seq).reshape(-1, seq).astype(np.int32))
self.layernorm = nn.LayerNorm((embedding_size, ))
def __init__(self, dim):
super().__init__()
self.layer_norm = nn.LayerNorm((dim, ), -1, -1)
self.xg2q = Dense(dim, dim, has_bias=False)
self.xg2k = Dense(dim, dim, has_bias=False)
self.xg2v = Dense(dim, dim, has_bias=False)
self.mul = P.Mul()
self.concat = P.Concat(-2)
def __init__(self, 词库总数, 向量维度, 层数, 头数, 丢弃率,辞数,最大长度=1024):
super(多层解码, self).__init__()
self.N = 层数
self.embed = 词向量印刻(词库总数, 向量维度)
self.embedP = 词向量印刻(最大长度, 向量维度)
self.decoders = nn.CellList([解码层(向量维度, 头数, 丢弃率) for i in range(层数)])
self.norm = nn.LayerNorm((向量维度,), epsilon=1e-6)
a = [i for i in range(辞数)]
b = np.array(a).reshape(1, 辞数)
self.po = Tensor(b, mindspore.int32)
self.dropout = nn.Dropout(1 - 丢弃率)
def __init__(self, config, layer_idx):
super(Block, self).__init__()
scale = 1 / math.sqrt(2.0 * config.num_layers)
if config.self_layernorm:
self.layernorm1 = LayerNorm((config.embedding_size,), config.dp).to_float(mstype.float32)
self.layernorm2 = LayerNorm((config.embedding_size,), config.dp).to_float(mstype.float32)
else:
self.layernorm1 = nn.LayerNorm(
(config.embedding_size,)).to_float(mstype.float32)
self.layernorm1.layer_norm.shard(((config.dp, 1, 1), (1,), (1,)))
self.layernorm2 = nn.LayerNorm(
(config.embedding_size,)).to_float(mstype.float32)
self.layernorm2.layer_norm.shard(((config.dp, 1, 1), (1,), (1,)))
self.attention = Attention(config, scale, layer_idx)
self.output = Output(config, scale)
self.post_layernorm_residual = config.post_layernorm_residual
self.add = P.TensorAdd().shard(((config.dp, 1, 1), (config.dp, 1, 1)))
self.add_last = P.TensorAdd().shard(((config.dp, 1, 1), (config.dp, 1, 1)))
self.add_last.recompute(False)
self.dtype = config.compute_dtype
def __init__(self,
in_channels,
out_channels,
initializer_range=0.02,
dropout_prob=0.1,
compute_type=mstype.float32):
super(BertOutput, self).__init__()
self.dense = nn.Dense(in_channels, out_channels,
weight_init=TruncatedNormal(initializer_range)).to_float(compute_type)
self.dropout = nn.Dropout(1 - dropout_prob)
self.add = P.TensorAdd()
self.layernorm = nn.LayerNorm(out_channels).to_float(compute_type)
self.cast = P.Cast()
def __init__(self, config):
super(GPT_Model, self).__init__()
self.get_attention_mask = AttentionMask(config)
self.word_embedding = EmbeddingLookup(config)
self.position_embedding = nn.Embedding(
config.seq_length,
config.embedding_size,
embedding_table=TruncatedNormal(0.02))
self.blocks = nn.CellList()
for i in range(config.num_layers):
self.blocks.append(Block(config, i + 1))
self.layernorm = nn.LayerNorm(
(config.embedding_size, )).to_float(config.compute_dtype)
self.use_past = config.use_past
self.past = tuple([None] * config.num_layers)
self.num_layers = config.num_layers
def __init__(self,
in_channels,
out_channels,
initializer_range=0.02,
dropout_prob=0.1,
compute_type=mstype.float32,
enable_fused_layernorm=False):
super(BertOutput, self).__init__()
self.dense = nn.Dense(in_channels, out_channels,
weight_init=TruncatedNormal(initializer_range)).to_float(compute_type)
self.dropout = nn.Dropout(1 - dropout_prob)
self.dropout_prob = dropout_prob
self.add = ops.TensorAdd()
if compute_type == mstype.float16:
self.layernorm = FusedLayerNorm((out_channels,),
use_batch_norm=enable_fused_layernorm).to_float(compute_type)
else:
self.layernorm = nn.LayerNorm((out_channels,)).to_float(compute_type)
self.cast = ops.Cast()
def __init__(self, config: Callable[..., None]) -> None:
super().__init__()
self.token_embeddings = nn.Embedding(
config.vocab_size,
config.hidden_size,
embedding_table=TruncatedNormal(config.initializer_range),
padding_idx=0,
)
self.position_embeddings = nn.Embedding(
config.max_position_embeddings,
config.hidden_size,
embedding_table=TruncatedNormal(config.initializer_range),
)
self.token_type_embeddings = nn.Embedding(
config.type_vocab_size,
config.hidden_size,
embedding_table=TruncatedNormal(config.initializer_range),
)
self.layer_norm = nn.LayerNorm((config.hidden_size, ), epsilon=1e-12)
self.dropout = nn.Dropout(1.0 - config.hidden_dropout_prob)
def __init__(self, config):
super(PANGUALPHA_ModelPipeline, self).__init__()
self.pangu_alpha_embedding = PANGUALPHA_EmbeddingPipeLine(config).set_comm_fusion(1)
self.pangu_alpha_embedding.stage = 0
self.pangu_alpha_mask = PANGUALPHA_Mask(config)
self.blocks = nn.CellList()
dropout_recompute = False
self.top_query_embedding = nn.Embedding(config.seq_length, config.embedding_size,
embedding_table=TruncatedNormal(0.02))
self.top_query_embedding.gather.shard(((1, 1), (config.dp,)))
self.top_query_embedding.expand.shard(((config.dp, 1),))
for i in range(config.num_layers):
if i == config.num_layers - 1:
self.top_query_embedding.set_comm_fusion(2)
self.top_query_embedding.stage = i * config.stage_num // config.num_layers
per_block = QueryLayer(config).set_comm_fusion(2)
else:
per_block = Block(config, i + 1).set_comm_fusion(2)
per_block.stage = i * config.stage_num // config.num_layers
per_block.recompute()
self.blocks.append(per_block)
if not dropout_recompute:
per_block.attention.dropout.dropout_gen_mask.recompute(False).add_prim_attr("_side_effect", True)
per_block.attention.prob_dropout.dropout_gen_mask.recompute(False).add_prim_attr("_side_effect", True)
per_block.output.dropout.dropout_gen_mask.recompute(False).add_prim_attr("_side_effect", True)
if config.self_layernorm:
self.layernorm = LayerNorm((config.embedding_size,), config.dp).to_float(mstype.float32)
else:
self.layernorm = nn.LayerNorm(
(config.embedding_size,)).to_float(mstype.float32)
self.layernorm.layer_norm.shard(((config.dp, 1, 1), (1,), (1,)))
self.layernorm.set_comm_fusion(2)
#self.layernorm.set_comm_fusion(3)
self.layernorm.stage = config.stage_num - 1
self.use_past = config.use_past
self.past = tuple([None] * config.num_layers)
self.dtype = config.compute_dtype
self.num_layers = config.num_layers
def __init__(self, config, is_training, num_tokens, dropout_prob=0.0, use_one_hot_embeddings=False):
super(BertPoetryModel, self).__init__()
self.bert = BertModel(config, is_training, use_one_hot_embeddings)
self.num_tokens = num_tokens
idx = np.arange(config.seq_length)
mask = idx[None, :] <= idx[:, None]
self.mask = Tensor([mask], mstype.float32)
self.MLM_Dense = nn.Dense(config.hidden_size, config.hidden_size,\
has_bias=True, weight_init=TruncatedNormal(0.02),\
activation='gelu').to_float(mstype.float16)
self.layer_norm = nn.LayerNorm((config.hidden_size,))
self.matmul = ops.MatMul(transpose_b=True)
self.biasadd = Parameter(initializer('zero', self.num_tokens), name='MLM_output_biasadd')
self.softmax = ops.Softmax(axis=-1)
self.seq_length = config.seq_length
self.hidden_size = config.hidden_size
self.cast = ops.Cast()
self.reshape = ops.Reshape()
self.batch_matmul = ops.BatchMatMul()
ones = np.ones(shape=(config.batch_size, config.seq_length, config.seq_length))
self.lower_triangle_mask = Tensor(np.tril(ones), dtype=mstype.float32)
self.multiply = ops.Mul()
def __init__(self,
use_relative_positions,
embedding_size,
embedding_shape,
use_token_type=False,
token_type_vocab_size=16,
use_one_hot_embeddings=False,
initializer_range=0.02,
max_position_embeddings=512,
dropout_prob=0.1):
super(EmbeddingPostprocessor, self).__init__()
self.use_token_type = use_token_type
self.token_type_vocab_size = token_type_vocab_size
self.use_one_hot_embeddings = use_one_hot_embeddings
self.max_position_embeddings = max_position_embeddings
self.token_type_embedding = nn.Embedding(
vocab_size=token_type_vocab_size,
embedding_size=embedding_size,
use_one_hot=use_one_hot_embeddings)
self.shape_flat = (-1, )
self.one_hot = P.OneHot()
self.on_value = Tensor(1.0, mstype.float32)
self.off_value = Tensor(0.1, mstype.float32)
self.array_mul = P.MatMul()
self.reshape = P.Reshape()
self.shape = tuple(embedding_shape)
self.dropout = nn.Dropout(1 - dropout_prob)
self.gather = P.Gather()
self.use_relative_positions = use_relative_positions
self.slice = P.StridedSlice()
_, seq, _ = self.shape
self.full_position_embedding = nn.Embedding(
vocab_size=max_position_embeddings,
embedding_size=embedding_size,
use_one_hot=False)
self.layernorm = nn.LayerNorm((embedding_size, ))
self.position_ids = Tensor(
np.arange(seq).reshape(-1, seq).astype(np.int32))
self.add = P.Add()
def __init__(self,
embedding_size,
embedding_shape,
use_relative_positions=False,
use_token_type=False,
token_type_vocab_size=16,
use_one_hot_embeddings=False,
initializer_range=0.02,
max_position_embeddings=512,
dropout_prob=0.1):
super(EmbeddingPostprocessor, self).__init__()
self.use_token_type = use_token_type
self.token_type_vocab_size = token_type_vocab_size
self.use_one_hot_embeddings = use_one_hot_embeddings
self.max_position_embeddings = max_position_embeddings
self.embedding_table = Parameter(initializer
(TruncatedNormal(initializer_range),
[token_type_vocab_size,
embedding_size]),
name='embedding_table')
self.shape_flat = (-1,)
self.one_hot = P.OneHot()
self.on_value = Tensor(1.0, mstype.float32)
self.off_value = Tensor(0.1, mstype.float32)
self.array_mul = P.MatMul()
self.reshape = P.Reshape()
self.shape = tuple(embedding_shape)
self.layernorm = nn.LayerNorm((embedding_size,))
self.dropout = nn.Dropout(1 - dropout_prob)
self.gather = P.GatherV2()
self.use_relative_positions = use_relative_positions
self.slice = P.StridedSlice()
self.full_position_embeddings = Parameter(initializer
(TruncatedNormal(initializer_range),
[max_position_embeddings,
embedding_size]),
name='full_position_embeddings')
def __init__(self,
in_channels,
out_channels,
initializer_range=0.02,
dropout_prob=0.1,
compute_type=mstype.float32,
enable_fused_layernorm=False):
super(BertOutput, self).__init__()
self.dense = Dense_Thor(in_channels=in_channels,
out_channels=out_channels,
weight_init=TruncatedNormal(initializer_range),
has_bias=True,
bias_init='zeros',
damping=damping,
loss_scale=loss_scale,
frequency=frequency,
activation=None,
batch_size=batch_size).to_float(compute_type)
self.dropout = nn.Dropout(1 - dropout_prob)
self.dropout_prob = dropout_prob
self.add = P.TensorAdd()
self.layernorm = nn.LayerNorm((out_channels, )).to_float(compute_type)
self.cast = P.Cast()
def __init__(self, in_channels=None):
super(LayerNorm, self).__init__()
self.layer_norm = nn.LayerNorm((in_channels, ))
self.cast = P.Cast()
self.get_dtype = P.DType()
def test_check_layer_norm_1():
x = Tensor(np.ones([20, 5, 10, 10]), mstype.float32)
shape1 = x.shape()[1:]
m = nn.LayerNorm(shape1, -1, 1)
with pytest.raises(NotImplementedError):
m(x)
def test_check_layer_norm_3():
x = Tensor(np.ones([20, 5, 10, 10]), mstype.float32)
shape1 = (10, 10)
m = nn.LayerNorm(shape1, begin_params_axis=2)
with pytest.raises(NotImplementedError):
m(x)
def __init__(self,
hidden_size: int,
hidden_dropout_prob: int = 0.1) -> None:
super().__init__()
self.layer_norm = nn.LayerNorm((hidden_size, ), epsilon=1e-12)
self.dropout = nn.Dropout(1.0 - hidden_dropout_prob)
def __init__(self, config):
super(PANGUALPHA_Model, self).__init__()
self.get_attention_mask = AttentionMask(config)
self.word_embedding = EmbeddingLookup(config).set_comm_fusion(1)
self.eod_reset = config.eod_reset
if config.load_ckpt_path:
# Loading the embedding table from the ckpt path:
embedding_path = os.path.join(config.load_ckpt_path, 'position_embedding.npy')
if os.path.exists(embedding_path):
p_table = np.load(embedding_path)
position_table_param = Tensor(p_table, mstype.float32)
else:
raise ValueError(f"{embedding_path} file not exits, please check whether position_embedding file exit.")
else:
position_table_param = TruncatedNormal(0.02)
self.position_embedding = nn.Embedding(
config.seq_length,
config.embedding_size,
embedding_table=position_table_param).set_comm_fusion(1)
self.word_embedding.embedding_table.parallel_optimizer = False
self.position_embedding.embedding_table.parallel_optimizer = False
self.position_embedding.gather.shard(((1, 1), (config.dp,)))
self.position_embedding.expand.shard(((config.dp, 1),))
self.blocks = nn.CellList()
fusion_group_num = 4
fusion_group_size = config.num_layers // fusion_group_num
fusion_group_size = max(fusion_group_size, 1)
num_layers = config.num_layers - 1
self.num_layers = num_layers
for i in range(num_layers):
per_block = Block(config, i + 1).set_comm_fusion(int(i / fusion_group_size) + 2)
per_block.recompute()
per_block.attention.dropout.dropout_gen_mask.recompute(False)
per_block.attention.prob_dropout.dropout_gen_mask.recompute(False)
per_block.output.dropout.dropout_gen_mask.recompute(False)
per_block.attention.dropout.dropout_gen_mask.add_prim_attr("_side_effect", True)
per_block.attention.prob_dropout.dropout_gen_mask.add_prim_attr("_side_effect", True)
per_block.output.dropout.dropout_gen_mask.add_prim_attr("_side_effect", True)
self.blocks.append(per_block)
if config.self_layernorm:
self.layernorm = LayerNorm((config.embedding_size,), config.dp).to_float(
mstype.float32).set_comm_fusion(
int((num_layers - 1) / fusion_group_size) + 2)
else:
self.layernorm = nn.LayerNorm((config.embedding_size,)).to_float(
mstype.float32).set_comm_fusion(
int((num_layers - 1) / fusion_group_size) + 2)
self.layernorm.layer_norm.shard(((config.dp, 1, 1), (1,), (1,)))
self.layernorm.gamma.parallel_optimizer = False
self.layernorm.beta.parallel_optimizer = False
self.use_past = config.use_past
self.past = tuple([None] * config.num_layers)
self.add = P.TensorAdd().shard(((config.dp, 1, 1), (config.dp, 1, 1)))
self.expand_dims = P.ExpandDims().shard(((config.dp, 1, 1),))
self.dtype = config.compute_dtype
self.dropout = nn.Dropout(1 - config.dropout_rate)
self.dropout.dropout_gen_mask.shard(((config.dp, 1, 1),))
self.dropout.dropout_do_mask.shard(((config.dp, 1, 1),))
if config.load_ckpt_path:
# Loading the embedding table from the ckpt path:
embedding_path = os.path.join(config.load_ckpt_path, 'top_query_embedding.npy')
if os.path.exists(embedding_path):
top_query_table = np.load(embedding_path)
top_query_table_param = Tensor(top_query_table, mstype.float32)
else:
raise ValueError(f"{embedding_path} file not exits, please check whether top_query_embedding file exist.")
else:
top_query_table_param = TruncatedNormal(0.02)
self.top_query_embedding = nn.Embedding(config.seq_length, config.embedding_size, \
embedding_table=top_query_table_param).set_comm_fusion(
int((config.num_layers - 1) / fusion_group_num) + 2)
self.top_query_embedding.embedding_table.parallel_optimizer = False
self.top_query_embedding.gather.shard(((1, 1), (config.dp,)))
self.top_query_embedding.expand.shard(((config.dp, 1),))
self.top_query_layer = QueryLayer(config)
self.top_query_layer.recompute()
self.top_query_layer.output.dropout.dropout_gen_mask.recompute(False)
self.top_query_layer.attention.dropout.dropout_gen_mask.recompute(False)
self.top_query_layer.attention.prob_dropout.dropout_gen_mask.recompute(False)
self.top_query_layer.output.dropout.dropout_gen_mask.add_prim_attr("_side_effect", True)
self.top_query_layer.attention.dropout.dropout_gen_mask.add_prim_attr("_side_effect", True)
self.top_query_layer.attention.prob_dropout.dropout_gen_mask.add_prim_attr("_side_effect", True)
self.top_query_layer.set_comm_fusion(int((config.num_layers - 1) / fusion_group_num) + 2)
