比较与torch.nn.TransformerEncoderLayer的功能差异

torch.nn.TransformerEncoderLayer(d_model, nhead, dim_feedforward=2048, dropout=0.1, activation=<function relu>, layer_norm_eps=1e-05, batch_first=False, norm_first=False, device=None, dtype=None)

class mindspore.nn.transformer.TransformerEncoderLayer(batch_size, hidden_size, ffn_hidden_size, num_heads, seq_length, attention_dropout_rate=0.1, hidden_dropout_rate=0.1, post_layernorm_residual=False, layernorm_compute_type=mstype.float32, softmax_compute_type=mstype.float32, param_init_type=mstype.float32, hidden_act="gelu", use_past=False, moe_config=default_moe_config, parallel_config=default_dpmp_config)(
    x, input_mask, init_reset=True, batch_valid_length=None
)

import numpy as np
from mindspore import dtype as mstype
from mindspore.nn.transformer import TransformerEncoderLayer
from mindspore import Tensor
model = TransformerEncoderLayer(batch_size=32, hidden_size=512,
                                ffn_hidden_size=2048, seq_length=10, num_heads=8)
encoder_input_value = Tensor(np.random.rand(32, 10, 512), mstype.float32)
encoder_input_mask = Tensor(np.ones((32, 10, 10)), mstype.float16)
output, past = model(encoder_input_value, encoder_input_mask)
print(output.shape)
# output:
# (32, 10, 512)

import torch
encoder_layer = torch.nn.TransformerEncoderLayerLayer(d_model=512, nhead=8)
transformer_encoder = torch.nn.TransformerEncoder(encoder_layer, num_layers=2)
src = torch.rand(10, 32, 512)
output = transformer_encoder(src)
print(output.shape)
# output:
# torch.Size([10, 32, 512])