79 lines
2.8 KiB
Python
79 lines
2.8 KiB
Python
import torch
|
|
from torch import nn
|
|
from torch.nn.utils.rnn import pad_sequence
|
|
import PytorchBoot.stereotype as stereotype
|
|
|
|
|
|
@stereotype.module("transformer_seq_encoder")
|
|
class TransformerSequenceEncoder(nn.Module):
|
|
def __init__(self, config):
|
|
super(TransformerSequenceEncoder, self).__init__()
|
|
self.config = config
|
|
embed_dim = config["pts_embed_dim"] + config["pose_embed_dim"]
|
|
encoder_layer = nn.TransformerEncoderLayer(
|
|
d_model=embed_dim,
|
|
nhead=config["num_heads"],
|
|
dim_feedforward=config["ffn_dim"],
|
|
batch_first=True,
|
|
)
|
|
self.transformer_encoder = nn.TransformerEncoder(
|
|
encoder_layer, num_layers=config["num_layers"]
|
|
)
|
|
self.fc = nn.Linear(embed_dim, config["output_dim"])
|
|
|
|
def encode_sequence(self, pts_embedding_list_batch, pose_embedding_list_batch):
|
|
# Combine features and pad sequences
|
|
combined_features_batch = []
|
|
lengths = []
|
|
|
|
for pts_embedding_list, pose_embedding_list in zip(pts_embedding_list_batch, pose_embedding_list_batch):
|
|
combined_features = [
|
|
torch.cat((pts_embed, pose_embed), dim=-1)
|
|
for pts_embed, pose_embed in zip(pts_embedding_list, pose_embedding_list)
|
|
]
|
|
combined_features_batch.append(torch.stack(combined_features))
|
|
lengths.append(len(combined_features))
|
|
|
|
combined_tensor = pad_sequence(combined_features_batch, batch_first=True) # Shape: [batch_size, max_seq_len, embed_dim]
|
|
|
|
# Prepare mask for padding
|
|
max_len = max(lengths)
|
|
padding_mask = torch.tensor([([0] * length + [1] * (max_len - length)) for length in lengths], dtype=torch.bool).to(combined_tensor.device)
|
|
# Transformer encoding
|
|
transformer_output = self.transformer_encoder(combined_tensor, src_key_padding_mask=padding_mask)
|
|
|
|
# Mean pooling
|
|
final_feature = transformer_output.mean(dim=1)
|
|
|
|
# Fully connected layer
|
|
final_output = self.fc(final_feature)
|
|
|
|
return final_output
|
|
|
|
|
|
if __name__ == "__main__":
|
|
config = {
|
|
"pts_embed_dim": 1024,
|
|
"pose_embed_dim": 256,
|
|
"num_heads": 4,
|
|
"ffn_dim": 256,
|
|
"num_layers": 3,
|
|
"output_dim": 2048,
|
|
}
|
|
|
|
encoder = TransformerSequenceEncoder(config)
|
|
seq_len = [5, 8, 9, 4]
|
|
batch_size = 4
|
|
|
|
pts_embedding_list_batch = [
|
|
torch.randn(seq_len[idx], config["pts_embed_dim"]) for idx in range(batch_size)
|
|
]
|
|
pose_embedding_list_batch = [
|
|
torch.randn(seq_len[idx], config["pose_embed_dim"]) for idx in range(batch_size)
|
|
]
|
|
output_feature = encoder.encode_sequence(
|
|
pts_embedding_list_batch, pose_embedding_list_batch
|
|
)
|
|
print("Encoded Feature:", output_feature)
|
|
print("Feature Shape:", output_feature.shape)
|