Spaces:
Running on Zero
Running on Zero
File size: 5,410 Bytes
2680bd5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 | from typing import Dict
import torch
import torch.nn as nn
from torch import Tensor
import numpy as np
from einops import repeat
class PositionalEncoding(nn.Module):
def __init__(self, d_model, dropout=0.1, max_len=5000, batch_first=False) -> None:
super().__init__()
self.batch_first = batch_first
self.dropout = nn.Dropout(p=dropout)
pe = torch.zeros(max_len, d_model)
position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
div_term = torch.exp(
torch.arange(0, d_model, 2).float() * (-np.log(10000.0) / d_model)
)
pe[:, 0::2] = torch.sin(position * div_term)
pe[:, 1::2] = torch.cos(position * div_term)
pe = pe.unsqueeze(0).transpose(0, 1)
self.register_buffer("pe", pe, persistent=False)
def forward(self, x: Tensor) -> Tensor:
if self.batch_first:
x = x + self.pe.permute(1, 0, 2)[:, : x.shape[1], :]
else:
x = x + self.pe[: x.shape[0], :]
return self.dropout(x)
class ACTORStyleEncoder(nn.Module):
# Similar to ACTOR but "action agnostic" and more general
def __init__(
self,
nfeats: int,
vae: bool,
latent_dim: int = 256,
ff_size: int = 1024,
num_layers: int = 4,
num_heads: int = 4,
dropout: float = 0.1,
activation: str = "gelu",
) -> None:
super().__init__()
self.nfeats = nfeats
self.projection = nn.Linear(nfeats, latent_dim)
self.vae = vae
self.nbtokens = 2 if vae else 1
self.tokens = nn.Parameter(torch.randn(self.nbtokens, latent_dim))
self.sequence_pos_encoding = PositionalEncoding(
latent_dim, dropout=dropout, batch_first=True
)
seq_trans_encoder_layer = nn.TransformerEncoderLayer(
d_model=latent_dim,
nhead=num_heads,
dim_feedforward=ff_size,
dropout=dropout,
activation=activation,
batch_first=True,
)
self.seqTransEncoder = nn.TransformerEncoder(
seq_trans_encoder_layer, num_layers=num_layers
)
def forward(self, x_dict: Dict) -> Tensor:
x = x_dict["x"]
mask = x_dict["mask"] # (B, L)
x = self.projection(x)
device = x.device
bs = len(x)
tokens = repeat(self.tokens, "nbtoken dim -> bs nbtoken dim", bs=bs)
xseq = torch.cat((tokens, x), 1)
token_mask = torch.ones((bs, self.nbtokens), dtype=bool, device=device)
aug_mask = torch.cat((token_mask, mask), 1)
# add positional encoding
xseq = self.sequence_pos_encoding(xseq)
final = self.seqTransEncoder(xseq, src_key_padding_mask=~aug_mask)
return final[:, : self.nbtokens]
class ACTORStyleDecoder(nn.Module):
# Similar to ACTOR Decoder
def __init__(
self,
nfeats: int,
latent_dim: int = 256,
ff_size: int = 1024,
num_layers: int = 4,
num_heads: int = 4,
dropout: float = 0.1,
activation: str = "gelu",
) -> None:
super().__init__()
output_feats = nfeats
self.nfeats = nfeats
self.sequence_pos_encoding = PositionalEncoding(
latent_dim, dropout, batch_first=True
)
seq_trans_decoder_layer = nn.TransformerDecoderLayer(
d_model=latent_dim,
nhead=num_heads,
dim_feedforward=ff_size,
dropout=dropout,
activation=activation,
batch_first=True,
)
self.seqTransDecoder = nn.TransformerDecoder(
seq_trans_decoder_layer, num_layers=num_layers
)
self.final_layer = nn.Linear(latent_dim, output_feats)
def forward(self, z_dict: Dict) -> Tensor:
z = z_dict["z"]
mask = z_dict["mask"]
latent_dim = z.shape[1]
bs, nframes = mask.shape
z = z[:, None] # sequence of 1 element for the memory
# Construct time queries
time_queries = torch.zeros(bs, nframes, latent_dim, device=z.device)
time_queries = self.sequence_pos_encoding(time_queries)
# Pass through the transformer decoder
# with the latent vector for memory
output = self.seqTransDecoder(
tgt=time_queries, memory=z, tgt_key_padding_mask=~mask
)
output = self.final_layer(output)
# zero for padded area
output[~mask] = 0
return output
if __name__ == "__main__":
from collections import OrderedDict
checkpoint_path = "/u/zzhang48/Codes/PythonProjects/MyHandTMR/outputs/tmr_snap_snapmotion/lightning_logs/imfa2q9i/checkpoints/epoch-epoch=24.ckpt"
full_checkpoint:OrderedDict = torch.load(checkpoint_path, map_location='cpu')['state_dict']
checkpoint = OrderedDict()
for key, value in full_checkpoint.items():
if key.startswith("motion_encoder."):
new_key = key[len("motion_encoder."):]
checkpoint[new_key] = value
# print(set([key.split(".")[0] for key in checkpoint.keys()]))
motion_encoder = ACTORStyleEncoder(
vae=True,
latent_dim=256,
ff_size=1024,
num_layers=6,
num_heads=4,
dropout=0.1,
activation='gelu',
nfeats=126
)
motion_encoder.load_state_dict(checkpoint, strict=True)
|