backup wip

这个提交包含在:
Alexander Soare
2024-04-02 19:13:49 +01:00
父节点 2b928eedd4
当前提交 65ef8c30d0
共有 3 个文件被更改,包括 80 次插入 和 19 次删除

查看文件

@ -130,27 +130,87 @@ def make_offline_buffer(
# (Pdb) stats['observation']['state']['mean'] # (Pdb) stats['observation']['state']['mean']
# tensor([-0.0071, -0.6293, 1.0351, -0.0517, -0.4642, -0.0754, 0.4751, -0.0373, # tensor([-0.0071, -0.6293, 1.0351, -0.0517, -0.4642, -0.0754, 0.4751, -0.0373,
# -0.3324, 0.9034, -0.2258, -0.3127, -0.2412, 0.6866]) # -0.3324, 0.9034, -0.2258, -0.3127, -0.2412, 0.6866])
stats['observation', 'state', 'mean'] = torch.tensor([-0.00740268, -0.63187766, 1.0356655 , -0.05027218, -0.46199223, stats["observation", "state", "mean"] = torch.tensor(
-0.07467502, 0.47467607, -0.03615446, -0.33203387, 0.9038929 , [
-0.22060776, -0.31011587, -0.23484458, 0.6842416 ]) -0.00740268,
-0.63187766,
1.0356655,
-0.05027218,
-0.46199223,
-0.07467502,
0.47467607,
-0.03615446,
-0.33203387,
0.9038929,
-0.22060776,
-0.31011587,
-0.23484458,
0.6842416,
]
)
# (Pdb) stats['observation']['state']['std'] # (Pdb) stats['observation']['state']['std']
# tensor([0.0022, 0.0520, 0.0291, 0.0092, 0.0267, 0.0145, 0.0563, 0.0179, 0.0494, # tensor([0.0022, 0.0520, 0.0291, 0.0092, 0.0267, 0.0145, 0.0563, 0.0179, 0.0494,
# 0.0326, 0.0476, 0.0535, 0.0956, 0.0513]) # 0.0326, 0.0476, 0.0535, 0.0956, 0.0513])
stats['observation', 'state', 'std'] = torch.tensor([0.01219023, 0.2975381 , 0.16728032, 0.04733803, 0.1486037 , stats["observation", "state", "std"] = torch.tensor(
0.08788499, 0.31752336, 0.1049916 , 0.27933604, 0.18094037, [
0.26604933, 0.30466506, 0.5298686 , 0.25505227]) 0.01219023,
0.2975381,
0.16728032,
0.04733803,
0.1486037,
0.08788499,
0.31752336,
0.1049916,
0.27933604,
0.18094037,
0.26604933,
0.30466506,
0.5298686,
0.25505227,
]
)
# (Pdb) stats['action']['mean'] # (Pdb) stats['action']['mean']
# tensor([-0.0075, -0.6346, 1.0353, -0.0465, -0.4686, -0.0738, 0.3723, -0.0396, # tensor([-0.0075, -0.6346, 1.0353, -0.0465, -0.4686, -0.0738, 0.3723, -0.0396,
# -0.3184, 0.8991, -0.2065, -0.3182, -0.2338, 0.5593]) # -0.3184, 0.8991, -0.2065, -0.3182, -0.2338, 0.5593])
stats['action']['mean'] = torch.tensor([-0.00756444, -0.6281845 , 1.0312834 , -0.04664314, -0.47211358, stats["action"]["mean"] = torch.tensor(
-0.074527 , 0.37389806, -0.03718753, -0.3261143 , 0.8997205 , [
-0.21371077, -0.31840396, -0.23360962, 0.551947]) -0.00756444,
-0.6281845,
1.0312834,
-0.04664314,
-0.47211358,
-0.074527,
0.37389806,
-0.03718753,
-0.3261143,
0.8997205,
-0.21371077,
-0.31840396,
-0.23360962,
0.551947,
]
)
# (Pdb) stats['action']['std'] # (Pdb) stats['action']['std']
# tensor([0.0023, 0.0514, 0.0290, 0.0086, 0.0263, 0.0143, 0.0593, 0.0185, 0.0510, # tensor([0.0023, 0.0514, 0.0290, 0.0086, 0.0263, 0.0143, 0.0593, 0.0185, 0.0510,
# 0.0328, 0.0478, 0.0531, 0.0945, 0.0794]) # 0.0328, 0.0478, 0.0531, 0.0945, 0.0794])
stats['action']['std'] = torch.tensor([0.01252818, 0.2957442 , 0.16701928, 0.04584508, 0.14833844, stats["action"]["std"] = torch.tensor(
0.08763024, 0.30665937, 0.10600077, 0.27572668, 0.1805853 , [
0.26304692, 0.30708534, 0.5305411 , 0.38381037]) 0.01252818,
0.2957442,
0.16701928,
0.04584508,
0.14833844,
0.08763024,
0.30665937,
0.10600077,
0.27572668,
0.1805853,
0.26304692,
0.30708534,
0.5305411,
0.38381037,
]
)
transforms.append(NormalizeTransform(stats, in_keys, mode=normalization_mode)) transforms.append(NormalizeTransform(stats, in_keys, mode=normalization_mode))
offline_buffer.set_transform(transforms) offline_buffer.set_transform(transforms)

查看文件

@ -2,7 +2,6 @@ import numpy as np
import torch import torch
from torch import nn from torch import nn
from torch.autograd import Variable from torch.autograd import Variable
from transformers import DetrForObjectDetection
from .backbone import build_backbone from .backbone import build_backbone
from .transformer import TransformerEncoder, TransformerEncoderLayer, build_transformer from .transformer import TransformerEncoder, TransformerEncoderLayer, build_transformer
@ -134,7 +133,9 @@ class ActionChunkingTransformer(nn.Module):
pos_embed = self.pos_table.clone().detach() pos_embed = self.pos_table.clone().detach()
pos_embed = pos_embed.permute(1, 0, 2) # (seq+1, 1, hidden_dim) pos_embed = pos_embed.permute(1, 0, 2) # (seq+1, 1, hidden_dim)
# query model # query model
vae_encoder_output = self.vae_encoder(vae_encoder_input, pos=pos_embed) # , src_key_padding_mask=is_pad) vae_encoder_output = self.vae_encoder(
vae_encoder_input, pos=pos_embed
) # , src_key_padding_mask=is_pad)
vae_encoder_output = vae_encoder_output[0] # take cls output only vae_encoder_output = vae_encoder_output[0] # take cls output only
latent_info = self.latent_proj(vae_encoder_output) latent_info = self.latent_proj(vae_encoder_output)
mu = latent_info[:, : self.latent_dim] mu = latent_info[:, : self.latent_dim]