File size: 4,539 Bytes
c94c8c9 |
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 |
import torch
import torch.nn as nn
import json
from pathlib import Path
import clip
from transformers import BertConfig, BertModel, BertTokenizer
from einops import rearrange
from model.build import MODEL_REGISTRY, BaseModel
from modules.layers.pointnet import PointNetPP
from modules.utils import get_mlp_head
from optim.utils import no_decay_param_group
@MODEL_REGISTRY.register()
class ObjCls(BaseModel):
def __init__(self, cfg):
super().__init__(cfg)
self.cfg = cfg
self.model_name = cfg.model.get("model_name", "pointnext")
self.language_type = cfg.model.get("language_type", "clip")
self.pre_extract_path = cfg.model.get("pre_extract_path", None)
cls_in_channel = 512 if self.language_type == "clip" else 768
self.point_feature_extractor = PointNetPP(
sa_n_points=[32, 16, None],
sa_n_samples=[32, 32, None],
sa_radii=[0.2, 0.4, None],
sa_mlps=[[3, 64, 64, 128], [128, 128, 128, 256], [256, 256, 512, cls_in_channel]],
)
if cfg.num_gpu > 1:
self.point_feature_extractor = torch.nn.SyncBatchNorm.convert_sync_batchnorm(self.point_feature_extractor)
if not cfg.model.open_vocab:
cls_hidden = cfg.model.get("cls_hidden", 1024)
num_classes = cfg.model.num_classes
self.cls_head = get_mlp_head(cls_in_channel, cls_hidden, num_classes)
else:
if self.pre_extract_path is not None:
file_name = f"scannet_607_{'clip-ViT-B16' if self.language_type == 'clip' else 'bert-base-uncased'}_id.pth"
self.register_buffer("text_embeds", torch.load(Path(self.pre_extract_path) / file_name).float())
else:
self.int2cat = json.load(open(cfg.model.vocab_path, "r"))
if self.language_type == "clip":
self.clip_head = clip.load("ViT-B/16")
self.text_embeds = self.clip_head.encode_text(clip.tokenize(self.int2cat)).detach()
elif self.language_type == "bert":
self.tokenizer = BertTokenizer.from_pretrained("bert-base-uncased", do_lower_case=True)
self.bert_config = BertConfig(
hidden_size=768, num_hidden_layers=3, num_attention=12, type_vocab_size=2
)
self.model = BertModel.from_pretrained("bert-base-uncased", config=self.bert_config)
self.encoded_input = self.tokenizer(
self.int2cat, padding=True, truncation=True, add_special_tokens=True, return_tensors="pt"
)
self.text_embeds = self.model(**self.encoded_input).last_hidden_state
self.text_embeds = self.text_embeds.detach()
else:
raise NotImplementedError
self.dropout = nn.Dropout(0.1)
def forward(self, data_dict):
# prepare dict
if 'cur_step' not in data_dict.keys():
data_dict['cur_step'] = 1
data_dict['total_steps'] = 1
obj_pcds = data_dict["obj_fts"]
batch_size, num_objs, _, _ = obj_pcds.size()
if self.model_name == "pointnext":
obj_locs = rearrange(obj_pcds[..., :3], 'b o p d -> (b o) p d')
obj_fts = rearrange(obj_pcds[..., 3:], 'b o p d -> (b o) d p').contiguous()
obj_embeds = self.point_feature_extractor(obj_locs, obj_fts, type="cls")
elif self.model_name == "pointnet++":
obj_pcds = rearrange(obj_pcds, 'b o p d -> (b o) p d')
obj_embeds = self.point_feature_extractor(obj_pcds)
elif self.model_name == "pointmlp":
obj_pcds = rearrange(obj_pcds, 'b o p d -> (b o) p d')
obj_embeds = self.point_feature_extractor(obj_pcds)
obj_embeds = self.dropout(obj_embeds)
if self.cfg.model.open_vocab:
logits = obj_embeds @ self.text_embeds.t()
data_dict["obj_logits"] = rearrange(logits, '(b o) c -> b o c', b=batch_size)
else:
data_dict["obj_logits"] = rearrange(self.cls_head(obj_embeds), '(b o) d -> b o d', b=batch_size)
return data_dict
def get_opt_params(self):
optimizer_grouped_parameters = []
optimizer_grouped_parameters.append({
"params": self.parameters(),
"weight_decay": self.cfg.solver.get("weight_decay", 0.0),
"lr": self.cfg.solver.lr
})
return optimizer_grouped_parameters
|