Spaces:
Running
on
T4
Running
on
T4
File size: 6,000 Bytes
6e445f1 |
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 |
# ------------------------------------------------------------------------------
# Reference: https://github.com/facebookresearch/Mask2Former/blob/main/mask2former/modeling/meta_arch/mask_former_head.py
# Modified by Jitesh Jain (https://github.com/praeclarumjj3)
# ------------------------------------------------------------------------------
import logging
from copy import deepcopy
from typing import Callable, Dict, List, Optional, Tuple, Union
import fvcore.nn.weight_init as weight_init
from torch import nn
from torch.nn import functional as F
from detectron2.config import configurable
from detectron2.layers import Conv2d, ShapeSpec, get_norm
from detectron2.modeling import SEM_SEG_HEADS_REGISTRY
from ..pixel_decoder.fpn import build_pixel_decoder
from ..transformer_decoder.oneformer_transformer_decoder import build_transformer_decoder
@SEM_SEG_HEADS_REGISTRY.register()
class OneFormerHead(nn.Module):
_version = 2
def _load_from_state_dict(
self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs
):
version = local_metadata.get("version", None)
if version is None or version < 2:
# Do not warn if train from scratch
scratch = True
logger = logging.getLogger(__name__)
for k in list(state_dict.keys()):
newk = k
if "sem_seg_head" in k and not k.startswith(prefix + "predictor"):
newk = k.replace(prefix, prefix + "pixel_decoder.")
# logger.debug(f"{k} ==> {newk}")
if newk != k:
state_dict[newk] = state_dict[k]
del state_dict[k]
scratch = False
if not scratch:
logger.warning(
f"Weight format of {self.__class__.__name__} have changed! "
"Please upgrade your models. Applying automatic conversion now ..."
)
@configurable
def __init__(
self,
input_shape: Dict[str, ShapeSpec],
*,
num_classes: int,
pixel_decoder: nn.Module,
loss_weight: float = 1.0,
ignore_value: int = -1,
# extra parameters
transformer_predictor: nn.Module,
transformer_in_feature: str,
):
"""
NOTE: this interface is experimental.
Args:
input_shape: shapes (channels and stride) of the input features
num_classes: number of classes to predict
pixel_decoder: the pixel decoder module
loss_weight: loss weight
ignore_value: category id to be ignored during training.
transformer_predictor: the transformer decoder that makes prediction
transformer_in_feature: input feature name to the transformer_predictor
"""
super().__init__()
input_shape = sorted(input_shape.items(), key=lambda x: x[1].stride)
self.in_features = [k for k, v in input_shape]
feature_strides = [v.stride for k, v in input_shape]
feature_channels = [v.channels for k, v in input_shape]
self.ignore_value = ignore_value
self.common_stride = 4
self.loss_weight = loss_weight
self.pixel_decoder = pixel_decoder
self.predictor = transformer_predictor
self.transformer_in_feature = transformer_in_feature
self.num_classes = num_classes
@classmethod
def from_config(cls, cfg, input_shape: Dict[str, ShapeSpec]):
# figure out in_channels to transformer predictor
if cfg.MODEL.ONE_FORMER.TRANSFORMER_IN_FEATURE == "transformer_encoder":
transformer_predictor_in_channels = cfg.MODEL.SEM_SEG_HEAD.CONVS_DIM
elif cfg.MODEL.ONE_FORMER.TRANSFORMER_IN_FEATURE == "pixel_embedding":
transformer_predictor_in_channels = cfg.MODEL.SEM_SEG_HEAD.MASK_DIM
elif cfg.MODEL.ONE_FORMER.TRANSFORMER_IN_FEATURE == "multi_scale_pixel_decoder":
transformer_predictor_in_channels = cfg.MODEL.SEM_SEG_HEAD.CONVS_DIM
else:
transformer_predictor_in_channels = input_shape[cfg.MODEL.ONE_FORMER.TRANSFORMER_IN_FEATURE].channels
return {
"input_shape": {
k: v for k, v in input_shape.items() if k in cfg.MODEL.SEM_SEG_HEAD.IN_FEATURES
},
"ignore_value": cfg.MODEL.SEM_SEG_HEAD.IGNORE_VALUE,
"num_classes": cfg.MODEL.SEM_SEG_HEAD.NUM_CLASSES,
"pixel_decoder": build_pixel_decoder(cfg, input_shape),
"loss_weight": cfg.MODEL.SEM_SEG_HEAD.LOSS_WEIGHT,
"transformer_in_feature": cfg.MODEL.ONE_FORMER.TRANSFORMER_IN_FEATURE,
"transformer_predictor": build_transformer_decoder(
cfg,
transformer_predictor_in_channels,
mask_classification=True,
),
}
def forward(self, features, tasks, mask=None):
return self.layers(features, tasks, mask)
def layers(self, features, tasks, mask=None):
mask_features, transformer_encoder_features, multi_scale_features, _, _ = self.pixel_decoder.forward_features(features)
if self.transformer_in_feature == "multi_scale_pixel_decoder":
predictions = self.predictor(multi_scale_features, mask_features, tasks, mask)
else:
if self.transformer_in_feature == "transformer_encoder":
assert (
transformer_encoder_features is not None
), "Please use the TransformerEncoderPixelDecoder."
predictions = self.predictor(transformer_encoder_features, mask_features, mask)
elif self.transformer_in_feature == "pixel_embedding":
predictions = self.predictor(mask_features, mask_features, mask)
else:
predictions = self.predictor(features[self.transformer_in_feature], mask_features, mask)
return predictions
|