https://github.com/facebookresearch/maskrcnn-benchmark/issues/521
ps：如果只需要fintune的话，直接看第二部分，是可以查看model的各个键值的，然后根据需要进行删减。
ps：后面有个乱七八糟的，是我记录的，后期删改
ps：还有个是可以冻结残差网络的训练层的，eg：resnet50，我冻结了最大层，也就是：FREEZE_CONV_BODY_AT = 4
ResNet50FPNStagesTo5 = tuple(
StageSpec(index=i, block_count=c, return_features=r)
for (i, c, r) in ((1, 3, True), (2, 4, True), (3, 6, True), (4, 3, True))
)
FREEZE_CONV_BODY_AT = 1 train resnet only 16 last layers
FREEZE_CONV_BODY_AT = 2 train resnet only 13 last layers
FREEZE_CONV_BODY_AT = 3 train resnet only 9 last layers
FREEZE_CONV_BODY_AT = 4 train resnet only 3 last layers
一、不同数据fintune，不同类别。
Steps1) COCO format将自己的数据集按照coco官方形式进行标注。这样会好一点，测试时候直接用coco官方api就可以了。可用labelme等等，再进行转换，还有对应的json文件。official coco format 2) Creating a
|
18
2
3
二、网络的fintune
需要的步骤是：
1 -为特定模型创建cfg
2 -使用load_c2_format函数，将为您提供包含模型字段的dict。在那里，可以通过删除字段等来执行您想要的模型手术
3 - 使用 torch.save保存目标,结构应该还是按dict(model=state_dict).保存，并且，请注意扩展名是.pth而不是pkl
4 - 更改 MODEL.WEIGHT指向更改后的.pth文件地址。
前三步分别对应下面的trim_detectron_model.py脚本，注意别忘了第4步说明如下：
pretrained_path：就是从MODEL_ZOO下载的作者训练好的结果，（后缀.pkl）
save_path：自己修改后用于fintune的模型地址，请注意扩展名是.pth
cfg：与模型对应的配置文件。
可以看到自己网络里的参数，比如我的是(截取部分，对应删除就好了)：[ # FAST RCNN层，2个FC和分类回归层 'fc1000.bias', 'fc1000.weight', 'fc6.bias', 'fc6.weight', 'fc7.bias', 'fc7.weight', 'bbox_pred.bias', 'bbox_pred.weight', 'cls_score.bias', 'cls_score.weight', # Mask层： 'mask_fcn1.bias', 'mask_fcn1.weight', 'mask_fcn2.bias', 'mask_fcn2.weight', 'mask_fcn3.bias', 'mask_fcn3.weight', 'mask_fcn4.bias', 'mask_fcn4.weight', 'conv5_mask.bias', 'conv5_mask.weight', 'mask_fcn_logits.bias', 'mask_fcn_logits.weight', # RPN层 'rpn.head.conv.bias', 'rpn.head.conv.weight', 'rpn.head.bbox_pred.bias', 'rpn.head.bbox_pred.weight', 'rpn.head.cls_logits.bias', 'rpn.head.cls_logits.weight' # FPN层 'fpn_inner1.bias', 'fpn_inner1.weight', 'fpn_inner2.bias', 'fpn_inner2.weight', 'fpn_inner3.bias', 'fpn_inner3.weight', 'fpn_inner4.bias', 'fpn_inner4.weight', 'fpn_layer1.bias', 'fpn_layer1.weight', 'fpn_layer2.bias','fpn_layer2.weight', 'fpn_layer3.bias', 'fpn_layer3.weight', 'fpn_layer4.bias', 'fpn_layer4.weight', # 残差层 ps：为啥只有bn1和bn2？？？？ 'conv1.bias', 'conv1.weight', 'bn1.bias', 'bn1.weight', 'layer1.0.downsample.0.bias', 'layer1.0.downsample.1.bias', 'layer1.0.downsample.1.weight', 'layer1.0.downsample.0.weight', 'layer1.0.conv1.bias', 'layer1.0.bn1.bias', 'layer1.0.bn1.weight', 'layer1.0.conv1.weight', 'layer1.0.conv2.bias', 'layer1.0.bn2.bias', 'layer1.0.bn2.weight', 'layer1.0.conv2.weight', ................. 'layer4.0.downsample.0.bias', 'layer4.0.downsample.1.bias', 'layer4.0.downsample.1.weight', 'layer4.0.downsample.0.weight', 'layer4.0.conv1.bias', 'layer4.0.bn1.bias', 'layer4.0.bn1.weight', 'layer4.0.conv1.weight', ................. 'layer4.2.conv3.bias', 'layer4.2.bn3.bias', 'layer4.2.bn3.weight', 'layer4.2.conv3.weight', ] '''
import os
import torch
import argparse
from maskrcnn_benchmark.config import cfg
from maskrcnn_benchmark.utils.c2_model_loading import load_c2_format
def removekey(d, listofkeys):
r = dict(d)
for key in listofkeys:
print('key: {} is removed'.format(key))
r.pop(key)
return r
parser = argparse.ArgumentParser(description="Trim Detection weights and save in PyTorch format.")
parser.add_argument(
"--pretrained_path",
default="~/.torch/models/_detectron_35858933_12_2017_baselines_e2e_mask_rcnn_R-50-FPN_1x.yaml.01_48_14.DzEQe4wC_output_train_coco_2014_train%3Acoco_2014_valminusminival_generalized_rcnn_model_final.pkl",
help="path to detectron pretrained weight(.pkl)",
type=str,
)
parser.add_argument(
"--save_path",
default="./pretrained_model/mask_rcnn_R-50-FPN_1x_detectron_no_last_layers.pth",
help="path to save the converted model",
type=str,
)
parser.add_argument(
"--cfg",
default="configs/e2e_mask_rcnn_R_50_FPN_1x.yaml",
help="path to config file",
type=str,
)
args = parser.parse_args()
#
DETECTRON_PATH = os.path.expanduser(args.pretrained_path)
print('detectron path: {}'.format(DETECTRON_PATH))
cfg.merge_from_file(args.cfg)
_d = load_c2_format(cfg, DETECTRON_PATH)
print([key for key in _d]) #['model'] 只有这一个键
print([so for so in newdict["model"]]) #或者执行keys = [k for k in _d['model'].keys()]也可以
newdict = _d
#把后面的值从前面移除
newdict['model'] = removekey(_d['model'],
['cls_score.bias', 'cls_score.weight', 'bbox_pred.bias', 'bbox_pred.weight'])
torch.save(newdict, args.save_path)
print('saved to {}.'.format(args.save_path))
# 检查是否按设计输出：
w = torch.load(args.save_path)
print([k for k in w['model'].keys()])
记录一个错误：
w = torch.load("X-101-32x8d.pth")
然而，发生错误：UnicodeDecodeError：'ascii'编解码器无法解码位置2中的字节0xad：序数不在范围内（128）
可以通过使用pickle来解决这个错误：
import pickle
with open("X-101-32x8d.pkl", "rb") as f: w = pickle.load(f, encoding='latin1')
[loaded_state_dict]函数可以检查函数中的哪些键 ，可以在这里进行操作（maskrcnn-benchmark/maskrcnn_benchmark/utils/model_serialization.py )
不会影响repo中的其他部分。大概是
model_dict = model.state_dict()
loaded_state_dict = {k: v for k, v in loaded_state_dict.items() if k in model_dict and "roi_heads" not in k}
model_dict.update(loaded_state_dict)
如果是这样，则无需经过4个步骤。只用编辑 [load_state_dict method]
align_and_update_state_dicts(model_state_dict, loaded_state_dict)
**步骤1**。运行demo.py以测试您的软件包是否已成功安装。同时，您可以保存预先训练的模型，该模型已从pkl格式转换为pth格式，可以通过pytorch模型加载。该模型是演示中COCODEmo对象的数据，因此您可以保存预训练模型的权重，如下所示
pretrained_model_path = "./pretrained_model/maskRCNN.pth"
check_point = {'state': coco_demo.model.state_dict()}
torch.save(check_point, pretrained_model_path)
这样，无需查看[_load_file fun](https://github.com/facebookresearch/maskrcnn-benchmark/blob/f25c6cff92d32d92abe8965d68401004e90c8bee/maskrcnn_benchmark/utils/checkpoint.py#L117).另外还有个保存权重到OrderedDict的方法：(https://github.com/pytorch/pytorch/blob/828cb18fa35c7c132ab920de1c0dc6d859f152d6/torch/nn/modules/module.py#L602)
retrained_model = torch.load(checkpoint_path)['model']
model_dict = self.model.state_dict()
pretrained_dict = {k: v for k, v in pretrained_model.items() if k in model_dict and "roi_heads" not in k}
model_dict.update(pretrained_dict)
**step 2**. 使用configs文件夹中的yaml文件创建fastercnn或maskRCNN模型。执行此操作时，您需要调用cfg.merge_from_list方法来更改[_C.MODEL.ROI_BOX_HEAD.NUM_CLASSES = 81]到数据集中的类数（后台在此实现中不计算）。(https://github.com/facebookresearch/maskrcnn-benchmark/blob/f25c6cff92d32d92abe8965d68401004e90c8bee/maskrcnn_benchmark/config/defaults.py#L182)
**step 3**.需要创建两个类。
第一个类：[torchvision.datasets.coco.CocoDetection] (
maskrcnn-benchmark/maskrcnn_benchmark/data/datasets/coco.py
class COCODataset(torchvision.datasets.coco.CocoDetection):
获取一个图像及其对应的bbox和标签，而不将Bbox和标签捆绑在BoxList对象中。也就是 return img, target, idx
此类与[COCODataset](https://github.com/facebookresearch/maskrcnn-benchmark/blob/f25c6cff92d32d92abe8965d68401004e90c8bee/maskrcnn_benchmark/data/datasets/coco.py#L9). 在每次迭代中，该类生成一个元组（img，bbox，label）,其中img由Image.open(...).convert('RGB')生成;
bbox 是 2d numpy 数组， label是1d 数组(mask的话就是掩码信息).
第二类继承自第一类，这个类的主要职责是将一个图像的真实包装在一个BoxList对象中，类似于[line]
| img, anno = super(COCODataset, self).__getitem__(idx) |
(https://github.com/facebookresearch/maskrcnn-benchmark/blob/f25c6cff92d32d92abe8965d68401004e90c8bee/maskrcnn_benchmark/data/datasets/coco.py#L43).
**step 4**. 对于pytorch预训练模型，权重在字典中被分类，其中键是layer的名称。因此，您可以排除不存在的layer或者您想要从头开始训练，如下所示：pretrained_model = torch.load(checkpoint_path)['state']
城市景观数据集cityscapes与coco
def clip_weights_from_pretrain_of_coco_to_cityscapes(f, out_file):
""""""
from maskrcnn_benchmark.config.paths_catalog import COCO_CATEGORIES
from maskrcnn_benchmark.config.paths_catalog import CITYSCAPES_FINE_CATEGORIES
coco_cats = COCO_CATEGORIES
cityscapes_cats = CITYSCAPES_FINE_CATEGORIES
coco_cats_to_inds = dict(zip(coco_cats, range(len(coco_cats))))
cityscapes_cats_to_inds = dict(
zip(cityscapes_cats, range(len(cityscapes_cats)))
)
checkpoint = torch.load(f)
m = checkpoint['model']
weight_names = {
"cls_score": "module.roi_heads.box.predictor.cls_score.weight",
"bbox_pred": "module.roi_heads.box.predictor.bbox_pred.weight",
"mask_fcn_logits": "module.roi_heads.mask.predictor.mask_fcn_logits.weight",
}
bias_names = {
"cls_score": "module.roi_heads.box.predictor.cls_score.bias",
"bbox_pred": "module.roi_heads.box.predictor.bbox_pred.bias",
"mask_fcn_logits": "module.roi_heads.mask.predictor.mask_fcn_logits.bias",
}
representation_size = m[weight_names["cls_score"]].size(1)
cls_score = nn.Linear(representation_size, len(cityscapes_cats))
nn.init.normal_(cls_score.weight, std=0.01)
nn.init.constant_(cls_score.bias, 0)
representation_size = m[weight_names["bbox_pred"]].size(1)
class_agnostic = m[weight_names["bbox_pred"]].size(0) != len(coco_cats) * 4
num_bbox_reg_classes = 2 if class_agnostic else len(cityscapes_cats)
bbox_pred = nn.Linear(representation_size, num_bbox_reg_classes * 4)
nn.init.normal_(bbox_pred.weight, std=0.001)
nn.init.constant_(bbox_pred.bias, 0)
dim_reduced = m[weight_names["mask_fcn_logits"]].size(1)
mask_fcn_logits = Conv2d(dim_reduced, len(cityscapes_cats), 1, 1, 0)
nn.init.constant_(mask_fcn_logits.bias, 0)
nn.init.kaiming_normal_(
mask_fcn_logits.weight, mode="fan_out", nonlinearity="relu"
)
def _copy_weight(src_weight, dst_weight):
for ix, cat in enumerate(cityscapes_cats):
if cat not in coco_cats:
continue
jx = coco_cats_to_inds[cat]
dst_weight[ix] = src_weight[jx]
return dst_weight
def _copy_bias(src_bias, dst_bias, class_agnostic=False):
if class_agnostic:
return dst_bias
return _copy_weight(src_bias, dst_bias)
m[weight_names["cls_score"]] = _copy_weight(
m[weight_names["cls_score"]], cls_score.weight
)
m[weight_names["bbox_pred"]] = _copy_weight(
m[weight_names["bbox_pred"]], bbox_pred.weight
)
m[weight_names["mask_fcn_logits"]] = _copy_weight(
m[weight_names["mask_fcn_logits"]], mask_fcn_logits.weight
)
m[bias_names["cls_score"]] = _copy_bias(
m[bias_names["cls_score"]], cls_score.bias
)
m[bias_names["bbox_pred"]] = _copy_bias(
m[bias_names["bbox_pred"]], bbox_pred.bias, class_agnostic
)
m[bias_names["mask_fcn_logits"]] = _copy_bias(
m[bias_names["mask_fcn_logits"]], mask_fcn_logits.bias
)
print("f: {}\nout_file: {}".format(f, out_file))
torch.save(m, out_file)
对最后一层进行手术的两种示例：
Download the model from MODEL_ZOO
wget https://download.pytorch.org/models/maskrcnn/e2e_mask_rcnn_X_101_32x8d_FPN_1x.pth
// Load the model in python
python
import torch
model = torch.load("e2e_mask_rcnn_X_101_32x8d_FPN_1x.pth ")
// Remove the previous training parameters.
del model['iteration']
del model['scheduler']
del model['optimizer']
// Remove the output layers in COCO, these are the mismatched layers you saw.
//Second stage prediction
del model["model"]["module.roi_heads.box.predictor.cls_score.weight"]
del model["model"]["module.roi_heads.box.predictor.cls_score.bias"]
del model["model"]["module.roi_heads.box.predictor.bbox_pred.weight"]
del model["model"]["module.roi_heads.box.predictor.bbox_pred.bias"]
//mask prediction
del model["model"]["module.roi_heads.mask.predictor.mask_fcn_logits.weight"]
del model["model"]["module.roi_heads.mask.predictor.mask_fcn_logits.bias"]
// RPN
del model["model"]["module.rpn.head.cls_logits.weight"]
del model["model"]["module.rpn.head.cls_logits.bias"]
del model["model"]["module.rpn.head.bbox_pred.weight"]
del model["model"]["module.rpn.head.bbox_pred.bias"]
//save the model
torch.save(model, "modified_model.pth")
Then use modified_model.pth in your MODEL.WEIGHT
def delete_net_weights_for_finetune(
model_file,
out_file,
rpn_final_convs=False,
bbox_final_fcs=True,
mask_final_conv=True
):
del_keys = []
checkpoint = torch.load(model_file)
print("keys: {}".format(checkpoint.keys()))
m = checkpoint['model']
if rpn_final_convs:
# 'module.rpn.anchor_generator.cell_anchors.0',
# 'module.rpn.anchor_generator.cell_anchors.1',
# 'module.rpn.anchor_generator.cell_anchors.2',
# 'module.rpn.anchor_generator.cell_anchors.3',
# 'module.rpn.anchor_generator.cell_anchors.4'
# 'module.rpn.head.cls_logits.weight',
# 'module.rpn.head.cls_logits.bias',
# 'module.rpn.head.bbox_pred.weight',
# 'module.rpn.head.bbox_pred.bias',
del_keys.extend([
k for k in m.keys() if k.find("rpn.anchor_generator") is not -1
])
del_keys.extend([
k for k in m.keys() if k.find("rpn.head.cls_logits") is not -1
])
del_keys.extend([
k for k in m.keys() if k.find("rpn.head.bbox_pred") is not -1
])
if bbox_final_fcs:
# 'module.roi_heads.box.predictor.cls_score.weight',
# 'module.roi_heads.box.predictor.cls_score.bias',
# 'module.roi_heads.box.predictor.bbox_pred.weight',
# 'module.roi_heads.box.predictor.bbox_pred.bias',
del_keys.extend([
k for k in m.keys() if k.find(
"roi_heads.box.predictor.cls_score"
) is not -1
])
del_keys.extend([
k for k in m.keys() if k.find(
"roi_heads.box.predictor.bbox_pred"
) is not -1
])
if mask_final_conv:
# 'module.roi_heads.mask.predictor.mask_fcn_logits.weight',
# 'module.roi_heads.mask.predictor.mask_fcn_logits.bias',
del_keys.extend([
k for k in m.keys() if k.find(
"roi_heads.mask.predictor.mask_fcn_logits"
) is not -1
])
for k in del_keys:
print("del k: {}".format(k))
del m[k]
# checkpoint['model'] = m
print("f: {}\nout_file: {}".format(f, out_file))
recursively_mkdirs(os.path.dirname(out_file))
torch.save({"model": m}, out_file)