search

Nathan-Li123 / SMOTer

[ECCV 2024] Beyond MOT: Semantic Multi-Object Tracking
Apache License 2.0
35 stars 0 forks source link

公布的cfg文件或者checkpoint文件存在问题 #3

Closed sove45 closed 3 months ago

sove45 commented 3 months ago

作者您好,我遵循github上的教程采样CH_FPN_1x.pth

其中cfg文件内容为 # BASE: "Base-CenterNet.yaml" MODEL: META_ARCHITECTURE: "BYTERCNN" CENTERNET: USE_DEFORMABLE: True NMS_TH_TEST: 0.6 ROI_HEADS: NAME: GRiTROIHeads PROPOSAL_APPEND_GT: False NO_BOX_HEAD: True WEIGHTS: '/15342518312/MOT/wegt/CH_FPN_1x.pth' ASSO_ON: True ASSO_HEAD: NEG_UNMATCHED: True ASSO_THRESH: 0.3 ASSO_THRESH_TEST: 0.55 ASSO_WEIGHT: 1.0 NUM_WEIGHT_LAYERS: 0 NO_POS_EMB: True SOLVER: USE_CUSTOM_SOLVER: True MAX_ITER: 70000 LR_SCHEDULER_NAME: "WarmupCosineLR" BASE_LR: 0.00005 OPTIMIZER: ADAMW IMS_PER_BATCH: 4 CLIP_GRADIENTS: ENABLED: True CLIP_TYPE: "full_model" CLIP_VALUE: 0.1 NORM_TYPE: 2.0 BACKBONE_MULTIPLIER: 0.1

CUSTOM_MULTIPLIER: 2.0

CUSTOM_MULTIPLIER_NAME: ["summary_text_decoder", "classifier"]

INPUT: FORMAT: RGB CUSTOM_AUG: EfficientDetResizeCrop TRAIN_SIZE: 1280 NOT_CLAMP_BOX: True TEST_INPUT_TYPE: square TEST_SIZE: 1560 VIDEO: TRAIN_LEN: 100 TEST_LEN: 100 VIDEO_INPUT: True VIDEO_TEST: MAX_CENTER_DIST: 1.0 OVERLAP_THRESH: 0.2 DECAY_TIME: 0.9 WITH_IOU: True NOT_MULT_THRESH: True DATASETS:

TRAIN: ("mot17_halftrain","crowdhuman_train")

TRAIN: ("bensmot_train",) TEST: ("bensmot_val",) FIND_UNUSED_PARAM: True DATALOADER: SAMPLER_TRAIN: "MultiDatasetSampler" SOURCE_AWARE: True DATASET_RATIO: [1] 采样的ckeckpoint文件为公开的CH_FPN_1x.pth 但是在测试过程中出现了大量的 WARNING [08/12 07:39:52 fvcore.common.checkpoint]: The checkpoint state_dict contains keys that are not used by the model: WARNING [08/12 07:39:52 fvcore.common.checkpoint]: Some model parameters or buffers are not found in the checkpoint: 警告,这说明cfg中设置的模型与ckpt文件的模型不相同, 相反在我根据此cfg文件重新训练的模型ckpt文件的时候并没有出现这种情况,这进一步说明了cfg文件与ckpt文件不匹配

sove45 commented 3 months ago

同样,在利用ckpt文件与公开的cfg文件进行测试的过程中还出现了输出的results, summary, caption, relation均为空集的情况

sove45 commented 3 months ago

同样,在利用ckpt文件与公开的cfg文件进行测试的过程中还出现了输出的results, summary, caption, relation均为空集的情况

我猜测这和cfg文件与ckpt文件不匹配有关

Nathan-Li123 commented 3 months ago

这是因为CH_FPN_1x.pth只是训练时的用到的只包含部分参数的预训练权重，而不是完整的权重文件。

sove45 commented 3 months ago

这是因为CH_FPN_1x.pth只是训练时的用到的只包含部分参数的预训练权重，而不是完整的权重文件。

好的,非常感谢您的帮助我将带上这些预训练权重来进行训练