feymanpriv / dolg Goto Github PK

您好，很感谢您提供了这么棒的工作。我最近在尝试复现您的工作，但是我遇到了一些问题，希望能得到您的回复。
关于您仓库中提供的 Roxford 和 Rpairs 数据集的 pickle 文件，我尝试使用 Roxford 和 Rpairs 数据集的图像和您提供的模型权重进行复现，并测试结果。

这是我用您的模型文件和您提供的roxford_R101_512_crop_test5.pickle文件在 Roxford 上的测试结果。

二者在 E M H 上mAP 都有不小的 drop。
所以想问下您，您提供的pkl文件是用您放出的模型权重 [[R-50-DOLG] 或者 [R-101-DOLG]生成的吗？
此外，我发现您提供的模型权重并未启用 ASPP 多尺度空洞卷积模块，期待您的回复。

@feymanpriv Thanks for your great work! When I fine-tune the DOLG model on my customized dataset using your model weight, the desc_top1_err and desc_top5_err are always 100. I am not sure if there is any mistake on my side, you can find my attached config below:

MODEL:
  TYPE: resnet
  DEPTH: 101
  NUM_CLASSES: 3847
  HEADS:
    IN_FEAT: 2048
    REDUCTION_DIM: 512
    MARGIN: 0.15
    SCALE: 30
RESNET:
  TRANS_FUN: bottleneck_transform
  NUM_GROUPS: 1
  WIDTH_PER_GROUP: 64
  STRIDE_1X1: False
BN:
  ZERO_INIT_FINAL_GAMMA: True
OPTIM:
  BASE_LR: 0.01
  LR_POLICY: cos
  STEPS: [0, 30, 60, 90]
  LR_MULT: 0.1
  MAX_EPOCH: 100
  MOMENTUM: 0.9
  NESTEROV: True
  WEIGHT_DECAY: 0.0001
  WARMUP_EPOCHS: 5
TRAIN:
  DATASET: GSV_imgs_bldg_v1
  SPLIT: GSV_imgs_bldg_v1_train_stratify.txt
  BATCH_SIZE: 36
  IM_SIZE: 224
  EVAL_PERIOD: 100
TEST:
  DATASET: GSV_imgs_bldg_v1
  SPLIT: GSV_imgs_bldg_v1_val_stratify.txt
  BATCH_SIZE: 36
  IM_SIZE: 256
NUM_GPUS: 6
DATA_LOADER:
  NUM_WORKERS: 4 
CUDNN:
  BENCHMARK: True
OUT_DIR: ./GSV_imgs_bldg_v1_output

and the training command is:

python train.py --cfg configs/resnet101_delg_4gpu_GSV.yaml OUT_DIR ./GSV_imgs_bldg_v1_output NUM_GPUS 6 TRAIN.BATCH_SIZE 36 TEST.BATCH_SIZE 36 PORT 13005 TRAIN.WEIGHTS ./weights/r101_dolg_512.pyth

Could you please help me with it? I really appreciate it.

@feymanpriv
Hello, thanks for your contribution.
But when I use the train.py of this rep to train model, it shows the log:

...
[trainer.py: 170]: Start epoch: 1
[meters.py: 155]: json_stats: {"_type": "train_iter", "desc_loss": NaN, "desc_top1_err": 100.0000, "desc_top5_err": 100.0000, "epoch": "1/100", "eta": "03,07:41:39", "iter": "10/625", "lr": 0.0001, "mem": 10812, "time_avg": 4.5911, "time_diff": 3.7708}
[meters.py: 155]: json_stats: {"_type": "train_iter", "desc_loss": NaN, "desc_top1_err": 100.0000, "desc_top5_err": 100.0000, "epoch": "1/100", "eta": "03,00:24:35", "iter": "20/625", "lr": 0.0001, "mem": 10812, "time_avg": 4.1721, "time_diff": 3.7448}
[meters.py: 155]: json_stats: {"_type": "train_iter", "desc_loss": NaN, "desc_top1_err": 100.0000, "desc_top5_err": 100.0000, "epoch": "1/100", "eta": "02,20:39:31", "iter": "30/625", "lr": 0.0001, "mem": 10812, "time_avg": 3.9566, "time_diff": 3.3038}
[meters.py: 155]: json_stats: {"_type": "train_iter", "desc_loss": NaN, "desc_top1_err": 100.0000, "desc_top5_err": 100.0000, "epoch": "1/100", "eta": "02,17:26:54", "iter": "40/625", "lr": 0.0001, "mem": 10812, "time_avg": 3.7723, "time_diff": 2.9340}
...

Why is the "desc_loss" still NaN? Can you give some advises?

Besides, I use your https://github.com/feymanpriv/DELG to train my same dataset, the value of "desc_loss" is normal.
Looking forward to your reply. Thanks

RT

ReadMe里面让安装requirements.txt
但是为啥工程里面没这个文件？

错误堆栈为：

Traceback (most recent call last):
File "evaler/infer.py", line 119, in main_multicard
data = extract(im, model)
File "evaler/infer.py", line 62, in extract
global_feature = model(input_data)
File "/home/tiger/workspace/anaconda/envs/DOLG/lib/python3.6/site-packages/torch/nn/modules/module.py", line 532, in call
result = self.forward(*input, **kwargs)
TypeError: forward() missing 1 required positional argument: 'targets'

Hello, I want to know how to download the oxford5k and the rparis6k dataset, which contain the images and labels.

paper 中提到训练的时候， 图像最终resize 到512x512，但是工程中是224x224.

工程中给出的weight 是在512x512下训练的，还是在224x224下的？

Hi, thanks for sharing your re-implementation of DELG in pytorch, that helps me a lot!
By the way, is the performance of DELG-R101(GLDv2-clean) in Table 1 of your paper obtained via query crop or not? I find that query crop actually decreased the test performance on RParis6k and ROxford5k.

按照代码里的配置完成了训练，发现过程中学习率最高只有BASE_LR参数设定的0.01，对比论文4.1节中提及的initial learning rate为0.05，请问是否应将BASE_LR修改为0.05？论文所说的initial learning rate是指warm up前第一个epoch的学习率，还是warm up后第6个epoch的最大学习率？

Hi,

I like your work and I am interested in testing the pre-trained models, but I seem not able to download them from Baidu via the provided links. Is it possible to share the models weights some other way?

Thanks, Tomas