Reproduce result of usa-airports about gcc HOT 5 CLOSED

Hi @larry2020626 ,

Glad to help you reproduce the result. Please kindly provide the following information:

1. Which model are you evaluating? (moco/e2e, downloaded/pretrained);
2. Your device (cpu/cuda);
3. Your pytorch/dgl version;
4. Your obtained result on airport.

In addition, did you try other datasets/tasks? Please feel free to provide more datapoints/screenshots.

Thanks!

from gcc.

Thanks very much for your reply！

I adopt E2E model, running on CUDA 10.0
My Device:
40 Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz
3b:00.0 3D controller: NVIDIA Corporation Device 1eb8 (rev a1)
5e:00.0 3D controller: NVIDIA Corporation Device 1eb8 (rev a1)
pytorch version: 1.4.0+cu100
dgl version '0.4.3post2'
current result is {'Micro-F1': 0.6218}

here is what I do:

1. to download small.bin pretrain dataset

python scripts/download.py --url https://cloud.tsinghua.edu.cn/f/b37eed70207c468ba367/?dl=1 --path data --fname small.bin

2.to pretrain:

bash scripts/pretrain.sh 0 --batch-size 256

3.For downstream tasks:

python scripts/download.py --url https://cloud.tsinghua.edu.cn/f/2535437e896c4b73b6bb/?dl=1 --path data --fname downstream.tar.gz
bash scripts/generate.sh 0 Pretrain_moco_False_dgl_gin_layer_5_lr_0.005_decay_1e-05_bsz_256_hid_64_samples_2000_nce_t_0.07_nce_k_32_rw_hops_256_restart_prob_0.8_aug_1st_ft_False_deg_16_pos_32_momentum_0.999 usa_airport
bash scripts/node_classification/ours.sh saved/Pretrain_moco_False_dgl_gin_layer_5_lr_0.005_decay_1e-05_bsz_256_hid_64_samples_2000_nce_t_0.07_nce_k_32_rw_hops_256_restart_prob_0.8_aug_1st_ft_False_deg_16_pos_32_momentum_0.999 64 usa_airport

Thanks!
And I will try to use downloaded models to evaluate on usa-airports dataset to check whether it is because my pre-training effect is not good enough.

from gcc.

Hi @larry2020626 ,

Thanks for your reply. The main reason is that your experiments correspond to the GCC (E2E, freeze) model, while "68.3" is the result of GCC (E2E, full). For the difference, please see the Freezing vs. full fine-tuning paragraph at the beginning of page 6.

To obtain the "full" (finetuning) result, please run python train.py --exp FT --model-path saved --tb-path tensorboard --tb-freq 5 --gpu 0 --dataset usa_airport --finetune --epochs 30 --resume saved/Pretrain_moco_False_dgl_gin_layer_5_lr_0.005_decay_1e-05_bsz_256_hid_64_samples_2000_nce_t_0.07_nce_k_32_rw_hops_256_restart_prob_0.8_aug_1st_ft_False_deg_16_pos_32_momentum_0.999/current.pth --cv. This will start finetuning the model on usa_airport on 10 splits, and the final mean/std of the accuracy will be reported. The result on my side is 66.7±4.1.

Besides, you can use batch size 1024 in paper instead of 256 but that will take even longer for pretraining. In that case, the result is 68.3±2.8.

from gcc.

Thanks！@qibinc , really appreciate your help.

from gcc.

Hi Larry, I'm glad it worked for you. Feel free to raise more issues if you encounter other problems. Closing this.

from gcc.