amirassov / kaggle-imaterialist Goto Github PK

It is a wonderful work! Can you provide the agmentation code. I don't know how to increase data augmentation in mmdetection. Waiting for your reply. Thank you for your help.
Thanks!

hi, congratulations on the kaggle win!

just wanted to ask if you can release the trained model, as I dont have the resources to train the model myself.

cheers

I want to use your TTA code. Where can I find it ?

I do not find it in the official edition mmdetection

#!/usr/bin/env bash

mkdir /dumps
wget https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/htc/htc_dconv_c3-c5_mstrain_400_1400_x101_64x4d_fpn_20e_20190408-0e50669c.pth -O /dumps/htc_dconv_c3-c5_mstrain_400_1400_x101_64x4d_fpn_20e_20190408-0e50669c.pth
python /kaggle-imaterialist/src/prune.py
--weights=/dumps/htc_dconv_c3-c5_mstrain_400_1400_x101_64x4d_fpn_20e_20190408-0e50669c.pth
--output=/dumps/htc_dconv_c3-c5_mstrain_400_1400_x101_64x4d_fpn_20e_20190408-0e50669c_prune.pth

The link in line 4 is not working i.e. https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/htc/htc_dconv_c3-c5_mstrain_400_1400_x101_64x4d_fpn_20e_20190408-0e50669c.pth is not working. So please provide an alternative.

Looks like configuration was changed.

I have a quick question about 'parameter tuning'

Parameter tuning:
After the 12th epoch with the default parameters, the metric on LB was 0.21913. Next, I tuned postprocessing thresholds using validation data:

rcnn=dict(
score_thr=0.5,
nms=dict(type='nms', iou_thr=0.3),
max_per_img=100,
mask_thr_binary=0.45
)

Does it mean that you use this for 'hard_overlaps_suppression' ?
def hard_overlaps_suppression(binary_mask, scores)

1. only keep mask whose mask_thr > 0.45
2. number of binary_mask is 100