Pretrained model for universenet in Nightowls dataset

[rsj007]

Hi，shinya7y!

Is it the pretrained model for universenet in Nightowls dataset available？I can't find the file "load_from = '../data/checkpoints/universenet50_fp16_8x2_lr0001_mstrain_640_1280_7e_waymo_open_20200526_080330/epoch_7_for_nightowls.pth'" in your project.

Thanks a lot！

[shinya7y]

Please follow this Notes, because the file was trained on Waymo Open (and converted using this script).

[rsj007]

Sorry. I have received Registration Confirmed email from Waymo dataset.
Please give me a download link of "universenet50_fp16_8x2_lr0001_mstrain_640_1280_7e_waymo_open_20200526_080330/epoch_7_for_nightowls.pth". Or you can contact with me by email---736720216@qq.com.
Thanks!

[shinya7y]

I will paste some scripts for testing. Please edit them on your own if they are outdated.

Splitting nightowls_test_imageids.json to 10 shards.

import json

json_path = '../data/nightowls/annotations/nightowls_test_imageids.json'  # noqa
output_path_f = '../data/nightowls/annotations/nightowls_test_imageids_shard{:02d}.json'  # noqa

with open(json_path) as f:
    all_info = json.load(f)

print(all_info.keys())
print(all_info['annotations'])
print(all_info['categories'])
print(all_info['poses'])

num_images = len(all_info['images'])
print('num_images:', num_images)

images_list = [[], [], [], [], [], [], [], [], [], []]
for image_idx, image in enumerate(all_info['images']):
    shard_index = int(image_idx * 10 / num_images)
    images_list[shard_index].append(image)

for i in range(10):
    print(len(images_list[i]))
    shard_info = {}
    shard_info['images'] = images_list[i]
    shard_info['categories'] = all_info['categories']
    shard_info['poses'] = all_info['poses']

    output_path = output_path_f.format(i)
    with open(output_path, 'w') as f:
        json.dump(shard_info, f, indent=None)

[shinya7y]

Inference by mmdetection.

for SHARD_ID in $(seq -w 00 09); do
    echo "shard${SHARD_ID}"
    JOB_ID=universenet50_fp16_4x4_lr0001_mstrain_640_1280_from_waymo_7e_nightowls_test
    LOG_DIR=${HOME}/logs/nightowls/universenet50_fp16_4x4_lr0001_mstrain_640_1280_from_waymo_7e_nightowls_20200608_093807
    ./tools/dist_test.sh \
        configs/nightowls/${JOB_ID}.py \
        ${LOG_DIR}/epoch_2.pth \
        4 \
        --cfg-options \
        "data.test.ann_file=data/nightowls/annotations/nightowls_test_imageids_shard${SHARD_ID}.json" \
        "data.test.img_prefix=data/nightowls/nightowls_test/" \
        --options "jsonfile_prefix=${LOG_DIR}/results_test_800_960_flip_epoch_2_shard${SHARD_ID}" \
        --format-only
done

[shinya7y]

Concatenating detections and extracting pedestrian category for submission.

import json


def dump_rounded_json(obj_to_dump, f, rounding=7):
    json.dump(
        json.loads(
            json.dumps(obj_to_dump),
            parse_float=lambda x: round(float(x), rounding)),
        f,
        indent=None,
        sort_keys=False)


json_path_f = '../logs/nightowls/universenet50_fp16_4x4_lr0001_mstrain_640_1280_from_waymo_7e_nightowls_20200608_093807/results_test_800_960_flip_epoch_2_shard{:02d}.bbox.json'  # noqa
output_path = '../logs/nightowls/universenet50_fp16_4x4_lr0001_mstrain_640_1280_from_waymo_7e_nightowls_20200608_093807/results_test_800_960_flip_epoch_2_concat_s.bbox.json'  # noqa
output_path_ped = '../logs/nightowls/universenet50_fp16_4x4_lr0001_mstrain_640_1280_from_waymo_7e_nightowls_20200608_093807/results_test_800_960_flip_epoch_2_concat_s_ped.bbox.json'  # noqa
score_thresh_ped = 0.03

dets = []
for shard_id in range(10):
    with open(json_path_f.format(shard_id)) as f:
        shard_dets = json.load(f)
        print(len(shard_dets))
        dets.extend(shard_dets)
print("dets", len(dets))

dets_ped = []
dets_all = []
for det in dets:
    if det["category_id"] == 1:
        if det["score"] >= score_thresh_ped:
            dets_ped.append(det)
            dets_all.append(det)
    else:
        dets_all.append(det)
print("dets_ped", len(dets_ped))
print("dets_all", len(dets_all))

with open(output_path, mode='w') as f:
    dump_rounded_json(dets_all, f)
with open(output_path_ped, mode='w') as f:
    dump_rounded_json(dets_ped, f)

[rsj007]

Hi,shinya7y!
I have two questions.
1.Why do you spilt the testing into 10 shards? Can I test as a whole?
2.Besides, which code language is about the “for loop” you mentioned MMdetection inference? Whether I will run the code for ten times?

Thanks

[shinya7y]

1. The sharding is mainly for avoiding CPU out of memory. You can first try to test as a whole.
2. bash

I can't support more details since these scripts are not related to the main code.

[rsj007]

Hi,shinya7y！
I get the MR equals -100.0% when I do as you said. I change the testing set path in the base/datasets/nightowls_detection.py.
Do you know what's the problem?

[shinya7y]

No. I cannot support any more.