Scalar writing frequency too high will lead to data loss?

[chivee]

Hi @lanpa ,

Thanks for your great work. I really enjoy the tensorboardx journey.
However, the data loss issues are haunting me, which was that if the Summary writer writing too many scalars. there will be a data loss issue.

do you have any comment on this?

[lanpa]

I just saw a bug report on missing text event. What did you mean by too many and how did you spot the missing value? It will be good to provide some code for me to debug.

[chivee]

import torch
import random
import torchvision.models as models
from tensorboardX import SummaryWriter

writer = SummaryWriter()
writer_list = []

for i in range(10):
    writer_list.append(SummaryWriter('test/{0}'.format(i)))

writer_item = []
for i in range(10):
    writer_item.append("data/test_name_{0}".format(i))

for n_iter in range(100000):
    for i in range(10):
        writer = writer_list[i]
        for j in range(10):
            rnd_value = random.uniform(0.9,1.1)
            writer.add_scalar('{0}'.format(writer_item[j]), rnd_value, n_iter)

# export scalar data to JSON for external processing
writer.export_scalars_to_json("./all_scalars.json")
writer.close()

[chivee]

@lanpa these code could reproduce the issue

[lanpa]

I didn't run the code, but I guess the data points should be there. see #44

[lanpa]

Confirms that it should working and the size of event file is 53M per writer, which implies 53B per each record, reasonable.