Logs not syncing after switching from tensorboardx to torch.utils.tensorboard

Metrics for my run are not showing up: https://app.wandb.ai/yaroslavvb/test-graphs_test/runs/lbuht9lu?workspace=user-yaroslavvb

Train Epoch: 100 [44928/60000 (75%)]	Loss: 7.943522, Accuracy: 0/44928 (0.00%), Elapsed Time: 225.7s
Train Epoch: 100 [51328/60000 (86%)]	Loss: 7.655457, Accuracy: 0/51328 (0.00%), Elapsed Time: 225.9s
Train Epoch: 100 [57728/60000 (96%)]	Loss: 6.832315, Accuracy: 0/57728 (0.00%), Elapsed Time: 226.1s

wandb: Waiting for W&B process to finish, PID 98062
wandb: Program ended successfully.
wandb: Syncing files in wandb/run-20190821_231347-lbuht9lu:
wandb:   upstream_diff_692e4cc15b4b6acd32f627215194be23445e9d69.patch
wandb: plus 7 W&B file(s) and 0 media file(s)
wandb:                                                                                
wandb: Synced result: https://app.wandb.ai/yaroslavvb/test-graphs_test/runs/lbuht9lu

Process finished with exit code 0

Debug log – debug.log

My current work-around is to use local tensorboard to visualize the run, it looks something like this:

[ira7bar]

I have experienced the same problem (using from torch.utils.tensorboard import SummaryWriter) and solved it by first initializing wandb and only then initializing the tensorboard writer. There was no need for manual patching before wandb init.

[ariG23498]

I tried reproducing the problem with the following code snippet:

import wandb
from tensorboardX import SummaryWriter
wandb.tensorboard.patch(root_logdir="exp")
wandb.init(entity="repro", project="tensorboard")
writer = SummaryWriter("exp")
scalars = {
    'scalar_1': 1.0,
    'scalar_2': 2.0,
    'scalar_3': 3.0
}
writer.add_scalars("test", scalars, global_step=1)
writer.close()
wandb.finish()

Somehow wandb is unable to patch the tensorboard to the logdir I suggest here.

On the other hand I tried modifying the code and used torch.utils.tensorboard.SummaryWriter()

import wandb
from torch.utils.tensorboard import SummaryWriter
wandb.tensorboard.patch(root_logdir="exp")
wandb.init(entity="repro", project="tensorboard")
writer = SummaryWriter("exp")
scalars = {
    'scalar_1': 1.0,
    'scalar_2': 2.0,
    'scalar_3': 3.0
}
writer.add_scalars("test", scalars, global_step=1)
writer.close()
wandb.finish()

This worked fine for me. My best guesses would be some complication with the tensorboardX integration.