Controlling the global step in TrainResult.log and EvalResult.log

I can find a way to control the global step of logging metrics to TrainResult using the log (or log_dict) functions? What is the proper way to use these functions so the logged metrics will show in tensorboard once per epoch? Currently, the steps showing on my tensorboard are (63*i - 1) for i=1 and so on. This is my training_step function (validation_step is similar using pl.EvalResult)

def training_step(self, batch, batch_idx) -> pl.TrainResult:
    x, mask = batch
    pred = self(x)
    loss = self.loss_function(pred, mask)
    result = pl.TrainResult(loss)
    result.log("Trainer/cross_entropy_loss", self.cross_entropy_loss(pred, mask))
    return result

I tried to set most of TrainResult.log parameters manually (like on_epoch, logger, sync_dist, reduce_fx and so on)

• OS: Ubuntu 18.04.4 Nvidia driver version: 440.33.01 CUDA versions available: cuda-10.0 cuda-10.1 cuda-10.2 cuda-9.0 cuda-9.2 Default CUDA version is 10.0
• torch==1.5.1
• pytorch-lightning==0.9.0

Thank you a lot in advance

[ydcjeff]

Hi @shalgi-beyond setting trainer = pl.Trainer(gpus=1, log_save_interval=1, row_log_interval=1) would do the trick. Since I am not quite familiar with tensorboard, I have created the question on the forum for you https://forums.pytorchlightning.ai/t/log-save-interval-and-row-log-interval/135

[stale[bot]]

This issue has been automatically marked as stale because it hasn’t had any recent activity. This issue will be closed in 7 days if no further activity occurs. Thank you for your contributions, Pytorch Lightning Team!