def write_weight_statitsics(writer: torch.utils.tensorboard.SummaryWriter,
module: torch.nn.Module, epoch: int):
# log the weights-norm for the parameters of the model
for weight_name in module.state_dict().keys():
w = module.state_dict()[weight_name]
norm = w.norm().item()
writer.add_scalar(f'Norm/{weight_name}', norm, epoch)
avg = w.abs().mean().item()
writer.add_scalar(f'avg/{weight_name}', avg, epoch)
writer.add_histogram(f'hist/{weight_name}', w, epoch)
def log_dict_with_writer(y_true: torch.Tensor,
y_pred: torch.Tensor,
summary_writer: torch.utils.tensorboard.SummaryWriter,
thr=None,
ood_label=0,
global_step=None):
"""
Log metrics to tensorboard with summary writer
:param y_true: true labels of the objects, shape=(N,)
:param torch.Tensor y_pred: logits, predictions of the model BEFORE the softmax function,
shape=(N, n_classes + 1). Note that ood label is not the one on which we train.
:param SummaryWriter summary_writer: a writer for logging metrics to tensorboard
:param float thr: Value of the maximum probability below which we consider an object as ood
:param ood_label: label which corresponds to an ood object
:return:
"""
metrics_dict = get_metrics_dict(y_true, y_pred, thr, ood_label)
for names, hist in metrics_dict["hist"].items():
summary_writer.add_histogram(names + str(global_step), hist)
for names, scalars in metrics_dict["scalar"].items():
summary_writer.add_histogram(names, scalars)