def visualize(self,
train_loader,
val_loader,
tensorboard=None,
epoch=None,
**kwargs):
visualizations = {}
# visualize pred
fig, _ = vis.plot_predictions(self, train_loader, key="pred")
visualizations["predictions/pred-train"] = fig
if val_loader is not None:
fig, _ = vis.plot_predictions(self, val_loader, key="pred")
visualizations["predictions/pred-val"] = fig
return visualizations
def visualize(self,
train_loader,
val_loader,
tensorboard=None,
epoch=None,
**kwargs):
visualizations = {}
# gradient norm tensorboard histograms
if tensorboard is not None:
vis.ce_gradient_norm_histogram(self,
train_loader,
tensorboard,
epoch,
name='train-ce-grad')
if val_loader is not None:
vis.ce_gradient_norm_histogram(self,
val_loader,
tensorboard,
epoch,
name='val-ce-grad')
# add gradient pair plots
if train_loader.dataset.dataset_name == 'mnist':
for p in [(0, 1), (4, 9)]:
fig, _ = vis.ce_gradient_pair_scatter(self,
train_loader,
d1=p[0],
d2=p[1])
visualizations['gradients/train-ce-scatter-{}-{}'.format(
p[0], p[1])] = fig
if val_loader is not None:
fig, _ = vis.ce_gradient_pair_scatter(self,
val_loader,
d1=p[0],
d2=p[1])
visualizations['gradients/val-ce-scatter-{}-{}'.format(
p[0], p[1])] = fig
# visualize pred
fig, _ = vis.plot_predictions(self, train_loader, key='pred')
visualizations['predictions/pred-train'] = fig
if val_loader is not None:
fig, _ = vis.plot_predictions(self, val_loader, key='pred')
visualizations['predictions/pred-val'] = fig
return visualizations
def visualize(self,
train_loader,
val_loader,
tensorboard=None,
epoch=None,
**kwargs):
visualizations = super(PredictGradOutput,
self).visualize(train_loader, val_loader,
tensorboard, epoch)
# visualize q_label_pred
fig, _ = vis.plot_predictions(self, train_loader, key='q_label_pred')
visualizations['predictions/q-label-pred-train'] = fig
if val_loader is not None:
fig, _ = vis.plot_predictions(self, val_loader, key='q_label_pred')
visualizations['predictions/q-label-pred-val'] = fig
return visualizations
def visualize(self,
train_loader,
val_loader,
tensorboard=None,
epoch=None,
**kwargs):
visualizations = super(PredictGradOutputFixedFormWithConfusion,
self).visualize(train_loader, val_loader,
tensorboard, epoch)
# visualize the confusion matrix
if tensorboard is not None:
fig, _ = vis.plot_confusion_matrix(
torch.softmax(self.Q_logits, dim=1))
visualizations['confusion-matrix'] = fig
# visualize q_label_pred
fig, _ = vis.plot_predictions(self, train_loader, key='q_label_pred')
visualizations['predictions/q-label-pred-train'] = fig
if val_loader is not None:
fig, _ = vis.plot_predictions(self, val_loader, key='q_label_pred')
visualizations['predictions/q-label-pred-val'] = fig
return visualizations