Interesting behaviour/interaction between matplotlib and validation_epoch_end(self, outputs)

[nick-torenvliet]

I am modifying some code taken out of the lightning docs and training a VAE to learn a represenation of time series thought to contain structure. The validation outputs are reconstructions of the time series. The following code is written with the intention of grabbing a random output and plotting it.

Two things happen that I don't understand.

1 - It seems the plt object from the previous epoch is still in memory and the new data is plotted alongside the old. I have tried deleting the data explicitly but to no avail...
2 - Somehow both the input and the output are getting plotted (in the attached plot you see the learned noise, and the well formed actual singal I am trying to reconstruct. I have no idea where the input signal is coming from - is it embedded in the output tensor somewhere?

def validation_epoch_end(self, outputs):
    if not self.save_images:
        return
    if not os.path.exists(self.save_path):
        os.makedirs(self.save_path)
    choice = random.choice(outputs)
    output_sample = random.choice(choice[0])
    plt.plot(output_sample.cpu())
    plt.savefig(f"{self.save_path}/epoch_{self.current_epoch}.png")
    print('this is def interesting')

[semaphore-egg]

Creating a new figure before plotting?

fig, ax = plt.subplots()
ax.set_xlim(0, 10)
ax.plot([1, 5], [0, 10])

[nick-torenvliet]

Yeah - that will work... I'm just suprised the plot object, which is only referenced from inside a
function call, is persistent across function calls.