Saving/restoring latent

[KiudLyrl]

it works by dumping/restoring the whole EMA and ADAM optimiser objects

exemple of how to restore :

model = Imagine(
text = TEXT,
save_every = SAVE_EVERY,
lr = LEARNING_RATE,
iterations = ITERATIONS,
save_progress = SAVE_PROGRESS,
out_folder = out_folder,
save_latents = True,
saved_latents_filepath = r"F:\my\path\magic_man.35.backup",
)

[wolfgangmeyers]

can you also update cli.py so the parameter can be passed through the command prompt?

[wolfgangmeyers]

also should add dill to requirements.txt

[wolfgangmeyers]

I ran this and hit the following error:

>>> import big_sleep
>>> a = big_sleep.Imagine(num_cutouts=32, iterations=50, epochs=5, save_latents=True, text="a blue glass orb", save_every=5)
>>> a()
Imagining "a_blue_glass_orb" ...
loss: -17.00:   8%|██████████████████                                                                                                                                                                                                                | 4/50 [00:01<00:15,  2.91it/s]
      epochs:   0%|                                                                                                                                                                                                                                           | 0/5 [00:01<?, ?it/s]
Traceback (most recent call last):                                                                                                                                                                                                                 | 1/50.0 [00:01<01:07,  1.38s/it]
  File "<stdin>", line 1, in <module>
  File "C:\Python39\lib\site-packages\torch\nn\modules\module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "C:\Users\wolfg\Documents\big-sleep\big_sleep\big_sleep.py", line 526, in forward
    out, loss = self.train_step(epoch, i, image_pbar)
  File "C:\Users\wolfg\Documents\big-sleep\big_sleep\big_sleep.py", line 500, in train_step
    dill.dump(current_state_backup, file = open(f'./{self.text_path}.{num}{self.seed_suffix}.backup', "wb"))
UnboundLocalError: local variable 'num' referenced before assignment

[wolfgangmeyers]

I suggest moving num above this if block so it can be used if save_progress hasn't been set to True

[wolfgangmeyers]

So I've tried loading saved latents and changing the text to something completely different, but regardless of the text the image seems to develop in the exact same way. This was "a blue glass orb" after 5 epochs of 50 iterations each:

After reloading the latents and running the same epochs and iterations again with text="a red fox" and text_min="a blue glass orb" this is the result:

Seems to change in the same way regardless of the prompt though.

[wolfgangmeyers]

It would be nice to be able to resume with the saved latents either with the same optimizer (continue the same image but refine it) or with a new one (transform the existing image into something different)

[wolfgangmeyers]

I've added the suggested changes and submitted a second PR - #89