Finetuning and Config used for HPS

[anonymous-atom]

@mihirp1998
I was tryin to finetune Stable Diffusion 1.5 using your HPS reward function and the hps.sh training script, I used batch size of 1 but still the training seems to get completed very quickly, 50 epochs just took 2-4 minutes.

def from_file(path, low=None, high=None):
    prompts = _load_lines(path)[low:high]
    return random.choice(prompts), {}

And here you are trying to just use batch_size number of prompts ? I am using batch_size of 2 on 1 A100 GPU to test the script.

    def _generate_samples(self, batch_size, with_grad=True, prompts=None):
        """
        Generate samples from the model

        Args:
            batch_size (int): Batch size to use for sampling
            with_grad (bool): Whether the generated RGBs should have gradients attached to it.

        Returns:
            prompt_image_pairs (Dict[Any])
        """
        prompt_image_pairs = {}

        sample_neg_prompt_embeds = self.neg_prompt_embed.repeat(batch_size, 1, 1)

        if prompts is None:
            prompts, prompt_metadata = zip(*[self.prompt_fn() for _ in range(batch_size)])

Your help will mean a lot!

[mihirp1998]

Can u check whether the training is actually getting run ? if not why is it skipping the training loop

[anonymous-atom]

It's working but I think it's only using few examples from the prompt file.

[mihirp1998]

i think random.choice is uniform over all prompts, i'm not sure what's the bug here. If you find it let me know.

[anonymous-atom]

Yeah Sure!

[anonymous-atom]

Also while I was tryin to train with a custom loss function, the models seem to collapse very early unless I adjust the learning rate. is this the expected behaviour ?

[anonymous-atom]

@mihirp1998 Sorry to tag you again, but can you let me know how much time it took per epoch on your 4 A100 GPU's ?

Clear Skies!

[mihirp1998]

On Aesthetics 2-3 minutes per epoch

…

On Sat, Nov 9, 2024 at 12:07 AM Karun ***@***.***> wrote: @mihirp1998 <https://github.com/mihirp1998> Sorry to tag you again, but can you let me know how much time it took per epoch on your 4 A100 GPU's ? Clear Skies! — Reply to this email directly, view it on GitHub <#17 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AE4C5LG33CWOFGYKS3SESDTZ7WJ7TAVCNFSM6AAAAABROVLMA2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDINRWGA2TENJXGM> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[anonymous-atom]

I wanted to confirm are you using all 750 prompts in hps_v2_all.txt file for 1 single epoch ?

[anonymous-atom]

Hi @mihirp1998 , so this is where I am confused:

So you used step() function to do 1 training step/ 1 epoch ?

    def train(self, epochs: Optional[int] = None):
        """
        Train the model for a given number of epochs
        """
        global_step = 0
        if epochs is None:
            epochs = self.config.num_epochs
        for epoch in range(self.first_epoch, epochs):
            global_step = self.step(epoch, global_step)

And here in step() function, it only seems to finetine on num_gpus * batch_size * train_gradient_accumulation_steps number of images, am I missing something ? What if someone used just 1 GPU to train ?

    def step(self, epoch: int, global_step: int):
   
        info = defaultdict(list)
        print(f"Epoch: {epoch}, Global Step: {global_step}")

        self.sd_pipeline.unet.train()

        for _ in range(self.config.train_gradient_accumulation_steps):
            with self.accelerator.accumulate(self.sd_pipeline.unet), self.autocast(), torch.enable_grad():
                prompt_image_pairs = self._generate_samples(
                    batch_size=self.config.train_batch_size,
                )