Stable diffusion uses an Exponential Moving Average of the model's weights to improve quality of resulting images and avoid overfitting to the most recently trained images. It also gives us a more stable EMA validation loss to use to pick the best checkpoints. See Understanding the use of EMA in Diffusion models on reddit for some discussion.
I want to skip to the code
This optimization, along with some others, are in this fork of InvokeAI: github:LunNova/InvokeAI-nyoom.
What is an Exponential Moving Average anyway?
An EMA is a moving average which can be calculated only by knowing the last EMA value, and the current value. The EMA doesn't require retaining the last N data points, making it quite memory efficient.
def next_ema(val: float, last_ema: float, decay: float) -> float:
return last_ema * decay + val * (1-decay)
You've probably seen this before, even if you didn't know the name. It's a simple idea and has been independently reinvented many times! I can remember using this sort of average as far back as 2013.
How do you implement an EMA for a machine learning model?
It's very similar, except rather than single floats as inputs you have to do this for every model parameter that you want to be part of the EMA. With PyTorch modules you can use the named_parameters() iterator to access all parameters.
EMA in Stable Diffusion
Stable Diffusion includes an implementation of an EMA called LitEma, found at ldm/modules/ema.py.
The implementation is very short, and most of the work involved is managing dicts of tensors.
Where are the EMA weights kept?
In this implementation, a device is never explicitly set, so the EMA weights end up on the GPU like everything else. This doubles the memory required to store the model parameters!
How can we move them to the CPU?
Unfortunately, Pytorch Lightning tries quite hard to put buffers on the GPU automatically. This is normally nice, and part of "hardware agnostic training".
Some other users have ran into similar issues.
We can get around this by avoiding using register_buffers and manually creating Tensors on the CPU, and storing them inside a list to avoid Lightning detecting the tensor.
self.cpu_shadow_params = [{}] if ema_on_cpu else None
for name, p in model.named_parameters():
if p.requires_grad:
# remove as '.'-character is not allowed in buffers
s_name = name.replace('.', '')
self.m_name2s_name.update({name: s_name})
cloned = p.detach().clone().data
cloned.requires_grad = False
# Added this check here
if ema_on_cpu:
self.cpu_shadow_params[0][s_name] = cloned.cpu()
else:
self.register_buffer(s_name, cloned)
Let's try it
After making the above change, and a few more bodges further down to copy tensors on/off the GPU and handle restoring checkpoints with the same state_dict keys, it works!
Unfortunately it's now taking over 2 seconds per batch because copying tensors from the GPU to CPU is slow.
This is too slow! We need a workaround
Yep, it's taking an extra second per batch to update on the CPU. Saving VRAM isn't that much use if we're halving the speed.
Fortunately for us, in our particular context we don't need to update the weights on every single batch. Machine learning model weights drift slowly, so we can apply an EMA update only on every Nth batch and still get mostly the same results, as long as N isn't set too high!
We'll also need to set the initial decay value to math.pow(decay, update_every) so we still make a big enough change.
What's going on with collected_params?
Glad you spotted that too, it does look a little odd.
store saves the current weights of the model in self.collected_params, and restore restores those back to the original model. This allows temporarily using the EMA weights as the main weights, and is used to log images during training. Here's how it gets used in ldm/models/diffusion/ddpm.py.
The original code isn't doing anything with self.collected_params after restoring so it's leaking an entire copy of the model after the first time store gets called!
For bonus points, let's fix that leak and move those tensors onto the CPU.
All together now
Here are our final changes: LunNova/InvokeAI-nyoom ema: move EMA weights to CPU and only update weights every N steps
This reduced training memory usage at a batch size of two from OOMing at 32GB to using only 27GB of VRAM.
Before this change it took 27GB to train a batch size of one so I'm quite happy with it!