docs(framework) Add how-to guide for designing stateful ClientApp objects

[jafermarq]

Shows how to make use of the Context's state to store metrics and model parameters.

Content formatted with https://github.com/dzhu/rstfmt

[Dawitkiros]

Hi,

I am not sure if I should create a new issue or comment here. I discussed with @jafermarq on how to implement a stateful client on Slack, and he gave me great tips and a link to this repository to provide my feedback on whether everything is working as expected. I followed the steps provided to store the client parameters in the context, like the following in the fit function right after training:

# Save the parameters after training to the context for the next round
p_record = ParametersRecord()
for k, v in self.net.state_dict().items():
    # Convert to NumPy, then to Array. Add to record
    p_record[k] = ndarray_to_array(v.cpu().numpy())
# Add to context
self.client_state.parameters_records["prev_round"] = p_record

I loaded the context in set_parameters as follows:

# Here, I am hoping to get the previous state dict from the client_state
prev_state_dict = {}
# Extract record from context
p_record = self.client_state.parameters_records["prev_round"]
# Deserialize arrays
for k, v in p_record.items():
    prev_state_dict[k] = torch.from_numpy(np.copy(basic_array_deserialization(v)))
prev_state_dict_list = segment_resnet_parameters(flatten_resnet_parameters(prev_state_dict))

Everything is working as expected, but I encountered the following warning:

UserWarning: The given NumPy array is not writable, and PyTorch does not support non-writable tensors. This means writing to this tensor will result in undefined behavior. You may want to copy the array to protect its data or make it writable before converting it to a tensor. This type of warning will be suppressed for the rest of this program.

I found a way to make it work using np.copy(), though I'm not sure if it's the right solution.

Thanks.

[jafermarq]

Hi @Dawitkiros , sorry for the delay. Could you try this:

from flwr.common import  array_from_numpy
p_record = ParametersRecord()
for k, v in self.net.state_dict().items():
    # Convert to NumPy, then to Array. Add to record
    # p_record[k] = ndarray_to_array(v.cpu().numpy())
    p_record[k] = array_from_numpy(v.cpu().numpy()) # use utility method 
    
...

for k, v in p_record.items():
    # prev_state_dict[k] = torch.from_numpy(np.copy(basic_array_deserialization(v)))
    prev_state_dict[k] = torch.from_numpy(v.numpy()) # instead of the line above

This shows how to do it with the built-int serialization from numpy to Array. Then, we can revert such process with the .numpy() method Array objects have.

flower/src/py/flwr/common/record/parametersrecord.py

Line 59 in b857425

def numpy(self) -> NDArray:

[Dawitkiros]

Hi @jafermarq,

Your suggestion works!

[jafermarq]

Hi @jafermarq,

Your suggestion works!

amazing! thanks for confirming. If you have some suggestions on how to improve this doc page. We are very happy to incorporate your feedback. We might be merging this PR soon.