def bufferize(worker: AbstractWorker, state: "State") -> StatePB:
"""
Serialize the State to Protobuf message
"""
protobuf_state = StatePB()
protobuf_placeholders = [
sy.serde.protobuf.serde._bufferize(worker, placeholder)
for placeholder in state.state_placeholders
]
protobuf_state.placeholders.extend(protobuf_placeholders)
state_tensors = []
for tensor in state.tensors():
protobuf_tensor = sy.serde.protobuf.serde._bufferize(
worker, tensor)
state_tensor = StateTensorPB()
if type(protobuf_tensor) == ParameterPB:
state_tensor.torch_param.CopyFrom(
sy.serde.protobuf.serde._bufferize(worker, tensor))
else:
state_tensor.torch_tensor.CopyFrom(
sy.serde.protobuf.serde._bufferize(worker, tensor))
state_tensors.append(state_tensor)
protobuf_state.tensors.extend(state_tensors)
return protobuf_state
def _object2proto(self) -> StatePB:
"""Returns a protobuf serialization of self.
As a requirement of all objects which inherit from Serializable,
this method transforms the current object into the corresponding
Protobuf object so that it can be further serialized.
:return: returns a protobuf object
:rtype: ObjectWithID_PB
.. note::
This method is purely an internal method. Please use object.serialize() or one of
the other public serialization methods if you wish to serialize an
object.
"""
proto = StatePB()
protobuf_placeholders = [
serialize(placeholder) for placeholder in self.state_placeholders
]
proto.placeholders.extend(protobuf_placeholders)
state_tensors = []
for tensor in self.tensors():
state_tensor = StateTensorPB()
state_tensor.torch_tensor.CopyFrom(serialize_tensor(tensor))
state_tensors.append(state_tensor)
proto.tensors.extend(state_tensors)
return proto
def unserialize_model_params(bin: bin):
"""Unserializes model or checkpoint or diff stored in db to list of tensors"""
state = StatePB()
state.ParseFromString(bin)
worker = sy.VirtualWorker(hook=None)
state = protobuf.serde._unbufferize(worker, state)
model_params = state.tensors()
return model_params
async def main():
#auth_token = jwt.encode({}, private_key, algorithm='RS256').decode('ascii')
#print(auth_token)
auth_request = {
"type": "model-centric/authenticate",
"data": {
"model_name": name,
"model_version": version,
#"auth_token": auth_token,
}
}
print('Auth request: ', json.dumps(auth_request, indent=2))
auth_response = await sendWsMessage(auth_request)
print('Auth response: ', json.dumps(auth_response, indent=2))
cycle_request = {
"type": "model-centric/cycle-request",
"data": {
"worker_id": auth_response['data']['worker_id'],
"model": name,
"version": version,
"ping": 1,
"download": 10000,
"upload": 10000,
}
}
cycle_response = await sendWsMessage(cycle_request)
print('Cycle response:', json.dumps(cycle_response, indent=2))
worker_id = auth_response['data']['worker_id']
request_key = cycle_response['data']['request_key']
model_id = cycle_response['data']['model_id']
training_plan_id = cycle_response['data']['plans']['training_plan']
# Model
req = requests.get(
f"http://{gridAddress}/model-centric/get-model?worker_id={worker_id}&request_key={request_key}&model_id={model_id}"
)
model_data = req.content
pb = StatePB()
pb.ParseFromString(req.content)
model_params_downloaded = protobuf.serde._unbufferize(
hook.local_worker, pb)
print("Params shapes:",
[p.shape for p in model_params_downloaded.tensors()])
# Plan "list of ops"
req = requests.get(
f"http://{gridAddress}/model-centric/get-plan?worker_id={worker_id}&request_key={request_key}&plan_id={training_plan_id}&receive_operations_as=list"
)
pb = PlanPB()
pb.ParseFromString(req.content)
plan_ops = protobuf.serde._unbufferize(hook.local_worker, pb)
print(plan_ops.code)
print(plan_ops.torchscript)
# Plan "torchscript"
req = requests.get(
f"http://{gridAddress}/model-centric/get-plan?worker_id={worker_id}&request_key={request_key}&plan_id={training_plan_id}&receive_operations_as=torchscript"
)
pb = PlanPB()
pb.ParseFromString(req.content)
plan_ts = protobuf.serde._unbufferize(hook.local_worker, pb)
print(plan_ts.code)
print(plan_ts.torchscript.code)
# Plan "tfjs"
req = requests.get(
f"http://{gridAddress}/model-centric/get-plan?worker_id={worker_id}&request_key={request_key}&plan_id={training_plan_id}&receive_operations_as=tfjs"
)
pb = PlanPB()
pb.ParseFromString(req.content)
plan_tfjs = protobuf.serde._unbufferize(hook.local_worker, pb)
print(plan_tfjs.code)
},
}
cycle_response = sendWsMessage(cycle_request)
print("Cycle response:", json.dumps(cycle_response, indent=2))
worker_id = auth_response["data"]["worker_id"]
request_key = cycle_response["data"]["request_key"]
model_id = cycle_response["data"]["model_id"]
training_plan_id = cycle_response["data"]["plans"]["training_plan"]
# Let's download Model and Training Plan (in various trainslations) and check they are actually workable.
req = requests.get(
f"http://{config.GRID_ADDRESS}/model-centric/get-model?worker_id={worker_id}&request_key={request_key}&model_id={model_id}"
)
model_data = req.content
pb = StatePB()
pb.ParseFromString(req.content)
model_params_downloaded = protobuf.serde._unbufferize(hook.local_worker, pb)
print("Params shapes:", [p.shape for p in model_params_downloaded.tensors()])
# Plan "list of ops"
req = requests.get(
f"http://{config.GRID_ADDRESS}/model-centric/get-plan?worker_id={worker_id}&request_key={request_key}&plan_id={training_plan_id}&receive_operations_as=list"
)
pb = PlanPB()
pb.ParseFromString(req.content)
plan_ops = protobuf.serde._unbufferize(hook.local_worker, pb)
print(plan_ops.code)
print(plan_ops.torchscript)
# Plan "torchscript"