minor fix and doc for cross backend fl

federatedscope/core/auxiliaries/trainer_builder.py

@@ -45,7 +45,8 @@ def get_trainer(model=None,
                                           only_for_eval=only_for_eval,
                                           monitor=monitor)
         elif config.backend == 'tensorflow':
-            from federatedscope.core.trainers import GeneralTFTrainer
+            from federatedscope.core.trainers.tf_trainer import \
+                GeneralTFTrainer
             trainer = GeneralTFTrainer(model=model,
                                        data=data,
                                        device=device,

federatedscope/core/proto/gRPC_comm_manager_pb2_grpc.py

@@ -2,7 +2,7 @@
 """Client and server classes corresponding to protobuf-defined services."""
 import grpc
 
-import federatedscope.core.proto.gRPC_comm_manager_pb2 \
+from federatedscope.core.proto import gRPC_comm_manager_pb2 \
     as gRPC__comm__manager__pb2
 
 

federatedscope/core/workers/client.py

@@ -104,8 +104,15 @@ def __init__(self,
         else:
             self.comp_speed = None
             self.comm_bandwidth = None
-        self.model_size = sys.getsizeof(pickle.dumps(
-            self.model)) / 1024.0 * 8.  # kbits
+
+        if self._cfg.backend == 'torch':
+            self.model_size = sys.getsizeof(pickle.dumps(
+                self.model)) / 1024.0 * 8.  # kbits
+        else:
+            # TODO: calculate model size for TF Model
+            self.model_size = 1.0
+            logger.warning(f'The calculation of model size in backend:'
+                           f'{self._cfg.backend} is not provided.')
 
         # Initialize communication manager
         self.server_id = server_id

federatedscope/core/workers/server.py

@@ -764,8 +764,15 @@ def trigger_for_start(self):
                     for client_index in np.arange(1, self.client_num + 1)
                 ]
             else:
-                model_size = sys.getsizeof(pickle.dumps(
-                    self.model)) / 1024.0 * 8.
+                if self._cfg.backend == 'torch':
+                    model_size = sys.getsizeof(pickle.dumps(
+                        self.model)) / 1024.0 * 8.
+                else:
+                    # TODO: calculate model size for TF Model
+                    model_size = 1.0
+                    logger.warning(f'The calculation of model size in backend:'
+                                   f'{self._cfg.backend} is not provided.')
+
                 client_resource = [
                     model_size / float(x['communication']) +
                     float(x['computation']) / 1000.

federatedscope/cross_backends/README.md

@@ -0,0 +1,30 @@
+## Cross-Backend Federated Learning
+
+We provide an example for constructing cross-backend (Tensorflow and PyTorch) federated learning, which trains an LR model on the synthetic toy data.
+
+The server runs with Tensorflow, and clients run with PyTorch (the suggested version of Tensorflow is 1.12.0):
+```shell script
+# Generate toy data
+python ../../scripts/distributed_scripts/gen_data.py
+# Server
+python ../main.py --cfg distributed_tf_server.yaml
+
+# Clients
+python ../main.py --cfg ../../scripts/distributed_scripts/distributed_configs/distributed_client_1.yaml
+python ../main.py --cfg ../../scripts/distributed_scripts/distributed_configs/distributed_client_2.yaml
+python ../main.py --cfg ../../scripts/distributed_scripts/distributed_configs/distributed_client_3.yaml
+```
+
+One of the client runs with Tensorflow, and the server and other clients run with PyTorch:
+```shell script
+# Generate toy data
+python ../../scripts/distributed_scripts/gen_data.py
+# Server
+python ../main.py --cfg ../../scripts/distributed_scripts/distributed_configs/distributed_server.yaml
+
+# Clients with Pytorch
+python ../main.py --cfg ../../scripts/distributed_scripts/distributed_configs/distributed_client_1.yaml
+python ../main.py --cfg ../../scripts/distributed_scripts/distributed_configs/distributed_client_2.yaml
+# Clients with Tensorflow
+python ../main.py --cfg distributed_tf_client_3.yaml
+```

federatedscope/cross_backends/__init__.py

@@ -0,0 +1,4 @@
+from federatedscope.cross_backends.tf_lr import LogisticRegression
+from federatedscope.cross_backends.tf_aggregator import FedAvgAggregator
+
+__all__ = ['LogisticRegression', 'FedAvgAggregator']

federatedscope/cross_backends/distributed_tf_client_3.yaml

@@ -0,0 +1,24 @@
+use_gpu: False
+backend: 'tensorflow'
+federate:
+  client_num: 3
+  mode: 'distributed'
+  total_round_num: 20
+  make_global_eval: False
+  online_aggr: False
+distribute:
+  use: True
+  server_host: '127.0.0.1'
+  server_port: 50051
+  client_host: '127.0.0.1'
+  client_port: 50054
+  role: 'client'
+  data_file: 'toy_data/client_3_data'
+trainer:
+  type: 'general'
+eval:
+  freq: 10
+data:
+  type: 'toy'
+model:
+  type: 'lr'

federatedscope/cross_backends/distributed_tf_server.yaml

@@ -0,0 +1,22 @@
+use_gpu: False
+backend: 'tensorflow'
+federate:
+  client_num: 3
+  mode: 'distributed'
+  total_round_num: 20
+  make_global_eval: True
+  online_aggr: False
+distribute:
+  use: True
+  server_host: '127.0.0.1'
+  server_port: 50051
+  role: 'server'
+  data_file: 'toy_data/server_data'
+trainer:
+  type: 'general'
+eval:
+  freq: 10
+data:
+  type: 'toy'
+model:
+  type: 'lr'