Merge pull request #492 from FedML-AI/dev/v0.7.0

Dev/v0.7.0

doc/en/starter/installation.md

@@ -11,7 +11,16 @@ FedML supports Linux, MacOS, Windows, and Android.
 ```
 pip install fedml
 ```
-(Note: please use python 3.6, 3.7, 3.8, 3.9. We will support 3.10 systematically in the next iteration.)
+
+The default machine learning engine is `PyTorch`. FedML also supports `TensorFlow`, `Jax`, and `MXNet`.
+You can install related engine as follows:
+```
+pip install "fedml[tensorflow]"
+pip install "fedml[jax]"
+pip install "fedml[mxnet]"
+```
+Note that the commands above only install the CPU version. 
+If you need GPU/TPU version, please follow TensorFlow/Jax/MXNet official guidance.
 
 ## Installing FedML with Anaconda
 

python/fedml/model/model_hub.py

@@ -72,11 +72,14 @@ def create(args, output_dim):
         model = (gen, disc)
     elif model_name == "lenet" and hasattr(args, "deeplearning_backend") and args.deeplearning_backend == "mnn":
         from .mobile.mnn_lenet import create_mnn_lenet5_model
+
         create_mnn_lenet5_model(args.global_model_file_path)
+        model = None  # for server MNN, the model is saved as computational graph and then send it to clients.
     elif model_name == "resnet20" and hasattr(args, "deeplearning_backend") and args.deeplearning_backend == "mnn":
         from .mobile.mnn_resnet import create_mnn_resnet20_model
 
         create_mnn_resnet20_model(args.global_model_file_path)
+        model = None  # for server MNN, the model is saved as computational graph and then send it to clients.
     else:
         raise Exception("no such model definition, please check the argument spelling or customize your own model")
     return model

python/quick_start/beehive/README.md

@@ -22,8 +22,8 @@ brew install android-platform-tools
 
 [https://github.com/FedML-AI/FedML/tree/master/python/quick_start/beehive](https://github.com/FedML-AI/FedML/tree/master/python/quick_start/beehive)
 
-5. For local debugging of cross-device server, please try 
-
+5. For local debugging of cross-device server, please try
+(Cross device server on Windows is not working because MNN lib can't work on Windows correctly, so you should run cross-device server on linux or other OS in which MNN lib can be working correctly)
 ```
 sh run_server.sh
 ```

python/quick_start/beehive/config/fedml_config.yaml

@@ -9,7 +9,7 @@ environment_args:
 
 data_args:
   dataset: "mnist"
-  data_cache_dir: "../../../data/mnist"
+  data_cache_dir: ~/fedml_data
   partition_method: "hetero"
   partition_alpha: 0.5
   train_size: 10000
@@ -18,8 +18,8 @@ data_args:
 model_args:
   model: "lenet"
   deeplearning_backend: "mnn"
-  model_file_cache_folder: "./model_file_cache" # will be filled by the server automatically
-  global_model_file_path: "./model_file_cache/global_model.mnn"
+  model_file_cache_folder: "model_file_cache" # will be filled by the server automatically
+  global_model_file_path: "model_file_cache/global_model.mnn"
 
 train_args:
   federated_optimizer: "FedAvg"

python/quick_start/beehive/my_dataset.py

@@ -5,11 +5,11 @@
 
 
 class MnistDataset(MNN.data.Dataset):
-    def __init__(self, training_dataset=True):
+    def __init__(self, training_dataset=True, root_path="./data"):
         super(MnistDataset, self).__init__()
         self.is_training_dataset = training_dataset
-        trainset = MNIST(root="./data", train=True, download=True)
-        testset = MNIST(root="./data", train=False, download=True)
+        trainset = MNIST(root=root_path, train=True, download=True)
+        testset = MNIST(root=root_path, train=False, download=True)
         if self.is_training_dataset:
             self.data = trainset.data / 255.0
             self.labels = trainset.targets

python/quick_start/beehive/torch_server.py

@@ -1,3 +1,6 @@
+import os
+from os.path import expanduser
+
 import MNN
 
 import fedml
@@ -12,16 +15,16 @@
     device = fedml.device.get_device(args)
 
     # load data
-    train_dataset = MnistDataset(True)
-    test_dataset = MnistDataset(False)
+    train_dataset = MnistDataset(True, root_path=args.data_cache_dir)
+    test_dataset = MnistDataset(False, root_path=args.data_cache_dir)
     train_dataloader = MNN.data.DataLoader(train_dataset, batch_size=64, shuffle=True)
     test_dataloader = MNN.data.DataLoader(
         test_dataset, batch_size=args.batch_size, shuffle=False
     )
     class_num = 10
 
     # load model
-    model = fedml.model.create(args, output_dim=class_num)
+    fedml.model.create(args, output_dim=class_num)
 
     # start training
     server = ServerMNN(

python/setup.py

@@ -54,7 +54,7 @@ def finalize_options(self):
 ]
 
 requirements_extra_jax = [
-    "jax[cuda]",
+    "jax[cpu]",
     "dm-haiku",
     "optax",
     "jaxlib"