Update to minimal working version, so many 'features'

.dockerignore

@@ -1,5 +1,4 @@
 venv
-default_models
 data_loaders
 data/cifar-10-batches-py
 data/cifar-100-python.tar.gz

charts/templates/client-slow.yaml

@@ -31,11 +31,11 @@ spec:
         - name: PYTHONUNBUFFERED
           value: "1"
         - name: RANK
-          value: {{ $rank }}
+          value: {{ quote $rank }}
         - name: TP_SOCKET_IFNAME
-          value: eth0
+          value: "eth0"
         - name: WORLD_SIZE
-          value: {{ $worldsize }}
+          value: {{ quote $worldsize }}
       image: gcr.io/cs4290-dml/fltk:latest
       name: client-slow
       resources:

charts/templates/fl-server-pod.yaml

@@ -20,9 +20,9 @@ spec:
         - name: RANK
           value: "0"
         - name: TP_SOCKET_IFNAME
-          value: eth0
+          value: "eth0"
         - name: WORLD_SIZE
-          value: {{ (.Values.fltk.worldsize | int) }}
+          value: {{ quote (.Values.fltk.worldsize | int) }}
       image: gcr.io/cs4290-dml/fltk:latest
       name: federation-lab-server
       ports:

configs/cloud_experiment.yaml

@@ -1,10 +1,11 @@
 ---
 # Experiment configuration
+# Experiment configuration
 total_epochs: 5
 epochs_per_cycle: 1
 wait_for_clients: true
-net: Cifar10CNN
-dataset: cifar10
+net: FashionMNISTCNN
+dataset: fashion-mnist
 sampler: "uniform"
 sampler_args:
   - 0.5     # p degree
@@ -14,20 +15,22 @@ cuda: false
 experiment_prefix: 'experiment_single_machine'
 output_location: 'output'
 tensor_board_active: true
-clients_per_round: 1
+clients_per_round: 20
 system:
   federator:
-    hostname: '10.128.0.2'
-    nic: 'ens4'
+    # Use the SERVICE provided by the fl-server to connect
+    hostname: 'fl-server.default.svc.cluster.local'
+    # Default NIC is eth0
+    nic: 'eth0'
   clients:
-    # TODO set this in environment variable
-    amount: 1
+    amount: 20
 # For a simple config is provided in configs/poison.example.yaml
 poison:
-  experiments:
-    min_ratio: 0.0
-    max_ratio: 0.25
-    steps: 2
+  seed: 420
+  ratio: 0.2
   attack:
-    scenario: "fraction"
-    type: "LABEL_FLIP"
+    type: "flip"
+    config:
+      - 4: 6
+      - 6: 4
+

configs/local_experiment.yaml

@@ -1,7 +1,7 @@
 ---
 # Experiment configuration
 # Experiment configuration
-total_epochs: 5
+total_epochs: 2
 epochs_per_cycle: 1
 wait_for_clients: true
 net: Cifar10CNN
@@ -15,13 +15,13 @@ cuda: false
 experiment_prefix: 'experiment_single_machine'
 output_location: 'output'
 tensor_board_active: true
-clients_per_round: 1
+clients_per_round: 2
 system:
   federator:
     hostname: '172.18.0.2'
     nic: 'eth0'
   clients:
-    amount: 1
+    amount: 2
 # For a simple config is provided in configs/poison.example.yaml
 poison:
   seed: 420

docker-compose-gcloud.yml

@@ -25,7 +25,7 @@ services:
       - MASTER_PORT=5000
       - PYTHONUNBUFFERED=1
       - RANK=1
-      - WORLD_SIZE=2
+      - WORLD_SIZE=20
       - GLOO_SOCKET_IFNAME=eth0
       - TP_SOCKET_IFNAME=eth0
     depends_on:
@@ -44,7 +44,26 @@ services:
       - MASTER_PORT=5000
       - PYTHONUNBUFFERED=1
       - RANK=2
-      - WORLD_SIZE=2
+      - WORLD_SIZE=20
+      - GLOO_SOCKET_IFNAME=eth0
+      - TP_SOCKET_IFNAME=eth0
+    depends_on:
+      - fl_server
+    deploy:
+      resources:
+        limits:
+          cpus: '0.5'
+          memory: 1024M
+
+  client_slow_3:
+    restart: 'no'
+    image: gcr.io/cs4290-dml/fltk:latest
+    command: python3 -m fltk single configs/local_experiment.yaml --rank=3
+    environment:
+      - MASTER_PORT=5000
+      - PYTHONUNBUFFERED=1
+      - RANK=3
+      - WORLD_SIZE=20
       - GLOO_SOCKET_IFNAME=eth0
       - TP_SOCKET_IFNAME=eth0
     depends_on:
@@ -54,3 +73,24 @@ services:
         limits:
           cpus: '0.5'
           memory: 1024M
+
+  client_slow_4:
+    restart: 'no'
+    image: gcr.io/cs4290-dml/fltk:latest
+    command: python3 -m fltk single configs/local_experiment.yaml --rank=3
+    environment:
+      - MASTER_PORT=5000
+      - PYTHONUNBUFFERED=1
+      - RANK=4
+      -
+      - WORLD_SIZE=20
+      - GLOO_SOCKET_IFNAME=eth0
+      - TP_SOCKET_IFNAME=eth0
+    depends_on:
+      - fl_server
+    deploy:
+      resources:
+        limits:
+          cpus: '0.5'
+          memory: 1024M
+

fltk/__main__.py

@@ -78,7 +78,8 @@ def main():
             yaml_data = yaml.load(config_file, Loader=yaml.FullLoader)
             cfg.merge_yaml(yaml_data)
             if args.mode == 'poison':
-                perform_poison_experiment(args, cfg, parser, yaml_data)
+                for ratio in [0.0, 0.05, 0.1, 0.15, 0.2]:
+                    perform_poison_experiment(args, cfg, parser, yaml_data, ratio)
             elif args.mode == 'single':
                 perform_single_experiment(args, cfg, parser, yaml_data)
             else:
@@ -106,7 +107,7 @@ def perform_single_experiment(args, cfg, parser, yaml_data):
 
 
 
-def perform_poison_experiment(args, cfg, parser, yaml_data):
+def perform_poison_experiment(args, cfg, parser, yaml_data, ratio=None):
     """
     Function to start poisoned experiment.
     """
@@ -129,6 +130,9 @@ def perform_poison_experiment(args, cfg, parser, yaml_data):
     if not nic:
         nic = yaml_data['system']['federator']['nic']
     print(f'rank={args.rank}, world_size={world_size}, host={master_address}, args=cfg, nic={nic}')
+    if ratio:
+        print(f'Setting ratio to {ratio}')
+        attack.ratio = ratio
     run_single(rank=args.rank, world_size=world_size, host=master_address, args=cfg, nic=nic, attack=attack)
 
 

fltk/client.py

@@ -80,7 +80,32 @@ def init_device(self):
         else:
             return torch.device("cpu")
 
+    def reset_model(self):
+        """
+        Function to reset the learning process. In addition, reset the loss function and the
+        optimizer, in case this uses certain decay according to some internal counter.
+        @return: None
+        @rtype: None
+        """
+        # Load the default model
+        self.set_net(self.load_default_model())
+        self.loss_function = self.args.get_loss_function()()
+        self.optimizer = torch.optim.SGD(self.net.parameters(),
+                                         lr=self.args.get_learning_rate(),
+                                         momentum=self.args.get_momentum())
+        self.scheduler = MinCapableStepLR(self.args.get_logger(), self.optimizer,
+                                          self.args.get_scheduler_step_size(),
+                                          self.args.get_scheduler_gamma(),
+                                          self.args.get_min_lr())
+        # Set the epoch counter
+        self.epoch_counter = 0
+
     def ping(self):
+        """
+        Aliveness checker for the federator during initialization.
+        @return: String to the important question, `ping?', which is pong.
+        @rtype: str
+        """
         return 'pong'
 
     def rpc_test(self):
@@ -300,7 +325,7 @@ def run_epochs(self, num_epoch, pill: PoisonPill = None):
         # Copy GPU tensors to CPU
         for k, v in weights.items():
             weights[k] = v.cpu()
-        return data, weights
+        return data, self.net.state_dict()
 
     def save_model(self, epoch, suffix):
         """

fltk/datasets/distributed/cifar10.py

@@ -1,24 +1,24 @@
+import logging
+
+from torch.utils.data import DataLoader
 from torchvision import datasets
 from torchvision import transforms
-from torch.utils.data import DataLoader, DistributedSampler
 
 from fltk.datasets.distributed.dataset import DistDataset
 from fltk.strategy.data_samplers import get_sampler
-import logging
-
 from fltk.util.poison.poisonpill import PoisonPill
 
 
 class DistCIFAR10Dataset(DistDataset):
 
     def __init__(self, args, pill: PoisonPill = None):
         super(DistCIFAR10Dataset, self).__init__(args, pill)
+        self.get_args().get_logger().debug(f"Instantiated CIFAR10 train data, with pill: {pill}")
         self.init_train_dataset()
         self.init_test_dataset()
-        if pill:
-            self.ingest_pill(pill)
 
-    def init_train_dataset(self, pill: PoisonPill = None):
+
+    def init_train_dataset(self):
         dist_loader_text = "distributed" if self.args.get_distributed() else ""
         self.get_args().get_logger().debug(f"Loading '{dist_loader_text}' CIFAR10 train data")
         normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
@@ -28,12 +28,11 @@ def init_train_dataset(self, pill: PoisonPill = None):
             transforms.ToTensor(),
             normalize
         ])
+
         self.train_dataset = datasets.CIFAR10(root=self.get_args().get_data_path(), train=True, download=True,
-                                         transform=transform)
+                                              transform=transform,
+                                              target_transform=None if not self.pill else self.pill.poison_targets())
 
-        # Poison
-        if pill:
-            self.ingest_pill(pill)
         self.train_sampler = get_sampler(self.train_dataset, self.args)
         self.train_loader = DataLoader(self.train_dataset, batch_size=16, sampler=self.train_sampler)
         logging.info("this client gets {} samples".format(len(self.train_sampler)))
@@ -46,8 +45,8 @@ def init_test_dataset(self):
             transforms.ToTensor(),
             normalize
         ])
+        # TODO: decide on whether to poison test or not target_transform=None if not self.pill else self.pill.poison_targets()
         self.test_dataset = datasets.CIFAR10(root=self.get_args().get_data_path(), train=False, download=True,
-                                        transform=transform)
+                                             transform=transform)
         self.test_sampler = get_sampler(self.test_dataset, self.args)
         self.test_loader = DataLoader(self.test_dataset, batch_size=16, sampler=self.test_sampler)
-

fltk/datasets/distributed/cifar100.py

@@ -24,7 +24,7 @@ def init_train_dataset(self):
             normalize
         ])
         self.train_dataset = datasets.CIFAR100(root=self.get_args().get_data_path(), train=True, download=True,
-                                              transform=transform)
+                                              transform=transform, target_transform=self.pill.poison_targets())
         self.train_sampler = get_sampler(self.train_dataset, self.args)
         self.train_loader = DataLoader(self.train_dataset, batch_size=16, sampler=self.train_sampler)
 

fltk/datasets/distributed/dataset.py

@@ -139,14 +139,3 @@ def init_test_dataset(self):
 	# 	:return: tuple
 	# 	"""
 	# 	return (next(iter(data_loader))[0].numpy(), next(iter(data_loader))[1].numpy())
-	def ingest_pill(self, pill: PoisonPill):
-		"""
-		Drink the CoolAid, apply poison to the input regarding the pill. Note that the pill may implement a noop,
-		meaning that this has no real result.
-		@param pill:
-		@type pill:
-		@return:
-		@rtype:
-		"""
-		self.train_dataset.targets = pill.poison_targets(self.train_dataset.targets)
-		self.test_dataset.targets = pill.poison_targets(self.test_dataset)

fltk/datasets/distributed/fashion_mnist.py

@@ -1,15 +1,16 @@
-from fltk.datasets.distributed import DistDataset
+from torch.utils.data import DataLoader
 from torchvision import datasets
 from torchvision import transforms
-from torch.utils.data import DataLoader, DistributedSampler
 
+from fltk.datasets.distributed import DistDataset
 from fltk.strategy.data_samplers import get_sampler
+from fltk.util.poison.poisonpill import PoisonPill
 
 
 class DistFashionMNISTDataset(DistDataset):
 
-    def __init__(self, args):
-        super(DistFashionMNISTDataset, self).__init__(args)
+    def __init__(self, args, pill: PoisonPill = None):
+        super(DistFashionMNISTDataset, self).__init__(args, pill)
         self.init_train_dataset()
         self.init_test_dataset()
 
@@ -18,22 +19,26 @@ def init_train_dataset(self):
         self.get_args().get_logger().debug(f"Loading '{dist_loader_text}' Fashion MNIST train data")
 
         self.train_dataset = datasets.FashionMNIST(root=self.get_args().get_data_path(), train=True, download=True,
-                                              transform=transforms.Compose([transforms.ToTensor()]))
+                                                   transform=transforms.Compose([transforms.ToTensor()]),
+                                                   target_transform=None if not self.pill else self.pill.poison_targets())
         self.train_sampler = get_sampler(self.train_dataset, self.args)
         self.train_loader = DataLoader(self.train_dataset, batch_size=16, sampler=self.train_sampler)
 
     def init_test_dataset(self):
         dist_loader_text = "distributed" if self.args.get_distributed() else ""
         self.get_args().get_logger().debug(f"Loading '{dist_loader_text}' Fashion MNIST test data")
         self.test_dataset = datasets.FashionMNIST(root=self.get_args().get_data_path(), train=False, download=True,
-                                             transform=transforms.Compose([transforms.ToTensor()]))
+                                                  transform=transforms.Compose([transforms.ToTensor()]),
+                                                  target_transform=None if not self.pill else self.pill.poison_targets())
         self.test_sampler = get_sampler(self.test_dataset, self.args)
         self.test_loader = DataLoader(self.test_dataset, batch_size=16, sampler=self.test_sampler)
 
     def load_train_dataset(self):
         self.get_args().get_logger().debug("Loading Fashion MNIST train data")
 
-        train_dataset = datasets.FashionMNIST(self.get_args().get_data_path(), train=True, download=True, transform=transforms.Compose([transforms.ToTensor()]))
+        train_dataset = datasets.FashionMNIST(self.get_args().get_data_path(), train=True, download=True,
+                                              transform=transforms.Compose([transforms.ToTensor()]),
+                                              target_transform=None if not self.pill else self.pill.poison_targets())
         train_loader = DataLoader(train_dataset, batch_size=len(train_dataset))
 
         train_data = self.get_tuple_from_data_loader(train_loader)
@@ -45,7 +50,9 @@ def load_train_dataset(self):
     def load_test_dataset(self):
         self.get_args().get_logger().debug("Loading Fashion MNIST test data")
 
-        test_dataset = datasets.FashionMNIST(self.get_args().get_data_path(), train=False, download=True, transform=transforms.Compose([transforms.ToTensor()]))
+        # TODO: target_transform=None if not self.pill else self.pill.poison_targets()
+        test_dataset = datasets.FashionMNIST(self.get_args().get_data_path(), train=False, download=True,
+                                             transform=transforms.Compose([transforms.ToTensor()]))
         test_loader = DataLoader(test_dataset, batch_size=len(test_dataset))
 
         test_data = self.get_tuple_from_data_loader(test_loader)