small-changes

spacy/cli/ray_param_server.py

@@ -5,11 +5,13 @@ from wasabi import msg
 from .. import util
 
 class OptimizerWorker:
-    def __init__(self, config_path):
+    def __init__(self, config_path, world_size, sync=True):
         self.optimizer = _create_optimizer(config_path)
         self.weights_dict = {}
+        self.world_size = world_size
+        self.sync = sync
 
-    def call(self, key, weights, gradient, *, lr_scale=1.0):
+    def call(self, rank, key, weights, gradient, *, lr_scale=1.0):
         if key not in self.weights_dict:
             self.weights_dict[key] = weights.copy()
         new_weights, new_grads = self.optimizer(
@@ -26,18 +28,19 @@ class OptimizerWorker:
 class RayOptimizer:
     local_optimizer = None
 
-    def __init__(self, config_path, use_gpu):
+    def __init__(self, config_path, use_gpu, rank):
         RemoteOptimizer = ray.remote(OptimizerWorker)
         if use_gpu >= 0:
             RemoteOptimizer = RemoteOptimizer.options(num_gpus=0.1)
         self.optimizer = RemoteOptimizer.remote(config_path)
+        self.rank = rank
         self.sync()
 
     def sync(self):
         self.local_optimizer = ray.get(self.optimizer.fetch.remote())
 
     def __call__(self, *args, **kwargs):
-        weights, grads = ray.get(self.optimizer.call.remote(*args, **kwargs))
+        weights, grads = ray.get(self.optimizer.call.remote(self.rank, *args, **kwargs))
         return weights.copy(), grads.copy()
 
     def __getattr__(self, name):

spacy/cli/ray_utils.py

@@ -69,7 +69,7 @@ class AllreduceOptimizer:
             weights = self.allreduce(weights) / self.communicator.size()
 
 
-        gradient = self.allreduce(gradient)
+        gradient = self.allreduce(gradient) / self.communicator.size()
         flat_weights, gradient = self.optimizer(key, weights, gradient, lr_scale=lr_scale)
         return flat_weights, gradient
 

spacy/cli/train_from_config.py

@@ -198,7 +198,7 @@ def train_cli(
 
     if num_workers and num_workers > 1:
         import ray
-        ray.init()
+        ray.init(address="auto")
         if strategy == "ps":
             from spacy.cli.ray_param_server import RayOptimizer
             remote_train = ray.remote(setup_and_train)
@@ -401,8 +401,12 @@ def train(
     msg.info(f"Training. Initial learn rate: {optimizer.learn_rate}")
     print_row = setup_printer(training, nlp)
 
+    tqdm_args = dict(total=training["eval_frequency"], leave=False)
+    global world_rank
+    if world_rank is not None:
+        tqdm_args["disable"] = bool(world_rank != 0)
     try:
-        progress = tqdm.tqdm(total=training["eval_frequency"], leave=False)
+        progress = tqdm.tqdm(**tqdm_args)
         for batch, info, is_best_checkpoint in training_step_iterator:
             progress.update(1)
             if is_best_checkpoint is not None:
@@ -411,7 +415,7 @@ def train(
                 if is_best_checkpoint and output_path is not None:
                     update_meta(training, nlp, info)
                     nlp.to_disk(output_path / "model-best")
-                progress = tqdm.tqdm(total=training["eval_frequency"], leave=False)
+                progress = tqdm.tqdm(**tqdm_args)
             # Clean up the objects to faciliate garbage collection.
             for eg in batch:
                 eg.doc = None
@@ -437,6 +441,10 @@ def train(
 
 def create_train_batches(nlp, corpus, cfg):
     epochs_todo = cfg.get("max_epochs", 0)
+    if world_rank is not None:
+        for i in range(world_rank):
+            # Increment random seed
+            random.random()
     while True:
         train_examples = list(
             corpus.train_dataset(
@@ -452,16 +460,18 @@ def create_train_batches(nlp, corpus, cfg):
             raise ValueError(Errors.E988)
         random.shuffle(train_examples)
 
-        if world_size is not None:
+        # # TODO: with large batches, this can be bad.
-            # Taken from https://github.com/pytorch/pytorch/blob/master/torch/utils/data/distributed.py
+        # if world_size is not None:
-            num_samples = int(math.ceil(len(train_examples) * 1.0 / world_size))
+        #     # Taken from https://github.com/pytorch/pytorch/blob/master/torch/utils/data/distributed.py
-            total_size = num_samples * world_size  # expected to overflow
+        #     num_samples = int(math.ceil(len(train_examples) * 1.0 / world_size))
-            train_examples += train_examples[:(total_size - len(train_examples))]
+        #     total_size = num_samples * world_size  # expected to overflow
-            assert len(train_examples) == total_size
+        #     train_examples += train_examples[:(total_size - len(train_examples))]
+        #     assert len(train_examples) == total_size
 
-            # subsample
+        #     # subsample
-            train_examples = train_examples[world_rank:total_size:world_size]
+        #     train_examples = train_examples[world_rank:total_size:world_size]
-            assert len(train_examples) == num_samples
+        #     assert len(train_examples) == num_samples
+        #     print(f"Reset epoch: Only using {num_samples} out of {total_size} samples")
 
         batches = util.minibatch_by_words(
             train_examples,
@@ -474,7 +484,7 @@ def create_train_batches(nlp, corpus, cfg):
             yield first
         except StopIteration:
             raise ValueError(Errors.E986)
-        for batch in batches:
+        for i, batch in enumerate(batches):
             yield batch
         epochs_todo -= 1
         # We intentionally compare exactly to 0 here, so that max_epochs < 1