Implement loss aware sparsifier

pytext/optimizer/sparsifiers/blockwise_sparsifier.py

@@ -91,6 +91,28 @@ def from_config(cls, config: Config):
             config.layerwise_pruning,
         )
 
+    def get_sparsifiable_params(self, model, requires_name=False):
+        sparsifiable_params = [
+            p
+            for n, p in model.named_parameters()
+            if p.requires_grad and len(p.shape) == 2
+        ]
+        sparsifiable_params_name = [
+            n
+            for n, p in model.named_parameters()
+            if p.requires_grad and len(p.shape) == 2
+        ]
+        if requires_name:
+            return sparsifiable_params_name, sparsifiable_params
+        else:
+            return sparsifiable_params
+
+    def get_current_sparsity(self, model):
+        sparsifiable_params = self.get_sparsifiable_params(model)
+        sparsifiable_params_count = sum(p.numel() for p in sparsifiable_params)
+        nonzero_params = sum(p.nonzero().size(0) for p in sparsifiable_params)
+        return (sparsifiable_params_count - nonzero_params) / sparsifiable_params_count
+
     def _padding_into_full_blocks(self, param):
         nrows, ncols = param.shape
         ncols_pad = math.ceil(ncols / self.block_size) * self.block_size

pytext/trainers/trainer.py

@@ -96,6 +96,9 @@ class Config(ConfigBase):
         target_time_limit_seconds: Optional[int] = None
         #: Whether to do evaluation and model selection based on it.
         do_eval: bool = True
+        #: if do_eval, do we load the best model state dict after training or just
+        # use the latest model state
+        load_best_model_after_train: bool = True
         #: Number of samples for logging training progress.
         num_samples_to_log_progress: int = 1000
         #: Number of forward & backward per batch before update gradients, the
@@ -465,7 +468,11 @@ def train_from_state(
             if should_update_model or train_config.save_all_checkpoints:
                 self.save_checkpoint(state, train_config)
         # Only bother loading the best model for master worker
-        if rank == 0 and state.best_model_state is not None:
+        if (
+            rank == 0
+            and state.best_model_state is not None
+            and self.config.load_best_model_after_train
+        ):
             self.load_best_model(state)
 
         return state.model, state.best_model_metric