batch packing for LM

pytext/data/data.py

@@ -190,6 +190,7 @@ class Data(Component):
     """
 
     __COMPONENT_TYPE__ = ComponentType.DATA_HANDLER
+    __EXPANSIBLE__ = True
 
     class Config(Component.Config):
         #: Specify where training/test/eval data come from. The default value

pytext/metric_reporters/language_model_metric_reporter.py

@@ -67,8 +67,12 @@ def add_batch_stats(
         self.batch_size.append(len(m_input[0]))
         self.aggregate_data(self.all_num_tokens, targets[1])
         now = time.time()
-        total_tokens = float(sum(targets[2]))
-        print(f"Tokens/s: {total_tokens / (now - self.time)}, ppl: {math.exp(loss)}")
+        if not n_batches % 1000:
+            total_tokens = float(sum(targets[2]))
+            print(
+                f"Tokens/s: {total_tokens / (now - self.time):.0f}, ppl: {math.exp(loss):.2f}",
+                flush=True,
+            )
         self.time = now
 
     def _reset(self):

pytext/utils/precision.py

@@ -115,8 +115,11 @@ def backward(optimizer, loss):
         else:
             # 1. Use automatic loss scaling to best use fp16 range
             # 2. Clear handle's cache of casted parameters
-            with optimizer.scale_loss(loss) as scaled_loss:
-                scaled_loss.backward()
+            if loss > 0:
+                with optimizer.scale_loss(loss) as scaled_loss:
+                    scaled_loss.backward()
+            else:
+                loss.backward()
     else:
         loss.backward()
 

pytext/workflow.py

@@ -102,7 +102,7 @@ def prepare_task(
 
     print("\nParameters: {}\n".format(config))
     _set_cuda(config.use_cuda_if_available, device_id, world_size)
-    _set_fp16(config.use_fp16 and world_size == 1)
+    _set_fp16(config.use_fp16)
     if config.random_seed is not None:
         set_random_seeds(config.random_seed)