Improve spacy pretrain (#4393)

* Support bilstm_depth arg in spacy pretrain

* Add option to ignore zero vectors in get_cossim_loss

* Use cosine loss in Cloze multitask

spacy/_ml.py

@@ -953,16 +953,24 @@ class CharacterEmbed(Model):
         return output, backprop_character_embed
 
 
-def get_cossim_loss(yh, y):
+def get_cossim_loss(yh, y, ignore_zeros=False):
+    xp = get_array_module(yh)
+    # Find the zero vectors
+    if ignore_zeros:
+        zero_indices = xp.abs(y).sum(axis=1) == 0
     # Add a small constant to avoid 0 vectors
     yh = yh + 1e-8
     y = y + 1e-8
     # https://math.stackexchange.com/questions/1923613/partial-derivative-of-cosine-similarity
-    xp = get_array_module(yh)
     norm_yh = xp.linalg.norm(yh, axis=1, keepdims=True)
     norm_y = xp.linalg.norm(y, axis=1, keepdims=True)
     mul_norms = norm_yh * norm_y
     cosine = (yh * y).sum(axis=1, keepdims=True) / mul_norms
     d_yh = (y / mul_norms) - (cosine * (yh / norm_yh ** 2))
-    loss = xp.abs(cosine - 1).sum()
+    losses = xp.abs(cosine - 1)
+    if ignore_zeros:
+        # If the target was a zero vector, don't count it in the loss.
+        d_yh[zero_indices] = 0
+        losses[zero_indices] = 0
+    loss = losses.sum()
     return loss, -d_yh

spacy/cli/pretrain.py

@@ -35,6 +35,7 @@ from .train import _load_pretrained_tok2vec
     output_dir=("Directory to write models to on each epoch", "positional", None, str),
     width=("Width of CNN layers", "option", "cw", int),
     depth=("Depth of CNN layers", "option", "cd", int),
+    bilstm_depth=("Depth of BiLSTM layers (requires PyTorch)", "option", "lstm", int),
     embed_rows=("Number of embedding rows", "option", "er", int),
     loss_func=(
         "Loss function to use for the objective. Either 'L2' or 'cosine'",
@@ -80,6 +81,7 @@ def pretrain(
     output_dir,
     width=96,
     depth=4,
+    bilstm_depth=2,
     embed_rows=2000,
     loss_func="cosine",
     use_vectors=False,
@@ -116,6 +118,10 @@ def pretrain(
     util.fix_random_seed(seed)
 
     has_gpu = prefer_gpu()
+    if has_gpu:
+        import torch
+
+        torch.set_default_tensor_type("torch.cuda.FloatTensor")
     msg.info("Using GPU" if has_gpu else "Not using GPU")
 
     output_dir = Path(output_dir)
@@ -151,7 +157,7 @@ def pretrain(
             embed_rows,
             conv_depth=depth,
             pretrained_vectors=pretrained_vectors,
-            bilstm_depth=0,  # Requires PyTorch. Experimental.
+            bilstm_depth=bilstm_depth,  # Requires PyTorch. Experimental.
             cnn_maxout_pieces=3,  # You can try setting this higher
             subword_features=True,  # Set to False for Chinese etc
         ),

spacy/pipeline/pipes.pyx

@@ -29,7 +29,7 @@ from .._ml import Tok2Vec, build_tagger_model, cosine, get_cossim_loss
 from .._ml import build_text_classifier, build_simple_cnn_text_classifier
 from .._ml import build_bow_text_classifier, build_nel_encoder
 from .._ml import link_vectors_to_models, zero_init, flatten
-from .._ml import masked_language_model, create_default_optimizer
+from .._ml import masked_language_model, create_default_optimizer, get_cossim_loss
 from ..errors import Errors, TempErrors, user_warning, Warnings
 from .. import util
 
@@ -880,8 +880,7 @@ class ClozeMultitask(Pipe):
         # and look them up all at once. This prevents data copying.
         ids = self.model.ops.flatten([doc.to_array(ID).ravel() for doc in docs])
         target = vectors[ids]
-        gradient = (prediction - target) / prediction.shape[0]
+        loss, gradient = get_cossim_loss(prediction, target, ignore_zeros=True)
-        loss = (gradient**2).sum()
         return float(loss), gradient
 
     def update(self, docs, golds, drop=0., sgd=None, losses=None):