Various updates/additions to CLI scripts (#5362)

* `debug-data`: determine coverage of provided vectors

* `evaluate`: support `blank:lg` model to make it possible to just evaluate
tokenization

* `init-model`: add option to truncate vectors to N most frequent vectors
from word2vec file

* `train`:

  * if training on GPU, only run evaluation/timing on CPU in the first
    iteration

  * if training is aborted, exit with a non-0 exit status

spacy/cli/debug_data.py

@@ -108,9 +108,11 @@ def debug_data(
     msg.good("Corpus is loadable")
 
     # Create all gold data here to avoid iterating over the train_docs constantly
-    gold_train_data = _compile_gold(train_docs, pipeline)
+    gold_train_data = _compile_gold(train_docs, pipeline, nlp)
-    gold_train_unpreprocessed_data = _compile_gold(train_docs_unpreprocessed, pipeline)
+    gold_train_unpreprocessed_data = _compile_gold(
-    gold_dev_data = _compile_gold(dev_docs, pipeline)
+        train_docs_unpreprocessed, pipeline, nlp
+    )
+    gold_dev_data = _compile_gold(dev_docs, pipeline, nlp)
 
     train_texts = gold_train_data["texts"]
     dev_texts = gold_dev_data["texts"]
@@ -182,6 +184,16 @@ def debug_data(
                 nlp.vocab.vectors_length,
             )
         )
+        n_missing_vectors = sum(gold_train_data["words_missing_vectors"].values())
+        msg.warn(
+            "{} words in training data without vectors ({:0.2f}%)".format(
+                n_missing_vectors,
+                n_missing_vectors / gold_train_data["n_words"],
+            ),
+        )
+        msg.text(
+            "10 most common words without vectors: {}".format(_format_labels(gold_train_data["words_missing_vectors"].most_common(10), counts=True)), show=verbose,
+        )
     else:
         msg.info("No word vectors present in the model")
 
@@ -562,7 +574,7 @@ def _load_file(file_path, msg):
     )
 
 
-def _compile_gold(train_docs, pipeline):
+def _compile_gold(train_docs, pipeline, nlp):
     data = {
         "ner": Counter(),
         "cats": Counter(),
@@ -574,6 +586,7 @@ def _compile_gold(train_docs, pipeline):
         "punct_ents": 0,
         "n_words": 0,
         "n_misaligned_words": 0,
+        "words_missing_vectors": Counter(),
         "n_sents": 0,
         "n_nonproj": 0,
         "n_cycles": 0,
@@ -586,6 +599,10 @@ def _compile_gold(train_docs, pipeline):
         data["n_words"] += len(valid_words)
         data["n_misaligned_words"] += len(gold.words) - len(valid_words)
         data["texts"].add(doc.text)
+        if len(nlp.vocab.vectors):
+            for word in valid_words:
+                if nlp.vocab.strings[word] not in nlp.vocab.vectors:
+                    data["words_missing_vectors"].update([word])
         if "ner" in pipeline:
             for i, label in enumerate(gold.ner):
                 if label is None:
@@ -636,7 +653,11 @@ def _format_labels(labels, counts=False):
 def _get_examples_without_label(data, label):
     count = 0
     for doc, gold in data:
-        labels = [label.split("-")[1] for label in gold.ner if label not in ("O", "-")]
+        labels = [
+            label.split("-")[1]
+            for label in gold.ner
+            if label is not None and label not in ("O", "-")
+        ]
         if label not in labels:
             count += 1
     return count

spacy/cli/evaluate.py

@@ -2,6 +2,7 @@
 from __future__ import unicode_literals, division, print_function
 
 import plac
+import spacy
 from timeit import default_timer as timer
 from wasabi import msg
 
@@ -43,6 +44,9 @@ def evaluate(
     if displacy_path and not displacy_path.exists():
         msg.fail("Visualization output directory not found", displacy_path, exits=1)
     corpus = GoldCorpus(data_path, data_path)
+    if model.startswith("blank:"):
+        nlp = spacy.blank(model.replace("blank:", ""))
+    else:
         nlp = util.load_model(model)
     dev_docs = list(corpus.dev_docs(nlp, gold_preproc=gold_preproc))
     begin = timer()

spacy/cli/init_model.py

@@ -35,6 +35,12 @@ DEFAULT_OOV_PROB = -20
     jsonl_loc=("Location of JSONL-formatted attributes file", "option", "j", Path),
     clusters_loc=("Optional location of brown clusters data", "option", "c", str),
     vectors_loc=("Optional vectors file in Word2Vec format", "option", "v", str),
+    truncate_vectors=(
+        "Optional number of vectors to truncate to when reading in vectors file",
+        "option",
+        "t",
+        int,
+    ),
     prune_vectors=("Optional number of vectors to prune to", "option", "V", int),
     vectors_name=(
         "Optional name for the word vectors, e.g. en_core_web_lg.vectors",
@@ -51,6 +57,7 @@ def init_model(
     clusters_loc=None,
     jsonl_loc=None,
     vectors_loc=None,
+    truncate_vectors=0,
     prune_vectors=-1,
     vectors_name=None,
     model_name=None,
@@ -88,7 +95,7 @@ def init_model(
         nlp = create_model(lang, lex_attrs, name=model_name)
     msg.good("Successfully created model")
     if vectors_loc is not None:
-        add_vectors(nlp, vectors_loc, prune_vectors, vectors_name)
+        add_vectors(nlp, vectors_loc, truncate_vectors, prune_vectors, vectors_name)
     vec_added = len(nlp.vocab.vectors)
     lex_added = len(nlp.vocab)
     msg.good(
@@ -169,7 +176,7 @@ def create_model(lang, lex_attrs, name=None):
     return nlp
 
 
-def add_vectors(nlp, vectors_loc, prune_vectors, name=None):
+def add_vectors(nlp, vectors_loc, truncate_vectors, prune_vectors, name=None):
     vectors_loc = ensure_path(vectors_loc)
     if vectors_loc and vectors_loc.parts[-1].endswith(".npz"):
         nlp.vocab.vectors = Vectors(data=numpy.load(vectors_loc.open("rb")))
@@ -179,7 +186,7 @@ def add_vectors(nlp, vectors_loc, prune_vectors, name=None):
     else:
         if vectors_loc:
             with msg.loading("Reading vectors from {}".format(vectors_loc)):
-                vectors_data, vector_keys = read_vectors(vectors_loc)
+                vectors_data, vector_keys = read_vectors(vectors_loc, truncate_vectors)
             msg.good("Loaded vectors from {}".format(vectors_loc))
         else:
             vectors_data, vector_keys = (None, None)
@@ -199,9 +206,11 @@ def add_vectors(nlp, vectors_loc, prune_vectors, name=None):
         nlp.vocab.prune_vectors(prune_vectors)
 
 
-def read_vectors(vectors_loc):
+def read_vectors(vectors_loc, truncate_vectors=0):
     f = open_file(vectors_loc)
     shape = tuple(int(size) for size in next(f).split())
+    if truncate_vectors >= 1:
+        shape = (truncate_vectors, shape[1])
     vectors_data = numpy.zeros(shape=shape, dtype="f")
     vectors_keys = []
     for i, line in enumerate(tqdm(f)):
@@ -212,6 +221,8 @@ def read_vectors(vectors_loc):
             msg.fail(Errors.E094.format(line_num=i, loc=vectors_loc), exits=1)
         vectors_data[i] = numpy.asarray(pieces, dtype="f")
         vectors_keys.append(word)
+        if i == truncate_vectors - 1:
+            break
     return vectors_data, vectors_keys
 
 

spacy/cli/train.py

@@ -454,6 +454,9 @@ def train(
                         cpu_wps = nwords / (end_time - start_time)
                     else:
                         gpu_wps = nwords / (end_time - start_time)
+                        # Only evaluate on CPU in the first iteration (for
+                        # timing) if GPU is enabled
+                        if i >= 1:
                             with Model.use_device("cpu"):
                                 nlp_loaded = util.load_model_from_path(epoch_model_path)
                                 for name, component in nlp_loaded.pipeline:
@@ -550,7 +553,8 @@ def train(
     except Exception as e:
         msg.warn(
             "Aborting and saving the final best model. "
-            "Encountered exception: {}".format(e)
+            "Encountered exception: {}".format(e),
+            exits=1,
         )
     finally:
         best_pipes = nlp.pipe_names

website/docs/api/cli.md

@@ -547,7 +547,8 @@ $ python -m spacy init-model [lang] [output_dir] [--jsonl-loc] [--vectors-loc]
 | `output_dir`            | positional | Model output directory. Will be created if it doesn't exist.                                                                                                                                                                                           |
 | `--jsonl-loc`, `-j`     | option     | Optional location of JSONL-formatted [vocabulary file](/api/annotation#vocab-jsonl) with lexical attributes.                                                                                                                                           |
 | `--vectors-loc`, `-v`   | option     | Optional location of vectors. Should be a file where the first row contains the dimensions of the vectors, followed by a space-separated Word2Vec table. File can be provided in `.txt` format or as a zipped text file in `.zip` or `.tar.gz` format. |
-| `--prune-vectors`, `-V` | flag       | Number of vectors to prune the vocabulary to. Defaults to `-1` for no pruning.                                                                                                                                                                         |
+| `--truncate-vectors`, `-t` | option  | Number of vectors to truncate to when reading in vectors file. Defaults to `0` for no truncation.                                                                                                                                                      |
+| `--prune-vectors`, `-V` | option     | Number of vectors to prune the vocabulary to. Defaults to `-1` for no pruning.                                                                                                                                                                         |
 | `--vectors-name`, `-vn` | option     | Name to assign to the word vectors in the `meta.json`, e.g. `en_core_web_md.vectors`.                                                                                                                                                                  |
 | **CREATES**             | model      | A spaCy model containing the vocab and vectors.                                                                                                                                                                                                        |