I am trying to run the spaCy CLI but my training data and dev data seem somehow to be incorrect as seen when I run debug:
| => python3 -m spacy debug-data en
./CLI_train_randsplit_anno191022.json ./CLI_dev_randsplit_anno191022.json --pipeline ner --verbose
=========================== Data format validation ===========================
✔ Corpus is loadable
=============================== Training stats ===============================
Training pipeline: ner
Starting with blank model 'en'
0 training docs
0 evaluation docs
✔ No overlap between training and evaluation data
✘ Low number of examples to train from a blank model (0)
It's recommended to use at least 2000 examples (minimum 100)
============================== Vocab & Vectors ==============================
ℹ 0 total words in the data (0 unique)
10 most common words:
ℹ No word vectors present in the model
========================== Named Entity Recognition ==========================
ℹ 0 new labels, 0 existing labels
0 missing values (tokens with '-' label)
✔ Good amount of examples for all labels
✔ Examples without occurrences available for all labels
✔ No entities consisting of or starting/ending with whitespace
================================== Summary ==================================
✔ 5 checks passed
✘ 1 error
Trying to train anyway yields:
| => python3 -m spacy train en ./models/CLI_1 ./CLI_train_randsplit_anno191022.json ./CLI_dev_randsplit_anno191022.json -n 150 -p 'ner' --verbose
dropout_from = 0.2 by default
dropout_to = 0.2 by default
dropout_decay = 0.0 by default
batch_from = 100.0 by default
batch_to = 1000.0 by default
batch_compound = 1.001 by default
Training pipeline: ['ner']
Starting with blank model 'en'
beam_width = 1 by default
beam_density = 0.0 by default
beam_update_prob = 1.0 by default
Counting training words (limit=0)
learn_rate = 0.001 by default
optimizer_B1 = 0.9 by default
optimizer_B2 = 0.999 by default
optimizer_eps = 1e-08 by default
L2_penalty = 1e-06 by default
grad_norm_clip = 1.0 by default
parser_hidden_depth = 1 by default
subword_features = True by default
conv_depth = 4 by default
bilstm_depth = 0 by default
parser_maxout_pieces = 2 by default
token_vector_width = 96 by default
hidden_width = 64 by default
embed_size = 2000 by default
Itn NER Loss NER P NER R NER F Token % CPU WPS
--- --------- ------ ------ ------ ------- -------
✔ Saved model to output directory
models/CLI_1/model-final
Traceback (most recent call last):
File "/usr/local/lib/python3.7/site-packages/spacy/cli/train.py", line 389, in train
scorer = nlp_loaded.evaluate(dev_docs, verbose=verbose)
File "/usr/local/lib/python3.7/site-packages/spacy/language.py", line 673, in evaluate
docs, golds = zip(*docs_golds)
ValueError: not enough values to unpack (expected 2, got 0)
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/usr/local/Cellar/python/3.7.4_1/Frameworks/Python.framework/Versions/3.7/lib/python3.7/runpy.py", line 193, in _run_module_as_main
"__main__", mod_spec)
File "/usr/local/Cellar/python/3.7.4_1/Frameworks/Python.framework/Versions/3.7/lib/python3.7/runpy.py", line 85, in _run_code
exec(code, run_globals)
File "/usr/local/lib/python3.7/site-packages/spacy/__main__.py", line 35, in <module>
plac.call(commands[command], sys.argv[1:])
File "/usr/local/lib/python3.7/site-packages/plac_core.py", line 328, in call
cmd, result = parser.consume(arglist)
File "/usr/local/lib/python3.7/site-packages/plac_core.py", line 207, in consume
return cmd, self.func(*(args + varargs + extraopts), **kwargs)
File "/usr/local/lib/python3.7/site-packages/spacy/cli/train.py", line 486, in train
best_model_path = _collate_best_model(meta, output_path, nlp.pipe_names)
File "/usr/local/lib/python3.7/site-packages/spacy/cli/train.py", line 548, in _collate_best_model
bests[component] = _find_best(output_path, component)
File "/usr/local/lib/python3.7/site-packages/spacy/cli/train.py", line 567, in _find_best
accs = srsly.read_json(epoch_model / "accuracy.json")
File "/usr/local/lib/python3.7/site-packages/srsly/_json_api.py", line 50, in read_json
file_path = force_path(location)
File "/usr/local/lib/python3.7/site-packages/srsly/util.py", line 21, in force_path
raise ValueError("Can't read file: {}".format(location))
ValueError: Can't read file: models/CLI_1/model0/accuracy.json
My training and dev docs were generated using spacy.gold.docs_to_json(), saved as json files using the function:
def make_CLI_json(mock_docs, CLI_out_file_path):
CLI_json = docs_to_json(mock_docs)
with open(CLI_out_file_path, 'w') as json_file:
json.dump(CLI_json, json_file)
I verified them both to be valid json at http://www.jsonlint.com.
I created the docs from which these json originated using the function:
def import_from_doccano(jx_in_file_path, view=True):
annotations = load_jsonl(jx_in_file_path)
mock_nlp = English()
sentencizer = mock_nlp.create_pipe("sentencizer")
unlabeled = 0
DATA = []
mock_docs = []
for anno in annotations:
# get DATA (as used in spacy inline training)
if "label" in anno.keys():
ents = [tuple([label[0], label[1], label[2]])
for label in anno["labels"]]
else:
ents = []
DATUM = (anno["text"], {"entities": ents})
DATA.append(DATUM)
# mock a doc for viz in displacy
mock_doc = mock_nlp(anno["text"])
if "labels" in anno.keys():
entities = anno["labels"]
if not entities:
unlabeled += 1
ents = [(e[0], e[1], e[2]) for e in entities]
spans = [mock_doc.char_span(s, e, label=L) for s, e, L in ents]
mock_doc.ents = _cleanup_spans(spans)
sentencizer(mock_doc)
if view:
displacy.render(mock_doc, style='ent')
mock_docs.append(mock_doc)
print(f'Unlabeled: {unlabeled}')
return DATA, mock_docs
I wrote the function above to return the examples in both the format required for inline training (e.g. as shown at https://github.com/explosion/spaCy/blob/master/examples/training/train_ner.py) as well as to form these kind of “mock” docs so that I can use displacy and/or the CLI. For the latter purpose, I followed the code shown at https://github.com/explosion/spaCy/blob/master/spacy/cli/converters/jsonl2json.py with a couple of notable differences. The _cleanup_spans() function is identical to the one in the example. I did not use the minibatch() but made a separate doc for each of my labeled annotations. Also, (perhaps critically?) I found that using the sentencizer ruined many of my annotations, possibly because the spans get shifted in a way that the _cleanup_spans() function fails to repair properly. Removing the sentencizer causes the docs_to_json() function to throw an error. In my function (unlike in the linked example) I therefore run the sentencizer on each doc after the entities are written to them, which preserves my annotations properly and allows the docs_to_json() function to run without complaints.
The function load_jsonl called within import_from_doccano() is defined as:
def load_jsonl(input_path):
data = []
with open(input_path, 'r', encoding='utf-8') as f:
for line in f:
data.append(json.loads(line.replace('\n|\r',''), strict=False))
print('Loaded {} records from {}'.format(len(data), input_path))
print()
return data
My annotations are each of length ~10000 characters or less. They are exported from doccano
(https://doccano.herokuapp.com/) as JSONL using the format:
{"id": 1, "text": "EU rejects ...", "labels": [[0,2,"ORG"], [11,17, "MISC"], [34,41,"ORG"]]}
{"id": 2, "text": "Peter Blackburn", "labels": [[0, 15, "PERSON"]]}
{"id": 3, "text": "President Obama", "labels": [[10, 15, "PERSON"]]}
...
The data are split into train and test sets using the function:
def test_train_split(DATA, mock_docs, n_train):
L = list(zip(DATA, mock_docs))
random.shuffle(L)
DATA, mock_docs = zip(*L)
DATA = [i for i in DATA]
mock_docs = [i for i in mock_docs]
TRAIN_DATA = DATA[:n_train]
train_docs = mock_docs[:n_train]
TEST_DATA = DATA[n_train:]
test_docs = mock_docs[n_train:]
return TRAIN_DATA, TEST_DATA, train_docs, test_docs
And finally each is written to json using the following function:
def make_CLI_json(mock_docs, CLI_out_file_path):
CLI_json = docs_to_json(mock_docs)
with open(CLI_out_file_path, 'w') as json_file:
json.dump(CLI_json, json_file)
I do not understand why the debug shows 0 training docs and 0 development docs, or why the train command fails. The JSON look correct as far as I can tell. Is my data formatted incorrectly, or is there something else going on? Any help or insights would be greatly appreciated.
This is my first question on SE- apologies in advance if I've failed to follow some or other guideline. There are a lot of components involved so I'm not sure how I might produce a minimal code example that would replicate my problem.
Mac OS 10.15 Catalina Everything is pip3 installed into user path No virtual environment
| => python3 -m spacy info --markdown
## Info about spaCy
* **spaCy version:** 2.2.1
* **Platform:** Darwin-19.0.0-x86_64-i386-64bit
* **Python version:** 3.7.4
This is a legitimately confusing aspect of the API. For internal/historical reasons, spacy.gold.docs_to_json() produces a dict that still needs to be wrapped in list to get to the final training format. Try:
srsly.write_json(filename, [spacy.gold.docs_to_json(docs)])
spacy debug-data doesn't have proper schema checks yet, so this is more frustrating/confusing than it should be.