def predict(model_name: str) -> Response: # pylint: disable=unused-variable
"""make a prediction using the specified model and return the results"""
if request.method == "OPTIONS":
return Response(response="", status=200)
model = app.predictors.get(model_name.lower())
if model is None:
raise ServerError("unknown model: {}".format(model_name),
status_code=400)
data = request.get_json()
log_blob = {
"model": model_name,
"inputs": data,
"cached": False,
"outputs": {}
}
# See if we hit or not. In theory this could result in false positives.
pre_hits = _caching_prediction.cache_info().hits # pylint: disable=no-value-for-parameter
try:
if cache_size > 0:
# lru_cache insists that all function arguments be hashable,
# so unfortunately we have to stringify the data.
prediction = _caching_prediction(model, json.dumps(data))
else:
# if cache_size is 0, skip caching altogether
prediction = model.predict_json(data)
except KeyError as err:
raise ServerError("Required JSON field not found: " + err.args[0],
status_code=400)
post_hits = _caching_prediction.cache_info().hits # pylint: disable=no-value-for-parameter
# Add to database and get permalink
if demo_db is not None:
try:
perma_id = demo_db.add_result(headers=dict(request.headers),
model_name=model_name,
inputs=data,
outputs=prediction)
if perma_id is not None:
slug = int_to_slug(perma_id)
prediction["slug"] = slug
log_blob["slug"] = slug
except Exception: # pylint: disable=broad-except
# TODO(joelgrus): catch more specific errors
logger.exception("Unable to add result to database",
exc_info=True)
if post_hits > pre_hits:
# Cache hit, so insert an artifical pause
log_blob["cached"] = True
time.sleep(0.25)
# The model predictions are extremely verbose, so we only log the most human-readable
# parts of them.
if model_name == "machine-comprehension":
log_blob["outputs"]["best_span_str"] = prediction["best_span_str"]
elif model_name == "coreference-resolution":
log_blob["outputs"]["clusters"] = prediction["clusters"]
log_blob["outputs"]["document"] = prediction["document"]
elif model_name == "textual-entailment":
log_blob["outputs"]["label_probs"] = prediction["label_probs"]
elif model_name == "named-entity-recognition":
log_blob["outputs"]["tags"] = prediction["tags"]
elif model_name == "semantic-role-labeling":
verbs = []
elif model_name == "crf_srl":
verbs = []
elif model_name == "frame_crf_srl":
verbs = []
for verb in prediction["verbs"]:
# Don't want to log boring verbs with no semantic parses.
good_tags = [tag for tag in verb["tags"] if tag != "0"]
if len(good_tags) > 1:
verbs.append({
"verb": verb["verb"],
"description": verb["description"]
})
log_blob["outputs"]["verbs"] = verbs
logger.info("prediction: %s", json.dumps(log_blob))
print(log_blob)
return jsonify(prediction)
def predict(model_name: str) -> Response: # pylint: disable=unused-variable
"""make a prediction using the specified model and return the results"""
if request.method == "OPTIONS":
return Response(response="", status=200)
# Do log if no argument is specified
record_to_database = request.args.get("record", "true").lower() != "false"
# Do use the cache if no argument is specified
use_cache = request.args.get("cache", "true").lower() != "false"
model = app.predictors.get(model_name.lower())
if model is None:
raise ServerError("unknown model: {}".format(model_name), status_code=400)
data = request.get_json()
log_blob = {"model": model_name, "inputs": data, "cached": False, "outputs": {}}
# Record the number of cache hits before we hit the cache so we can tell whether we hit or not.
# In theory this could result in false positives.
pre_hits = _caching_prediction.cache_info().hits # pylint: disable=no-value-for-parameter
try:
if use_cache and cache_size > 0:
# lru_cache insists that all function arguments be hashable,
# so unfortunately we have to stringify the data.
prediction = _caching_prediction(model, json.dumps(data))
else:
# if cache_size is 0, skip caching altogether
prediction = model.predict_json(data)
except KeyError as err:
raise ServerError("Required JSON field not found: " + err.args[0], status_code=400)
post_hits = _caching_prediction.cache_info().hits # pylint: disable=no-value-for-parameter
if record_to_database and demo_db is not None:
try:
perma_id = None
perma_id = demo_db.add_result(headers=dict(request.headers),
model_name=model_name,
inputs=data,
outputs=prediction)
if perma_id is not None:
slug = int_to_slug(perma_id)
prediction["slug"] = slug
log_blob["slug"] = slug
except Exception: # pylint: disable=broad-except
# TODO(joelgrus): catch more specific errors
logger.exception("Unable to add result to database", exc_info=True)
if use_cache and post_hits > pre_hits:
# Cache hit, so insert an artifical pause
log_blob["cached"] = True
time.sleep(0.25)
# The model predictions are extremely verbose, so we only log the most human-readable
# parts of them.
if model_name == "machine-comprehension":
log_blob["outputs"]["best_span_str"] = prediction["best_span_str"]
elif model_name == "coreference-resolution":
log_blob["outputs"]["clusters"] = prediction["clusters"]
log_blob["outputs"]["document"] = prediction["document"]
elif model_name == "textual-entailment":
log_blob["outputs"]["label_probs"] = prediction["label_probs"]
elif model_name == "named-entity-recognition":
log_blob["outputs"]["tags"] = prediction["tags"]
elif model_name == "semantic-role-labeling":
verbs = []
for verb in prediction["verbs"]:
# Don't want to log boring verbs with no semantic parses.
good_tags = [tag for tag in verb["tags"] if tag != "0"]
if len(good_tags) > 1:
verbs.append({"verb": verb["verb"], "description": verb["description"]})
log_blob["outputs"]["verbs"] = verbs
elif model_name == "constituency-parsing":
log_blob["outputs"]["trees"] = prediction["trees"]
logger.info("prediction: %s", json.dumps(log_blob))
print(log_blob)
return jsonify(prediction)
