def test_simple_gradient_coref(self):
inputs = {
"document":
"This is a single string document about a test. Sometimes it "
"contains coreferent parts."
}
archive = load_archive(FIXTURES_ROOT / "coref" / "serialization" /
"model.tar.gz")
predictor = Predictor.from_archive(archive, "coreference-resolution")
interpreter = SimpleGradient(predictor)
interpretation = interpreter.saliency_interpret_from_json(inputs)
assert interpretation is not None
assert "instance_1" in interpretation
assert "grad_input_1" in interpretation["instance_1"]
grad_input_1 = interpretation["instance_1"]["grad_input_1"]
# There are 16 tokens in the input. This gets translated into 22 wordpieces, but we need to
# compute gradients on whatever the model considered its "input", which in this case is
# tokens, because the model uses a mismatched tokenizer / embedder.
assert len(grad_input_1) == 16
# two interpretations should be identical for gradient
repeat_interpretation = interpreter.saliency_interpret_from_json(
inputs)
repeat_grad_input_1 = repeat_interpretation["instance_1"][
"grad_input_1"]
for grad, repeat_grad in zip(grad_input_1, repeat_grad_input_1):
assert grad == approx(repeat_grad)
def test_interpret_fails_when_embedding_layer_not_found(self):
inputs = {"sentence": "It was the ending that I hated"}
vocab = Vocabulary()
vocab.add_tokens_to_namespace(
[w for w in inputs["sentence"].split(" ")])
model = FakeModelForTestingInterpret(
vocab, max_tokens=len(inputs["sentence"].split(" ")))
predictor = TextClassifierPredictor(model,
TextClassificationJsonReader())
interpreter = SimpleGradient(predictor)
with raises(RuntimeError):
interpreter.saliency_interpret_from_json(inputs)
def test_simple_gradient_masked_lm(self):
inputs = {
"sentence":
"This is a single string [MASK] about a test . Sometimes it "
"contains coreferent parts ."
}
archive = load_archive(self.FIXTURES_ROOT / "masked_language_model" /
"serialization" / "model.tar.gz")
predictor = Predictor.from_archive(archive, "masked_language_model")
interpreter = SimpleGradient(predictor)
interpretation = interpreter.saliency_interpret_from_json(inputs)
assert interpretation is not None
assert "instance_1" in interpretation
assert "grad_input_1" in interpretation["instance_1"]
grad_input_1 = interpretation["instance_1"]["grad_input_1"]
assert len(grad_input_1) == 16 # 16 words in input
def __init__(self, model_path=None, cuda_device=1):
# model_path = model_path or LSTM_MODEL_PATH
model_path = model_path or ROBERTA_MODEL_PATH
self.predictor = Predictor.from_path(model_path,
cuda_device=cuda_device)
_tokenizer = PretrainedTransformerTokenizer(
model_name="roberta-base", max_length=TRANSFORMER_WORDPIECE_LIMIT)
class_name_mapper = {"0": "Negative", "1": "Positive"}
_model = self.predictor._model
_label_namespace = _model._label_namespace
class_names = [
class_name_mapper[_model.vocab.get_index_to_token_vocabulary(
_label_namespace).get(0)],
class_name_mapper[_model.vocab.get_index_to_token_vocabulary(
_label_namespace).get(1)]
]
# reset the tokenizer to remove separators
self.tokenizer = lambda s: [
t.text.replace("Ġ", "").replace('Ċ', '').replace('ĉ', "")
for t in _tokenizer.tokenize(s)
][1:-1]
self.explainer_lime = LimeTextExplainer(
class_names=class_names, split_expression=self.tokenizer)
self.explainer_integrate = IntegratedGradient(self.predictor)
self.explainer_simple = SimpleGradient(self.predictor)
def load_interpreters(self) -> Mapping[str, SaliencyInterpreter]:
"""
Returns a mapping of interpreters keyed by a unique identifier. Requests to
`/interpret/:id` will invoke the interpreter with the provided `:id`. Override this method
to add or remove interpreters.
"""
return { "simple": SimpleGradient(self.predictor),
"smooth": SmoothGradient(self.predictor),
"integrated": IntegratedGradient(self.predictor) }
def test_interpret_works_with_custom_embedding_layer(self):
inputs = {"sentence": "It was the ending that I hated"}
vocab = Vocabulary()
vocab.add_tokens_to_namespace(
[w for w in inputs["sentence"].split(" ")])
model = FakeModelForTestingInterpret(
vocab, max_tokens=len(inputs["sentence"].split(" ")))
predictor = FakePredictorForTestingInterpret(
model, TextClassificationJsonReader())
interpreter = SimpleGradient(predictor)
interpretation = interpreter.saliency_interpret_from_json(inputs)
assert interpretation is not None
assert "instance_1" in interpretation
assert "grad_input_1" in interpretation["instance_1"]
grad_input_1 = interpretation["instance_1"]["grad_input_1"]
assert len(grad_input_1) == 7 # 7 words in input
def test_gradient_visualization():
from allennlp.predictors.predictor import Predictor
predictor = Predictor.from_path(
"https://storage.googleapis.com/allennlp-public-models/sst-roberta-large-2020.06.08.tar.gz"
)
sentence = "a very well-made, funny and entertaining picture."
inputs = {"sentence": sentence}
from allennlp.interpret.saliency_interpreters import SimpleGradient
simple_gradient_interpreter = SimpleGradient(predictor)
simple_gradient_interpretation = simple_gradient_interpreter.saliency_interpret_from_json(
inputs
)
gradients = simple_gradient_interpretation["instance_1"]["grad_input_1"]
assert max(gradients) - min(gradients) < 0.75
def test_simple_gradient_coref(self):
inputs = {
"document": "This is a single string document about a test. Sometimes it "
"contains coreferent parts."
}
archive = load_archive(self.FIXTURES_ROOT / "coref" / "serialization" / "model.tar.gz")
predictor = Predictor.from_archive(archive, "coreference-resolution")
interpreter = SimpleGradient(predictor)
interpretation = interpreter.saliency_interpret_from_json(inputs)
assert interpretation is not None
assert "instance_1" in interpretation
assert "grad_input_1" in interpretation["instance_1"]
grad_input_1 = interpretation["instance_1"]["grad_input_1"]
assert len(grad_input_1) == 16 # 16 words in input
# two interpretations should be identical for gradient
repeat_interpretation = interpreter.saliency_interpret_from_json(inputs)
repeat_grad_input_1 = repeat_interpretation["instance_1"]["grad_input_1"]
for grad, repeat_grad in zip(grad_input_1, repeat_grad_input_1):
assert grad == approx(repeat_grad)
def test_simple_gradient_basic_text(self):
inputs = {"sentence": "It was the ending that I hated"}
archive = load_archive(
self.FIXTURES_ROOT / "basic_classifier" / "serialization" / "model.tar.gz"
)
predictor = Predictor.from_archive(archive, "text_classifier")
interpreter = SimpleGradient(predictor)
interpretation = interpreter.saliency_interpret_from_json(inputs)
assert interpretation is not None
assert "instance_1" in interpretation
assert "grad_input_1" in interpretation["instance_1"]
grad_input_1 = interpretation["instance_1"]["grad_input_1"]
assert len(grad_input_1) == 7 # 7 words in input
# two interpretations should be identical for gradient
repeat_interpretation = interpreter.saliency_interpret_from_json(inputs)
repeat_grad_input_1 = repeat_interpretation["instance_1"]["grad_input_1"]
for grad, repeat_grad in zip(grad_input_1, repeat_grad_input_1):
assert grad == approx(repeat_grad)
def load_interpreters(self) -> Dict[str, SaliencyInterpreter]:
"""
Returns a mapping of interpreters keyed by a unique identifier. Requests to
`/interpret/:id` will invoke the interpreter with the provided `:id`. Override this method
to add or remove interpreters.
"""
interpreters: Dict[str, SaliencyInterpreter] = {}
if "simple_gradient" in self.model.interpreters:
interpreters["simple_gradient"] = SimpleGradient(self.predictor)
if "smooth_gradient" in self.model.interpreters:
interpreters["smooth_gradient"] = SmoothGradient(self.predictor)
if "integrated_gradient" in self.model.interpreters:
interpreters["integrated_gradient"] = IntegratedGradient(
self.predictor)
return interpreters
class AllenNLPExplainer(Explainer):
def __init__(self, model):
self.model = model
self.tokenizer = model.tokenizer
self.device = model.device
self.predictor = model.predictor
self.exp = SimpleGradient(self.predictor)
def explain_instance(self, x):
'''
x - 1 input instance
returns - list of top tokens/importance weights
'''
explanation = self.exp.saliency_interpret_from_json({"sentence": x})
grad = explanation['instance_1']['grad_input_1']
label = self.predictor.predict(x)['label']
return grad, label
@overrides
def explain_instances(self, X):
'''
X - array of input sentences
'''
grad_list = []
label_list = []
for s in X:
grad, label = self.explain_instance(s)
grad_list.append(grad)
label_list.append(label)
return grad_list, label_list
def __init__(self, model):
self.model = model
self.tokenizer = model.tokenizer
self.device = model.device
self.predictor = model.predictor
self.exp = SimpleGradient(self.predictor)
def make_app(models: Dict[str, DemoModel],
demo_db: Optional[DemoDatabase] = None,
cache_size: int = 128,
interpret_cache_size: int = 500,
attack_cache_size: int = 500) -> Flask:
app = Flask(__name__) # pylint: disable=invalid-name
start_time = datetime.now(pytz.utc)
start_time_str = start_time.strftime("%Y-%m-%d %H:%M:%S %Z")
app.predictors = {}
app.max_request_lengths = {
} # requests longer than these will be rejected to prevent OOME
app.attackers = defaultdict(dict)
app.interpreters = defaultdict(dict)
app.wsgi_app = ProxyFix(
app.wsgi_app) # sets the requester IP with the X-Forwarded-For header
for name, demo_model in models.items():
if demo_model is not None:
logger.info(f"loading {name} model")
predictor = demo_model.predictor()
app.predictors[name] = predictor
app.max_request_lengths[name] = demo_model.max_request_length
if name in supported_interpret_models:
app.interpreters[name]['simple_gradient'] = SimpleGradient(
predictor)
app.interpreters[name][
'integrated_gradient'] = IntegratedGradient(predictor)
app.interpreters[name]['smooth_gradient'] = SmoothGradient(
predictor)
app.attackers[name]["input_reduction"] = InputReduction(
predictor)
if name == 'masked-lm':
app.attackers[name]["hotflip"] = Hotflip(predictor, 'bert')
elif name == "next-token-lm":
app.attackers[name]["hotflip"] = Hotflip(predictor, 'gpt2')
elif 'named-entity-recognition' in name:
# We haven't implemented hotflip for NER.
continue
elif name == 'textual-entailment':
# The SNLI model only has ELMo embeddings, which don't work with hotflip on
# their own.
continue
else:
app.attackers[name]["hotflip"] = Hotflip(predictor)
app.attackers[name]["hotflip"].initialize()
# Disable caching for HTML documents and API responses so that clients
# always talk to the source (this server).
@app.after_request
def set_cache_headers(resp: Response) -> Response:
if resp.mimetype == "text/html" or resp.mimetype == "application/json":
return with_no_cache_headers(resp)
else:
return resp
@app.errorhandler(ServerError)
def handle_invalid_usage(error: ServerError) -> Response: # pylint: disable=unused-variable
response = jsonify(error.to_dict())
response.status_code = error.status_code
return response
@lru_cache(maxsize=cache_size)
def _caching_prediction(model: Predictor, data: str) -> JsonDict:
"""
Just a wrapper around ``model.predict_json`` that allows us to use a cache decorator.
"""
return model.predict_json(json.loads(data))
@lru_cache(maxsize=interpret_cache_size)
def _caching_interpret(interpreter: SaliencyInterpreter,
data: str) -> JsonDict:
"""
Just a wrapper around ``model.interpret_from_json`` that allows us to use a cache decorator.
"""
return interpreter.saliency_interpret_from_json(json.loads(data))
@lru_cache(maxsize=attack_cache_size)
def _caching_attack(attacker: Attacker, data: str,
input_field_to_attack: str, grad_input_field: str,
target: str) -> JsonDict:
"""
Just a wrapper around ``model.attack_from_json`` that allows us to use a cache decorator.
"""
return attacker.attack_from_json(
inputs=json.loads(data),
input_field_to_attack=input_field_to_attack,
grad_input_field=grad_input_field,
target=json.loads(target))
@app.route('/')
def index() -> str: # pylint: disableunused-variable
loaded_modules = {}
for n, m in models.items():
loaded_modules[n] = m.__dict__
return jsonify({"allennlp_version": VERSION, "models": loaded_modules})
@app.route('/permadata/<model_name>', methods=['POST', 'OPTIONS'])
def permadata(model_name: str) -> Response: # pylint: disable=unused-variable
"""
If the user requests a permalink, the front end will POST here with the payload
{ slug: slug }
which we convert to an integer id and use to retrieve saved results from the database.
"""
# This is just CORS boilerplate.
if request.method == "OPTIONS":
return Response(response="", status=200)
# If we don't have a database configured, there are no permalinks.
if demo_db is None:
raise ServerError('Permalinks are not enabled', 400)
# Convert the provided slug to an integer id.
slug = request.get_json()["slug"]
perma_id = slug_to_int(slug)
if perma_id is None:
# Malformed slug
raise ServerError("Unrecognized permalink: {}".format(slug), 400)
# Fetch the results from the database.
try:
permadata = demo_db.get_result(perma_id)
except psycopg2.Error:
logger.exception(
"Unable to get results from database: perma_id %s", perma_id)
raise ServerError('Database trouble', 500)
if permadata is None:
# No data found, invalid id?
raise ServerError("Unrecognized permalink: {}".format(slug), 400)
return jsonify({
"modelName": permadata.model_name,
"requestData": permadata.request_data,
"responseData": permadata.response_data
})
@app.route('/predict/<model_name>', methods=['POST', 'OPTIONS'])
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"
lowered_model_name = model_name.lower()
model = app.predictors.get(lowered_model_name)
if model is None:
raise ServerError("unknown model: {}".format(model_name),
status_code=400)
max_request_length = app.max_request_lengths[lowered_model_name]
data = request.get_json()
serialized_request = json.dumps(data)
if len(serialized_request) > max_request_length:
raise ServerError(
f"Max request length exceeded for model {model_name}! " +
f"Max: {max_request_length} Actual: {len(serialized_request)}")
logger.info("request: %s",
json.dumps({
"model": model_name,
"inputs": data
}))
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
if record_to_database and demo_db is not None:
try:
perma_id = None
perma_id = demo_db.insert_request(
headers=dict(request.headers),
requester=request.remote_addr,
model_name=model_name,
inputs=data)
except Exception: # pylint: disable=broad-except
# TODO(joelgrus): catch more specific errors
logger.exception("Unable to add request to database",
exc_info=True)
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)
post_hits = _caching_prediction.cache_info().hits # pylint: disable=no-value-for-parameter
if record_to_database and demo_db is not None and perma_id is not None:
try:
demo_db.update_response(perma_id=perma_id, outputs=prediction)
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 response 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 "comprehension" in model_name:
if 'best_span_str' in prediction:
answer = prediction['best_span_str']
else:
answer = prediction['answer']
log_blob["outputs"]["answer"] = answer
elif model_name == "nmn-drop":
answer = prediction['answer']
log_blob["outputs"]["answer"] = answer
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 == "sentiment-analysis":
log_blob["outputs"]["probs"] = prediction["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"]
elif model_name == "wikitables-parser":
log_blob['outputs']['logical_form'] = prediction['logical_form']
log_blob['outputs']['answer'] = prediction['answer']
elif model_name == "nlvr-parser":
log_blob['outputs']['logical_form'] = prediction['logical_form'][0]
log_blob['outputs']['answer'] = prediction['denotations'][0][0]
elif model_name == "atis-parser":
log_blob['outputs']['predicted_sql_query'] = prediction[
'predicted_sql_query']
# TODO(brendanr): Add event2mind log_blob here?
logger.info("prediction: %s", json.dumps(log_blob))
return jsonify(prediction)
@app.route('/attack/<model_name>', methods=['POST', 'OPTIONS'])
def attack(model_name: str) -> Response:
"""
Modify input to change prediction of model
"""
if request.method == "OPTIONS":
return Response(response="", status=200)
# Do use the cache if no argument is specified
use_cache = request.args.get("cache", "true").lower() != "false"
lowered_model_name = model_name.lower()
data = request.get_json()
attacker_name = data.pop("attacker")
input_field_to_attack = data.pop("inputToAttack")
grad_input_field = data.pop("gradInput")
target = data.pop("target", None)
model_attackers = app.attackers.get(lowered_model_name)
if model_attackers is None:
raise ServerError("unknown model: {}".format(model_name),
status_code=400)
attacker = model_attackers.get(attacker_name)
if attacker is None:
raise ServerError("unknown attacker for model: {} {}".format(
attacker_name, model_name),
status_code=400)
max_request_length = app.max_request_lengths[lowered_model_name]
serialized_request = json.dumps(data)
if len(serialized_request) > max_request_length:
raise ServerError(
f"Max request length exceeded for model {model_name}! " +
f"Max: {max_request_length} Actual: {len(serialized_request)}")
pre_hits = _caching_attack.cache_info().hits # pylint: disable=no-value-for-parameter
if use_cache and attack_cache_size > 0:
# lru_cache insists that all function arguments be hashable,
# so unfortunately we have to stringify the data.
attack = _caching_attack(attacker, json.dumps(data),
input_field_to_attack, grad_input_field,
json.dumps(target))
else:
# if cache_size is 0, skip caching altogether
attack = attacker.attack_from_json(
inputs=data,
input_field_to_attack=input_field_to_attack,
grad_input_field=grad_input_field,
target=target)
post_hits = _caching_attack.cache_info().hits # pylint: disable=no-value-for-parameter
if use_cache and post_hits > pre_hits:
# Cache hit, so insert an artifical pause
time.sleep(0.25)
return jsonify(attack)
@app.route('/interpret/<model_name>', methods=['POST', 'OPTIONS'])
def interpret(model_name: str) -> Response:
"""
Interpret prediction of the model
"""
if request.method == "OPTIONS":
return Response(response="", status=200)
# Do use the cache if no argument is specified
use_cache = request.args.get("cache", "true").lower() != "false"
lowered_model_name = model_name.lower()
data = request.get_json()
interpreter_name = data.pop("interpreter")
model_interpreters = app.interpreters.get(lowered_model_name)
if model_interpreters is None:
raise ServerError(
"no interpreters for model: {}".format(model_name),
status_code=400)
interpreter = model_interpreters.get(interpreter_name)
if interpreter is None:
raise ServerError("unknown interpreter for model: {} {}".format(
interpreter_name, model_name),
status_code=400)
max_request_length = app.max_request_lengths[lowered_model_name]
serialized_request = json.dumps(data)
if len(serialized_request) > max_request_length:
raise ServerError(
f"Max request length exceeded for interpreter {model_name}! " +
f"Max: {max_request_length} Actual: {len(serialized_request)}")
pre_hits = _caching_interpret.cache_info().hits # pylint: disable=no-value-for-parameter
if use_cache and interpret_cache_size > 0:
# lru_cache insists that all function arguments be hashable,
# so unfortunately we have to stringify the data.
interpretation = _caching_interpret(interpreter, json.dumps(data))
else:
# if cache_size is 0, skip caching altogether
interpretation = interpreter.saliency_interpret_from_json(data)
post_hits = _caching_prediction.cache_info().hits # pylint: disable=no-value-for-parameter
if use_cache and post_hits > pre_hits:
# Cache hit, so insert an artifical pause
time.sleep(0.25)
return jsonify(interpretation)
@app.route('/models')
def list_models() -> Response: # pylint: disable=unused-variable
"""list the available models"""
return jsonify({"models": list(app.predictors.keys())})
@app.route('/info')
def info() -> Response: # pylint: disable=unused-variable
"""List metadata about the running webserver"""
uptime = str(datetime.now(pytz.utc) - start_time)
git_version = os.environ.get('ALLENNLP_DEMO_SOURCE_COMMIT') or ""
return jsonify({
"start_time":
start_time_str,
"uptime":
uptime,
"git_version":
git_version,
"peak_memory_mb":
peak_memory_mb(),
"githubUrl":
"http://github.com/allenai/allennlp-demo/commit/" + git_version
})
@app.route('/health')
def health() -> Response: # pylint: disable=unused-variable
return "healthy"
# As an SPA, we need to return index.html for /model-name and /model-name/permalink
def return_page(permalink: str = None) -> Response: # pylint: disable=unused-argument, unused-variable
"""return the page"""
return send_file(os.path.join(build_dir, 'index.html'))
for model_name in models:
logger.info(f"setting up default routes for {model_name}")
app.add_url_rule(f"/{model_name}", view_func=return_page)
app.add_url_rule(f"/{model_name}/<permalink>", view_func=return_page)
return app