def action_serve(argc, argv):
global prj, app, classes, num_outputs
args = parse_args(argv)
prj = Project(args.path)
err = prj.load()
if err is not None:
log.error("error while loading project: %s", err)
quit()
elif not prj.is_trained():
log.error("no trained Keras model found for this project")
quit()
if args.classes is None:
num_outputs = prj.model.output.shape[1]
if prj.classes is None:
classes = ["class_%d" % i for i in range(num_outputs)]
else:
classes = [prj.classes[i] for i in range(num_outputs)]
else:
classes = [s.strip() for s in args.classes.split(',') if s.strip() != ""]
num_outputs = len(classes)
if args.profile:
from werkzeug.contrib.profiler import ProfilerMiddleware
args.debug = True
app.config['PROFILE'] = True
app.wsgi_app = ProfilerMiddleware(app.wsgi_app, restrictions=[args.restrictions])
app.run(host=args.address, port=args.port, debug=args.debug)
def action_to_fdeep(argc, argv):
args = parse_args(argv)
prj = Project(args.path)
err = prj.load()
if err is not None:
log.error("error while loading project: %s", err)
quit()
elif not prj.is_trained():
log.error("no trained model found for this project")
quit()
convert(prj.weights_path, prj.fdeep_path, args.no_tests, args.metadata)
def action_serve(argc, argv):
global prj, app
args = parse_args(argv)
prj = Project(args.path)
err = prj.load()
if err is not None:
log.error("error while loading project: %s", err)
quit()
elif not prj.is_trained():
log.error("no trained Keras model found for this project")
quit()
app.run(host=args.host, port=args.port, debug=args.debug)
def action_to_fdeep(args):
prj = Project(args.project_path)
err = prj.load()
if err is not None:
log.error("error while loading project: %s", err)
quit()
elif not prj.is_trained():
log.error("no trained Keras model found for this project")
quit()
log.info("converting %s to %s ...", prj.weights_path, prj.fdeep_path)
convert(prj.model, prj.fdeep_path)
def action_to_tf(argc, argv):
args = parse_args(argv)
prj = Project(args.path)
err = prj.load()
if err is not None:
log.error("error while loading project: %s", err)
quit()
elif not prj.is_trained():
log.error("no trained Keras model found for this projec")
quit()
frozen_graph = freeze_session(K.get_session(),
output_names=[out.op.name for out in prj.model.outputs])
log.info("saving protobuf to %s ...", os.path.join(prj.path, 'model.pb'))
tf.train.write_graph(frozen_graph, prj.path, "model.pb", as_text=False)
def action_relevance(argc, argv):
args = parse_args(argv)
prj = Project(args.path)
err = prj.load()
if err is not None:
log.error("error while loading project: %s", err)
quit()
elif not prj.is_trained():
log.error("no trained Keras model found for this project")
quit()
prj.prepare(args.dataset, 0.0, 0.0)
X, y = prj.dataset.subsample(args.ratio)
nrows, ncols = X.shape if prj.dataset.is_flat else (X[0].shape[0], len(X))
attributes = get_attributes(args.attributes, ncols)
log.info("computing relevance of %d attributes on %d samples ...", ncols,
nrows)
start = time.time()
ref_accu, ref_cm = prj.accuracy_for(X, y, repo_as_dict=True)
deltas = []
tot = 0
speed = (1.0 / (time.time() - start)) * nrows
for col in range(0, ncols):
log.info(
"[%.2f evals/s] computing relevance for attribute [%d/%d] %s ...",
speed, col + 1, ncols, attributes[col])
backup = zeroize_feature(X, col, prj.dataset.is_flat)
start = time.time()
accu, cm = prj.accuracy_for(X, y, repo_as_dict=True)
speed = (1.0 / (time.time() - start)) * nrows
delta = ref_accu['weighted avg']['precision'] - accu['weighted avg'][
'precision']
tot += delta
deltas.append((col, delta))
restore_feature(X, col, backup, prj.dataset.is_flat)
deltas = sorted(deltas, key=lambda x: abs(x[1]), reverse=True)
rels = []
num_zero = 0
table = [("Column", "Feature", "Relevance")]
for delta in deltas:
col, d = delta
colname = attributes[col]
rel = {"attribute": colname, "index": col, "relevance": 0.0}
if d != 0.0:
relevance = (d / tot) * 100.0
row = ("%d" % col, attributes[col], "%.2f%%" % relevance)
row = ["\033[31m%s\033[0m" % e
for e in row] if relevance < 0.0 else row
table.append(row)
rel['relevance'] = relevance
else:
num_zero += 1
rels.append(rel)
print("")
print(AsciiTable(table).table)
print("")
if num_zero > 0:
log.info("%d features have 0 relevance.", num_zero)
if args.to_json is not None:
print("")
log.info("creating %s ...", args.to_json)
with open(args.to_json, 'w+') as fp:
json.dump(rels, fp, default=default)
def action_relevance(argc, argv):
global prj, deltas, tot, start, speed, nrows, ncols, attributes
args = parse_args(argv)
prj = Project(args.path)
err = prj.load()
if err is not None:
log.error("error while loading project: %s", err)
quit()
elif not prj.is_trained():
log.error("no trained Keras model found for this project")
quit()
prj.prepare(args.dataset, 0.0, 0.0)
# one single worker in blocking mode = serial
if args.workers == 0:
args.workers = 1
X, y = prj.dataset.subsample(args.ratio)
nrows, ncols = X.shape if prj.dataset.is_flat else (X[0].shape[0], len(X))
attributes = get_attributes(args.attributes, ncols)
queue = TaskQueue('relevance', num_workers=args.workers, blocking=True)
if args.workers == 1:
log.info("computing relevance of %d attributes on %d samples using '%s' metric (slow mode) ...", ncols, nrows, args.metric)
else:
log.info("computing relevance of %d attributes on %d samples using '%s' metric (parallel with %d workers) ...", ncols, nrows, args.metric, queue.num_workers)
start = time.time()
ref_accu, ref_cm = prj.accuracy_for(X, y, repo_as_dict = True)
speed = (1.0 / (time.time() - start)) * nrows
for col in range(0, ncols):
queue.add_task( run_inference_without,
X, y, col, prj.dataset.is_flat,
ref_accu['weighted avg'][args.metric],
args.metric)
# wait for all inferences to finish
queue.join()
# sort relevances by absolute value
deltas = sorted(deltas, key = lambda x: abs(x[1]), reverse = True)
rels = []
num_zero = 0
table = [("Column", "Feature", "Relevance")]
for delta in deltas:
col, d = delta
colname = attributes[col]
rel = {
"attribute": colname,
"index": col,
"relevance": 0.0
}
if d != 0.0:
relevance = (d / tot) * 100.0
row = ("%d" % col, attributes[col], "%.2f%%" % relevance)
row = ["\033[31m%s\033[0m" % e for e in row] if relevance < 0.0 else row
table.append(row)
rel['relevance'] = relevance
else:
num_zero += 1
rels.append(rel)
print("")
print(AsciiTable(table).table)
print("")
if num_zero > 0:
log.info("%d features have 0 relevance.", num_zero)
if args.to_json is not None:
print("")
log.info("creating %s ...", args.to_json)
with open(args.to_json, 'w+') as fp:
json.dump(rels, fp, default=default)