Use python3-pycoral instead of python3-edgetpu
Change-Id: I008dfc6c334dee1963c8fdeaf94a5e25263f0aea
diff --git a/debian/changelog b/debian/changelog
index a8bf97f..2205fbb 100644
--- a/debian/changelog
+++ b/debian/changelog
@@ -1,3 +1,9 @@
+edgetpuvision (7-1) mendel-eagle; urgency=medium
+
+ * Switching from python3-edgetpu to python3-pycoral.
+
+ -- Coral <coral-support@google.com> Mon, 09 Nov 2020 16:24:25 -0800
+
edgetpuvision (6-1) mendel-day; urgency=medium
* Cut for day release.
diff --git a/debian/control b/debian/control
index 4b4e6b5..8914388 100644
--- a/debian/control
+++ b/debian/control
@@ -20,7 +20,6 @@
gstreamer1.0-plugins-ugly,
gstreamer1.0-python3-plugin-loader,
python3-cairo,
- python3-edgetpu,
python3-gi,
python3-gi-cairo,
python3-gst-1.0,
@@ -28,8 +27,8 @@
python3-numpy,
python3-pil,
python3-protobuf,
+ python3-pycoral,
weston
Recommends: imx-gst1.0-plugin,
Description: EdgeTPU camera API
API to run inference on image data coming from the camera.
-
diff --git a/edgetpuvision/classify.py b/edgetpuvision/classify.py
index 408ecda..165806f 100644
--- a/edgetpuvision/classify.py
+++ b/edgetpuvision/classify.py
@@ -25,7 +25,8 @@
import itertools
import time
-from edgetpu.classification.engine import ClassificationEngine
+from pycoral.adapters import classify
+from pycoral.utils import edgetpu
from . import svg
from . import utils
@@ -109,15 +110,15 @@
fps_counter = utils.avg_fps_counter(30)
- engines, titles = utils.make_engines(args.model, ClassificationEngine)
- assert utils.same_input_image_sizes(engines)
- engines = itertools.cycle(engines)
- engine = next(engines)
+ interpreters, titles = utils.make_interpreters(args.model)
+ assert utils.same_input_image_sizes(interpreters)
+ interpreters = itertools.cycle(interpreters)
+ interpreter = next(interpreters)
labels = utils.load_labels(args.labels)
draw_overlay = True
- yield utils.input_image_size(engine)
+ yield utils.input_image_size(interpreter)
output = None
while True:
@@ -126,15 +127,17 @@
inference_rate = next(fps_counter)
if draw_overlay:
start = time.monotonic()
- results = engine.classify_with_input_tensor(tensor, threshold=args.threshold, top_k=args.top_k)
+ edgetpu.run_inference(interpreter, tensor)
inference_time = time.monotonic() - start
- results = [(labels[i], score) for i, score in results]
+ classes = classify.get_classes(interpreter, top_k=args.top_k,
+ score_threshold=args.threshold)
+ results = [(labels[class_id], score) for class_id, score in classes]
results = acc.send(results)
if args.print:
print_results(inference_rate, results)
- title = titles[engine]
+ title = titles[interpreter]
output = overlay(title, results, inference_time, inference_rate, layout)
else:
output = None
@@ -142,7 +145,7 @@
if command == 'o':
draw_overlay = not draw_overlay
elif command == 'n':
- engine = next(engines)
+ interpreter = next(interpreters)
def add_render_gen_args(parser):
parser.add_argument('--model', required=True,
diff --git a/edgetpuvision/detect.py b/edgetpuvision/detect.py
index 8908617..96ec8f4 100644
--- a/edgetpuvision/detect.py
+++ b/edgetpuvision/detect.py
@@ -27,12 +27,12 @@
"""
import argparse
-import collections
import colorsys
import itertools
import time
-from edgetpu.detection.engine import DetectionEngine
+from pycoral.adapters import detect
+from pycoral.utils import edgetpu
from . import svg
from . import utils
@@ -44,15 +44,6 @@
'.bbox': svg.Style(fill_opacity=0.0,
stroke_width='0.1em')}))
-BBox = collections.namedtuple('BBox', ('x', 'y', 'w', 'h'))
-BBox.area = lambda self: self.w * self.h
-BBox.scale = lambda self, sx, sy: BBox(x=self.x * sx, y=self.y * sy,
- w=self.w * sx, h=self.h * sy)
-BBox.__str__ = lambda self: 'BBox(x=%.2f y=%.2f w=%.2f h=%.2f)' % self
-
-Object = collections.namedtuple('Object', ('id', 'label', 'score', 'bbox'))
-Object.__str__ = lambda self: 'Object(id=%d, label=%s, score=%.2f, %s)' % self
-
def size_em(length):
return '%sem' % str(0.6 * (length + 1))
@@ -72,7 +63,7 @@
return lambda obj_id: 'white'
-def overlay(title, objs, get_color, inference_time, inference_rate, layout):
+def overlay(title, objs, get_color, labels, inference_time, inference_rate, layout):
x0, y0, width, height = layout.window
font_size = 0.03 * height
@@ -86,14 +77,15 @@
for obj in objs:
percent = int(100 * obj.score)
- if obj.label:
- caption = '%d%% %s' % (percent, obj.label)
+ if labels:
+ caption = '%d%% %s' % (percent, labels[obj.id])
else:
caption = '%d%%' % percent
- x, y, w, h = obj.bbox.scale(*layout.size)
color = get_color(obj.id)
-
+ inference_width, inference_height = layout.inference_size
+ bbox = obj.bbox.scale(1.0 / inference_width, 1.0 / inference_height).scale(*layout.size)
+ x, y, w, h = bbox.xmin, bbox.ymin, bbox.width, bbox.height
doc += svg.Rect(x=x, y=y, width=w, height=h,
style='stroke:%s' % color, _class='bbox')
doc += svg.Rect(x=x, y=y+h ,
@@ -125,26 +117,18 @@
return str(doc)
-
-def convert(obj, labels):
- x0, y0, x1, y1 = obj.bounding_box.flatten().tolist()
- return Object(id=obj.label_id,
- label=labels[obj.label_id] if labels else None,
- score=obj.score,
- bbox=BBox(x=x0, y=y0, w=x1 - x0, h=y1 - y0))
-
def print_results(inference_rate, objs):
print('\nInference (rate=%.2f fps):' % inference_rate)
for i, obj in enumerate(objs):
- print(' %d: %s, area=%.2f' % (i, obj, obj.bbox.area()))
+ print(' %d: %s, area=%.2f' % (i, obj, obj.bbox.area))
def render_gen(args):
fps_counter = utils.avg_fps_counter(30)
- engines, titles = utils.make_engines(args.model, DetectionEngine)
- assert utils.same_input_image_sizes(engines)
- engines = itertools.cycle(engines)
- engine = next(engines)
+ interpreters, titles = utils.make_interpreters(args.model)
+ assert utils.same_input_image_sizes(interpreters)
+ interpreters = itertools.cycle(interpreters)
+ interpreter = next(interpreters)
labels = utils.load_labels(args.labels) if args.labels else None
filtered_labels = set(l.strip() for l in args.filter.split(',')) if args.filter else None
@@ -152,7 +136,8 @@
draw_overlay = True
- yield utils.input_image_size(engine)
+ width, height = utils.input_image_size(interpreter)
+ yield width, height
output = None
while True:
@@ -161,27 +146,28 @@
inference_rate = next(fps_counter)
if draw_overlay:
start = time.monotonic()
- objs = engine .detect_with_input_tensor(tensor, threshold=args.threshold, top_k=args.top_k)
+ edgetpu.run_inference(interpreter, tensor)
inference_time = time.monotonic() - start
- objs = [convert(obj, labels) for obj in objs]
+ objs = detect.get_objects(interpreter, args.threshold)[:args.top_k]
if labels and filtered_labels:
- objs = [obj for obj in objs if obj.label in filtered_labels]
+ objs = [obj for obj in objs if labels[obj.id] in filtered_labels]
- objs = [obj for obj in objs if args.min_area <= obj.bbox.area() <= args.max_area]
+ objs = [obj for obj in objs \
+ if args.min_area <= obj.bbox.scale(1.0 / width, 1.0 / height).area <= args.max_area]
if args.print:
print_results(inference_rate, objs)
- title = titles[engine]
- output = overlay(title, objs, get_color, inference_time, inference_rate, layout)
+ title = titles[interpreter]
+ output = overlay(title, objs, get_color, labels, inference_time, inference_rate, layout)
else:
output = None
if command == 'o':
draw_overlay = not draw_overlay
elif command == 'n':
- engine = next(engines)
+ interpreter = next(interpreters)
def add_render_gen_args(parser):
parser.add_argument('--model',
diff --git a/edgetpuvision/utils.py b/edgetpuvision/utils.py
index df0892d..8a06c05 100644
--- a/edgetpuvision/utils.py
+++ b/edgetpuvision/utils.py
@@ -17,6 +17,9 @@
import re
import time
+from pycoral.adapters import common
+from pycoral.utils import edgetpu
+
LABEL_PATTERN = re.compile(r'\s*(\d+)(.+)')
def load_labels(path):
@@ -24,13 +27,11 @@
lines = (LABEL_PATTERN.match(line).groups() for line in f.readlines())
return {int(num): text.strip() for num, text in lines}
+def input_image_size(interpreter):
+ return common.input_size(interpreter)
-def input_image_size(engine):
- _, h, w, _ = engine.get_input_tensor_shape()
- return w, h
-
-def same_input_image_sizes(engines):
- return len({input_image_size(engine) for engine in engines}) == 1
+def same_input_image_sizes(interpreters):
+ return len({input_image_size(interpreter) for interpreter in interpreters}) == 1
def avg_fps_counter(window_size):
window = collections.deque(maxlen=window_size)
@@ -43,14 +44,15 @@
prev = curr
yield len(window) / sum(window)
-def make_engines(models, engine_class):
- engines, titles = [], {}
+def make_interpreters(models):
+ interpreters, titles = [], {}
for model in models.split(','):
if '@' in model:
model_path, title = model.split('@')
else:
model_path, title = model, os.path.basename(os.path.normpath(model))
- engine = engine_class(model_path)
- engines.append(engine)
- titles[engine] = title
- return engines, titles
+ interpreter = edgetpu.make_interpreter(model_path)
+ interpreter.allocate_tensors()
+ interpreters.append(interpreter)
+ titles[interpreter] = title
+ return interpreters, titles
