opencv/modules/gapi/misc/python/test/test_gapi_streaming.py

560 lines
19 KiB
Python
Raw Normal View History

#!/usr/bin/env python
import numpy as np
import cv2 as cv
import os
import sys
import unittest
import time
from tests_common import NewOpenCVTests
try:
if sys.version_info[:2] < (3, 0):
raise unittest.SkipTest('Python 2.x is not supported')
@cv.gapi.op('custom.delay', in_types=[cv.GMat], out_types=[cv.GMat])
class GDelay:
Merge pull request #23597 from dmatveev:dm/gapi_onnx_py_integration G-API: Integration branch for ONNX & Python-related changes #23597 # Changes overview ## 1. Expose ONNX backend's Normalization and Mean-value parameters in Python * Since Python G-API bindings rely on `Generic` infer to express Inference, the `Generic` specialization of `onnx::Params` was extended with new methods to control normalization (`/255`) and mean-value; these methods were exposed in the Python bindings * Found some questionable parts in the existing API which I'd like to review/discuss (see comments) UPD: 1. Thanks to @TolyaTalamanov normalization inconsistencies have been identified with `squeezenet1.0-9` ONNX model itself; tests using these model were updated to DISABLE normalization and NOT using mean/value. 2. Questionable parts were removed and tests still pass. ### Details (taken from @TolyaTalamanov's comment): `squeezenet1.0.*onnx` - doesn't require scaling to [0,1] and mean/std because the weights of the first convolution already scaled. ONNX documentation is broken. So the correct approach to use this models is: 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 but without normalization step: ``` # DON'T DO IT: # mean_vec = np.array([0.485, 0.456, 0.406]) # stddev_vec = np.array([0.229, 0.224, 0.225]) # norm_img_data = np.zeros(img_data.shape).astype('float32') # for i in range(img_data.shape[0]): # norm_img_data[i,:,:] = (img_data[i,:,:]/255 - mean_vec[i]) / stddev_vec[i] # # add batch channel # norm_img_data = norm_img_data.reshape(1, 3, 224, 224).astype('float32') # return norm_img_data # INSTEAD return img_data.reshape(1, 3, 224, 224) ``` 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', False) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. --- `squeezenet1.1.*onnx` - requires scaling to [0,1] and mean/std - onnx documentation is correct. 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', True) // default net.cfgMeanStd('data_0', [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. ## 2. Expose Fluid & kernel package-related functionality in Python * `cv::gapi::combine()` * `cv::GKernelPackage::size()` (mainly for testing purposes) * `cv::gapi::imgproc::fluid::kernels()` Added a test for the above. ## 3. Fixed issues with Python stateful kernel handling Fixed error message when `outMeta()` of custom python operation fails. ## 4. Fixed various issues in Python tests 1. `test_gapi_streaming.py` - fixed behavior of Desync test to avoid sporadic issues 2. `test_gapi_infer_onnx.py` - fixed model lookup (it was still using the ONNX Zoo layout but was NOT using the proper env var we use to point to one). ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake
2023-05-30 22:52:17 +08:00
"""Delay for 50 ms."""
@staticmethod
def outMeta(desc):
return desc
@cv.gapi.kernel(GDelay)
class GDelayImpl:
"""Implementation for GDelay operation."""
@staticmethod
def run(img):
Merge pull request #23597 from dmatveev:dm/gapi_onnx_py_integration G-API: Integration branch for ONNX & Python-related changes #23597 # Changes overview ## 1. Expose ONNX backend's Normalization and Mean-value parameters in Python * Since Python G-API bindings rely on `Generic` infer to express Inference, the `Generic` specialization of `onnx::Params` was extended with new methods to control normalization (`/255`) and mean-value; these methods were exposed in the Python bindings * Found some questionable parts in the existing API which I'd like to review/discuss (see comments) UPD: 1. Thanks to @TolyaTalamanov normalization inconsistencies have been identified with `squeezenet1.0-9` ONNX model itself; tests using these model were updated to DISABLE normalization and NOT using mean/value. 2. Questionable parts were removed and tests still pass. ### Details (taken from @TolyaTalamanov's comment): `squeezenet1.0.*onnx` - doesn't require scaling to [0,1] and mean/std because the weights of the first convolution already scaled. ONNX documentation is broken. So the correct approach to use this models is: 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 but without normalization step: ``` # DON'T DO IT: # mean_vec = np.array([0.485, 0.456, 0.406]) # stddev_vec = np.array([0.229, 0.224, 0.225]) # norm_img_data = np.zeros(img_data.shape).astype('float32') # for i in range(img_data.shape[0]): # norm_img_data[i,:,:] = (img_data[i,:,:]/255 - mean_vec[i]) / stddev_vec[i] # # add batch channel # norm_img_data = norm_img_data.reshape(1, 3, 224, 224).astype('float32') # return norm_img_data # INSTEAD return img_data.reshape(1, 3, 224, 224) ``` 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', False) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. --- `squeezenet1.1.*onnx` - requires scaling to [0,1] and mean/std - onnx documentation is correct. 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', True) // default net.cfgMeanStd('data_0', [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. ## 2. Expose Fluid & kernel package-related functionality in Python * `cv::gapi::combine()` * `cv::GKernelPackage::size()` (mainly for testing purposes) * `cv::gapi::imgproc::fluid::kernels()` Added a test for the above. ## 3. Fixed issues with Python stateful kernel handling Fixed error message when `outMeta()` of custom python operation fails. ## 4. Fixed various issues in Python tests 1. `test_gapi_streaming.py` - fixed behavior of Desync test to avoid sporadic issues 2. `test_gapi_infer_onnx.py` - fixed model lookup (it was still using the ONNX Zoo layout but was NOT using the proper env var we use to point to one). ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake
2023-05-30 22:52:17 +08:00
time.sleep(0.05)
return img
def convertNV12p2BGR(in_nv12):
shape = in_nv12.shape
y_height = shape[0] // 3 * 2
uv_shape = (shape[0] // 3, shape[1])
new_uv_shape = (uv_shape[0], uv_shape[1] // 2, 2)
return cv.cvtColorTwoPlane(in_nv12[:y_height, :],
in_nv12[ y_height:, :].reshape(new_uv_shape),
cv.COLOR_YUV2BGR_NV12)
class test_gapi_streaming(NewOpenCVTests):
def test_image_input(self):
sz = (1280, 720)
in_mat = np.random.randint(0, 100, sz).astype(np.uint8)
# OpenCV
expected = cv.medianBlur(in_mat, 3)
# G-API
g_in = cv.GMat()
g_out = cv.gapi.medianBlur(g_in, 3)
c = cv.GComputation(g_in, g_out)
ccomp = c.compileStreaming(cv.gapi.descr_of(in_mat))
ccomp.setSource(cv.gin(in_mat))
ccomp.start()
_, actual = ccomp.pull()
# Assert
self.assertEqual(0.0, cv.norm(expected, actual, cv.NORM_INF))
def test_video_input(self):
ksize = 3
path = self.find_file('cv/video/768x576.avi', [os.environ['OPENCV_TEST_DATA_PATH']])
# OpenCV
cap = cv.VideoCapture(path)
# G-API
g_in = cv.GMat()
g_out = cv.gapi.medianBlur(g_in, ksize)
c = cv.GComputation(g_in, g_out)
ccomp = c.compileStreaming()
source = cv.gapi.wip.make_capture_src(path)
ccomp.setSource(cv.gin(source))
ccomp.start()
# Assert
max_num_frames = 10
proc_num_frames = 0
while cap.isOpened():
has_expected, expected = cap.read()
has_actual, actual = ccomp.pull()
self.assertEqual(has_expected, has_actual)
if not has_actual:
break
self.assertEqual(0.0, cv.norm(cv.medianBlur(expected, ksize), actual, cv.NORM_INF))
proc_num_frames += 1
if proc_num_frames == max_num_frames:
break
def test_video_split3(self):
path = self.find_file('cv/video/768x576.avi', [os.environ['OPENCV_TEST_DATA_PATH']])
# OpenCV
cap = cv.VideoCapture(path)
# G-API
g_in = cv.GMat()
b, g, r = cv.gapi.split3(g_in)
c = cv.GComputation(cv.GIn(g_in), cv.GOut(b, g, r))
ccomp = c.compileStreaming()
source = cv.gapi.wip.make_capture_src(path)
ccomp.setSource(cv.gin(source))
ccomp.start()
# Assert
max_num_frames = 10
proc_num_frames = 0
while cap.isOpened():
has_expected, frame = cap.read()
has_actual, actual = ccomp.pull()
self.assertEqual(has_expected, has_actual)
if not has_actual:
break
expected = cv.split(frame)
for e, a in zip(expected, actual):
self.assertEqual(0.0, cv.norm(e, a, cv.NORM_INF))
proc_num_frames += 1
if proc_num_frames == max_num_frames:
break
def test_video_add(self):
sz = (576, 768, 3)
in_mat = np.random.randint(0, 100, sz).astype(np.uint8)
path = self.find_file('cv/video/768x576.avi', [os.environ['OPENCV_TEST_DATA_PATH']])
# OpenCV
cap = cv.VideoCapture(path)
# G-API
g_in1 = cv.GMat()
g_in2 = cv.GMat()
out = cv.gapi.add(g_in1, g_in2)
c = cv.GComputation(cv.GIn(g_in1, g_in2), cv.GOut(out))
ccomp = c.compileStreaming()
source = cv.gapi.wip.make_capture_src(path)
ccomp.setSource(cv.gin(source, in_mat))
ccomp.start()
# Assert
max_num_frames = 10
proc_num_frames = 0
while cap.isOpened():
has_expected, frame = cap.read()
has_actual, actual = ccomp.pull()
self.assertEqual(has_expected, has_actual)
if not has_actual:
break
expected = cv.add(frame, in_mat)
self.assertEqual(0.0, cv.norm(expected, actual, cv.NORM_INF))
proc_num_frames += 1
if proc_num_frames == max_num_frames:
break
def test_video_good_features_to_track(self):
path = self.find_file('cv/video/768x576.avi', [os.environ['OPENCV_TEST_DATA_PATH']])
# NB: goodFeaturesToTrack configuration
max_corners = 50
quality_lvl = 0.01
min_distance = 10
block_sz = 3
use_harris_detector = True
k = 0.04
mask = None
# OpenCV
cap = cv.VideoCapture(path)
# G-API
g_in = cv.GMat()
g_gray = cv.gapi.RGB2Gray(g_in)
g_out = cv.gapi.goodFeaturesToTrack(g_gray, max_corners, quality_lvl,
min_distance, mask, block_sz, use_harris_detector, k)
c = cv.GComputation(cv.GIn(g_in), cv.GOut(g_out))
ccomp = c.compileStreaming()
source = cv.gapi.wip.make_capture_src(path)
ccomp.setSource(cv.gin(source))
ccomp.start()
# Assert
max_num_frames = 10
proc_num_frames = 0
while cap.isOpened():
has_expected, frame = cap.read()
has_actual, actual = ccomp.pull()
self.assertEqual(has_expected, has_actual)
if not has_actual:
break
# OpenCV
frame = cv.cvtColor(frame, cv.COLOR_RGB2GRAY)
expected = cv.goodFeaturesToTrack(frame, max_corners, quality_lvl,
min_distance, mask=mask,
blockSize=block_sz, useHarrisDetector=use_harris_detector, k=k)
for e, a in zip(expected, actual):
# NB: OpenCV & G-API have different output shapes:
# OpenCV - (num_points, 1, 2)
# G-API - (num_points, 2)
self.assertEqual(0.0, cv.norm(e.flatten(),
np.array(a, np.float32).flatten(),
cv.NORM_INF))
proc_num_frames += 1
if proc_num_frames == max_num_frames:
break
def test_gapi_streaming_meta(self):
path = self.find_file('cv/video/768x576.avi', [os.environ['OPENCV_TEST_DATA_PATH']])
# G-API
g_in = cv.GMat()
g_ts = cv.gapi.streaming.timestamp(g_in)
g_seqno = cv.gapi.streaming.seqNo(g_in)
g_seqid = cv.gapi.streaming.seq_id(g_in)
c = cv.GComputation(cv.GIn(g_in), cv.GOut(g_ts, g_seqno, g_seqid))
ccomp = c.compileStreaming()
source = cv.gapi.wip.make_capture_src(path)
ccomp.setSource(cv.gin(source))
ccomp.start()
# Assert
max_num_frames = 10
curr_frame_number = 0
while True:
has_frame, (ts, seqno, seqid) = ccomp.pull()
if not has_frame:
break
self.assertEqual(curr_frame_number, seqno)
self.assertEqual(curr_frame_number, seqid)
curr_frame_number += 1
if curr_frame_number == max_num_frames:
break
def test_desync(self):
path = self.find_file('cv/video/768x576.avi', [os.environ['OPENCV_TEST_DATA_PATH']])
# G-API
g_in = cv.GMat()
g_out1 = cv.gapi.copy(g_in)
des = cv.gapi.streaming.desync(g_in)
g_out2 = GDelay.on(des)
c = cv.GComputation(cv.GIn(g_in), cv.GOut(g_out1, g_out2))
kernels = cv.gapi.kernels(GDelayImpl)
ccomp = c.compileStreaming(args=cv.gapi.compile_args(kernels))
source = cv.gapi.wip.make_capture_src(path)
ccomp.setSource(cv.gin(source))
ccomp.start()
# Assert
Merge pull request #23597 from dmatveev:dm/gapi_onnx_py_integration G-API: Integration branch for ONNX & Python-related changes #23597 # Changes overview ## 1. Expose ONNX backend's Normalization and Mean-value parameters in Python * Since Python G-API bindings rely on `Generic` infer to express Inference, the `Generic` specialization of `onnx::Params` was extended with new methods to control normalization (`/255`) and mean-value; these methods were exposed in the Python bindings * Found some questionable parts in the existing API which I'd like to review/discuss (see comments) UPD: 1. Thanks to @TolyaTalamanov normalization inconsistencies have been identified with `squeezenet1.0-9` ONNX model itself; tests using these model were updated to DISABLE normalization and NOT using mean/value. 2. Questionable parts were removed and tests still pass. ### Details (taken from @TolyaTalamanov's comment): `squeezenet1.0.*onnx` - doesn't require scaling to [0,1] and mean/std because the weights of the first convolution already scaled. ONNX documentation is broken. So the correct approach to use this models is: 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 but without normalization step: ``` # DON'T DO IT: # mean_vec = np.array([0.485, 0.456, 0.406]) # stddev_vec = np.array([0.229, 0.224, 0.225]) # norm_img_data = np.zeros(img_data.shape).astype('float32') # for i in range(img_data.shape[0]): # norm_img_data[i,:,:] = (img_data[i,:,:]/255 - mean_vec[i]) / stddev_vec[i] # # add batch channel # norm_img_data = norm_img_data.reshape(1, 3, 224, 224).astype('float32') # return norm_img_data # INSTEAD return img_data.reshape(1, 3, 224, 224) ``` 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', False) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. --- `squeezenet1.1.*onnx` - requires scaling to [0,1] and mean/std - onnx documentation is correct. 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', True) // default net.cfgMeanStd('data_0', [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. ## 2. Expose Fluid & kernel package-related functionality in Python * `cv::gapi::combine()` * `cv::GKernelPackage::size()` (mainly for testing purposes) * `cv::gapi::imgproc::fluid::kernels()` Added a test for the above. ## 3. Fixed issues with Python stateful kernel handling Fixed error message when `outMeta()` of custom python operation fails. ## 4. Fixed various issues in Python tests 1. `test_gapi_streaming.py` - fixed behavior of Desync test to avoid sporadic issues 2. `test_gapi_infer_onnx.py` - fixed model lookup (it was still using the ONNX Zoo layout but was NOT using the proper env var we use to point to one). ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake
2023-05-30 22:52:17 +08:00
max_num_frames = 50
out_counter = 0
desync_out_counter = 0
none_counter = 0
while True:
has_frame, (out1, out2) = ccomp.pull()
if not has_frame:
break
if not out1 is None:
out_counter += 1
if not out2 is None:
desync_out_counter += 1
else:
none_counter += 1
Merge pull request #23597 from dmatveev:dm/gapi_onnx_py_integration G-API: Integration branch for ONNX & Python-related changes #23597 # Changes overview ## 1. Expose ONNX backend's Normalization and Mean-value parameters in Python * Since Python G-API bindings rely on `Generic` infer to express Inference, the `Generic` specialization of `onnx::Params` was extended with new methods to control normalization (`/255`) and mean-value; these methods were exposed in the Python bindings * Found some questionable parts in the existing API which I'd like to review/discuss (see comments) UPD: 1. Thanks to @TolyaTalamanov normalization inconsistencies have been identified with `squeezenet1.0-9` ONNX model itself; tests using these model were updated to DISABLE normalization and NOT using mean/value. 2. Questionable parts were removed and tests still pass. ### Details (taken from @TolyaTalamanov's comment): `squeezenet1.0.*onnx` - doesn't require scaling to [0,1] and mean/std because the weights of the first convolution already scaled. ONNX documentation is broken. So the correct approach to use this models is: 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 but without normalization step: ``` # DON'T DO IT: # mean_vec = np.array([0.485, 0.456, 0.406]) # stddev_vec = np.array([0.229, 0.224, 0.225]) # norm_img_data = np.zeros(img_data.shape).astype('float32') # for i in range(img_data.shape[0]): # norm_img_data[i,:,:] = (img_data[i,:,:]/255 - mean_vec[i]) / stddev_vec[i] # # add batch channel # norm_img_data = norm_img_data.reshape(1, 3, 224, 224).astype('float32') # return norm_img_data # INSTEAD return img_data.reshape(1, 3, 224, 224) ``` 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', False) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. --- `squeezenet1.1.*onnx` - requires scaling to [0,1] and mean/std - onnx documentation is correct. 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', True) // default net.cfgMeanStd('data_0', [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. ## 2. Expose Fluid & kernel package-related functionality in Python * `cv::gapi::combine()` * `cv::GKernelPackage::size()` (mainly for testing purposes) * `cv::gapi::imgproc::fluid::kernels()` Added a test for the above. ## 3. Fixed issues with Python stateful kernel handling Fixed error message when `outMeta()` of custom python operation fails. ## 4. Fixed various issues in Python tests 1. `test_gapi_streaming.py` - fixed behavior of Desync test to avoid sporadic issues 2. `test_gapi_infer_onnx.py` - fixed model lookup (it was still using the ONNX Zoo layout but was NOT using the proper env var we use to point to one). ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake
2023-05-30 22:52:17 +08:00
if out_counter == max_num_frames:
ccomp.stop()
break
Merge pull request #23597 from dmatveev:dm/gapi_onnx_py_integration G-API: Integration branch for ONNX & Python-related changes #23597 # Changes overview ## 1. Expose ONNX backend's Normalization and Mean-value parameters in Python * Since Python G-API bindings rely on `Generic` infer to express Inference, the `Generic` specialization of `onnx::Params` was extended with new methods to control normalization (`/255`) and mean-value; these methods were exposed in the Python bindings * Found some questionable parts in the existing API which I'd like to review/discuss (see comments) UPD: 1. Thanks to @TolyaTalamanov normalization inconsistencies have been identified with `squeezenet1.0-9` ONNX model itself; tests using these model were updated to DISABLE normalization and NOT using mean/value. 2. Questionable parts were removed and tests still pass. ### Details (taken from @TolyaTalamanov's comment): `squeezenet1.0.*onnx` - doesn't require scaling to [0,1] and mean/std because the weights of the first convolution already scaled. ONNX documentation is broken. So the correct approach to use this models is: 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 but without normalization step: ``` # DON'T DO IT: # mean_vec = np.array([0.485, 0.456, 0.406]) # stddev_vec = np.array([0.229, 0.224, 0.225]) # norm_img_data = np.zeros(img_data.shape).astype('float32') # for i in range(img_data.shape[0]): # norm_img_data[i,:,:] = (img_data[i,:,:]/255 - mean_vec[i]) / stddev_vec[i] # # add batch channel # norm_img_data = norm_img_data.reshape(1, 3, 224, 224).astype('float32') # return norm_img_data # INSTEAD return img_data.reshape(1, 3, 224, 224) ``` 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', False) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. --- `squeezenet1.1.*onnx` - requires scaling to [0,1] and mean/std - onnx documentation is correct. 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', True) // default net.cfgMeanStd('data_0', [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. ## 2. Expose Fluid & kernel package-related functionality in Python * `cv::gapi::combine()` * `cv::GKernelPackage::size()` (mainly for testing purposes) * `cv::gapi::imgproc::fluid::kernels()` Added a test for the above. ## 3. Fixed issues with Python stateful kernel handling Fixed error message when `outMeta()` of custom python operation fails. ## 4. Fixed various issues in Python tests 1. `test_gapi_streaming.py` - fixed behavior of Desync test to avoid sporadic issues 2. `test_gapi_infer_onnx.py` - fixed model lookup (it was still using the ONNX Zoo layout but was NOT using the proper env var we use to point to one). ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake
2023-05-30 22:52:17 +08:00
self.assertLess(0, out_counter)
self.assertLess(desync_out_counter, out_counter)
self.assertLess(0, none_counter)
def test_compile_streaming_empty(self):
g_in = cv.GMat()
comp = cv.GComputation(g_in, cv.gapi.medianBlur(g_in, 3))
comp.compileStreaming()
def test_compile_streaming_args(self):
g_in = cv.GMat()
comp = cv.GComputation(g_in, cv.gapi.medianBlur(g_in, 3))
comp.compileStreaming(cv.gapi.compile_args(cv.gapi.streaming.queue_capacity(1)))
def test_compile_streaming_descr_of(self):
g_in = cv.GMat()
comp = cv.GComputation(g_in, cv.gapi.medianBlur(g_in, 3))
img = np.zeros((3,300,300), dtype=np.float32)
comp.compileStreaming(cv.gapi.descr_of(img))
def test_compile_streaming_descr_of_and_args(self):
g_in = cv.GMat()
comp = cv.GComputation(g_in, cv.gapi.medianBlur(g_in, 3))
img = np.zeros((3,300,300), dtype=np.float32)
comp.compileStreaming(cv.gapi.descr_of(img),
cv.gapi.compile_args(cv.gapi.streaming.queue_capacity(1)))
def test_compile_streaming_meta(self):
g_in = cv.GMat()
comp = cv.GComputation(g_in, cv.gapi.medianBlur(g_in, 3))
img = np.zeros((3,300,300), dtype=np.float32)
comp.compileStreaming([cv.GMatDesc(cv.CV_8U, 3, (300, 300))])
def test_compile_streaming_meta_and_args(self):
g_in = cv.GMat()
comp = cv.GComputation(g_in, cv.gapi.medianBlur(g_in, 3))
img = np.zeros((3,300,300), dtype=np.float32)
comp.compileStreaming([cv.GMatDesc(cv.CV_8U, 3, (300, 300))],
cv.gapi.compile_args(cv.gapi.streaming.queue_capacity(1)))
def get_gst_source(self, gstpipeline):
# NB: Skip test in case gstreamer isn't available.
try:
return cv.gapi.wip.make_gst_src(gstpipeline)
except cv.error as e:
if str(e).find('Built without GStreamer support!') == -1:
raise e
else:
raise unittest.SkipTest(str(e))
def test_gst_source(self):
if not cv.videoio_registry.hasBackend(cv.CAP_GSTREAMER):
raise unittest.SkipTest("Backend is not available/disabled: GSTREAMER")
gstpipeline = """videotestsrc is-live=true pattern=colors num-buffers=10 !
videorate ! videoscale ! video/x-raw,width=1920,height=1080,
framerate=30/1 ! appsink"""
g_in = cv.GMat()
g_out = cv.gapi.copy(g_in)
c = cv.GComputation(cv.GIn(g_in), cv.GOut(g_out))
ccomp = c.compileStreaming()
source = self.get_gst_source(gstpipeline)
ccomp.setSource(cv.gin(source))
ccomp.start()
has_frame, output = ccomp.pull()
while has_frame:
self.assertTrue(output.size != 0)
has_frame, output = ccomp.pull()
def open_VideoCapture_gstreamer(self, gstpipeline):
try:
cap = cv.VideoCapture(gstpipeline, cv.CAP_GSTREAMER)
except Exception as e:
raise unittest.SkipTest("Backend GSTREAMER can't open the video; " +
"cause: " + str(e))
if not cap.isOpened():
raise unittest.SkipTest("Backend GSTREAMER can't open the video")
return cap
def test_gst_source_accuracy(self):
if not cv.videoio_registry.hasBackend(cv.CAP_GSTREAMER):
raise unittest.SkipTest("Backend is not available/disabled: GSTREAMER")
path = self.find_file('highgui/video/big_buck_bunny.avi',
[os.environ['OPENCV_TEST_DATA_PATH']])
gstpipeline = """filesrc location=""" + path + """ ! decodebin ! videoconvert !
videoscale ! video/x-raw,format=NV12 ! appsink"""
# G-API pipeline
g_in = cv.GMat()
g_out = cv.gapi.copy(g_in)
c = cv.GComputation(cv.GIn(g_in), cv.GOut(g_out))
ccomp = c.compileStreaming()
# G-API Gst-source
source = self.get_gst_source(gstpipeline)
ccomp.setSource(cv.gin(source))
ccomp.start()
# OpenCV Gst-source
cap = self.open_VideoCapture_gstreamer(gstpipeline)
# Assert
max_num_frames = 10
for _ in range(max_num_frames):
has_expected, expected = cap.read()
has_actual, actual = ccomp.pull()
self.assertEqual(has_expected, has_actual)
if not has_expected:
break
self.assertEqual(0.0, cv.norm(convertNV12p2BGR(expected), actual, cv.NORM_INF))
def get_gst_pipeline(self, gstpipeline):
# NB: Skip test in case gstreamer isn't available.
try:
return cv.gapi.wip.GStreamerPipeline(gstpipeline)
except cv.error as e:
if str(e).find('Built without GStreamer support!') == -1:
raise e
else:
raise unittest.SkipTest(str(e))
except SystemError as e:
raise unittest.SkipTest(str(e) + ", caused by " + str(e.__cause__))
def test_gst_multiple_sources(self):
if not cv.videoio_registry.hasBackend(cv.CAP_GSTREAMER):
raise unittest.SkipTest("Backend is not available/disabled: GSTREAMER")
gstpipeline = """videotestsrc is-live=true pattern=colors num-buffers=10 !
videorate ! videoscale !
video/x-raw,width=1920,height=1080,framerate=30/1 !
appsink name=sink1
videotestsrc is-live=true pattern=colors num-buffers=10 !
videorate ! videoscale !
video/x-raw,width=1920,height=1080,framerate=30/1 !
appsink name=sink2"""
g_in1 = cv.GMat()
g_in2 = cv.GMat()
g_out = cv.gapi.add(g_in1, g_in2)
c = cv.GComputation(cv.GIn(g_in1, g_in2), cv.GOut(g_out))
ccomp = c.compileStreaming()
pp = self.get_gst_pipeline(gstpipeline)
src1 = cv.gapi.wip.get_streaming_source(pp, "sink1")
src2 = cv.gapi.wip.get_streaming_source(pp, "sink2")
ccomp.setSource(cv.gin(src1, src2))
ccomp.start()
has_frame, out = ccomp.pull()
while has_frame:
self.assertTrue(out.size != 0)
has_frame, out = ccomp.pull()
def test_gst_multiple_sources_accuracy(self):
if not cv.videoio_registry.hasBackend(cv.CAP_GSTREAMER):
raise unittest.SkipTest("Backend is not available/disabled: GSTREAMER")
path = self.find_file('highgui/video/big_buck_bunny.avi',
[os.environ['OPENCV_TEST_DATA_PATH']])
gstpipeline1 = """filesrc location=""" + path + """ ! decodebin ! videoconvert !
videoscale ! video/x-raw,format=NV12 ! appsink"""
gstpipeline2 = """filesrc location=""" + path + """ ! decodebin !
videoflip method=clockwise ! videoconvert ! videoscale !
video/x-raw,format=NV12 ! appsink"""
gstpipeline_gapi = gstpipeline1 + ' name=sink1 ' + gstpipeline2 + ' name=sink2'
# G-API pipeline
g_in1 = cv.GMat()
g_in2 = cv.GMat()
g_out1 = cv.gapi.copy(g_in1)
g_out2 = cv.gapi.copy(g_in2)
c = cv.GComputation(cv.GIn(g_in1, g_in2), cv.GOut(g_out1, g_out2))
ccomp = c.compileStreaming()
# G-API Gst-source
pp = self.get_gst_pipeline(gstpipeline_gapi)
src1 = cv.gapi.wip.get_streaming_source(pp, "sink1")
src2 = cv.gapi.wip.get_streaming_source(pp, "sink2")
ccomp.setSource(cv.gin(src1, src2))
ccomp.start()
# OpenCV Gst-source
cap1 = self.open_VideoCapture_gstreamer(gstpipeline1)
cap2 = self.open_VideoCapture_gstreamer(gstpipeline2)
# Assert
max_num_frames = 10
for _ in range(max_num_frames):
has_expected1, expected1 = cap1.read()
has_expected2, expected2 = cap2.read()
has_actual, (actual1, actual2) = ccomp.pull()
self.assertEqual(has_expected1, has_expected2)
has_expected = has_expected1 and has_expected2
self.assertEqual(has_expected, has_actual)
if not has_expected:
break
self.assertEqual(0.0, cv.norm(convertNV12p2BGR(expected1), actual1, cv.NORM_INF))
self.assertEqual(0.0, cv.norm(convertNV12p2BGR(expected2), actual2, cv.NORM_INF))
except unittest.SkipTest as e:
message = str(e)
class TestSkip(unittest.TestCase):
def setUp(self):
self.skipTest('Skip tests: ' + message)
def test_skip():
pass
pass
if __name__ == '__main__':
NewOpenCVTests.bootstrap()