def CIFAR10Test(loopCnt=600, isFull=False, barLen=105):
pyb.LED(1).off()
sensor.reset() # Reset and initialize the sensor.
sensor.set_contrast(3)
sensor.set_pixformat(
sensor.RGB565) # Set pixel format to RGB565 (or GRAYSCALE)
sensor.set_framesize(sensor.VGA) # Set frame size to QVGA (320x240)
sensor.set_windowing((192, 192)) # Set window
sensor.skip_frames(time=300) # Wait for settings take effect.
sensor.set_auto_gain(False)
#sensor.set_framerate(0<<9|1<<12)
if isFull:
net = nn.load('/cifar10.network')
else:
net = nn.load('/cifar10_fast.network')
labels = [
'plane', 'auto', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship',
'truck'
]
clock = time.clock()
tAvg = 0.0
startTick = time.ticks()
while (True):
if time.ticks() - startTick > loopCnt:
break
clock.tick()
img = sensor.snapshot()
t0 = time.ticks()
lst = net.search(img, threshold=0.640, min_scale=1, scale_mul=0.8, \
x_overlap=-1, y_overlap=-1, contrast_threshold=0.5)
t1 = time.ticks() - t0
tAvg = tAvg * 0.9 + t1 * 0.1
img.draw_string(
4,
8,
'CIFAR-10: classify:\nplane,auto,cat,dog,\ndeer,horse,frog,ship,\ntruck,horse',
color=(0, 0, 0))
lnLen = (barLen * (loopCnt - (time.ticks() - startTick))) // loopCnt
DrawPgsBar(img, barLen, loopCnt, startTick)
for obj in lst:
print(' %s - Confidence %f%%' %
(labels[obj.index()], obj.value()))
rc = obj.rect()
#img.draw_rectangle(rc, color=(255,255,255))
img.draw_rectangle(barLen + 10,
1,
50,
8,
fill=True,
color=(0, 0, 0))
img.draw_string(barLen + 10, 0, labels[obj.index()])
print('algo time cost : %.2f ms' % (tAvg))
def trainAll(params,
trainOpt,
trainInput,
trainTarget,
trainInputWeights,
validInput,
validTarget,
validInputWeights,
initWeights=None):
model = nn.load(params['modelFilename'])
if initWeights is not None:
model.loadWeights(initWeights)
trainer = nn.Trainer(name=params['name'],
model=model,
trainOpt=trainOpt,
outputFolder=params['outputFolder'])
# Combine train & valid set
allInput, allTarget, allInputWeights = combineInputs(
trainInput, trainTarget, trainInputWeights, validInput, validTarget,
validInputWeights)
trainer.train(trainInput=allInput,
trainTarget=allTarget,
trainInputWeights=allInputWeights)
return model, trainer
def trainValid(params,
trainOpt,
trainInput,
trainTarget,
trainInputWeights,
validInput,
validTarget,
validInputWeights,
initWeights=None):
model = nn.load(params['modelFilename'])
if initWeights is not None:
model.loadWeights(initWeights)
trainer = nn.Trainer(
name=params['name']+\
('-v' if params['validDataFilename'] is not None else ''),
model=model,
trainOpt=trainOpt,
outputFolder=params['outputFolder']
)
# Validation training
trainer.train(trainInput=trainInput,
trainTarget=trainTarget,
trainInputWeights=trainInputWeights,
validInput=validInput,
validTarget=validTarget,
validInputWeights=validInputWeights)
return model, trainer
def trainValid(
params,
trainOpt,
trainInput,
trainTarget,
trainInputWeights,
validInput,
validTarget,
validInputWeights,
initWeights=None,
):
model = nn.load(params["modelFilename"])
if initWeights is not None:
model.loadWeights(initWeights)
trainer = nn.Trainer(
name=params["name"] + ("-v" if params["validDataFilename"] is not None else ""),
model=model,
trainOpt=trainOpt,
outputFolder=params["outputFolder"],
)
# Validation training
trainer.train(
trainInput=trainInput,
trainTarget=trainTarget,
trainInputWeights=trainInputWeights,
validInput=validInput,
validTarget=validTarget,
validInputWeights=validInputWeights,
)
return model, trainer
def loadModel(
taskId,
resultsFolder):
print 'Loading model...'
modelSpecFile = '%s/%s/%s.model.yml' % (resultsFolder, taskId, taskId)
modelWeightsFile = '%s/%s/%s.w.npy' % (resultsFolder, taskId, taskId)
model = nn.load(modelSpecFile)
model.loadWeights(np.load(modelWeightsFile))
return model
def loadEnsemble(
taskIds,
resultsFolder):
"""
Load class specific models.
"""
models = []
for taskId in taskIds:
taskFolder = os.path.join(resultsFolder, taskId)
modelSpec = os.path.join(taskFolder, '%s.model.yml' % taskId)
modelWeights = os.path.join(taskFolder, '%s.w.npy' % taskId)
model = nn.load(modelSpec)
model.loadWeights(np.load(modelWeights))
models.append(model)
return models
def main(argv):
model = nn.load(argv[1])
gui = paint.gui()
print("past init")
while True:
time.sleep(2 - time.monotonic() % 2)
image = gui.get_image()
data = np.array(image).astype(np.float32)
data = skimage.measure.block_reduce(data,
block_size=(10, 10),
func=np.mean)
data = torch.from_numpy(data)
data = data[None, None, :, :] # correct the number of dimensions
print(data.size())
nn.predict(model, data)
def unittest(data_path, temp_path):
import os, nn, image
match = 0
test_dir = data_path+"/cifar10"
files = os.listdir(test_dir)
total = len(files)
labels = ['airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck']
net = nn.load(data_path+'/cifar10_fast.network')
for f in sorted(files):
img = image.Image(test_dir+"/"+f, copy_to_fb=True)
out = net.forward(img)
label = labels[out.index(max(out))]
if (f.split('-')[0] == label):
match += 1
return ((match / total * 100 ) > 90)
def unittest(data_path, temp_path):
import os, nn, image
match = 0
test_dir = data_path+"/cifar10"
files = os.listdir(test_dir)
total = len(files)
labels = ['airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck']
net = nn.load('/cifar10_fast.network')
for f in sorted(files):
img = image.Image(test_dir+"/"+f, copy_to_fb=True)
out = net.forward(img)
label = labels[out.index(max(out))]
if (f.split('-')[0] == label):
match += 1
return (total == 21 and match == 19)
def LENETTest(loopCnt=1200, barLen=60):
sensor.reset() # Reset and initialize the sensor.
sensor.set_contrast(3)
sensor.set_pixformat(
sensor.GRAYSCALE) # Set pixel format to RGB565 (or GRAYSCALE)
sensor.set_framesize(sensor.VGA) # Set frame size to QVGA (320x240)
sensor.set_windowing((84, 84)) # Set 128x128 window.
sensor.skip_frames(time=1400) # Wait for settings take effect.
sensor.set_auto_gain(False)
sensor.set_framerate(2 << 2)
#sensor.set_auto_whitebal(False)
#sensor.set_auto_exposure(False)
net = nn.load('/lenet.network')
labels = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']
clock = time.clock()
avg = 0.0
pyb.LED(1).on()
startTick = time.ticks()
while (True):
if time.ticks() - startTick > loopCnt:
break
clock.tick()
img = sensor.snapshot()
img.draw_string(3, 8, 'recg 0-9', color=(0, 0, 0))
t1 = time.ticks()
tmp_img = img.copy().binary([(120, 255)], invert=True)
lst = net.search(tmp_img, threshold=0.8, min_scale=1, scale_mul=0.8, \
x_overlap=-1, y_overlap=-1, contrast_threshold=0.5, softmax=False)
t2 = time.ticks() - t1
avg = avg * 0.95 + t2 * 0.05
lnLen = (barLen * (loopCnt - (time.ticks() - startTick))) // loopCnt
img.draw_rectangle(0, 2, barLen + 1, 3)
img.draw_rectangle(0, 3, lnLen, 1, fill=True)
for obj in lst:
print('Detected %s - Confidence %f%%' %
(labels[obj.index()], obj.value()))
img.draw_rectangle(obj.rect())
img.draw_string(barLen + 8,
2,
labels[obj.index()],
color=(0, 0, 0))
# print(clock.fps())
print('algo time cost : %.2f ms' % (avg))
def trainAll(
params,
trainOpt,
trainInput,
trainTarget,
trainInputWeights,
validInput,
validTarget,
validInputWeights,
initWeights=None,
):
model = nn.load(params["modelFilename"])
if initWeights is not None:
model.loadWeights(initWeights)
trainer = nn.Trainer(name=params["name"], model=model, trainOpt=trainOpt, outputFolder=params["outputFolder"])
# Combine train & valid set
allInput, allTarget, allInputWeights = combineInputs(
trainInput, trainTarget, trainInputWeights, validInput, validTarget, validInputWeights
)
trainer.train(trainInput=allInput, trainTarget=allTarget, trainInputWeights=allInputWeights)
return model, trainer
def LENetTest(loopCnt=600, isFull=False, barLen=80):
sensor.reset() # Reset and initialize the sensor.
sensor.set_contrast(3)
sensor.set_pixformat(
sensor.GRAYSCALE) # Set pixel format to RGB565 (or GRAYSCALE)
sensor.set_framesize(sensor.CIF) # Set frame size to QVGA (320x240)
sensor.set_windowing((96, 96)) # Set 128x128 window.
sensor.set_auto_gain(True)
sensor.set_auto_whitebal(False)
sensor.set_auto_exposure(False)
sensor.skip_frames(time=400) # Wait for settings take effect.
net = nn.load('/lenet.network')
labels = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']
clock = time.clock()
tAvg = 0.0
startTick = time.ticks()
while (True):
if time.ticks() - startTick > loopCnt:
break
clock.tick()
img = sensor.snapshot()
t0 = time.ticks()
tmp_img = img.copy().binary([(120, 255)], invert=True)
lst = net.search(tmp_img, threshold=0.8, min_scale=1.0, scale_mul=0.8, \
x_overlap=-1, y_overlap=-1, contrast_threshold=0.5, softmax=False)
t1 = time.ticks() - t0
tAvg = tAvg * 0.9 + t1 * 0.1
img.draw_string(4, 8, 'LENET', color=(0, 0, 0))
lnLen = (barLen * (loopCnt - (time.ticks() - startTick))) // loopCnt
DrawPgsBar(img, barLen, loopCnt, startTick)
for obj in lst:
print('Detected %s - Confidence %f%%' %
(labels[obj.index()], obj.value()))
img.draw_rectangle(obj.rect(), color=(255, 255, 255))
img.draw_string(4, 4, labels[obj.index()])
print('algo time cost : %.2f ms' % (tAvg))
for pseudolabel_images, pseudolabel_labels in ((args.data_pseudolabel_images, args.data_pseudolabel_labels),
(args.data_pseudolabel_images2, args.data_pseudolabel_labels2)):
plabel = PseudoLabel(pseudolabel_images, pseudolabel_labels)
splits = int(len(plabel) / len(datasets['train']) / args.data_pseudolabel_rate) # 30% normal images
if splits > 1:
plabel = RollSplitSet(plabel, splits)
datasets['train'] += plabel
#datasets = {mode: d if mode == 'train' else expand_valid_dataset(d) for mode, d in datasets.items()}
datasets_t = {mode: TransformedDataset(d, data_augs[mode]) for mode, d in datasets.items()}
classnames = datasets['train'].classnames[1:]
if args.model:
print('loading model {}...'.format(args.model))
model = nn.load(args.model)
else:
model = nn.create(args.version, args.basenet, classnames)
if args.weights:
print(f'loading weights {args.weights}')
weights = torch.load(args.weights, map_location='cpu')
weights = weights.get('state_dict', weights)
model.load_state_dict(weights, strict=True)
loss_functions = dict(bce=binary_cross_entropy_with_logits,
focal=binary_focal_loss_with_logits,
lovasz=lovasz_hinge,
lovasz_bce=lovasz_bce_with_logits,
lovasz_focal_symmetric=lovasz_focal_symmetric,
dice=dice_loss_with_logits,
dice_and_bce=dice_and_bce)
sensor.reset()
sensor.set_pixformat(sensor.GRAYSCALE)
sensor.set_framesize(sensor.QQSIF)
sensor.set_auto_gain(True)
sensor.set_auto_whitebal(True)
sensor.set_auto_exposure(True)
sensor.skip_frames(time=500)
sensor.set_contrast(3)
thresholds = (0, 50)
uart = UART(3, 9600, timeout_char=10)
uart.init(9600)
clock = time.clock()
ImageX = 88
ImageY = 60
net = nn.load('/fnt-chars74k.network')
labels = ['n/a', '0', '1', '2', '3', '4', '5', '6', '7', '8', '9']
for i in range(ord('A'), ord('Z') + 1):
labels.append(chr(i))
for i in range(ord('a'), ord('z') + 1):
labels.append(chr(i))
while (True):
clock.tick()
img = sensor.snapshot()
img.cartoon(size=1, seed_threshold=1)
img.binary([thresholds], zero=True)
img.rotation_corr(0, 0, 180, 0, 0, 1)
blobfound = False
for blob in img.find_blobs([thresholds], area_threshold=100, merge=False):
if (img.get_statistics(roi=blob.rect()).stdev() < 50):
if "-v-" in params["savedModelId"]:
# Train model
model, trainer = trainValid(
params,
trainOpt,
trainInput,
trainTarget,
trainInputWeights,
validInput,
validTarget,
validInputWeights,
initWeights=initWeights,
)
# Reload model
model = nn.load(params["modelFilename"])
model.loadWeights(np.load(trainer.modelFilename))
# Run tests
runTests(params, model, trainer)
# Re-train
if params["testDataFilename"] is not None and params["validDataFilename"] is not None:
# Setup options
trainOpt["needValid"] = False
print "Stopped score:", trainer.stoppedTrainScore
trainOpt["stopScore"] = trainer.stoppedTrainScore
# Train train+valid
model, trainer = trainAll(
trainData = np.load(trainDataFilename)
trainInput = trainData[0]
trainTarget = trainData[1]
trainInput, trainTarget = vt.shuffleData(
trainInput, trainTarget, np.random.RandomState(2))
with open(configFilename) as f:
trainOpt = yaml.load(f)
for i in range(0, 10):
trainInput_, trainTarget_, testInput_, testTarget_ = \
vt.splitData(trainInput, trainTarget, 0.1, i)
trainOpt['heldOutRatio'] = 0.1
trainOpt['xvalidNo'] = 0
trainOpt['needValid'] = True
model = nn.load(modelFilename)
trainer = nn.Trainer(
name=name + ('-%d-v' % i),
model=model,
trainOpt=trainOpt,
outputFolder=outputFolder
)
trainer.train(trainInput_, trainTarget_)
# Train again with all data, without validation
trainOpt['needValid'] = False
trainOpt['numEpoch'] = trainer.stoppedEpoch + 1
trainer = nn.Trainer(
name=name + ('-%d' % i),
model=model,
trainOpt=trainOpt,
# Untitled - By: Gehaha - 周四 3月 7 2019
# 笑脸识别历程
import sensor, time, image, os, nn
sensor.reset()
sensor.set_contrast(2)
sensor.set_pixformat(sensor.RGB565)
sensor.set_framesize(sensor.QVGA)
sensor.skip_frames(time=2000)
sensor.set_auto_gain(False)
#加载微笑检测网络
net = nn.load('/smile.network')
#load face haar cascade
face_cascade = image.HaarCascade("frontalface", stages=25)
print(face_cascade)
#FPS clock
clock = time.clock()
while (True):
clock.tick()
# capture snapshot
img = sensor.snapshot()
# 识别人脸
objects = img.find_features(face_cascade,
threshold=0.75,
scale_factor=1.25)
#检测微笑
#翻译和注释：01Studio
import sensor, image, time, os, nn
#摄像头初始化
sensor.reset() # Reset and initialize the sensor.
sensor.set_contrast(3)
sensor.set_pixformat(
sensor.GRAYSCALE) # Set pixel format to RGB565 (or GRAYSCALE)
sensor.set_framesize(sensor.QVGA) # Set frame size to QVGA (320x240)
sensor.set_windowing((128, 128)) # Set 128x128 window.
sensor.skip_frames(time=100)
sensor.set_auto_gain(False)
sensor.set_auto_exposure(False)
# 加载 lenet 神经网络模型
net = nn.load('/lenet.network')
labels = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']
clock = time.clock() # Create a clock object to track the FPS.
while (True):
clock.tick() # Update the FPS clock.
img = sensor.snapshot() # Take a picture and return the image.
#copy()表示创建一个图像副本储存在MicroPython堆中而不是帧缓冲区
#二值化是为了方便处理，阈值可以自己设定。
out = net.forward(img.copy().binary([(150, 255)], invert=True))
max_idx = out.index(max(out))
score = int(out[max_idx] * 100)
#大于70分认为识别可信
if (score < 70):
trainData = np.load(trainDataFilename)
trainInput = trainData[0]
trainTarget = trainData[1]
trainInput, trainTarget = vt.shuffleData(trainInput, trainTarget,
np.random.RandomState(2))
with open(configFilename) as f:
trainOpt = yaml.load(f)
for i in range(0, 10):
trainInput_, trainTarget_, testInput_, testTarget_ = \
vt.splitData(trainInput, trainTarget, 0.1, i)
trainOpt['heldOutRatio'] = 0.1
trainOpt['xvalidNo'] = 0
trainOpt['needValid'] = True
model = nn.load(modelFilename)
trainer = nn.Trainer(name=name + ('-%d-v' % i),
model=model,
trainOpt=trainOpt,
outputFolder=outputFolder)
trainer.train(trainInput_, trainTarget_)
# Train again with all data, without validation
trainOpt['needValid'] = False
trainOpt['numEpoch'] = trainer.stoppedEpoch + 1
trainer = nn.Trainer(name=name + ('-%d' % i),
model=model,
trainOpt=trainOpt,
outputFolder=outputFolder)
trainer.train(trainInput_, trainTarget_)
testOutput = nn.test(model, testInput_)
# CIFAR10 Example
import sensor, image, time, os, nn
sensor.reset() # Reset and initialize the sensor.
sensor.set_contrast(3)
sensor.set_pixformat(sensor.RGB565) # Set pixel format to RGB565
sensor.set_framesize(sensor.QVGA) # Set frame size to QVGA (320x240)
sensor.set_windowing((128, 128)) # Set 128x128 window.
sensor.skip_frames(time=1000)
sensor.set_auto_gain(False)
sensor.set_auto_exposure(False)
# Load cifar10 network
net = nn.load('/cifar10.network')
# Faster, smaller and less accurate.
#net = nn.load('/cifar10_fast.network')
labels = [
'airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse',
'ship', 'truck'
]
clock = time.clock() # Create a clock object to track the FPS.
while (True):
clock.tick() # Update the FPS clock.
img = sensor.snapshot() # Take a picture and return the image.
out = net.forward(img)
max_idx = out.index(max(out))
score = int(out[max_idx] * 100)
if (score < 70):
score_str = "??:??%"
else:
# LetNet Example
import sensor, image, time, os, nn
sensor.reset() # Reset and initialize the sensor.
sensor.set_contrast(3)
sensor.set_pixformat(sensor.GRAYSCALE) # Set pixel format to RGB565 (or GRAYSCALE)
sensor.set_framesize(sensor.QVGA) # Set frame size to QVGA (320x240)
sensor.set_windowing((128, 128)) # Set 128x128 window.
sensor.skip_frames(time=100)
sensor.set_auto_gain(False)
sensor.set_auto_exposure(False)
# Load lenet network
net = nn.load('/lenet.network')
labels = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']
clock = time.clock() # Create a clock object to track the FPS.
while(True):
clock.tick() # Update the FPS clock.
img = sensor.snapshot() # Take a picture and return the image.
out = net.forward(img.copy().binary([(150, 255)], invert=True))
max_idx = out.index(max(out))
score = int(out[max_idx]*100)
if (score < 70):
score_str = "??:??%"
else:
score_str = "%s:%d%% "%(labels[max_idx], score)
img.draw_string(0, 0, score_str)
print(clock.fps()) # Note: OpenMV Cam runs about half as fast when connected
# to the IDE. The FPS should increase once disconnected.
with open(params['configFilename']) as f:
trainOpt = yaml.load(f)
trainData = np.load(params['trainDataFilename'])
trainInput = trainData[0]
trainTarget = trainData[1]
if params['validDataFilename'] is not None:
validData = np.load(params['validDataFilename'])
validInput = validData[0]
validTarget = validData[1]
else:
validInput = None
validTarget = None
model = nn.load(params['modelFilename'])
trainer = nn.Trainer(
name=params['name']+\
('-v' if params['validDataFilename'] is not None else ''),
model=model,
trainOpt=trainOpt,
outputFolder=params['outputFolder']
)
trainer.train(trainInput, trainTarget, validInput, validTarget)
if params['testDataFilename'] is not None:
if params['imageqa']:
imageqa_test.testAll(
trainer.name, model, params['dataFolder'], params['outputFolder'])
else:
# Simle detection using Haar Cascade + CNN.
import sensor, time, image, os, nn
sensor.reset() # Reset and initialize the sensor.
sensor.set_contrast(2)
sensor.set_pixformat(sensor.RGB565) # Set pixel format to RGB565
sensor.set_framesize(sensor.QVGA) # Set frame size to QVGA (320x240)
sensor.skip_frames(time=2000)
sensor.set_auto_gain(False)
# Load smile detection network
net = nn.load('/smile.network')
# Load Face Haar Cascade
face_cascade = image.HaarCascade("frontalface", stages=25)
print(face_cascade)
# FPS clock
clock = time.clock()
while (True):
clock.tick()
# Capture snapshot
img = sensor.snapshot()
# Find faces.
objects = img.find_features(face_cascade, threshold=0.75, scale_factor=1.25)
# Detect smiles
for r in objects:
# Resize and center detection area
self.clear_button = QPushButton(self, text="Очистить поле")
self.clear_button.setGeometry(20, 10, 256, 50)
self.clear_button.clicked.connect(self.clear_input)
self.translate_button = QPushButton(self, text="Перевести в текст")
self.translate_button.setGeometry(360, 10, 256, 50)
self.translate_button.clicked.connect(self.translate_number)
self.show()
def clear_input(self):
self.input.image.fill(Qt.white)
self.repaint()
def translate_number(self):
img = np.array(recarray_view(self.input.image.scaled(28, 28)))
prediction, percent = nn.get_value(img)
self.info.setText(f'Вероятность: {round(percent*1000)/10}%')
self.output.setText(str(prediction))
if __name__ == "__main__":
app = QApplication([])
if not nn.load():
QMessageBox.warning(None, "Ошибка",
"При загрузке нейронной сети возникла ошибка. Пожалуйста, проверьте её наличие.")
else:
window = Window()
sys.exit(app.exec_())
# Batch info
batchEnd = min(N, batchStart + numExPerBat)
Ytmp = model.forward(X[batchStart:batchEnd], dropout=False)
if Y is None:
Yshape = np.copy(Ytmp.shape)
Yshape[0] = N
Y = np.zeros(Yshape)
Y[batchStart:batchEnd] = Ytmp
batchStart += numExPerBat
return Y
if __name__ == '__main__':
"""
Usage: test.py id -test test.npy
"""
taskId = sys.argv[1]
for i in range(2, len(sys.argv)):
if sys.argv[i] == '-test':
testDataFile = sys.argv[i + 1]
testAnswerFile = os.path.join('results/%s' % taskId, '%s.test.o.txt' % taskId)
testTruthFile = os.path.join('results/%s' % taskId, '%s.test.t.txt' % taskId)
modelFile = 'results/%s/%s.model.yml' % (taskId, taskId)
model = nn.load(modelFile)
model.loadWeights(np.load('results/%s/%s.w.npy' % (taskId, taskId)))
testData = np.load(testDataFile)
X = testData
Y = test(model, X)
np.savetxt(testAnswerFile,Y, delimiter=',')
import sensor, time, nn, image, os
sensor.reset() # Reset and initialize the sensor.
sensor.set_contrast(3)
sensor.set_pixformat(
sensor.RGB565) # Set pixel format to RGB565 (or GRAYSCALE)
sensor.set_framesize(sensor.QVGA) # Set frame size to QVGA (320x240)
sensor.set_windowing((192, 192)) # Set window
sensor.set_framerate(2 << 9 | 4 << 11)
sensor.skip_frames(time=100) # Wait for settings take effect.
# net = nn.load('/cifar10.network')
net = nn.load('/cifar10_fast.network')
labels = [
'plane', 'auto', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship',
'truck'
]
clock = time.clock()
tAvg = 40
while (True):
clock.tick()
img = sensor.snapshot()
t0 = time.ticks()
lst = net.search(img, threshold=0.595, min_scale=1, scale_mul=0.8, \
x_overlap=-1, y_overlap=-1, contrast_threshold=0.5, softamax=False)
t1 = time.ticks() - t0
tAvg = (tAvg * 200 + t1 * 56) >> 8
for obj in lst:
print('Detected %s - Confidence %f%%' %
(labels[obj.index()], obj.value()))
rc = obj.rect()
print(type(rc))
# In this example we slide the LeNet detector window over the image and get a list of activations
# where there might be an object. Note that use a CNN with a sliding window is extremely compute
# expensive so for an exhaustive search do not expect the CNN to be real-time.
import sensor, image, time, os, nn
sensor.reset() # Reset and initialize the sensor.
sensor.set_pixformat(sensor.RGB565) # Set pixel format to RGB565 (or GRAYSCALE)
sensor.set_framesize(sensor.QVGA) # Set frame size to QVGA (320x240)
sensor.set_windowing((128, 128)) # Set 128x128 window.
sensor.skip_frames(time=750) # Don't let autogain run very long.
sensor.set_auto_gain(False) # Turn off autogain.
sensor.set_auto_exposure(False) # Turn off whitebalance.
# Load cifar10 network (You can get the network from OpenMV IDE).
net = nn.load('/cifar10.network')
# Faster, smaller and less accurate.
# net = nn.load('/cifar10_fast.network')
labels = ['airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck']
clock = time.clock()
while(True):
clock.tick()
img = sensor.snapshot()
# net.search() will search an roi in the image for the network (or the whole image if the roi is not
# specified). At each location to look in the image if one of the classifier outputs is larger than
# threshold the location and label will be stored in an object list and returned. At each scale the
# detection window is moved around in the ROI using x_overlap (0-1) and y_overlap (0-1) as a guide.
# If you set the overlap to 0.5 then each detection window will overlap the previous one by 50%. Note
import nn
model = nn.load("model.NND")
print(model.layers[0].weights)