def main():
init_logging('webcamInferenceOnnx.log')
args = parse_args()
modelFilePath = args.modelFilePath
inference(modelFilePath)
def main():
init_logging('activeLearning.log')
args = parse_args()
datasetDir = args.datasetDir
trainingDir = args.trainingDir if args.trainingDir is not None else datasetDir
learningRate = args.learningRate
checkpointFilePath = args.checkpointFilePath
modelEncoder = args.modelEncoder
# efficientNet sizes: 224, 240, 260, 300, 380, 456, 528, 600
# efficientNet optimizer: RMSProp, decay 0.9 and momentum 0.9; batch norm momentum 0.99, weight decay 1e-5; initial learning rate 0.256 that decays by 0.97 every 2.4 epochs
# efficientNet special strategy: stochastic depth with survival probability 0.8, model size dependant dropout
optimizer = tf.keras.optimizers.Adam(
lr=learningRate,
# beta_1=0.9,
# beta_2=0.999,
# epsilon=1e-07,
# amsgrad=False,
# name='Adam'
)
model, preprocess_input = makeModel(optimizer, modelEncoder)
if checkpointFilePath is not None:
model.load_weights(checkpointFilePath)
logger.info('model weights from %s are loaded', checkpointFilePath)
humanDataset, activeLearningDataset, nonHumanDataset, valHumanDataset, valNonHumanDataset = openSegmentationDatasets(
datasetDir)
explicitTrain(model, preprocess_input, humanDataset, activeLearningDataset,
nonHumanDataset, valHumanDataset, valNonHumanDataset,
trainingDir, modelEncoder, args.batchSize, args.epochs,
args.startEpoch)
def main():
init_logging('saveModel.log')
args = parse_args()
checkpointFilePath = args.checkpointFilePath
modelFilePath = args.modelFilePath
model, preprocess_input = makeModel()
model.load_weights(checkpointFilePath)
logger.info('model weights from %s are loaded', checkpointFilePath)
model.save(modelFilePath)
logger.info('model saved to %s', modelFilePath)
def main():
init_logging('training.log')
logger.debug('gc enabled: {}'.format(gc.isenabled()))
# gc.set_debug(gc.DEBUG_LEAK)
args = parse_args()
datasetDir = args.datasetDir
trainingDir = args.trainingDir if args.trainingDir is not None else datasetDir
learningRate = args.learningRate
checkpointFilePath = args.checkpointFilePath
modelEncoder = args.modelEncoder
# efficientNet sizes: 224, 240, 260, 300, 380, 456, 528, 600
# efficientNet optimizer: RMSprop, decay 0.9 and momentum 0.9; batch norm momentum 0.99, weight decay 1e-5; initial learning rate 0.256 that decays by 0.97 every 2.4 epochs
# efficientNet special strategy: stochastic depth with survival probability 0.8, model size dependant dropout
# if modelEncoder == 'resnet18':
# optimizer = tfa.keras.optimizers.AdamW(
optimizer = tf.keras.optimizers.Adam(
lr=learningRate,
# beta_1=0.9,
# beta_2=0.999,
# epsilon=1e-07,
# amsgrad=False,
# name='Adam'
)
# else: #modelEncoder == 'efficientNet3':
# optimizer = tf.keras.optimizers.RMSprop(
# lr=learningRate,
# momentum=0.9
# )
model, preprocess_input = makeModel(optimizer, modelEncoder)
if checkpointFilePath is not None:
model.load_weights(checkpointFilePath)
# model = tf.saved_model.load(checkpointFilePath)
logger.info('model weights from %s are loaded', checkpointFilePath)
humanDataset, nonHumanDataset, valHumanDataset, valNonHumanDataset = openSegmentationDatasets(
datasetDir) #,train_nh_number=0)
explicitTrain(model,
preprocess_input,
humanDataset,
nonHumanDataset,
valHumanDataset,
valNonHumanDataset,
trainingDir,
modelEncoder,
imageSize=320,
batchSize=args.batchSize,
epochs=args.epochs,
startEpoch=args.startEpoch)
def main():
init_logging('quantizeOnnx.log')
args = parse_args()
modelFilePath = args.modelFilePath
quantizedModelFilePath = args.quantizedModelFilePath
datasetDir = args.datasetDir
# dr = ModelCalibrationDataReader(datasetDir, modelFilePath)
# quantize_static(modelFilePath, quantizedModelFilePath, dr)
quantize_dynamic(modelFilePath,
quantizedModelFilePath,
weight_type=QuantType.QInt8)
print('Calibrated and quantized model saved.')
def main():
init_logging('webcamInference.log')
args = parse_args()
modelFilePath = args.modelFilePath
model = tf.keras.models.load_model(
modelFilePath,
compile=False,
custom_objects={'iou_score': sm.metrics.iou_score}
)
# model = tf.saved_model.load(modelFilePath)
print("model summary\n{}".format(model.summary()))
inference(model, imageSize=320)
def main():
init_logging('tf2onnx.log')
args = parse_args()
checkpointFilePath = args.checkpointFilePath
modelFilePath = args.modelFilePath
onnxFilePath = args.onnxFilePath
if checkpointFilePath is not None:
model, preprocess_input = makeModel()
model.load_weights(checkpointFilePath)
logger.info('model weights from %s are loaded', checkpointFilePath)
model.save(modelFilePath)
logger.info('model saved to %s', modelFilePath)
# python -m tf2onnx.convert --saved-model tensorflow-model-path --opset 10 --output model.onnx
subprocess.run([
"python3.8", "-m", "tf2onnx.convert", "--saved-model", modelFilePath,
"--opset", "11", "--output", onnxFilePath
])
def main():
args = parse_args()
datasetDir = args.datasetDir
# datasetDir = '../data' # for debug assign your datafolder
init_logging('prepare.log')
humanDataset = CocoDatasetBuilder(
os.path.join(datasetDir, 'train2017'),
os.path.join(datasetDir,
'annotations/instances_train2017.json')).addClasses(
[1]).build()
CocoDataset.save(humanDataset,
os.path.join(datasetDir,
'human')) #, startIndex=(61600 + 2515))
nonHumanDataset = CocoDatasetBuilder(
os.path.join(datasetDir, 'train2017'),
os.path.join(datasetDir, 'annotations/instances_train2017.json')
).selectAll().filterNonClasses([1]).build(shuffle=True)
CocoDataset.save(nonHumanDataset,
os.path.join(datasetDir,
'nonHuman')) #, startIndex=28288)
valHumanDataset = CocoDatasetBuilder(
os.path.join(datasetDir, 'val2017'),
os.path.join(datasetDir,
'annotations/instances_val2017.json')).addClasses(
[1]).build()
CocoDataset.save(valHumanDataset, os.path.join(datasetDir, 'valHuman'))
valNonHumanDataset = CocoDatasetBuilder(
os.path.join(datasetDir, 'val2017'),
os.path.join(datasetDir, 'annotations/instances_val2017.json')
).selectAll().filterNonClasses([1]).build(shuffle=True)
CocoDataset.save(valNonHumanDataset, os.path.join(datasetDir,
'valNonHuman'))