Python Linknet_bottleneck_crop Exemples

Exemple #1

Fichier : keras_to_onnx.py Projet : vladostan/segmification_convert

with open(ann_files[0]) as json_file:
    data = json.load(json_file)
    tags = data['tags']
    objects = data['objects']
    
img_path = ann_files[0].replace('/ann/', '/img/').split('.json')[0]
label_path = ann_files[0].replace('/ann/', '/masks_machine/').split('.json')[0]

print("Images dtype: {}".format(get_image(img_path).dtype))
print("Labels dtype: {}\n".format(get_image(label_path, label = True).dtype))
print("Images shape: {}".format(get_image(img_path, resize = True if resize else False).shape))
print("Labels shape: {}\n".format(get_image(label_path, label = True, resize = True if resize else False).shape))

# In[]:
model = sm.Linknet_bottleneck_crop(backbone_name=backbone, input_shape=input_shape, classification_classes=classification_classes, segmentation_classes = segmentation_classes, classification_activation = 'sigmoid', segmentation_activation='sigmoid')
model.load_weights('weights/' + weights + '.hdf5')

# In[]:
from keras.utils import plot_model
plot_model(model, to_file='keras_model.png')

# In[]:
onnx_model = keras2onnx.convert_keras(model, model.name)

# In[]:
import onnx
onnx.save_model(onnx_model, 'model.onnx')

# In[]:
preprocessing_fn = sm.get_preprocessing(backbone)

Exemple #2

    input_shape = (256, 640, 3) if resize else (512, 1280, 3)

    classification_classes = 1
    segmentation_classes = 1

    classification_activation = 'sigmoid' if classification_classes == 1 else 'softmax'
    segmentation_activation = 'sigmoid' if segmentation_classes == 1 else 'softmax'

    backbone = 'resnet18'

    weights = "2019-09-30 17-32-13"

    model = sm.Linknet_bottleneck_crop(
        backbone_name=backbone,
        input_shape=(256, 640, 3),
        classification_classes=classification_classes,
        segmentation_classes=segmentation_classes,
        classification_activation=classification_activation,
        segmentation_activation=segmentation_activation)
    model.load_weights('weights/' + weights + '.hdf5')

    preprocessing_fn = sm.get_preprocessing(backbone)

    model._make_predict_function()

    print("START ZERO PREDICT")
    model.predict(np.zeros(((1, ) + input_shape), dtype=np.uint8))
    print("END ZERO PREDICT")

    font = cv2.FONT_HERSHEY_SIMPLEX
    fontScale = 1