def test_resnet50_architecture(self):
resnet50 = architectures.resnet50()
self.assertEqual(resnet50.input_shape, (224, 224, 3))
self.assertEqual(resnet50.output_shape, (1000,))
random_input = asfloat(np.random.random((7, 224, 224, 3)))
prediction = self.eval(resnet50.output(random_input))
self.assertEqual(prediction.shape, (7, 1000))
def test_resnet50_architecture(self):
resnet50 = architectures.resnet50()
self.assertEqual(resnet50.input_shape, (3, 224, 224))
self.assertEqual(resnet50.output_shape, (1000, ))
resnet50_predict = resnet50.compile()
random_input = asfloat(np.random.random((7, 3, 224, 224)))
prediction = resnet50_predict(random_input)
self.assertEqual(prediction.shape, (7, 1000))
def test_resnet50_spatial(self):
resnet50 = architectures.resnet50(
include_global_pool=False,
in_out_ratio=8,
)
self.assertEqual(resnet50.input_shape, (224, 224, 3))
self.assertEqual(resnet50.output_shape, (28, 28, 2048))
random_input = asfloat(np.random.random((7, 224, 224, 3)))
prediction = self.eval(resnet50.output(random_input))
self.assertEqual(prediction.shape, (7, 28, 28, 2048))
def test_resnet50_dilation_rates(self):
resnet50 = architectures.resnet50(
include_global_pool=False,
in_out_ratio=8,
)
layer = resnet50.layer('res4d_branch2b')
self.assertEqual(layer.dilation, (2, 2))
layer = resnet50.layer('res5a_branch2b')
self.assertEqual(layer.dilation, (2, 2))
layer = resnet50.layer('res5b_branch2b')
self.assertEqual(layer.dilation, (4, 4))
def prepare_image(fname):
with open(IMAGENET_MEAN_FILE, 'rb') as f:
# Mean values is the average image accros all training dataset.
# if dataset (1000, 3, 224, 224) then mean image shape
# is (3, 224, 224) and computes as data.mean(axis=0)
mean_values = pickle.load(f)
image = read_image(fname, image_size=(256, 256), crop_size=(224, 224))
# Convert RGB to BGR
image[:, (0, 1, 2), :, :] = image[:, (2, 1, 0), :, :]
return asfloat(image - mean_values)
environment.speedup()
resnet50 = architectures.resnet50()
if not os.path.exists(RESNET50_WEIGHTS_FILE):
download_file(
url="http://neupy.s3.amazonaws.com/imagenet-models/resnet50.pickle",
filepath=RESNET50_WEIGHTS_FILE,
description='Downloading weights')
storage.load(resnet50, RESNET50_WEIGHTS_FILE)
predict = resnet50.compile()
dog_image = prepare_image(DOG_IMAGE_PATH)
output = predict(dog_image)
print_top_n(output, n=5)
def test_resnet50_no_global_pooling(self):
resnet50 = architectures.resnet50(include_global_pool=False)
self.assertEqual(resnet50.input_shape, (224, 224, 3))
self.assertEqual(resnet50.output_shape, (7, 7, 2048))
def test_resnet50_exceptions(self):
with self.assertRaises(ValueError):
architectures.resnet50(in_out_ratio=2)
def create_deeplab_model(resnet50_weights=None,
deeplab_weights=None,
size=None):
print("Initializing ResNet-50 architecture...")
SamePadConv = partial(Convolution, bias=None, padding='same')
resnet50 = architectures.resnet50(
input_shape=(size, size, 3),
include_global_pool=False,
in_out_ratio=16,
)
if resnet50_weights is not None:
# Pre-trained ResNet-50 contains parameters for the final
# classification layer. We don't use this layer and for this reason
# we need to set ``ignore_missing=True``
print("Recovering ResNet-50 parameters...")
storage.load(resnet50, resnet50_weights, ignore_missing=True)
in_height, in_width, _ = resnet50.input_shape
out_height, out_width, _ = resnet50.output_shape
resnet50_input = resnet50.layers[0]
deeplab_input = Input(resnet50.output_shape, name='deeplab-input')
print("Initializing Deeplab architecture...")
deeplab = join(
deeplab_input,
# Atrous Spatial Pyramid Pooling
parallel(
SamePadConv((1, 1, 256)) > BatchNorm(),
SamePadConv((3, 3, 256), dilation=6) > BatchNorm(),
SamePadConv((3, 3, 256), dilation=12) > BatchNorm(),
SamePadConv((3, 3, 256), dilation=18) > BatchNorm(), [
GlobalPooling('avg'),
Reshape((1, 1, -1)),
SamePadConv((1, 1, 256)) > BatchNorm(),
IncludeResidualInputs(deeplab_input),
ResizeBilinear(),
]),
Concatenate(),
SamePadConv((1, 1, 256)) > BatchNorm(),
# Convert to the classification maps
Convolution((1, 1, 21), padding='same'),
IncludeResidualInputs(resnet50_input),
ResizeBilinear((in_height, in_width)),
Softmax(name='segmentation-proba'),
)
if deeplab_weights is not None:
print("Recovering Deeplab parameters...")
storage.load(deeplab, deeplab_weights, ignore_missing=True)
print("Patching layers...")
patches = {
BatchNorm: {
'alpha': 1 - 0.997,
'epsion': 1e-5,
}
}
patch_layers(deeplab, patches)
patch_layers(resnet50, patches)
return resnet50, deeplab