def test_caffe_model_forward_gradient(tmpdir):
import caffe
from caffe import layers as L
bounds = (0, 255)
channels = num_classes = 1000
net_spec = caffe.NetSpec()
net_spec.data = L.Input(name="data",
shape=dict(dim=[1, num_classes, 5, 5]))
net_spec.reduce_1 = L.Reduction(net_spec.data,
reduction_param={
"operation": 4,
"axis": 3
})
net_spec.output = L.Reduction(net_spec.reduce_1,
reduction_param={
"operation": 4,
"axis": 2
})
net_spec.label = L.Input(name="label", shape=dict(dim=[1]))
net_spec.loss = L.SoftmaxWithLoss(net_spec.output, net_spec.label)
wf = tmpdir.mkdir("test_models_caffe").join(
"test_caffe_model_gradient_proto_{}.prototxt".format(num_classes))
wf.write("force_backward: true\n" + str(net_spec.to_proto()))
preprocessing = (
np.arange(num_classes)[:, None, None],
np.random.uniform(size=(channels, 5, 5)) + 1,
)
net = caffe.Net(str(wf), caffe.TEST)
model = CaffeModel(net, bounds=bounds, preprocessing=preprocessing)
epsilon = 1e-2
np.random.seed(23)
test_images = np.random.rand(5, channels, 5, 5).astype(np.float32)
test_labels = [7] * 5
_, g1 = model.forward_and_gradient_one(test_images, test_labels)
l1 = model._loss_fn(test_images - epsilon / 2 * g1, test_labels)
l2 = model._loss_fn(test_images + epsilon / 2 * g1, test_labels)
assert np.all(1e4 * (l2 - l1) > 1)
# make sure that gradient is numerically correct
np.testing.assert_array_almost_equal(
1e4 * (l2 - l1),
1e4 * epsilon *
np.linalg.norm(g1.reshape(len(g1), -1, g1.shape[-1]), axis=(1, 2))**2,
decimal=1,
)
def bn_model_caffe(request, tmpdir):
"""Same as bn_model but with Caffe."""
import caffe
from caffe import layers as L
bounds = (0, 1)
num_classes = channels = getattr(request, "param", 1000)
net_spec = caffe.NetSpec()
net_spec.data = L.Input(name="data",
shape=dict(dim=[1, channels, 5, 5]))
net_spec.reduce_1 = L.Reduction(net_spec.data,
reduction_param={"operation": 4,
"axis": 3})
net_spec.output = L.Reduction(net_spec.reduce_1,
reduction_param={"operation": 4,
"axis": 2})
net_spec.label = L.Input(name="label", shape=dict(dim=[1]))
net_spec.loss = L.SoftmaxWithLoss(net_spec.output, net_spec.label)
wf = tmpdir.mkdir("test_models_caffe_fixture")\
.join("test_caffe_{}.prototxt".format(num_classes))
wf.write("force_backward: true\n" + str(net_spec.to_proto()))
net = caffe.Net(str(wf), caffe.TEST)
model = CaffeModel(net, bounds=bounds)
return model
def create_fmodel(cls, cfg):
model = cls(**cfg["cfg"])
bounds = tuple(cfg.get("bounds", [0, 1]))
preprocessing = tuple(cfg.get("preprocessing", [0, 255]))
fmodel = CaffeModel(model.net,
bounds=bounds,
channel_axis=1,
preprocessing=preprocessing,
data_blob_name=cfg.get("data_blob_name", "data"),
label_blob_name=cfg.get("label_blob_name",
"label"),
output_blob_name=cfg.get("output_blob_name",
"output"))
return fmodel
def test_caffe_model_preprocessing_shape_change(tmpdir):
import caffe
from caffe import layers as L
bounds = (0, 255)
channels = num_classes = 1000
net_spec = caffe.NetSpec()
net_spec.data = L.Input(name="data",
shape=dict(dim=[1, num_classes, 5, 5]))
net_spec.reduce_1 = L.Reduction(net_spec.data,
reduction_param={
"operation": 4,
"axis": 3
})
net_spec.output = L.Reduction(net_spec.reduce_1,
reduction_param={
"operation": 4,
"axis": 2
})
net_spec.label = L.Input(name="label", shape=dict(dim=[1]))
net_spec.loss = L.SoftmaxWithLoss(net_spec.output, net_spec.label)
wf = tmpdir.mkdir("test_models_caffe")\
.join("test_caffe_model_preprocessing_shape_change_{}.prototxt"
.format(num_classes))
wf.write("force_backward: true\n" + str(net_spec.to_proto()))
net = caffe.Net(str(wf), caffe.TEST)
model1 = CaffeModel(net, bounds=bounds)
def preprocessing2(x):
if x.ndim == 3:
x = np.transpose(x, axes=(2, 0, 1))
elif x.ndim == 4:
x = np.transpose(x, axes=(0, 3, 1, 2))
def grad(dmdp):
assert dmdp.ndim == 3
dmdx = np.transpose(dmdp, axes=(1, 2, 0))
return dmdx
return x, grad
model2 = CaffeModel(net, bounds=bounds, preprocessing=preprocessing2)
np.random.seed(22)
test_images_nhwc = np.random.rand(2, 5, 5, channels).astype(np.float32)
test_images_nchw = np.transpose(test_images_nhwc, (0, 3, 1, 2))
p1 = model1.forward(test_images_nchw)
p2 = model2.forward(test_images_nhwc)
assert np.all(p1 == p2)
p1 = model1.forward_one(test_images_nchw[0])
p2 = model2.forward_one(test_images_nhwc[0])
assert np.all(p1 == p2)
g1 = model1.gradient_one(test_images_nchw[0], 3)
assert g1.ndim == 3
g1 = np.transpose(g1, (1, 2, 0))
g2 = model2.gradient_one(test_images_nhwc[0], 3)
np.testing.assert_array_almost_equal(g1, g2)