def build(self, inputs, targets):
"""
Args:
inputs (torch.Tensor): feature matrix with shape (batch_size, feat_dim).
targets (torch.LongTensor): ground truth labels with shape (num_classes).
"""
n = inputs.shape[0]
dist = math.reduce_sum(math.pow(
inputs, flow.constant_like(inputs, 2, dtype=flow.float32)),
axis=1)
shape_tensor = flow.constant(value=0.0,
dtype=flow.float32,
shape=(n, n))
dist = flow.broadcast_like(dist, like=shape_tensor, broadcast_axes=[1])
dist = math.add(
dist, flow.transpose(dist, perm=(1, 0),
batch_axis_non_change=True))
temp1 = math.multiply(
-2,
flow.matmul(
inputs,
flow.transpose(inputs, perm=(1, 0),
batch_axis_non_change=True)))
dist = math.add(dist, temp1)
dist = math.sqrt(flow.clamp(dist, min_value=1e-12))
mask = math.equal(
flow.broadcast_like(targets, like=shape_tensor,
broadcast_axes=[1]),
flow.transpose(flow.broadcast_like(targets,
like=shape_tensor,
broadcast_axes=[1]),
perm=(1, 0),
batch_axis_non_change=True))
mask_rev = math.not_equal(
flow.broadcast_like(targets, like=shape_tensor,
broadcast_axes=[1]),
flow.transpose(flow.broadcast_like(targets,
like=shape_tensor,
broadcast_axes=[1]),
perm=(1, 0),
batch_axis_non_change=True))
dist_ap, dist_an = [], []
for i in range(n):
temp_dist = flow.slice_v2(dist, [(i, i + 1, 1)])
temp_mask = flow.slice_v2(mask, [(i, i + 1, 1)])
temp_mask_rev = flow.slice_v2(mask_rev, [(i, i + 1, 1)])
dist_ap.append(
math.reduce_max(
flow.gather_nd(temp_dist, flow.where(temp_mask))))
dist_an.append(
math.reduce_min(
flow.gather_nd(temp_dist, flow.where(temp_mask_rev))))
dist_ap = flow.concat(dist_ap, 0)
dist_an = flow.concat(dist_an, 0)
y = flow.ones_like(dist_an)
# return dist_an, dist_ap, y
return self._MarginRankingLoss(dist_an, dist_ap, y)
def slice(input_blob: oft.Numpy.Placeholder(shape=(2, 5, 4), dtype=flow.float)):
x = flow.get_variable(
shape=(2, 5, 4),
dtype=flow.float,
initializer=flow.random_uniform_initializer(0, 2),
name="variable",
)
x = flow.identity(x)
flow.watch_diff(x, slice_grad_cb)
y = flow.slice_v2(x, [(None, None, None), (2, -2, None)])
flow.losses.add_loss(y)
return y
def slice(input_blob: oft.Numpy.Placeholder(shape=(2, 5, 4),
dtype=flow.float)):
x = flow.get_variable(
shape=(2, 5, 4),
dtype=flow.float,
initializer=flow.random_uniform_initializer(0, 2),
name="variable",
)
x = flow.identity(x)
flow.watch_diff(x, slice_grad_cb)
y = flow.slice_v2(x, [(None, None, None), (2, -2, None)])
flow.optimizer.SGD(flow.optimizer.PiecewiseConstantScheduler([],
[1e-3]),
momentum=0).minimize(y)
return y
def slice_with_grad_job(x: otp.Numpy.Placeholder(
shape=input_shape, dtype=dtype)) -> otp.Numpy:
var = flow.get_variable(
shape=input_shape,
dtype=dtype,
initializer=flow.constant_initializer(0.0),
name="variable",
)
x = x + var
if callable(watch_diff_cb):
flow.watch_diff(x, watch_diff_cb)
y = flow.slice_v2(x, slice_tup_list, name="SliceWithGrad")
flow.optimizer.SGD(flow.optimizer.PiecewiseConstantScheduler([],
[1e-3]),
momentum=0).minimize(y)
return y
def do_slice(x, indices):
outputs = []
for slice_tup_list in indices:
output = flow.slice_v2(x, slice_tup_list)
outputs.append(output)
return outputs
def _do_slice(input, args, name=None):
outputs = []
for slice_tup_list in args:
output = flow.slice_v2(input, slice_tup_list, name)
outputs.append(output)
return outputs
def forward(self, inputs, targets):
n = inputs.shape[0]
# Compute pairwise distance, replace by the official when merged
tempname = datetime.datetime.now().strftime('%Y-%m-%d-%H-%M-%S.%f')
shape_tensor = flow.constant(value=0.0,
dtype=flow.float32,
shape=(n, n))
if self.distance == 'euclidean':
blob_2 = flow.get_variable(
"blob_2_" + tempname,
shape=inputs.shape,
initializer=flow.constant_initializer(2),
dtype=inputs.dtype)
dist = flow.math.pow(inputs, blob_2)
dist = flow.math.reduce_sum(dist, axis=1, keepdims=True)
dist = flow.broadcast_like(dist, shape_tensor)
tempdist = flow.transpose(dist)
dist = dist + tempdist
inputs_t = flow.transpose(inputs)
dist = addmm(dist, inputs, inputs_t, beta=1, alpha=-2)
dist = flow.clamp(dist, min_value=1e-12)
dist = flow.math.sqrt(dist)
elif self.distance == 'cosine':
#fnorm=flow.math.l2_normalize(inputs, axis=1)
fnorm = flow.math.reduce_mean(flow.math.divide(
inputs, flow.math.l2_normalize(inputs, axis=1)),
axis=1,
keepdims=True)
expand_fnorm = flow.broadcast_like(fnorm,
like=inputs,
broadcast_axes=[1])
l2norm = flow.math.divide(inputs, expand_fnorm)
l2norm_t = flow.transpose(l2norm, perm=(1, 0))
dist = flow.math.negative(flow.matmul(l2norm, l2norm_t))
# For each anchor, find the hardest positive and negative
mask = math.equal(
flow.broadcast_like(targets, like=shape_tensor,
broadcast_axes=[1]),
flow.transpose(flow.broadcast_like(targets,
like=shape_tensor,
broadcast_axes=[1]),
perm=(1, 0),
batch_axis_non_change=True))
mask_rev = math.not_equal(
flow.broadcast_like(targets, like=shape_tensor,
broadcast_axes=[1]),
flow.transpose(flow.broadcast_like(targets,
like=shape_tensor,
broadcast_axes=[1]),
perm=(1, 0),
batch_axis_non_change=True))
dist_ap, dist_an = [], []
for i in range(n):
temp_dist = flow.slice_v2(dist, [(i, i + 1, 1)])
temp_mask = flow.slice_v2(mask, [(i, i + 1, 1)])
temp_mask_rev = flow.slice_v2(mask_rev, [(i, i + 1, 1)])
temp_dist_ap = flow.expand_dims(
math.reduce_max(
flow.gather_nd(temp_dist, flow.where(temp_mask))), 0)
temp_dist_an = flow.expand_dims(
math.reduce_min(
flow.gather_nd(temp_dist, flow.where(temp_mask_rev))), 0)
dist_ap.append(temp_dist_ap)
dist_an.append(temp_dist_an)
dist_ap = flow.concat(dist_ap, 0)
dist_an = flow.concat(dist_an, 0)
y = flow.ones_like(dist_an)
return self._MarginRankingLoss(dist_an, dist_ap, y)