def test_shape(self):
with self.test_session(use_gpu=False, force_gpu=False):
input1 = np.empty((8,40,31))
input2 = np.empty((8,1,40,31))
input3 = np.empty((2,2,2,40,31))
inputs = (input1, input2, input3)
for i in inputs:
output_tensor = ops.replace_nonfinite( input=i )
out_shape = output_tensor.get_shape().as_list()
self.assertAllEqual(out_shape, i.shape)
def scinv_gradloss(output, gt,scope=None,weight=1.0):
with tf.name_scope(scope, 'scinv_gradloss', [output, gt]):
# convert from NHWC to NCHW
output=convertNHWC2NCHW(output)
gt=convertNHWC2NCHW(gt)
# compute mask and make zero areas NaN in order to remove them later
zero=tf.zeros_like(gt)
mask = tf.not_equal(gt, zero)
mask=tf.cast(mask,tf.float32)
n=tf.reduce_sum(mask)# number of elements for evaluation
gt_nan = tf.divide(gt,mask) # make NaN the areas where no need to eval
epsilon=0.000000001
# compute scale invariant grad loss
grad_output = ops.scale_invariant_gradient(input=output, deltas=[1,2,4,8,16], weights=[1,1,1,1,1], epsilon=epsilon)
grad_gt = ops.scale_invariant_gradient(input=gt_nan, deltas=[1,2,4,8,16], weights=[1,1,1,1,1], epsilon=epsilon)
diff = ops.replace_nonfinite(grad_gt-grad_output)
gradLoss=tf.reduce_sum(tf.sqrt(tf.reduce_sum(diff**2, axis=1)+epsilon))/n
tf.add_to_collection('losses', weight*gradLoss)
def scinv_gradloss(output, gt, batch_size, scope=None):
with tf.name_scope(scope, 'scinv_gradloss', [output, gt]):
#convert from NHWC to NCHW
output = convertNHWC2NCHW(output)
gt = convertNHWC2NCHW(gt)
#pad the tensor
paddings = [[0, 0], [1, 0], [2, 4], [
3, 4
]] # the 3d and 4th dimention pad with 4 zero from each side
output = tf.pad(output, paddings, 'CONSTANT')
gt = tf.pad(gt, paddings, 'CONSTANT')
# compute mask and make zero areas NaN in order to remove them later
zero = tf.zeros_like(gt)
mask = tf.not_equal(gt, zero)
mask = tf.cast(mask, tf.float32)
#mask=tf.div(mask,mask)# divide 0/0 gives us NaN values
#output_nan= tf.div(output,mask)
#gt_nan=tf.div(gt,mask)
# compute scale invariant grad loss
grad_output = ops.scale_invariant_gradient(input=output,
deltas=[1, 2, 4],
weights=[1, 0.5, 0.25],
epsilon=0.001)
grad_gt = ops.scale_invariant_gradient(input=gt,
deltas=[1, 2, 4],
weights=[1, 0.5, 0.25],
epsilon=0.001)
diff = ops.replace_nonfinite(grad_output - grad_gt)
#diff=tf.subtract(grad_output,grad_gt)
#apply mask
#mask_out=tf.concat(0, [mask,mask])# the grad has 2 time more matrices in tensor because of grad by x and by y
#diff=tf.multiply(mask_out,diff)
#remove NaN values
#diff=tf.select(tf.is_nan(diff),tf.zeros_like(diff),diff)
gradLoss = tf.reduce_sum(tf.square(diff)) / (2 * batch_size)
tf.add_to_collection(LOSSES_COLLECTION, gradLoss)
return gradLoss
def _test_grad(self,dtype):
A = np.random.rand(9).astype(dtype)
A[2] = np.nan
shape = A.shape
data = tf.constant(A)
output = ops.replace_nonfinite(input=data, value=123)
#print(A)
#print(output.eval())
err = tf.test.compute_gradient_error(
data, shape,
output, output.get_shape().as_list(),
x_init_value = A )
print('error',err,flush=True)
self.assertLess(err, 1e-3)
grad = tf.test.compute_gradient(
data, shape,
output, output.get_shape().as_list(),
x_init_value = A,
delta=0.1)
for g in grad:
print(g)
print(g.shape)
def _test_nonfinite(self,dtype):
value = 123
A = np.array([np.nan, np.inf, -np.inf, 100], dtype=dtype)
result = ops.replace_nonfinite(A, value=value).eval()
self.assertAllEqual(result, [value]*3+[100])