I'm learning Tensorflow 2.10, with Python 3.7.7.
I'm trying to use the tutorial "Tensorflow - Custom training: walkthrough" to use my own loss function.
This is my first version of loss function, and it works:
def loss(model, x, y):
output = model(x)
return tf.norm(y - output)
I have changed to try another one, and it doesn't work:
def my_loss(model, x, y):
output = model(x)
# Only valid values for output var are 0.0 and 1.0.
output_np = np.array(output)
output_np[output_np >= 0.5] = 1.0
output_np[output_np < 0.5] = 0.0
# Counts how many 1.0 are on y var.
unique, counts = np.unique(y, return_counts=True)
dict_wmh = dict(zip(unique, counts))
wmh_count = 0
if 1.0 in dict_wmh:
wmh_count = dict_wmh[1.0]
# Add y and output to get another array.
c = y + output_np
unique, counts = np.unique(c, return_counts=True)
dict_net = dict(zip(unique, counts))
# Counts how many 2.0 are on this new array.
net_count = 0
if 2.0 in dict_net:
net_count = dict_net[2.0]
# Return the different between the number of ones in the label and the network output.
return wmh_count - net_count
But I can use it because my new loss function "interrupts the gradient chain registered by the gradient tape".
So, I have tried to use only Tensorflow Tensor:
def my_loss_tensor(model, x, y):
output = model(x)
# Only valid values for output var are 0.0 and 1.0.
output = tf.math.round(output)
output = tf.clip_by_value(output, clip_value_min=0.0, clip_value_max=1.0)
# Counts how many 1.0 are on y var (WMH mask).
y_ele, y_idx, y_count = tf.unique_with_counts(y)
# Add output to WMH mask.
sum = tf.math.add(output, y)
# Counts how many 2.0 are on the sum.
sum_ele, sum_idx, sum_count = tf.unique_with_counts(sum)
return tf.math.subtract(sum_count[sum_ele == 1.0], y_count[y_ele == 2.0])
x is a tf.Tensor([[[[...]]]], shape=(1, 200, 200, 1), dtype=float32)
y is a tf.Tensor([[[[...]]]], shape=(1, 200, 200, 1), dtype=float32)
They are images (200x200x1).
I get the following error:
unique expects a 1D vector. [Op:UniqueWithCounts]
Any idea about how to count how many times appear a value on a Tensor?
The real image data are on the 200x200 dimensions, the other two are used on my CNN.
