I am updating codes from tensorflow 1.x to 2.1.0.
I changed tensorflow 1.x code
labels = tf.cast(labels, tf.int64)
predict = tf.argmax(input=logits, axis=1)
tf.metrics.accuracy(labels=labels, predictions=predict)
to tensorflow 2.1.0 code.
labels = tf.cast(labels, tf.int64)
predict = tf.argmax(input=logits, axis=1)
tf.keras.metrics.Accuracy.update_state(labels, predict) #updated code
But, when I run the updated code, I got the following error.
TypeError: update_state() missing 1 required positional argument: 'y_pred'
So, I checked the tensorflow 2.1.0 document, and parameters for tf.keras.metrics.Accuracy.update_state() seem to be a list (in form of [ , , , ]). Then, I searched for a way to convert tensor to a list, which is
labels = tf.make_tensor_proto(labels)
labels = tf.make_ndarray(labels)
After I run this code, it gives the following error.
TypeError: List of Tensors when single Tensor expected
So, I tried to turn a list of Tensors into Tensors with
labels = tf.stack(labels)
#or
labels = torch.stack(labels)
tf.stack() did not work, as it gave the same initial TypeError saying 'y_pred' is missing at the updated code.
torch.stack(), however, gave the following error.
TypeError: stack() : argument 'tensors' (position 1) must be tuple of Tensors, not Tensor
So, I am guessing torch.stack() only accepts a tuple, NOT a list.
But, tf.stack() seems to accept a list, but it does not turn it into a Tensor?
Are my labels and predict even a list of Tensors in the first place? If so, why would tf.stack() not turn them into Tensors? How can I correctly convert labels and predict so that they can be passed into tf.keras.metrics.Accuracy.update_state()?
I would very much appreciate if not using compat.v1. unless absolutely necessary.
try in this way:
labels = [0,1]
logits = np.asarray([[0.9,0.1],[0.1,0.9]])
labels = tf.cast(labels, tf.int64)
predict = tf.argmax(input=logits, axis=1)
acc = tf.keras.metrics.Accuracy()
acc = acc.update_state(y_true=labels, y_pred=predict)
acc