Keras 2 input model, incompatible Shapes-CodePudding

The

Here is the stack trace:

ValueError: in user code: File "/Users/**/code/venv/lib/python3.10/site-packages/keras/engine/training.py", line 1051, in train_function * return step_function(self, iterator) File "/Users/**/code/venv/lib/python3.10/site-packages/keras/engine/training.py", line 1040, in step_function ** outputs = model.distribute_strategy.run(run_step, args=(data,)) File "/Users/**/code/venv/lib/python3.10/site-packages/keras/engine/training.py", line 1030, in run_step ** outputs = model.train_step(data) File "/Users/**/code/venv/lib/python3.10/site-packages/keras/engine/training.py", line 890, in train_step loss = self.compute_loss(x, y, y_pred, sample_weight) File "/Users/**/code/venv/lib/python3.10/site-packages/keras/engine/training.py", line 948, in compute_loss return self.compiled_loss( File "/Users/**/code/venv/lib/python3.10/site-packages/keras/engine/compile_utils.py", line 201, in __call__ loss_value = loss_obj(y_t, y_p, sample_weight=sw) File "/Users/**/code/venv/lib/python3.10/site-packages/keras/losses.py", line 139, in __call__ losses = call_fn(y_true, y_pred) File "/Users/**/code/venv/lib/python3.10/site-packages/keras/losses.py", line 243, in call ** return ag_fn(y_true, y_pred, **self._fn_kwargs) File "/Users/**/code/venv/lib/python3.10/site-packages/keras/losses.py", line 1787, in categorical_crossentropy return backend.categorical_crossentropy( File "/Users/**/code/venv/lib/python3.10/site-packages/keras/backend.py", line 5119, in categorical_crossentropy target.shape.assert_is_compatible_with(output.shape) ValueError: Shapes (None, 1) and (None, 20) are incompatible

CodePudding user response：

I'm not certain this is the issue, but it does show that categorical_crossentropy can object to a tensor of shape (n) in a tensorflow dataset, as you seem to have, demanding instead a tensor of shape (1,n)

It's too long to post as a comment anyway, so I can only post it as an 'answer'

In the code below, I'm creating a tf dataset like yours with one row, and with Y.shape (n). The categorical_crossentropy loss won't accept it, but does work if Y.shape is (1,6)

(I don't, I'm afraid, know how to make the reshaping work without setting tf.config.run_functions_eagerly(True), but that may not be an issue for you it you are using a generator)

Here's my minimal model and dataset setup:

import tensorflow as tf import numpy as np tf.config.run_functions_eagerly(True) # tf.autograph objects to something about the reshape dataset = tf.data.Dataset.from_tensor_slices(((np.array([[0.]]),np.array([[1.]])),np.array([[8., 3., 0., 8., 2., 1.]]))) # Even though the original array had shape (1,6), the dataset element has shape (6) print(f"dataset y shape: {[row[1].shape for row in dataset]}") # -- create and compile model, 2 inputs, one output inp1 = tf.keras.Input(shape=(1,)) inp2 = tf.keras.Input(shape=(1,)) x = tf.concat([inp1, inp2], axis=1) y = tf.keras.layers.Dense(6)(x) model = tf.keras.models.Model(inputs=[inp1, inp2], outputs=y) model.compile(optimizer='Adam', loss='categorical_crossentropy')

Then this fails, with a similar error to yours:

history = model.fit(dataset, epochs=1)

but this succeeds, after reshaping y from (6) to (1,6):

def reshaper(x, y): return x, tf.expand_dims(y, 0) dataset2 = dataset.map(reshaper) print(f"dataset2 y shape: {[row[1].shape for row in dataset2]}") history = model.fit(dataset2, epochs=1)

CodePudding user response：

Hidden units in output layer should be 1.

z = Dense(1, activation="softmax")(z)