Accuracy is lost after save_weights/load_weights

[Pandaaaa906]

Keras version: 3

TensorFlow version

2.16.1

Current behavior?

1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 354ms/step - accuracy: 0.5000 - loss: 1.1560 [1.1560312509536743, 0.5] Epoch 1/10 1/1 ━━━━━━━━━━━━━━━━━━━━ 1s 596ms/step - accuracy: 0.5000 - loss: 1.1560 Epoch 2/10 1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 28ms/step - accuracy: 0.5000 - loss: 14.5018 Epoch 3/10 1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 30ms/step - accuracy: 0.5000 - loss: 9.9714 Epoch 4/10 1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 31ms/step - accuracy: 0.7500 - loss: 1.3363 Epoch 5/10 1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 26ms/step - accuracy: 1.0000 - loss: 8.9407e-08 Epoch 6/10 1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 29ms/step - accuracy: 1.0000 - loss: 4.7684e-07 Epoch 7/10 1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 31ms/step - accuracy: 0.7500 - loss: 0.2545 Epoch 8/10 1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 27ms/step - accuracy: 0.7500 - loss: 0.8729 Epoch 9/10 1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 28ms/step - accuracy: 1.0000 - loss: 9.1682e-04 Epoch 10/10 1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 25ms/step - accuracy: 1.0000 - loss: 2.6822e-07 1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 27ms/step - accuracy: 1.0000 - loss: 0.0000e+00 [0.0, 1.0] 1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 335ms/step - accuracy: 0.2500 - loss: 0.8475 # this should be acc 1.0 loss 0 [0.847506046295166, 0.25]

Standalone code to reproduce the issue

import tensorflow as tf

class CusModel(tf.keras.Model):
    def __init__(self):
        super().__init__()
        self.dense = tf.keras.layers.Dense(units=2, activation='softmax', name='output')

    def call(self, x):
        return self.dense(x)

dummy_data_x = tf.convert_to_tensor([[0, 0],
                [1, 0],
                [0, 1],
                [1, 1]])
dummy_data_y = tf.convert_to_tensor([0, 1, 0, 1])

model = CusModel()
model.compile(optimizer=tf.keras.optimizers.Adam(10.0),
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])
print(model.evaluate(x=dummy_data_x, y=dummy_data_y))
model.fit(x=dummy_data_x, y=dummy_data_y, epochs=10)
print(model.evaluate(x=dummy_data_x, y=dummy_data_y))
model.save_weights('test_model.weights.h5')

model = CusModel()
model.load_weights('test_model.weights.h5')
model.compile(optimizer=tf.keras.optimizers.Adam(10.0),
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])
print(model.evaluate(x=dummy_data_x, y=dummy_data_y))

[ghsanti]

There are a few ways to fix it.

method 1

build the layer in the Model subclass (add the method build):

    def build(self, x): # method within your Model.
      self.dense.build(x)

And call model.build((2,)) after compile, and before evaluate.

method 2

To force weights you can also evaluate twice, as done below:

print(model.evaluate(x=dummy_data_x, ...)) # creates weights
model.load_weights('test_model.weights.h5') # loads new
print(model.evaluate(x=dummy_data_x, ...)) # ok result

this works because the layer will be build for you.

[sonali-kumari1]

Hi @Pandaaaa906,

Thanks for reporting the issue.

Instead of saving and loading the weights, you can try model.save() and keras.models.load_model(). By doing this, you can ensure that model architecture,optimizer state and weights everything is saved.

Attaching gist for your reference.

[james77777778]

@Pandaaaa906 I have submitted a PR (which has been merged) to raise an error when calling save_weights and load_weights on an unbuilt model. You should consider adopting the methods provided by @ghsanti to resolve this issue.

Thanks @ghsanti for pointing this out.