load_weights can't work

session 1.

def build_model():
model = RetrievalModel(item_model, user_model)
learning_rate = 0.01
model.compile(optimizer=tf.keras.optimizers.Adagrad(learning_rate))
return model

model = build_model()
model.fit(behavior_dataset, epochs=30)
model.save_weights(save_path,overwrite=True )

compute_loss_args = {
"user_id" : tf.constant(["45"]),
"work_id" : tf.constant(["45"]),
"tags" : tf.constant([""]),
"work_uid" : tf.constant(["45"]),
"money_goods" : tf.constant([100]),
"category_id" : tf.constant(["2"]),
"bid_type" : tf.constant(["normal"]),
"is_rec" : tf.constant(["1"]),
"weights" : tf.constant([1]),
}
model(compute_loss_args)
s = model.load_weights(save_path ).expect_partial()

k = 100
user_id = '1035369'
index = tfrs.layers.factorized_top_k.BruteForce(model.user_model,k)
index.index_from_dataset(
work_dataset.shuffle(100_100).map(lambda x: (x["work_id"], model.item_model(x))) #注意，这里是全局可推荐列表
)
print(f"rec user_id :{user_id}")
print(user_id in unique_user_id)

is right。

session 2:

def build_model():
model = RetrievalModel(item_model, user_model)
learning_rate = 0.01
model.compile(optimizer=tf.keras.optimizers.Adagrad(learning_rate))
return model

• _model = build_model() #delete
• model.fit(behavior_dataset, epochs=30) #delete
• model.save_weights(save_path,overwrite=True )#delete_

compute_loss_args = {
"user_id" : tf.constant(["45"]),
"work_id" : tf.constant(["45"]),
"tags" : tf.constant([""]),
"work_uid" : tf.constant(["45"]),
"money_goods" : tf.constant([100]),
"category_id" : tf.constant(["2"]),
"bid_type" : tf.constant(["normal"]),
"is_rec" : tf.constant(["1"]),
"weights" : tf.constant([1]),
}
model(compute_loss_args)
s = model.load_weights(save_path ).expect_partial()

k = 100
user_id = '1035369'
index = tfrs.layers.factorized_top_k.BruteForce(model.user_model,k)
index.index_from_dataset(
work_dataset.shuffle(100_100).map(lambda x: (x["work_id"], model.item_model(x))) #注意，这里是全局可推荐列表
)
print(f"rec user_id :{user_id}")
print(user_id in unique_user_id)

is wrong 不准确。

tensorboard 2.15.2
keras 2.15.0

Start by reproducing the two sessions around build_model(), model.fit(), save_weights(), model(compute_loss_args), and load_weights(). Compare the resulting recommendation output and confirm whether the issue is weight restoration or model initialization; done means the cause is isolated and the expected behavior is documented with a minimal reproducible example.