这来自Keras文档示例：训练一个模型来计算priority_score以及转发电子邮件的部门。
我用另一种方式实现模型，我可以编译它，但我不能训练模型。我想这是一个模型I/O问题，即我需要提供正确的I/O数据格式。
ValueError: Failed to find data adapter that can handle input: (<class 'dict'> containing {"<class 'str'>"} keys and {"<class 'numpy.ndarray'>", '(<class 'list'> containing values of types {"<class 'str'>"})'} values), (<class 'dict'> containing {"<class 'str'>"} keys and {"<class 'numpy.ndarray'>"} values)
太长了，所以我没有把它放进这篇文章的标题里。
我的模型有3个输入：
- title_input：应该是单个字符串
- body_input：应该是单个字符串
- tags_input：12个0或1的数组。例如，[0,1,0,1,0,0,0,0,1]
输出为：
- 优先级：浮点
- 部门：4 0,1s的数组
问题
有人能告诉我我的代码出了什么问题吗？
一般来说，我应该如何看待模型的I/O？比如这个案例。我以为准备N个字符串，比如800个字符串和800个标签是可以的。但我总是出错。我解决了大部分问题，但没能克服这个问题。请分享你的经历。谢谢
附录
模型摘要
Model: "functional_1"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
tags_input (InputLayer) [(None, 12)] 0
__________________________________________________________________________________________________
flatten (Flatten) (None, 12) 0 tags_input[0][0]
__________________________________________________________________________________________________
title_input (InputLayer) [(None, 1)] 0
__________________________________________________________________________________________________
body_input (InputLayer) [(None, 1)] 0
__________________________________________________________________________________________________
dense (Dense) (None, 500) 6500 flatten[0][0]
__________________________________________________________________________________________________
text_vectorization (TextVectori (None, 500) 0 title_input[0][0]
__________________________________________________________________________________________________
text_vectorization_1 (TextVecto (None, 500) 0 body_input[0][0]
__________________________________________________________________________________________________
tf_op_layer_ExpandDims (TensorF [(None, 500, 1)] 0 dense[0][0]
__________________________________________________________________________________________________
embedding (Embedding) (None, 500, 100) 1000100 text_vectorization[0][0]
__________________________________________________________________________________________________
embedding_1 (Embedding) (None, 500, 100) 1000100 text_vectorization_1[0][0]
__________________________________________________________________________________________________
dense_1 (Dense) (None, 500, 100) 200 tf_op_layer_ExpandDims[0][0]
__________________________________________________________________________________________________
concatenate (Concatenate) (None, 500, 300) 0 embedding[0][0]
embedding_1[0][0]
dense_1[0][0]
__________________________________________________________________________________________________
priority (Dense) (None, 500, 1) 301 concatenate[0][0]
__________________________________________________________________________________________________
departments (Dense) (None, 500, 4) 1204 concatenate[0][0]
==================================================================================================
Total params: 2,008,405
Trainable params: 2,008,405
Non-trainable params: 0
__________________________________________________________________________________________________
完整代码
def MultiInputAndOutpt():
max_features = 10000
sequnce_length = 500
embedding_dims = 100
num_departments = 4
num_tags = 12
str = "hello"
title_vect = TextVectorization(max_tokens=max_features, output_mode="int", output_sequence_length=sequnce_length)
body_vect = TextVectorization(max_tokens=max_features, output_mode="int", output_sequence_length=sequnce_length)
title_input = keras.Input(shape=(1,), dtype=tf.string, name="title_input")
x1 = title_vect(title_input)
x1 = layers.Embedding(input_dim=max_features + 1, output_dim=embedding_dims)(x1)
body_input = keras.Input(shape=(1,), dtype=tf.string, name="body_input")
x2 = body_vect(body_input)
x2 = layers.Embedding(input_dim=max_features + 1, output_dim=embedding_dims)(x2)
tags_input = keras.Input(shape=(num_tags,), name="tags_input")
x3 = layers.Flatten()(tags_input)
x3 = layers.Dense(500)(x3)
x3 = tf.expand_dims(x3, axis=-1)
x3 = layers.Dense(100)(x3)
x = layers.concatenate([x1, x2, x3])
priority_score = layers.Dense(1)(x)
priority_score = tf.reshape(priority_score, (-1, 1), name="priority")
departments = layers.Dense(num_departments)(x)
departments = tf.reshape(departments, (-1, num_departments), name="departments")
model = keras.Model(inputs=[title_input, body_input, tags_input], outputs=[priority_score, departments])
model.summary()
model.compile(optimizer=keras.optimizers.Adam(learning_rate=0.1),
loss=[keras.losses.BinaryCrossentropy(from_logits=True),
keras.losses.CategoricalCrossentropy(from_logits=True)],
loss_weights=[1.0, 0.2],
)
# title_data = np.random.randint(num_words, size=(1280, 10))
# body_data = np.random.randint(num_words, size=(1280, 100))
alphabet = np.array(list(string.ascii_lowercase + ' '))
title_data = np.random.choice(alphabet, size=(800, 1000))
body_data = np.random.choice(alphabet, size=(800, 1000))
tags_data = np.random.randint(2, size=(800, num_tags)).astype("float32")
body_data = ["".join(body_data[i]) for i in range(len(body_data))]
title_data = ["".join(title_data[i]) for i in range(len(title_data))]
# Dummy target data
priority_targets = np.random.random(size=(800, 1))
dept_targets = np.random.randint(2, size=(800, num_departments))
model.fit(
{"title_input": title_data, "body_input": body_data, "input3": tags_data},
{"priority": priority_targets, "departments": dept_targets},
epochs=2,
batch_size=32, )
我自己想好了：
输入
模型的输入是正确的。不过我不需要压平。
tags_input (InputLayer) [(None, 12)] 0
__________________________________________________________________________________________________
title_input (InputLayer) [(None, 1)] 0
__________________________________________________________________________________________________
body_input (InputLayer) [(None, 1)] 0
__________________________________________________________________________________________________
输出
输出不正确，我不希望(None, 500, 1)和(None, 500, 4)作为输出。我只需要1优先级分数和4值的1部门列表。
要将形状从(None, 500, 1)更改为(None, 1)，我需要删除一些值。有很多方法可以做到这一点，在这里我选择了直接把中间调暗。
...
departments = layers.Dense(num_departments)(x) # Shape: (None, 500, 4)
departments = tf.slice(departments, [0, 0, 0], [-1, 1, 4]) # Shape (None, 1, 4)
departments = tf.squeeze(departments, [1]) # Shape (None, 4) but its not a squeeze type
departments = layers.Dense(num_departments, name="departments")(departments) # Shape (None, 4)
...
与priority_score输出相同。
现在输出变成
priority_score (Dense) (None, 1) 2 tf.compat.v1.squeeze[0][0]
__________________________________________________________________________________________________
departments (Dense) (None, 4) 20 tf.compat.v1.squeeze_1[0][0]
列车型号
下一步是准备训练数据。我们需要的是构建
- 标题数据：N个字符串，形状(N，1(。这里1表示一个python字符串
- body数据：与tiele数据相同
- 标记数据：浮点数的N数组，形状(N，12(
目标：
- 优先级_核心：N个浮点，形状(N，1(
- 部门：浮动的N阵列，形状(N，4(
其中N可以是任何数字。
然后我们调用拟合函数：
model.fit(
{"title_input": title_data, "body_input": body_data, "tags_input": tags_data},
{"priority_score": priority_targets, "departments": dept_targets, },
epochs=50,
batch_size=64, )
令人惊讶的是，损失不断增加：
Epoch 1/50
157/157 [==============================] - 5s 28ms/step - loss: 1.3467 - priority_score_loss: 0.6938 - departments_loss: 3.2644 - priority_score_acc: 0.0000e+00 - departments_acc: 0.1267
Epoch 2/50
157/157 [==============================] - 4s 27ms/step - loss: 4.6381 - priority_score_loss: 0.6976 - departments_loss: 19.7023 - priority_score_acc: 0.0000e+00 - departments_acc: 0.2483
Epoch 3/50
157/157 [==============================] - 4s 28ms/step - loss: 16.9411 - priority_score_loss: 0.6984 - departments_loss: 81.2137 - priority_score_acc: 0.0000e+00 - departments_acc: 0.1569
Epoch 4/50
157/157 [==============================] - 5s 29ms/step - loss: 23.8020 - priority_score_loss: 0.7075 - departments_loss: 115.4721 - priority_score_acc: 0.0000e+00 - departments_acc: 0.1427
Epoch 5/50
157/157 [==============================] - 5s 29ms/step - loss: 1.8650 - priority_score_loss: 0.7046 - departments_loss: 5.8019 - priority_score_acc: 0.0000e+00 - departments_acc: 0.1995
Epoch 6/50
157/157 [==============================] - 5s 30ms/step - loss: 3.0613 - priority_score_loss: 0.7025 - departments_loss: 11.7943 - priority_score_acc: 0.0000e+00 - departments_acc: 0.2472
Epoch 7/50
157/157 [==============================] - 5s 30ms/step - loss: 5.2455 - priority_score_loss: 0.7032 - departments_loss: 22.7114 - priority_score_acc: 0.0000e+00 - departments_acc: 0.2402
Epoch 8/50
157/157 [==============================] - 5s 30ms/step - loss: 6.0378 - priority_score_loss: 0.7013 - departments_loss: 26.6828 - priority_score_acc: 0.0000e+00 - departments_acc: 0.2418
Epoch 9/50
157/157 [==============================] - 5s 30ms/step - loss: 10.8300 - priority_score_loss: 0.7033 - departments_loss: 50.6334 - priority_score_acc: 0.0000e+00 - departments_acc: 0.2465
Epoch 10/50
157/157 [==============================] - 4s 27ms/step - loss: 12.1005 - priority_score_loss: 0.7019 - departments_loss: 56.9929 - priority_score_acc: 0.0000e+00 - departments_acc: 0.2627
Epoch 11/50
157/157 [==============================] - 4s 27ms/step - loss: 15.8248 - priority_score_loss: 0.6983 - departments_loss: 75.6328 - priority_score_acc: 0.0000e+00 - departments_acc: 0.2513
Epoch 12/50
157/157 [==============================] - 5s 29ms/step - loss: 19.3059 - priority_score_loss: 0.6940 - departments_loss: 93.0596 - priority_score_acc: 0.0000e+00 - departments_acc: 0.2386
Epoch 13/50
157/157 [==============================] - 5s 29ms/step - loss: 32.6499 - priority_score_loss: 0.6937 - departments_loss: 159.7808 - priority_score_acc: 0.0000e+00 - departments_acc: 0.2526
Epoch 14/50
157/157 [==============================] - 4s 28ms/step - loss: 31.1433 - priority_score_loss: 0.6936 - departments_loss: 152.2486 - priority_score_acc: 0.0000e+00 - departments_acc: 0.2499
Epoch 15/50
157/157 [==============================] - 5s 29ms/step - loss: 41.9199 - priority_score_loss: 0.6932 - departments_loss: 206.1338 - priority_score_acc: 0.0000e+00 - departments_acc: 0.2362
Epoch 16/50
157/157 [==============================] - 5s 30ms/step - loss: 40.2069 - priority_score_loss: 0.6931 - departments_loss: 197.5692 - priority_score_acc: 0.0000e+00 - departments_acc: 0.2300
Epoch 17/50
157/157 [==============================] - 5s 30ms/step - loss: 60.4129 - priority_score_loss: 0.6932 - departments_loss: 298.5986 - priority_score_acc: 0.0000e+00 - departments_acc: 0.2425
Epoch 18/50
157/157 [==============================] - 5s 30ms/step - loss: 75.8330 - priority_score_loss: 0.6932 - departments_loss: 375.6990 - priority_score_acc: 0.0000e+00 - departments_acc: 0.2332
Epoch 19/50
157/157 [==============================] - 5s 29ms/step - loss: 81.5731 - priority_score_loss: 0.6931 - departments_loss: 404.4002 - priority_score_acc: 0.0000e+00 - departments_acc: 0.2568
Epoch 20/50
157/157 [==============================] - 4s 28ms/step - loss: 103.4053 - priority_score_loss: 0.6932 - departments_loss: 513.5608 - priority_score_acc: 0.0000e+00 - departments_acc: 0.2409
Epoch 21/50
157/157 [==============================] - 4s 28ms/step - loss: 106.4842 - priority_score_loss: 0.6932 - departments_loss: 528.9552 - priority_score_acc: 0.0000e+00 - departments_acc: 0.2584
Epoch 22/50
157/157 [==============================] - 4s 28ms/step - loss: 121.2103 - priority_score_loss: 0.6932 - departments_loss: 602.5854 - priority_score_acc: 0.0000e+00 - departments_acc: 0.2332
Epoch 23/50
157/157 [==============================] - 5s 29ms/step - loss: 139.4970 - priority_score_loss: 0.6932 - departments_loss: 694.0189 - priority_score_acc: 0.0000e+00 - departments_acc: 0.2421
Epoch 24/50
157/157 [==============================] - 5s 29ms/step - loss: 180.7346 - priority_score_loss: 0.6933 - departments_loss: 900.2067 - priority_score_acc: 0.0000e+00 - departments_acc: 0.2449
Epoch 25/50
157/157 [==============================] - 4s 28ms/step - loss: 201.8011 - priority_score_loss: 0.6932 - departments_loss: 1005.5396 - priority_score_acc: 0.0000e+00 - departments_acc: 0.2420
Epoch 26/50
我想这是因为训练数据是随机生成的，并且模型构建得不好。无论如何，我们现在可以训练模型并用一些数据进行预测。
这是一次很好的学习经历。