TensorFlow报错InvalidArgumentError：需为float类型Placeholder张量赋值

Hey there! Let's work through this error you're facing—it's a super common issue when transitioning from image datasets like MNIST to text data in TensorFlow, so you're not alone here. Let's break down the likely causes and fixes step by step:

1. Mismatched Data Type Between Placeholder and Input Data #

This is the most frequent culprit. The error is telling you that the tensor named Placeholder expects float values, but you're feeding something else (like integers, or a numpy array with the wrong dtype).

• Check your placeholder definition: Make sure it's explicitly set to tf.float32 (or tf.float64 if that's what you need):

  # Example: Correct placeholder for text data (seq_len=10, emb_dim=100, plus channel dim)
  x_input = tf.placeholder(tf.float32, shape=[None, 10, 100, 1], name='Placeholder')
  

• Ensure your random training data is cast to float: When generating random data, don't forget to convert it to the matching dtype:

  rdm = RandomState(42)
  # Generate random data and cast to float32
  train_data = rdm.randn(batch_size, 10, 100).astype(np.float32)
  # Add a channel dimension (critical for CNNs, which expect 4D input: [batch, height, width, channels])
  train_data = np.expand_dims(train_data, axis=-1)
  

2. Incorrect Key in feed_dict #

Sometimes the error pops up not because of dtype, but because you're not feeding the right placeholder in your feed_dict.

• If you assigned the placeholder to a variable (e.g., x_input = tf.placeholder(...)), make sure you use that variable as the key in feed_dict, not just the string name:

  # Wrong: Using the tensor name directly might lead to issues (TensorFlow appends :0 by default)
  sess.run(train_op, feed_dict={'Placeholder': train_data})
  
  # Right: Use the variable you defined
  sess.run(train_op, feed_dict={x_input: train_data})
  

3. Missing Channel Dimension for CNN Input #

Unlike MNIST's 28x28 square (which becomes 28x28x1 for CNNs), your text data is 10x100—CNNs in TensorFlow expect 4D input tensors ([batch_size, height, width, channels]). For text, we usually add a single channel dimension (since we're dealing with 1D sequences in 2D tensor form).

• Use np.expand_dims to add that final dimension, as shown in step 1. Without it, even if dtype is correct, you might run into shape-related errors that could cascade into this placeholder issue.

Example Working Snippet #

Here's a condensed version of how your code might look with these fixes:

import tensorflow as tf
import numpy as np
from numpy.random import RandomState

batch_size = 16
seq_len = 10
emb_dim = 100
num_classes = 2

# Define placeholders with correct dtype and shape
x_input = tf.placeholder(tf.float32, shape=[None, seq_len, emb_dim, 1], name='Placeholder')
y_input = tf.placeholder(tf.int32, shape=[None, num_classes])

# Build a simple CNN for text
conv_layer = tf.layers.conv2d(
    inputs=x_input,
    filters=32,
    kernel_size=(3, emb_dim),  # Convolve over 3 words and full embedding dimension
    activation=tf.nn.relu
)
pool_layer = tf.layers.max_pooling2d(inputs=conv_layer, pool_size=(seq_len - 3 + 1, 1), strides=1)
flattened = tf.layers.flatten(pool_layer)
logits = tf.layers.dense(inputs=flattened, units=num_classes)

# Loss and optimizer
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels=y_input, logits=logits))
train_op = tf.train.AdamOptimizer(0.001).minimize(loss)

# Generate valid training data
rdm = RandomState(123)
train_x = rdm.randn(batch_size, seq_len, emb_dim).astype(np.float32)
train_x = np.expand_dims(train_x, axis=-1)
train_y = np.eye(num_classes)[rdm.randint(0, num_classes, size=batch_size)]

# Run training
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    # Feed the correct variables with matching dtype
    _, loss_val = sess.run([train_op, loss], feed_dict={x_input: train_x, y_input: train_y})
    print(f"First batch loss: {loss_val:.4f}")

Start by verifying the dtype match first—it's the easiest fix. If that doesn't work, double-check your feed_dict keys and input shape. You'll have this up and running in no time!

内容的提问来源于stack exchange，提问作者Miranda