TensorFlow 'InvalidArgumentError: Graph execution error' — Find the Real Cause

InvalidArgumentError: Graph execution error on its own tells you almost nothing — it's the outer wrapper. The message you actually need is nested further down the traceback. Here are the three causes behind most occurrences.

1. Shape mismatch

The classic is batching tensors of different shapes (e.g. images of varying sizes):

InvalidArgumentError: Cannot batch tensors with different shapes in
component 0. First element had shape [408,500,3] and element 1 had shape [360,343,3].

Resize/pad to a common shape before batching:

def preprocess(image, label):
    image = tf.image.resize(image, [400, 400])
    return image, label
 
ds = ds.map(preprocess).batch(32)

The same class of error appears when a layer's input/output shapes don't line up — check model.summary() and the shape feeding the failing layer.

2. Label out of range (very common)

With sparse_categorical_crossentropy, labels must be 0 .. num_classes-1. A label equal to num_classes (1-indexed labels, or a stray extra class) triggers:

InvalidArgumentError: Received a label value of 10 which is outside the
valid range of [0, 10).  Label values: ...

Fix the labels or the final layer's unit count so they agree:

import numpy as np
print(np.min(y), np.max(y))          # must be within [0, num_classes-1]
# e.g. 10 classes -> Dense(10), labels 0..9
outputs = tf.keras.layers.Dense(num_classes, activation="softmax")

3. Dtype mismatch

An op fed a dtype it doesn't expect (e.g. integer where float is required). Cast explicitly:

x = tf.cast(x, tf.float32)

How to debug any graph execution error

1. Scroll to the innermost error — ignore the outer wrapper.
2. Print shapes and dtypes of the tensors feeding the failing op.
3. For classification, print label min/max versus the number of output units.
4. Reproduce with one batch (ds.take(1)) to isolate it quickly.

Related guides

• AI/ML interview & engineering hub

Sources

• TensorFlow API docs