TF backend: keras.ops.histogram calls .numpy() in graph mode and crashes with AttributeError on symbolic tensors

[jacob-talroo]

Summary

On the TensorFlow backend, keras.ops.histogram crashes in graph mode because the TF implementation converts a tensor to a Python list via .numpy().tolist(). This is not graph-safe for symbolic tensors in a Keras Functional/Model context. The same code works on the JAX backend.

Minimal reproducible example

# Repro: Keras 3 + TensorFlow backend, graph mode, jit or not
import os
os.environ["KERAS_BACKEND"] = "tensorflow"

import tensorflow as tf
import keras

tf.config.run_functions_eagerly(False)  # ensure graph

class HistogramProbe(keras.layers.Layer):
    def __init__(self, bins=8, lo=0.0, hi=1.0, **kwargs):
        super().__init__(**kwargs)
        self.bins = int(bins)
        self.lo = float(lo)
        self.hi = float(hi)

    def call(self, x):
        flat = keras.ops.reshape(x, (-1,))
        # This triggers TF backend histogram implementation which does .numpy().tolist()
        counts, edges = keras.ops.histogram(flat, bins=self.bins, range=(self.lo, self.hi))
        return x  # identity on data path

inp = keras.Input(shape=(16,), name="inp")
h = keras.layers.Dense(8, activation="relu")(inp)
h = HistogramProbe(name="probe")(h)
out = keras.layers.Dense(1)(h)
model = keras.Model(inp, out)

model.compile(optimizer="adam", loss="mse", jit_compile=True)  # jit_compile can be True or False
x = tf.random.uniform((32, 16))
y = tf.zeros((32, 1))

# Fails with AttributeError in graph mode
model.fit(x, y, epochs=1, batch_size=8)

See https://colab.research.google.com/drive/19R8qt7UjmX6Qz4YNkmz7KWWe3wO27MEw?usp=sharing

Observed error (trimmed)

AttributeError: Exception encountered when calling HistogramProbe.call().

Could not automatically infer the output shape / dtype of 'probe' ...
Error encountered:
'SymbolicTensor' object has no attribute 'numpy'

Arguments received by HistogramProbe.call():
  • args=('<KerasTensor shape=(None, 8), dtype=float32, ...>',)

Root cause (source references)
In the TF backend implementation of keras.ops.histogram, the code does:

bin_edges = tf.linspace(min_val, max_val, bins + 1)
bin_edges_list = bin_edges.numpy().tolist()  # <-- requires eager tensor
bin_indices = tf.raw_ops.Bucketize(input=x, boundaries=bin_edges_list[1:-1])

https://github.com/keras-team/keras/blob/b491c860fc2750e2b6006b55358d3251dbb4a9f0/keras/src/backend/tensorflow/numpy.py#L2966C1-L2966C48

Calling .numpy() on a symbolic tensor in a graph causes the crash.

Expected behavior

keras.ops.histogram should be graph-safe on the TensorFlow backend (or clearly documented as eager-only). Ideally it should run inside a Keras model without errors.

Environment

Please replace the versions with your actual outputs:

keras: 3.11.3
tensorflow: 2.19.0
jax: 0.5.3
python: 3.12.11 (main, Jun  4 2025, 08:56:18) [GCC 11.4.0]
platform: Linux-6.6.97+-x86_64-with-glibc2.35
backend: tensorflow

Workarounds

Something like:

def tf_histogram_fixed_width_1d(x, bins, range):
    lo = range[0]
    hi = range[1]
    x = tf.reshape(x, [-1])
    x = tf.clip_by_value(x, lo, hi)
    counts = tf.histogram_fixed_width(values=x, value_range=[lo, hi], nbins=bins)
    return counts, tf.linspace(lo, hi, bins + 1)

However, I think there might be slight edge issues around lo and hi.

Notes

• The same repro runs fine on the JAX backend.
• The failure happens whether jit_compile=True or False, as long as the call happens in graph mode / symbolic context.

Thanks!

[jacob-talroo]

Slightly better alternative implementation:

def tf_ops_histogram(x, bins=10, range=None):
    import tensorflow as tf

    min_val = tf.math.reduce_min(x) if range is None else range[0]
    max_val = tf.math.reduce_max(x) if range is None else range[1]

    # Ignore out-of-range values (inclusive mask like Keras backend)
    in_range_mask = (x >= min_val) & (x <= max_val)
    in_range_x = tf.boolean_mask(x, in_range_mask)

    counts = tf.histogram_fixed_width(
        values=in_range_x,
        value_range=[min_val, max_val],
        nbins=bins,
    )

    return tf.cast(counts, x.dtype), tf.linspace(min_val, max_val, bins + 1)

It is still not exactly the same as keras.ops.histogram for some edge cases like NaNs, infinite and not flat values.

[sonali-kumari1]

Hi @jacob-talroo -
Thanks for the detailed information and the helpful workarounds!
I have reproduced this issue on my end with the latest version of Keras(3.11.3) and TensorFlow(2.20.0) and encountered the same 'SymbolicTensor' object has no attribute 'numpy' error in this gist. We will look into this issue and update you soon. Thanks again!