# Warning during online training leading to NaN loss

**URL:** <https://discuss.bayesflow.org/t/warning-during-online-training-leading-to-nan-loss/131>\
**Category:** General\
**Created:** [February 25, 2025, 12:56pm UTC](https://discuss.bayesflow.org/t/warning-during-online-training-leading-to-nan-loss/131 "2025-02-25T12:56:45Z")\
**Posts on this page:** 5\
**Page:** 1

<div class="post-metadata">

**Author:** ![philippreiser](https://avatars.discourse-cdn.com/v4/letter/p/f04885/32.png) [@philippreiser](https://discuss.bayesflow.org/u/philippreiser)\
**Post date:** [February 25, 2025, 12:56pm UTC](https://discuss.bayesflow.org/t/warning-during-online-training-leading-to-nan-loss/131/1 "2025-02-25T12:56:45Z")

</div>

Hi!

I’m currently working on the dev branch and training a FlowMatching Inference network using an OnlineDataset. However, I occasionally encounter the following warnings:

```python
/data/homes/reiser/.conda/envs/sabi_env/lib/python3.11/site-packages/bayesflow/utils/numpy_utils.py:7: RuntimeWarning: divide by zero encountered in log
  return np.log(x / (1 - x))
/data/homes/reiser/.conda/envs/sabi_env/lib/python3.11/site-packages/numpy/core/_methods.py:173: RuntimeWarning: invalid value encountered in subtract
  x = asanyarray(arr - arrmean)
/data/homes/reiser/.conda/envs/sabi_env/lib/python3.11/site-packages/bayesflow/adapters/transforms/standardize.py:86: RuntimeWarning: invalid value encountered in subtract  
return (data - mean) / std

```

This sometimes occurs after 71 epochs, and other times after 300 epochs, eventually leading to NaN values in the loss.

Do you have any suggestions on how to resolve this issue?  
Thanks in advance!

---

<div class="post-metadata">

**Author:** ![philippreiser](https://avatars.discourse-cdn.com/v4/letter/p/f04885/32.png) [@philippreiser](https://discuss.bayesflow.org/u/philippreiser)\
**Post date:** [February 25, 2025, 1:14pm UTC](https://discuss.bayesflow.org/t/warning-during-online-training-leading-to-nan-loss/131/2 "2025-02-25T13:14:54Z")

</div>

Here is an example where this issue occurred for me in Epoch 71:

```python
import os

if "KERAS_BACKEND" not in os.environ:
    # set this to "torch", "tensorflow", or "jax"
    os.environ["KERAS_BACKEND"] = "torch"

import bayesflow as bf
import torch as to
import keras
import numpy as np

# set random seed
to.manual_seed(0)
np.random.seed(0)

def prior():
    # beta: regression coefficients (intercept, slope)
    w = np.float32(np.random.uniform(0.6, 1.4))
    return dict(w=w)

def likelihood(w):
    N = 10
    # x: predictor variable
    x = np.float32(np.random.uniform(1, 200, N))
    # y: response variable
    y = np.random.normal(w * np.log(x) + 0.01 * x + 1 + np.sin(0.05 * x), 0.1, size=N)
    return dict(y=y, x=x)

simulator = bf.simulators.make_simulator([prior, likelihood])

adapter = (
    bf.Adapter()
    .constrain("w", lower=0.6, upper=1.4)
    .as_set(["x", "y"])
    .standardize()
    .concatenate(["w"], into="inference_variables")
    .concatenate(["x", "y"], into="summary_variables")
)
inference_network = bf.networks.FlowMatching()
summary_network = bf.networks.DeepSet(depth=2)
approximator = bf.ContinuousApproximator(
   inference_network=inference_network,
   summary_network=summary_network,
   adapter=adapter,
)
epochs = 500
num_batches = 100
batch_size = 64
optimizer = keras.optimizers.Adam(learning_rate=5e-4, clipnorm=1.0)
approximator.compile(optimizer=optimizer)
history = approximator.fit(
    epochs=epochs,
    num_batches=num_batches,
    batch_size=batch_size,
    simulator=simulator,
)

```

---

<div class="post-metadata">

**Author:** ![KLDivergence](https://yyz1.discourse-cdn.com/flex007/user_avatar/discuss.bayesflow.org/kldivergence/32/15_2.png) [@KLDivergence](https://discuss.bayesflow.org/u/KLDivergence)\
**Post date:** [February 25, 2025, 1:32pm UTC](https://discuss.bayesflow.org/t/warning-during-online-training-leading-to-nan-loss/131/3 "2025-02-25T13:32:32Z")

</div>

Hey Philipp, this seems to be coming from the Adapter. I suspect that the constrain method is to blame. Could you try removing it and see if the warnings disappear?

---

<div class="post-metadata">

**Author:** ![LarsKue](https://yyz1.discourse-cdn.com/flex007/user_avatar/discuss.bayesflow.org/larskue/32/100_2.png) [@LarsKue](https://discuss.bayesflow.org/u/LarsKue)\
**Post date:** [February 25, 2025, 1:40pm UTC](https://discuss.bayesflow.org/t/warning-during-online-training-leading-to-nan-loss/131/4 "2025-02-25T13:40:13Z")

</div>

Looks like the inverse sigmoid transform from the `adapter.constrain()` is at fault here. When the sampled `w` is sufficiently close to `0.6` or `1.4`, we get division by zero and infinite return values.

Since we don’t support making the boundaries inclusive yet, can you please add a small value to either end of the constraint and check if this helps:

```py3
epsilon = 1e-9
adapter.constrain("w", lower=0.6 - epsilon, upper=1.4 + epsilon)

```

---

<div class="post-metadata">

**Author:** ![philippreiser](https://avatars.discourse-cdn.com/v4/letter/p/f04885/32.png) [@philippreiser](https://discuss.bayesflow.org/u/philippreiser)\
**Post date:** [February 25, 2025, 4:00pm UTC](https://discuss.bayesflow.org/t/warning-during-online-training-leading-to-nan-loss/131/5 "2025-02-25T16:00:36Z")

</div>

Adding the small `epsilon` solved the problem. Thank you!