ValueError: Invalid dtype: float32_ref - SVD & SVD++

[TonciG]

Hi,
First of all, this is an amazing library and a real pleasure to use. Great job!
Unfortunately I have encountered this issue (which is due data incompatibility) when I try to optimise (using Optuna) hyperparameters for SVD and SVD++.
The issue seem to be linked to 'reg' hyperparameter of these algorithms. Namely, when I try to optimise it together with other hyperparameters (e.g. loss_type, embed_size, norm_embed (in case of SVD), lr, lr_decay, and batch_size), the issue arises. But when I set reg = 0, no problem whatsoever. So everything is okay unless you try to optimise reg hyperparameter.
Do you have any idea why this may be so and how it could be fixed?
Btw, my versions are:
Python == 3.11.13
numpy == 1.26.4
keras == 3.14.0
tensorflow == 2.21.0
optuna == 4.8.0
Thanks

[massquantity]

Hi, this looks like a TensorFlow compatibility issue. The error ValueError: Invalid dtype: float32_ref is specific to TensorFlow 1.x internals — float32_ref is how TF1 represents mutable variable dtypes internally, and it leaks out when variables are reused across Optuna trials in the same process.

Try downgrading to TensorFlow 2.15, which is the last version that ships with the TF1 compatibility layer (tf.compat.v1) intact and tends to handle TF1-style graphs more gracefully.

pip install tensorflow==2.15

If that alone doesn't solve it, also add tf.compat.v1.reset_default_graph() at the top of your Optuna objective function to ensure no graph state leaks between trials:

def objective(trial):
    tf.compat.v1.reset_default_graph()
    
    reg = trial.suggest_float("reg", 1e-5, 1e-1, log=True)
    # ... rest of your trial

The reason reg=0 works is that the regularization branch is likely skipped or short-circuited entirely when reg=0, so the problematic variable reference is never hit.

[TonciG]

Hi,
I would like to fix the problem so it does not require downgrading TensorFlow. This is my last resort.
And since I already use your second suggestion (clearing and resetting the default graph at the start of every trial), will try with gc_after_trial=True parameter of the study.optimize function.
In any case, many thanks for your help!

[TonciG]

Hi, I tried all of the options above (have been used #1 successfully until this issue has arisen) and no one is working. Even tried to downgrade TF to 2.15 but that requires downgrading Python from the version I am using now. So I've given up and decided to optimise SVD and SVD++ without reg. Cheers

…

On Wed, Apr 8, 2026 at 7:01 PM massquantity ***@***.***> wrote: *massquantity* left a comment (massquantity/LibRecommender#532) <#532 (comment)> Answer from Claude Sonnet 4.6 Good point! Since you can't modify the library, you need to work around it on your side. The key insight is that *Optuna is passing a Python float to the library, but somewhere the library's TF1 graph is getting confused across trials.* The most effective workarounds, in order of likelihood to work: ------------------------------ *1. Force a fresh TF graph per trial* This is the most likely fix — Optuna reuses the same Python process, so TF1 variables from previous trials bleed into new ones: import tensorflow as tf def objective(trial): # Reset the default graph before each trial tf.reset_default_graph() reg = trial.suggest_float("reg", 1e-5, 1e-1, log=True) # ... rest of your trial code ------------------------------ *2. Run each trial in an isolated session* def objective(trial): tf.reset_default_graph() with tf.Session() as sess: reg = trial.suggest_float("reg", 1e-5, 1e-1, log=True) # ... build and run model inside session ------------------------------ *3. Use Optuna's process-level isolation* Tell Optuna to run each trial in a *fresh subprocess*, so the TF graph is completely clean: study = optuna.create_study(direction="minimize")study.optimize( objective, n_trials=100, n_jobs=1, # Each trial gets a fresh process - no graph state leaks gc_after_trial=True ) gc_after_trial=True forces garbage collection between trials, which often resolves TF1 variable leakage even without full process isolation. ------------------------------ *4. Suggest reg as categorical with explicit floats* If the above don't work, constrain reg to a discrete set of *explicit Python floats* to avoid any type ambiguity: reg = trial.suggest_categorical("reg", [1e-5, 1e-4, 1e-3, 1e-2, 1e-1])# NOT [0, 0.001, ...] — the int 0 can cause mixed types ------------------------------ *Recommended combination* def objective(trial): tf.reset_default_graph() # 👈 most important line reg = trial.suggest_float("reg", 1e-5, 1e-1, log=True) loss_type = trial.suggest_categorical("loss_type", ["squared_loss", "log_loss"]) embed_size = trial.suggest_int("embed_size", 16, 256) lr = trial.suggest_float("lr", 1e-4, 1e-1, log=True) # ... model = SVD(reg=reg, loss_type=loss_type, ...) # your library call # ... train and evaluate return score study.optimize(objective, n_trials=100, gc_after_trial=True) tf.reset_default_graph() at the top of each trial is almost certainly the fix here. — Reply to this email directly, view it on GitHub <#532 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AUAJCJG7XH2PE73BZFMFW234U2HX7AVCNFSM6AAAAACXQUANKKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DEMBYGM3TKMBQGQ> . You are receiving this because you authored the thread.Message ID: ***@***.***>