Integration: Project renaming breaks TensorBoard loading
See original GitHub issueDescribe the bug I am using TF 2.3.1 and wandb 0.10.10., and Safari.
I log the data via the TensorBoard callback for keras, and set sync_tensorboard to True. With this settings, everything worked fine the last days. I did not change anything. As of today, I can not visualize the logged data in TB, this is also for some older logs I did (weeks ago). It says no data, but I can recall using this visualization on this exact run with success.
On my two runs I did yesterday and this morning, I also log data, but the TensorBoard screen stays empty. I can’t seem to reproduce this behavior, for the following toy example it works fine:
from tensorflow import keras
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
import itertools, io, sklearn.metrics, os
!pip install wandb
import wandb
wandb.login()
def do_stuff():
import tensorflow as tf
from tensorflow.keras.layers import Input, Dense, Reshape
from tensorflow.keras import Sequential
model = Sequential([
Input(shape=(128)),
Dense(20, name="test1"),
Dense(3)
])
wandb.init(sync_tensorboard=True)
wand_cb = wandb.keras.WandbCallback('val_sparse_categorical_crossentropy')
inp = tf.random.uniform([128, 128], dtype=tf.float32)
labels = tf.random.uniform([128], 0, 3, dtype=tf.int32)
labels = np.random.randint(0, 3, size=(128))
model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=["sparse_categorical_accuracy"])
#model.summary()
tb = tf.keras.callbacks.TensorBoard(log_dir="test")
model.fit(inp, labels, epochs=10, callbacks=[tb, wand_cb])
print(wandb.run.name)
wandb.finish()
I noticed that I exceed the storage limit by 8 GBs, can this infer? (Also suggesting to send users a notification if quota is reached).
Issue Analytics
- State:
- Created 3 years ago
- Comments:6 (2 by maintainers)
Top GitHub Comments
Update: Have it confirmed, renaming breaks the TensorBoard. Run below snippet in Colab once:
Afterwards, rename project “test_a” to e.g. “test_b” in the GUI. Now, check the TensorBoard feature for this run again in the GUI. It does not work anymore.
As a second step, re-run the code snippet again, changing the project in init to “test_b”. This visualization will fail too, albeit it was logged to the already renamed project. This seems like an issue on the side of data handling, not on the client side.
Hi @Yannik1337, this was fixed early this year. Please let us know if you’re still experiencing similar issues.