You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Have I written custom code (as opposed to using a stock example script provided in TensorFlow): No
OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Ubuntu 18.04
TensorFlow installed from (source or binary): built from source
TensorFlow version (use command below): 2.0.0 (i.e. the release)
Keras version: 2.3.0
Python version: 3.7 conda
Bazel version (if compiling Tensorflow from source): 0.26.1
GCC/Compiler version (if compiling Tensorflow from source): 7.4.0
CUDA/cuDNN version: 10 / 7.6.4
GPU model and memory: RTX 2080 Ti and Tesla V100 (tried on both. error occurs on both)
Describe the current behavior
I am going through Francois Chollet's book "Deep Learning with Python" and running the code in his Jupyter Notebooks with Tensorflow 2.0.0 as a backend to Keras 2.3.0. Notebook 6.3, (under the heading "1.6 Using recurrent dropout to fight overfitting") has a model with a tensorflow.keras.layers.GRU(32, dropout=0.2, recurrent_dropout=0.2, input_shape=(None, float_data.shape[-1])). The data is read earlier in the notebook from jena_climate_2009_2016.csv. I get a loss of 699013271268870062080.0000 after the first epoch and similar figures after subsequent epochs. This figure is simply wrong (see below). The original notebook (from Francois Chollet) is here: link to github and includes the correct output.
Describe the expected behavior
The loss after 1 or 2 epochs is supposed to be around 0.3
Code to reproduce the issue
Provide a reproducible test case that is the bare minimum necessary to generate the problem.
Download the data as follows:
cd ~
mkdir Datasets
cd ~/Datasets
mkdir jena_climate
cd jena_climate
wget https://s3.amazonaws.com/keras-datasets/jena_climate_2009_2016.csv.zip
unzip jena_climate_2009_2016.csv.zip
Run jupyter notebook and load the notebook in a Python 3.7 environment with tensorflow 2.0.0 as the backend and keras 2.30.
Run each cell from the beginning of the notebook so you load the data and create the generators before you get to the example under heading 1.6. Then try to run the example. You will find that the loss is terribly wrong.
The code under heading "1.7 Stacking recurrent layers" also runs incorrectly. The loss produced is "nan" and val_loss is "nan" (both should be around 0.3). I think it is the same problem with layers.GRU
I have reproduced this problem running tensorflow.keras in tensorflow 2.0.0.
The problem also occurs running Keras 2.3.0 with a tensorflow 1.1.4 backend.
The problem does not occur with tensorflow.keras in tensorflow 1.1.4.
The text was updated successfully, but these errors were encountered:
System information
Describe the current behavior
I am going through Francois Chollet's book "Deep Learning with Python" and running the code in his Jupyter Notebooks with Tensorflow 2.0.0 as a backend to Keras 2.3.0. Notebook 6.3, (under the heading "1.6 Using recurrent dropout to fight overfitting") has a model with a tensorflow.keras.layers.GRU(32, dropout=0.2, recurrent_dropout=0.2, input_shape=(None, float_data.shape[-1])). The data is read earlier in the notebook from jena_climate_2009_2016.csv. I get a loss of 699013271268870062080.0000 after the first epoch and similar figures after subsequent epochs. This figure is simply wrong (see below). The original notebook (from Francois Chollet) is here: link to github and includes the correct output.
Describe the expected behavior
The loss after 1 or 2 epochs is supposed to be around 0.3
Code to reproduce the issue
Provide a reproducible test case that is the bare minimum necessary to generate the problem.
Download the data as follows:
cd ~
mkdir Datasets
cd ~/Datasets
mkdir jena_climate
cd jena_climate
wget https://s3.amazonaws.com/keras-datasets/jena_climate_2009_2016.csv.zip
unzip jena_climate_2009_2016.csv.zip
Run jupyter notebook and load the notebook in a Python 3.7 environment with tensorflow 2.0.0 as the backend and keras 2.30.
Run each cell from the beginning of the notebook so you load the data and create the generators before you get to the example under heading 1.6. Then try to run the example. You will find that the loss is terribly wrong.
The code under heading "1.7 Stacking recurrent layers" also runs incorrectly. The loss produced is "nan" and val_loss is "nan" (both should be around 0.3). I think it is the same problem with layers.GRU
I have reproduced this problem running tensorflow.keras in tensorflow 2.0.0.
The problem also occurs running Keras 2.3.0 with a tensorflow 1.1.4 backend.
The problem does not occur with tensorflow.keras in tensorflow 1.1.4.
The text was updated successfully, but these errors were encountered: