Anaconda installs TensorFlow 1.15 instead of 2.0

Question

I've been trying to install Tensorflow 2.0 in Anaconda.

So far, tensorflow works fine (i can use the library in my code), but when i'm installing it, it says "installing version: 2.0", then i get version 1.15 instead.

The whole package results updatable (since versione 2.0 does exist and should be available), but even if i try to update it, i keep getting 1.15, which gets back to updatable, and i get back into the loop.

I tried both from the Anaconda Navigator interface and the prompt using conda update tensorflow, without any success.

Here's there is the link package: tensorflow 2.0 Anaconda

How can i solve the problem?

EDIT :

I successfully installed TensorFlow 2.0 using conda install -c anaconda tensorflow. Then i got back into Anaconda Navigator and i tried to install Keras 2.2.4.

Looking at the installed packages, seems that installing Keras overwrites the version 2.0 of TensorFlow with the 1.15, am i wrong? Is this a bug?

Here's the image:

Solution 1

conda install -c anaconda tensorflow-gpu

or

conda install -c anaconda tensorflow

Look for these:

---

Tip: completely shred any file w/ name substring "keras" and "tensorflow" in your Anaconda environment before installing - else stuff is likely to break.

1. Navigate to your site-packages
2. Search **keras** --> delete all results
3. Search **tensorflow** --> delete all results

Above should work even better from the environment's root folder (e.g. envs/tf2_env) -- the idea is, Anaconda caches whatever it's already installed, so if it isn't completely removed, you may find trouble getting latest vers of packages.

---

Related Q's answered in comments:

1. Why is conda reverting TF 2 to TF 1.15.0 when I try installing Keras 2.2.4?

   • TF2's only compatible w/ Keras 2.2.5+, so conda is taking care of compatibility

2. Why are some from keras imports no longer working and need from tensorflow.keras?

   • from keras is very different from from tensorflow.keras, implementation-wise. from keras is backend-neutral, whereas tf.keras is more tightly integrated w/ TF. There can also be significant performance differences. TF2 and onward, tf.keras is preferable, and development will proceed along it, as noted by F. Chollet
   • Also beware of tf.python.keras

Solution 2

Another alternative is to install Tensorflow with Miniconda.

I simply have used Miniconda3 for windows and I was able to install multiple versions of Tensorflow on different environment respectively.

Example 1

// create a new environment
conda create --name workflowone

// activate environment
conda activate workflowone

// Install desired package
conda install tensorflow=1.15

// deactivate environment
conda deactivate

Example 2

// create a new environment
conda create --name workflowtwo

// activate environment
conda activate workflowtwo

// Install desired package
conda install tensorflow  // If you do not specify the version, you will download the latest version of Tensorflow

// deactivate environment
conda deactivate