TensorFlow 2.6 updates and EfficientNet V2

[mrdbourke]

TensorFlow 2.6 recently got released: https://github.com/tensorflow/tensorflow/releases/tag/v2.6.0

Also, EfficientNetV2's are on TensorFlow Hub: https://tfhub.dev/google/collections/efficientnet_v2/1

I've created a notebook with some of the changes related to the course, see it here: https://colab.research.google.com/drive/1xUvtafkdJbvcqI7gyCO4rHAKLw-Che8T?usp=sharing

I'll put the major notes here too.

See the video replay of going through these updates on YouTube: https://youtu.be/B5qe_DDd8KA

TensorFlow 2.6

One (slightly) major change you'll notice throughout the course to using TensorFlow 2.6+ is the tf.keras.layers namespace.

From the update:

Import paths moved from tf.keras.layers.preprocessing.experimental to tf.keras.layers

For example,

# Old 
rescale_old = tf.keras.layers.experimental.preprocessing.Rescaling(scale=1/255.)

# New
rescale_new = tf.keras.layers.Rescaling(scale=1/255.)

Notice the shorter paths. This is the case for many layers which were previously in the tf.keras.layers.experimental.preprocessing module.

Course notebook 08 reflects and notes this change.

EfficientNetV2 pretrained models on TensorFlow Hub

In the course we use pretrained EfficientNet V1 for image classification projects (food vision).

Now EfficientNet V2 (faster and more accurate than V1) is available through TensorFlow Hub.

Note: In the experimental notebook linked above, using the same data with two different models (EfficientNetB0 V1 vs EfficientNetB0 V2), it looks like EfficientNetB0 V1 outperforms V2. I'm not sure why this is happening. Best to experiment with your own projects and see which is better.

Not much changes though, only the TensorFlow Hub URL used to load in a pretrained model:

# EfficientNetB0 v1 feature vector (old)
efficientnetb0_v1_url = "https://tfhub.dev/tensorflow/efficientnet/b0/feature-vector/1"

# EfficientNetB0 v2 feature vector (new)
efficientnetb0_v2_url = "https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet1k_b0/feature_vector/2"

I've updated course notebook 04 to reflect this change.

For more, see the resources linked above.