For MobilenetV2+ see this file mobilenet/README.md

MobileNets are small, low-latency, low-power models parameterized to meet the resource constraints of a variety of use cases. They can be built upon for classification, detection, embeddings and segmentation similar to how other popular large scale models, such as Inception, are used. MobileNets can be run efficiently on mobile devices with TensorFlow Lite.

MobileNets trade off between latency, size and accuracy while comparing favorably with popular models from the literature.

Choose the right MobileNet model to fit your latency and size budget. The size of the network in memory and on disk is proportional to the number of parameters. The latency and power usage of the network scales with the number of Multiply-Accumulates (MACs) which measures the number of fused Multiplication and Addition operations. These MobileNet models have been trained on the ILSVRC-2012-CLS image classification dataset. Accuracies were computed by evaluating using a single image crop.

Revisions to models:

• July 12, 2018: Update to TFLite models that fixes an accuracy issue resolved by making conversion support weights with narrow_range. We now report validation on the actual TensorFlow Lite model rather than the emulated quantization number of TensorFlow.
• August 2, 2018: Update to TFLite models that fixes an accuracy issue resolved by making sure the numerics of quantization match TF quantized training accurately.

The linked model tar files contain the following:

• Trained model checkpoints
• Eval graph text protos (to be easily viewed)
• Frozen trained models
• Info file containing input and output information
• Converted TensorFlow Lite flatbuffer model

Note that quantized model GraphDefs are still float models, they just have FakeQuantization operation embedded to simulate quantization. These are converted by TensorFlow Lite to be fully quantized. The final effect of quantization can be seen by comparing the frozen fake quantized graph to the size of the TFLite flatbuffer, i.e. The TFLite flatbuffer is about 1/4 the size. For more information on the quantization techniques used here, see here. There isn't any equivalent in TF2.x yet, more information can be found in this RFC

Here is an example of how to download the MobileNet_v1_1.0_224 checkpoint:

$ CHECKPOINT_DIR=/tmp/checkpoints
$ mkdir ${CHECKPOINT_DIR}
$ wget http://download.tensorflow.org/models/mobilenet_v1_2018_02_22/mobilenet_v1_1.0_224.tgz
$ tar -xvf mobilenet_v1_1.0_224.tgz
$ mv mobilenet_v1_1.0_224.ckpt.* ${CHECKPOINT_DIR}

This package contains scripts for training floating point and eight-bit fixed point TensorFlow models.

Quantization tools used are described here. There isn't any equivalent in TF2.x yet, more information can be found in this RFC

Conversion to fully quantized models for mobile can be done through TensorFlow Lite.

$ bazel build -c opt --config=cuda mobilenet_v1_{eval,train}

Train:

$ ./bazel-bin/mobilenet_v1_train --dataset_dir "path/to/dataset" --checkpoint_dir "path/to/checkpoints"

Eval:

$ ./bazel-bin/mobilenet_v1_eval --dataset_dir "path/to/dataset" --checkpoint_dir "path/to/checkpoints"

Train from preexisting float checkpoint:

$ ./bazel-bin/mobilenet_v1_train --dataset_dir "path/to/dataset" --checkpoint_dir "path/to/checkpoints" \
  --quantize=True --fine_tune_checkpoint=float/checkpoint/path

Train from scratch:

$ ./bazel-bin/mobilenet_v1_train --dataset_dir "path/to/dataset" --checkpoint_dir "path/to/checkpoints" --quantize=True

Eval:

$ ./bazel-bin/mobilenet_v1_eval --dataset_dir "path/to/dataset" --checkpoint_dir "path/to/checkpoints" --quantize=True

The resulting float and quantized models can be run on-device via TensorFlow Lite.