how-to-use-slim.md

How to use Slim

Install

$ pip install --upgrade tf_slim

Usage

import tf_slim as slim

How to train a model using the Slim tool

1. Clone the repository:

$ git clone https://github.com/tensorflow/models.git
$ cd models/research/slim

2. Download the Visual Wake Words dataset and covert it to the labeled data (TFRecords).

$ python3 download_and_convert_data.py \
    --dataset_name=visualwakewords \
    --dataset_dir=data/visualwakewords

3. Train the mobilenet_v1_025 model using the below command:

$ python3 train_image_classifier.py \
    --train_dir=vww_96_grayscale \
    --dataset_name=visualwakewords \
    --dataset_split_name=train \
    --dataset_dir=data/visualwakewords \
    --model_name=mobilenet_v1_025 \
    --preprocessing_name=mobilenet_v1 \
    --train_image_size=96 \
    --use_grayscale=True \
    --save_summaries_secs=300 \
    --learning_rate=0.045 \
    --label_smoothing=0.1 \
    --learning_rate_decay_factor=0.98 \
    --num_epochs_per_decay=2.5 \
    --moving_average_decay=0.9999 \
    --batch_size=96 \
    --max_number_of_steps=100000

If you use CPU only, specify --clone_on_cpu=True:

$ python3 train_image_classifier.py --clone_on_cpu=True\
    --train_dir=vww_96_grayscale \
    --dataset_name=visualwakewords \
    --dataset_split_name=train \
    --dataset_dir=data/visualwakewords \
    --model_name=mobilenet_v1_025 \
    --preprocessing_name=mobilenet_v1 \
    --train_image_size=96 \
    --use_grayscale=True \
    --save_summaries_secs=300 \
    --learning_rate=0.045 \
    --label_smoothing=0.1 \
    --learning_rate_decay_factor=0.98 \
    --num_epochs_per_decay=2.5 \
    --moving_average_decay=0.9999 \
    --batch_size=96 \
    --max_number_of_steps=100000

4. Monitor the training process using tensorboard:

$ tensorboard --logdir=vww_96_grayscale
...
TensorBoard 1.15.0 at http://mijin:6006/ (Press CTRL+C to quit)

You can see the experiment metrics like loss and accuracy by accessing the below links:

# On the local machine
http://mijin:6006

# On the remote machine (e.g., mijin's IP: 123.123.123.123)
http://123.123.123.123:6006

For example:

5. Evaluate the model during the training process using the below command.

Replace the model.ckpt-xxx value of --checkpoint_path with the correct value according to the prefix of the checkpoint files in your --train_dir. In my case, the prefix was model.ckpt-7505:

$ python3 eval_image_classifier.py \
--alsologtostderr \
--checkpoint_path=vww_96_grayscale/model.ckpt-7505 \
--dataset_dir=data/visualwakewords \
--dataset_name=visualwakewords \
--dataset_split_name=val \
--model_name=mobilenet_v1_025 \
--preprocessing_name=mobilenet_v1 \
--use_grayscale=True \
--train_image_size=96

...
INFO:tensorflow:Evaluation [406/406]
I0917 16:16:56.870146 139859572754176 evaluation.py:167] Evaluation [406/406]
eval/Accuracy[0.795936]
eval/Recall_5[1]
INFO:tensorflow:Finished evaluation at 2020-09-17-16:16:57
I0917 16:16:57.096140 139859572754176 evaluation.py:275] Finished evaluation at 2020-09-17-16:16:57

In the above case, the accuaracy is about 79.59% (0.795936).

Reference

• TensorFlow-Slim
• TensorFlow Model Garden
• TinyML Book