TF-slim
is a new lightweight high-level API of TensorFlow (tensorflow.contrib.slim)
for defining, training and evaluating complex
models. This directory contains
code for training and evaluating several widely used Convolutional Neural
Network (CNN) image classification models using TF-slim.
It contains scripts that will allow
you to train models from scratch or fine-tune them from pre-trained network
weights. It also contains code for downloading standard image datasets,
converting them
to TensorFlow's native TFRecord format and reading them in using TF-Slim's
data reading and queueing utilities. You can easily train any model on any of
these datasets, as we demonstrate below. We've also included a
jupyter notebook,
which provides working examples of how to use TF-Slim for image classification.
Maintainers of TF-slim:
- Nathan Silberman, github: nathansilberman
- Sergio Guadarrama, github: sguada
Installation and setup
Preparing the datasets
Using pre-trained models
Training from scratch
Fine tuning to a new task
Evaluating performance
Exporting Inference Graph
Troubleshooting
In this section, we describe the steps required to install the appropriate prerequisite packages.
TF-Slim is available as tf.contrib.slim via TensorFlow 1.0. To test that your
installation is working, execute the following command; it should run without
raising any errors.
python -c "import tensorflow.contrib.slim as slim; eval = slim.evaluation.evaluate_once"
To use TF-Slim for image classification, you also have to install the TF-Slim image models library, which is not part of the core TF library. To do this, check out the tensorflow/models repository as follows:
cd $HOME/workspace
git clone https://github.com/tensorflow/models/This will put the TF-Slim image models library in $HOME/workspace/models/slim.
(It will also create a directory called
models/inception,
which contains an older version of slim; you can safely ignore this.)
To verify that this has worked, execute the following commands; it should run without raising any errors.
cd $HOME/workspace/models/slim
python -c "from nets import cifarnet; mynet = cifarnet.cifarnet"
As part of this library, we've included scripts to download several popular image datasets (listed below) and convert them to slim format.
| Dataset | Training Set Size | Testing Set Size | Number of Classes | Comments |
|---|---|---|---|---|
| Flowers | 2500 | 2500 | 5 | Various sizes (source: Flickr) |
| Cifar10 | 60k | 10k | 10 | 32x32 color |
| MNIST | 60k | 10k | 10 | 28x28 gray |
| ImageNet | 1.2M | 50k | 1000 | Various sizes |
For each dataset, we'll need to download the raw data and convert it to TensorFlow's native TFRecord format. Each TFRecord contains a TF-Example protocol buffer. Below we demonstrate how to do this for the Flowers dataset.
$ DATA_DIR=/tmp/data/flowers
$ python download_and_convert_data.py \
--dataset_name=flowers \
--dataset_dir="${DATA_DIR}"When the script finishes you will find several TFRecord files created:
$ ls ${DATA_DIR}
flowers_train-00000-of-00005.tfrecord
...
flowers_train-00004-of-00005.tfrecord
flowers_validation-00000-of-00005.tfrecord
...
flowers_validation-00004-of-00005.tfrecord
labels.txtThese represent the training and validation data, sharded over 5 files each.
You will also find the $DATA_DIR/labels.txt file which contains the mapping
from integer labels to class names.
You can use the same script to create the mnist and cifar10 datasets. However, for ImageNet, you have to follow the instructions here. Note that you first have to sign up for an account at image-net.org. Also, the download can take several hours, and could use up to 500GB.
Once the TFRecord files have been created, you can easily define a Slim Dataset, which stores pointers to the data file, as well as various other pieces of metadata, such as the class labels, the train/test split, and how to parse the TFExample protos. We have included the TF-Slim Dataset descriptors for Cifar10, ImageNet, Flowers, and MNIST. An example of how to load data using a TF-Slim dataset descriptor using a TF-Slim DatasetDataProvider is found below:
import tensorflow as tf
from datasets import flowers
slim = tf.contrib.slim
# Selects the 'validation' dataset.
dataset = flowers.get_split('validation', DATA_DIR)
# Creates a TF-Slim DataProvider which reads the dataset in the background
# during both training and testing.
provider = slim.dataset_data_provider.DatasetDataProvider(dataset)
[image, label] = provider.get(['image', 'label'])Neural nets work best when they have many parameters, making them powerful function approximators. However, this means they must be trained on very large datasets. Because training models from scratch can be a very computationally intensive process requiring days or even weeks, we provide various pre-trained models, as listed below. These CNNs have been trained on the ILSVRC-2012-CLS image classification dataset.
In the table below, we list each model, the corresponding TensorFlow model file, the link to the model checkpoint, and the top 1 and top 5 accuracy (on the imagenet test set). Note that the VGG and ResNet V1 parameters have been converted from their original caffe formats (here and here), whereas the Inception and ResNet V2 parameters have been trained internally at Google. Also be aware that these accuracies were computed by evaluating using a single image crop. Some academic papers report higher accuracy by using multiple crops at multiple scales.
All 16 MobileNet Models reported in the MobileNet Paper can be found here.
(*): Results quoted from the paper. Here is an example of how to download the Inception V3 checkpoint:
$ CHECKPOINT_DIR=/tmp/checkpoints
$ mkdir ${CHECKPOINT_DIR}
$ wget http://download.tensorflow.org/models/inception_v3_2016_08_28.tar.gz
$ tar -xvf inception_v3_2016_08_28.tar.gz
$ mv inception_v3.ckpt ${CHECKPOINT_DIR}
$ rm inception_v3_2016_08_28.tar.gzWe provide an easy way to train a model from scratch using any TF-Slim dataset. The following example demonstrates how to train Inception V3 using the default parameters on the ImageNet dataset.
DATASET_DIR=/tmp/imagenet
TRAIN_DIR=/tmp/train_logs
python train_image_classifier.py \
--train_dir=${TRAIN_DIR} \
--dataset_name=imagenet \
--dataset_split_name=train \
--dataset_dir=${DATASET_DIR} \
--model_name=inception_v3This process may take several days, depending on your hardware setup. For convenience, we provide a way to train a model on multiple GPUs, and/or multiple CPUs, either synchrononously or asynchronously. See model_deploy for details.
Rather than training from scratch, we'll often want to start from a pre-trained
model and fine-tune it.
To indicate a checkpoint from which to fine-tune, we'll call training with
the --checkpoint_path flag and assign it an absolute path to a checkpoint
file.
When fine-tuning a model, we need to be careful about restoring checkpoint
weights. In particular, when we fine-tune a model on a new task with a different
number of output labels, we wont be able restore the final logits (classifier)
layer. For this, we'll use the --checkpoint_exclude_scopes flag. This flag
hinders certain variables from being loaded. When fine-tuning on a
classification task using a different number of classes than the trained model,
the new model will have a final 'logits' layer whose dimensions differ from the
pre-trained model. For example, if fine-tuning an ImageNet-trained model on
Flowers, the pre-trained logits layer will have dimensions [2048 x 1001] but
our new logits layer will have dimensions [2048 x 5]. Consequently, this
flag indicates to TF-Slim to avoid loading these weights from the checkpoint.
Keep in mind that warm-starting from a checkpoint affects the model's weights
only during the initialization of the model. Once a model has started training,
a new checkpoint will be created in ${TRAIN_DIR}. If the fine-tuning
training is stopped and restarted, this new checkpoint will be the one from
which weights are restored and not the ${checkpoint_path}$. Consequently,
the flags --checkpoint_path and --checkpoint_exclude_scopes are only used
during the 0-th global step (model initialization). Typically for fine-tuning
one only want train a sub-set of layers, so the flag --trainable_scopes allows
to specify which subsets of layers should trained, the rest would remain frozen.
Below we give an example of fine-tuning inception-v3 on flowers, inception_v3 was trained on ImageNet with 1000 class labels, but the flowers dataset only have 5 classes. Since the dataset is quite small we will only train the new layers.
$ DATASET_DIR=/tmp/flowers
$ TRAIN_DIR=/tmp/flowers-models/inception_v3
$ CHECKPOINT_PATH=/tmp/my_checkpoints/inception_v3.ckpt
$ python train_image_classifier.py \
--train_dir=${TRAIN_DIR} \
--dataset_dir=${DATASET_DIR} \
--dataset_name=flowers \
--dataset_split_name=train \
--model_name=inception_v3 \
--checkpoint_path=${CHECKPOINT_PATH} \
--checkpoint_exclude_scopes=InceptionV3/Logits,InceptionV3/AuxLogits \
--trainable_scopes=InceptionV3/Logits,InceptionV3/AuxLogitsTo evaluate the performance of a model (whether pretrained or your own), you can use the eval_image_classifier.py script, as shown below.
Below we give an example of downloading the pretrained inception model and evaluating it on the imagenet dataset.
CHECKPOINT_FILE = ${CHECKPOINT_DIR}/inception_v3.ckpt # Example
$ python eval_image_classifier.py \
--alsologtostderr \
--checkpoint_path=${CHECKPOINT_FILE} \
--dataset_dir=${DATASET_DIR} \
--dataset_name=imagenet \
--dataset_split_name=validation \
--model_name=inception_v3Saves out a GraphDef containing the architecture of the model.
To use it with a model name defined by slim, run:
$ python export_inference_graph.py \
--alsologtostderr \
--model_name=inception_v3 \
--output_file=/tmp/inception_v3_inf_graph.pb
$ python export_inference_graph.py \
--alsologtostderr \
--model_name=mobilenet_v1 \
--image_size=224 \
--output_file=/tmp/mobilenet_v1_224.pbIf you then want to use the resulting model with your own or pretrained checkpoints as part of a mobile model, you can run freeze_graph to get a graph def with the variables inlined as constants using:
bazel build tensorflow/python/tools:freeze_graph
bazel-bin/tensorflow/python/tools/freeze_graph \
--input_graph=/tmp/inception_v3_inf_graph.pb \
--input_checkpoint=/tmp/checkpoints/inception_v3.ckpt \
--input_binary=true --output_graph=/tmp/frozen_inception_v3.pb \
--output_node_names=InceptionV3/Predictions/Reshape_1The output node names will vary depending on the model, but you can inspect and estimate them using the summarize_graph tool:
bazel build tensorflow/tools/graph_transforms:summarize_graph
bazel-bin/tensorflow/tools/graph_transforms/summarize_graph \
--in_graph=/tmp/inception_v3_inf_graph.pbTo run the resulting graph in C++, you can look at the label_image sample code:
bazel build tensorflow/examples/label_image:label_image
bazel-bin/tensorflow/examples/label_image/label_image \
--image=${HOME}/Pictures/flowers.jpg \
--input_layer=input \
--output_layer=InceptionV3/Predictions/Reshape_1 \
--graph=/tmp/frozen_inception_v3.pb \
--labels=/tmp/imagenet_slim_labels.txt \
--input_mean=0 \
--input_std=255 \
--logtostderrSee Model Runs out of CPU memory.
The ImageNet dataset provided has an empty background class which can be used to fine-tune the model to other tasks. If you try training or fine-tuning the VGG or ResNet models using the ImageNet dataset, you might encounter the following error:
InvalidArgumentError: Assign requires shapes of both tensors to match. lhs shape= [1001] rhs shape= [1000]This is due to the fact that the VGG and ResNet V1 final layers have only 1000 outputs rather than 1001.
To fix this issue, you can set the --labels_offset=1 flag. This results in
the ImageNet labels being shifted down by one:
The preprocessing functions all take height and width as parameters. You
can change the default values using the following snippet:
image_preprocessing_fn = preprocessing_factory.get_preprocessing(
preprocessing_name,
height=MY_NEW_HEIGHT,
width=MY_NEW_WIDTH,
is_training=True)