inception-v3-cifar10

Based on TensorFlow-Slim.

Install

Pull Docker image

$ docker pull tensorflow/tensorflow:1.7.0-py3
$ docker run -it --name {docker-name} tensorflow/tensorflow:1.7.0-py3 /bin/bash

Pull GitHub repository

~# apt-get update
~# apt-get install git
~# git clone https://github.com/PJunhyuk/inception-v3-cifar10

Download dataset

~/inception-v3-cifar10# python download_and_convert_data.py --dataset_name=cifar10 --dataset_dir="/tmp/data/cifar10"

Usage

Train

~/inception-v3-cifar10# python train_image_classifier.py --train_dir=/tmp/models/inception_v3 --dataset_name=cifar10 --dataset_split_name=train --dataset_dir=/tmp/data/cifar10 --model_name=inception_v3 --clone_on_cpu=True

Evaluate

~/inception-v3-cifar10# python eval_image_classifier.py --alsologtostderr -checkpoint_path=/tmp/models/inception_v3/model.ckpt-14457 --dataset_dir=/tmp/data/cifar10 --dataset_name=cifar10 --dataset_split_name=test --model_name=inception_v3 --clone_on_cpu=True

Download&Unzip pre-trained model

Official

/tmp/models# wget http://download.tensorflow.org/models/inception_v3_2016_08_28.tar.gz
/tmp/models# tar -xvzf inception_v3_2016_08_28.tar.gz

Self trained (26583 steps)

/tmp/models# wget https://www.dropbox.com/s/19ckps2vp0b0t1l/model.ckpt-26583.tar.gz?dl=0
/tmp/models# tar -xvzf model.ckpt-26583.tar.gz

Fine-tuning

> python train_image_classifier.py --train_dir=/tmp/models/pre_inception_v3 --dataset_dir=/tmp/data/cifar10 --dataset_name=cifar10 --dataset_split_name=train --model_name=inception_v3 --checkpoint_path=/tmp/models/inception_v3.ckpt --checkpoint_exclude_scopes=InceptionV3/Logits,InceptionV3/AuxLogits --trainable_scopes=InceptionV3/Logits,InceptionV3/AuxLogits --clone_on_cpu=True

TensorBoard

> tensorboard --logdir=/tmp/models/inception_v3

Copy

Docker to server

$ /docker cp {docker-name}:/tmp/models/inception_v3 .

Server to local

$ scp -r {user-name}@{server-ip}:{server-directory} .

Check weights in trained checkpoint file

Check check.py

all tensors

from tensorflow.python.tools import inspect_checkpoint
inspect_checkpoint.print_tensors_in_checkpoint_file('/tmp/models/inception_v3/model.ckpt-20', tensor_name='', all_tensors=True, all_tensor_names=True)

specific tensor

from tensorflow.python.tools import inspect_checkpoint
inspect_checkpoint.print_tensors_in_checkpoint_file('/tmp/models/inception_v3/model.ckpt-20', tensor_name='InceptionV3/Conv2d_2a_3x3/BatchNorm/beta/RMSProp_1', all_tensors=False)

Results

Total required time

only CPU: Intel(R) Xeon(R) CPU E5-2687W v3 @ 3.10GHz

Train: 6.5xx sec/step
Fine-tuning: 1.3xx sec/step

only CPU: Intel(R) Core(TM) i7-8700K CPU @ 3.70GHz

Train: 25.xxx sec/step
Fine-tuning: 6.xxx sec/step

only CPU: Intel(R) Core(TM) i5-6600 CPU @ 3.30GHz

Train: 41.xxx sec/step

Accuracy

inception_v3 - train (14457 steps)

• eval/Accuracy[0.6713]
• eval/Recall_5[0.9622]

inception_v3 - train (26583 steps)

• eval/Accuracy[0.7976]
• eval/Recall_5[0.9904]

Descriptions

Tensorflow-slim

A lightweight high-level API of TensorFlow for defining, training and evaluating complex models.
TensorFlow-Slim.

Image Resize

Resize image size from 32x32x3 -> 299x299x3.

Network structure

| type | patch size/stride | input size | |:-:|:-:|:-:|:-:| | conv | 3x3/2 | 299x299x3 | | conv | 3x3/1 | 149x149x32 | | conv padded | 3x3/1 | 147x147x32 | | pool | 3x3/2 | 147x147x64 | | conv | 3x3/1 | 73x73x64 | | conv | 3x3/2 | 71x71x80 | | conv | 3x3/1 | 35x35x192 | | 3xInception | - | 35x35x288 | | 5xInception | - | 17x17x768 | | 2xInception | - | 8x8x1280 | | pool | 8x8 | 8x8x2048 | | linear | logits| 1x1x2048 | | softmax | classifier | 1x1x1000 |

You can check more detail descriptions in paper - Rethinking the Inception Architecture for Computer Vision.

You can check codes in inception_v3.py.

• padding in slim.arg_scope can get VALID or SAME(default). difference between 'SAME' and 'VALID'.
• Each Inception blocks has several branches, and it combined for net by using tf.concat function. Depth of net(combined) is sum of all branches' depth.
  • ex) In Mixed_5b, 256 = 64 + 48 + 96 + 32.

LICENCE

Apache License 2.0