Dice

Learning Tensorflow from Google and combining with Unity, and any other front end using TCP.

dice_first.MOV

The idea here is to get Unity to recognize numbers on dice. So I made a large one. Then I retrained the inception model with 60 webcam frames of each side, moving about and rotating. A python server listens to a port waiting for an image and returns a classification. Unity then brings forward that game object.

Training image samples

The webcam snaps images every quarter second for 60 images. This is a low number of images for retraining.

Using regular size dice (540 images per class)

ClassifyWebcam.cs

using System.Net;using System.Net.Sockets;using System.IO;using System.Threading;using UnityEngine;using System;
public class ClassifyWebcam : MonoBehaviour { public GameObject goOne; public GameObject goTwo; public GameObject goThree; public GameObject goFour; public GameObject goFive; public GameObject goSix; public GameObject goOther; byte[] data = new byte[0]; IPEndPoint ipep; Socket server; Texture2D texTest; Thread clientReceiveThread; WebCamTexture webcamTexture; bool ynDone = true; int cntFrames; public int fps; Vector3 posHome; Vector3 posPicked; public string classification; GameObject[] gos; Vector3[] looks; Vector3[] targets; int numGos = 7; int classificationN = -1; float smooth = .5f; float speed = .1f; float tolerance = .2f; GameObject goCube; float score; GameObject goScore; Vector3 posGoScore; Vector3 scaGoScore; GameObject goScoreBase; float scoreHeight = 1; float threshold = .5f;
void Start() { if (Application.platform == RuntimePlatform.IPhonePlayer) { ipep = new IPEndPoint(IPAddress.Parse("10.0.0.189"), 60000); } else { ipep = new IPEndPoint(IPAddress.Parse("127.0.0.1"), 60000); } goScore = GameObject.CreatePrimitive(PrimitiveType.Cube); goScore.transform.position = new Vector3(-.72f, .35f + .25f, -8.5f); posGoScore = goScore.transform.position; goScore.transform.localScale = new Vector3(.25f, .25f, .125f); scaGoScore = goScore.transform.localScale; goScore.transform.eulerAngles = new Vector3(0, 0, 180); goScore.GetComponent<Renderer>().material.color = Color.green; // goScoreBase = GameObject.CreatePrimitive(PrimitiveType.Cube); goScoreBase.transform.localScale = new Vector3(scaGoScore.x, scoreHeight, scaGoScore.z); goScoreBase.transform.position = new Vector3(posGoScore.x, posGoScore.y + scoreHeight / 2, posGoScore.z + .125f/2); // gos = new GameObject[numGos]; looks = new Vector3[numGos]; targets = new Vector3[numGos]; gos[0] = goOne; gos[1] = goTwo; gos[2] = goThree; gos[3] = goFour; gos[4] = goFive; gos[5] = goSix; gos[6] = goOther; float z = -8.5f; posHome = new Vector3(0, 1, z+2); posPicked = new Vector3(0, 1, z); Application.runInBackground = true; if (Application.platform == RuntimePlatform.IPhonePlayer) { WebCamDevice[] devices = WebCamTexture.devices; for (int n = 0; n < devices.Length; n++) { if (devices[n].isFrontFacing == false) { webcamTexture = new WebCamTexture(devices[n].name, 640, 480); break; } } } if (webcamTexture == null) { webcamTexture = new WebCamTexture(); } webcamTexture.Play(); goCube = GameObject.CreatePrimitive(PrimitiveType.Cube); goCube.transform.position = new Vector3(-.72f, .35f, -8.5f); goCube.transform.localScale = new Vector3(.25f, .25f, .25f); goCube.transform.eulerAngles = new Vector3(0, 0, 180); goCube.GetComponent<Renderer>().material.mainTexture = webcamTexture; InvokeRepeating("ShowFps", 1, 1); //InvokeRepeating("RandomClass", 2, 2); }
void ShowFps() { fps = cntFrames; cntFrames = 0; }
void RandomClass() { int n = UnityEngine.Random.Range(0, 6); if (n == 0) { classification = "one"; } if (n == 1) { classification = "two"; } if (n == 2) { classification = "three"; } if (n == 3) { classification = "four"; } if (n == 4) { classification = "five"; } if (n == 5) { classification = "six"; } if (n == 6) { classification = "other"; } Debug.Log(n + ": " + classification + "\n"); classificationN = n; UpdateTargets(); }
void UpdateTargets() { for (int n = 0; n < numGos; n++) { if (n != classificationN) { targets[n] = posHome; } else { targets[n] = posPicked; } } }
void UpdateGos() { for (int n = 0; n < numGos; n++) { float dist = Vector3.Distance(gos[n].transform.position, targets[n]); if (dist > tolerance) { Vector3 pos = smooth * looks[n] + (1 - smooth) * targets[n]; gos[n].transform.LookAt(pos); gos[n].transform.position += gos[n].transform.forward * speed; looks[n] = pos; } } }
void Update() { if (ynDone == true) { if (server != null) { server.Close(); PickDice(); UpdateScore(); UpdateTargets(); } ynDone = false; ProcessImage(); cntFrames++; } UpdateGos(); }
void UpdateScore() { float sy = score * scoreHeight; goScore.transform.localScale = new Vector3(scaGoScore.x, sy, scaGoScore.z); goScore.transform.position = new Vector3(posGoScore.x, posGoScore.y + sy/2, posGoScore.z); if (score >= threshold) { goScore.GetComponent<Renderer>().material.color = Color.green; } else { goScore.GetComponent<Renderer>().material.color = Color.red; } }
void ProcessImage() { GrabImage(); ConnectToServer(); int sent = SendVarData(texTest.EncodeToJPG()); }
void PickDice() { //if (score >= threshold) //{ if (classification == "one") { classificationN = 0; } if (classification == "two") { classificationN = 1; } if (classification == "three") { classificationN = 2; } if (classification == "four") { classificationN = 3; } if (classification == "five") { classificationN = 4; } if (classification == "six") { classificationN = 5; } if (classification == "other") { classificationN = 6; } //} }
void ConnectToServer() { server = new Socket(AddressFamily.InterNetwork, SocketType.Stream, ProtocolType.Tcp); try { server.Connect(ipep); } catch (SocketException e) { Debug.Log(e.ToString()); } StartListening(); }
void StartListening() { try { clientReceiveThread = new Thread (new ThreadStart(Listen)); clientReceiveThread.IsBackground = true; clientReceiveThread.Start(); } catch (Exception e) { Debug.Log("On client connect exception " + e); } }
void GrabImage() { Texture2D snap = new Texture2D(webcamTexture.width, webcamTexture.height); snap.SetPixels(webcamTexture.GetPixels()); snap.Apply(); texTest = snap; }
private int SendVarData(byte[] data) { int total = 0; int size = data.Length; int dataleft = size; int sent;
byte[] datasize = new byte[0]; datasize = BitConverter.GetBytes(size); sent = server.Send(datasize);
while (total < size) { sent = server.Send(data, total, dataleft, SocketFlags.None); total += sent; dataleft -= sent; } return total; }
private void Listen() { byte[] dataIn = new byte[1024]; int dataReceivedLen = server.Receive(dataIn); string txtResponse = System.Text.Encoding.ASCII.GetString(dataIn, 0, dataReceivedLen); string[] stuff = txtResponse.Split('|'); string[] stuffClass = stuff[0].Trim().Split(':'); classification = stuffClass[0].Trim(); score = float.Parse(stuffClass[1].Trim()); Debug.Log(stuff[0] + "\n"); ynDone = true; }
private void OnApplicationQuit() { if (server != null) { server.Close(); } }}

DataAugment.cs

using System.Collections;using System.Collections.Generic;using UnityEngine;using System.IO;
public class DataAugmentScale : MonoBehaviour { public Camera cam; Texture2D texGrab; public int numImages = 60; public int startWith = 0; public int cnt = 0; RenderTexture renderTex; GameObject goCube; AudioSource audioSource; public ClassificationType classificationType; public Texture2D image; bool ynDone; string path; string pathDice; public bool ynGrayScale; public string prefix = "small_";
public enum ClassificationType { one, two, three, four, five, six, other };
// Use this for initialization void Start () { path = "/Users/amreamer/Documents/dice/dice/Retrain/augment/" + classificationType + "/"; pathDice = "/Users/amreamer/Documents/dice/dice/Retrain/Dice/" + classificationType + "/"; string filename = prefix + classificationType.ToString(); image = Resources.Load<Texture2D>(filename); renderTex = new RenderTexture(Screen.width, Screen.height, 24); cam.targetTexture = renderTex; // Application.runInBackground = true; // goCube = GameObject.CreatePrimitive(PrimitiveType.Cube); goCube.transform.position = new Vector3(0, 1, -8.5f); goCube.transform.eulerAngles = new Vector3(0, 0, 180); goCube.GetComponent<Renderer>().material.mainTexture = image; // audioSource = GetComponent<AudioSource>(); // InvokeRepeating("AugmentGrabAndSaveImage", 3, .25f); Invoke("PlaySound", 2.5f); startWith = FindHighestFileNumber() + 1; cnt = startWith; Debug.Log("path:" + path + "\n"); Debug.Log("startWith:" + startWith + "\n"); }
void PlaySound() { audioSource.Play(); } void AugmentData() { //float roll = Random.Range(0, 360f); //goCube.transform.eulerAngles = new Vector3(0, 0, roll); float sca = Random.Range(.25f, 1); goCube.transform.localScale = new Vector3(sca, sca, sca); float ang = 10; float pitch = Random.Range(-ang, ang); float yaw = Random.Range(-ang, ang); float roll = Random.Range(-90, 90f); goCube.transform.eulerAngles = new Vector3(pitch, yaw, roll); } void AugmentGrabAndSaveImage() { if (ynDone == false) { AugmentData(); GrabImage(); SaveTextureToFile(); } }
void GrabImage() { DestroyImmediate(texGrab); RenderTexture.active = renderTex; texGrab = new Texture2D(renderTex.width, renderTex.height); texGrab.ReadPixels(new Rect(0, 0, renderTex.width, renderTex.height), 0, 0); texGrab.Apply(); }
int FindHighestFileNumber () { int maxN = 0; foreach (string file in System.IO.Directory.GetFiles(pathDice)) { if (!file.Contains("_Store")) { string[] stuff = file.Split('_'); string[] fileStuff = stuff[stuff.Length - 1].Split('.'); string answer = fileStuff[0]; int n = int.Parse(answer); if (n > maxN) { maxN = n; } } } return maxN; }
void SaveTextureToFile() { int limit = startWith + numImages; if (cnt <= limit) { if (cnt < limit) { if (ynGrayScale == true) { MakeTexGrabGrayScale(); } string file = path + "picture_" + cnt + ".jpg"; System.IO.File.WriteAllBytes(file, texGrab.EncodeToJPG()); Debug.Log(file + "\n"); } else { if (cnt == limit) { PlaySound(); } } cnt++; } else { ynDone = true; } }
void MakeTexGrabGrayScale() { for (int x = 0; x < texGrab.width; x++) { for (int y = 0; y < texGrab.height; y++) { Color c = texGrab.GetPixel(x, y); float sum = c.r + c.g + c.b; float g = sum / 3; Color colorGray = new Color(g, g, g); texGrab.SetPixel(x, y, colorGray); } } }
}

Retrain.cs

using System.Collections;using System.Collections.Generic;using UnityEngine;using System.IO;
public class Retrain : MonoBehaviour { Texture2D texGrab; public int numImages = 60; public string className = "near"; public int startWith = 0; public int cnt = 0; RenderTexture renderTex; WebCamTexture webcamTexture; GameObject goCube;
// Use this for initialization void Start () { cnt = startWith; Application.runInBackground = true; // if (webcamTexture == null) { webcamTexture = new WebCamTexture(); } webcamTexture.Play(); // goCube = GameObject.CreatePrimitive(PrimitiveType.Cube); goCube.transform.position = new Vector3(0, 1, -8.5f); goCube.transform.eulerAngles = new Vector3(0, 0, 180); goCube.GetComponent<Renderer>().material.mainTexture = webcamTexture; // InvokeRepeating("GrabAndSaveImage", 3, .25f); } // Update is called once per frame void Update () { }
void GrabAndSaveImage() { GrabImage(); SaveTextureToFile(); }
void GrabImage() { DestroyImmediate(texGrab); texGrab = new Texture2D(webcamTexture.width, webcamTexture.height); texGrab.SetPixels(webcamTexture.GetPixels()); texGrab.Apply(); }
void SaveTextureToFile() { if (cnt < (startWith + numImages)) { string path = "/Users/amreamer/Documents/dice/dice/Retrain/dice/" + className + "/"; System.IO.File.WriteAllBytes(path + "picture_" + cnt + ".jpg", texGrab.EncodeToJPG()); cnt++; } }
}

client_send_image_tcp.py

# client.py
import sysimport timeimport socket # Import socket module
s = socket.socket() # Create a socket objecthost = socket.gethostname() # Get local machine nameport = 60000 # Reserve a port for your service.
s.connect((host, port))#s.setsockopt(socket.SOL_SOCKET, socket.SO_SNDBUF, 1024)print(sys.argv[0])f = open(sys.argv[1], 'rb')data = f.read()length = len(data)s.send(length.to_bytes(4, 'little'))
f.seek(0,0)while True: chunk = f.read(1024) if not chunk: break s.send(chunk)
print('sent data: %', length)data = s.recv(1024)s.close()print('data received: ' + data.decode("utf-8"))exit()

classify_server.py

import socketimport tensorflow as tf
# to shutdown server# lsof -i :60000# kill -9 PID
label_lines = [line.rstrip() for line in tf.gfile.GFile("./output_labels.txt")]with tf.gfile.FastGFile("./output_graph.pb", 'rb') as f: graph_def = tf.GraphDef() graph_def.ParseFromString(f.read()) _ = tf.import_graph_def(graph_def, name='')with tf.Session() as sess: softmax_tensor = sess.graph.get_tensor_by_name('final_result:0') HOST = '' PORT = 60000 ADDR = (HOST,PORT) BUFSIZE = 4096 serv = socket.socket(socket.AF_INET, socket.SOCK_STREAM) serv.bind(ADDR) serv.listen(5) print('listening ...') while True: conn, addr = serv.accept() length_data = conn.recv(4) length = int.from_bytes(length_data, byteorder='little', signed=False) total = 0 while True: data = conn.recv(BUFSIZE) if not data: break if total == 0: data_all = data else: data_all = data_all + data total = total + len(data) if total >= length: break data = data_all predictions = sess.run(softmax_tensor, {'DecodeJpeg/contents:0': data}) top_k = predictions[0].argsort()[-len(predictions[0]):][::-1] response = '' for node_id in top_k: human_string = label_lines[node_id] score = predictions[0][node_id] response = response + ' %s : %.5f |' % (human_string, score) #response = "a : 0 | b : 1 | c : 2" #response = 'length: ' + str(length) conn.sendall(response.encode('utf-8')) conn.close() stuff = response.split('|') print(stuff[0])

output_labels.txt

threeonetwosixfivefour

folders

Comparing 60 images per classification to 120 images per classification

Data Augmentation

Rotating content by random angles

Now with 120 images per side x 6 = 720 images

Other Classification

Adding a "other" category of 120 background helps reduce false positives. Also confidence bar is red when confidence is under 50%

log_first.txt

retrain.py (downloaded and modified)

# Copyright 2015 The TensorFlow Authors. All Rights Reserved.## Licensed under the Apache License, Version 2.0 (the "License");# you may not use this file except in compliance with the License.# You may obtain a copy of the License at## http://www.apache.org/licenses/LICENSE-2.0## Unless required by applicable law or agreed to in writing, software# distributed under the License is distributed on an "AS IS" BASIS,# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.# See the License for the specific language governing permissions and# limitations under the License.# =============================================================================="""Simple transfer learning with an Inception v3 architecture model whichdisplays summaries in TensorBoard.
This example shows how to take a Inception v3 architecture model trained onImageNet images, and train a new top layer that can recognize other classes ofimages.
The top layer receives as input a 2048-dimensional vector for each image. Wetrain a softmax layer on top of this representation. Assuming the softmax layercontains N labels, this corresponds to learning N + 2048*N model parameterscorresponding to the learned biases and weights.
Here's an example, which assumes you have a folder containing class-namedsubfolders, each full of images for each label. The example folder flower_photosshould have a structure like this:
~/flower_photos/daisy/photo1.jpg~/flower_photos/daisy/photo2.jpg...~/flower_photos/rose/anotherphoto77.jpg...~/flower_photos/sunflower/somepicture.jpg
The subfolder names are important, since they define what label is applied toeach image, but the filenames themselves don't matter. Once your images areprepared, you can run the training with a command like this:
bazel build third_party/tensorflow/examples/image_retraining:retrain && \bazel-bin/third_party/tensorflow/examples/image_retraining/retrain \--image_dir ~/flower_photos
You can replace the image_dir argument with any folder containing subfolders ofimages. The label for each image is taken from the name of the subfolder it'sin.
This produces a new model file that can be loaded and run by any TensorFlowprogram, for example the label_image sample code.

To use with TensorBoard:
By default, this script will log summaries to /tmp/retrain_logs directory
Visualize the summaries with this command:
tensorboard --logdir /tmp/retrain_logs
"""from __future__ import absolute_importfrom __future__ import divisionfrom __future__ import print_function
import argparsefrom datetime import datetimeimport hashlibimport os.pathimport randomimport reimport structimport sysimport tarfile
import numpy as npfrom six.moves import urllibimport tensorflow as tf
from tensorflow.python.framework import graph_utilfrom tensorflow.python.framework import tensor_shapefrom tensorflow.python.platform import gfilefrom tensorflow.python.util import compat
FLAGS = Nonecurrent_dir_path = os.path.dirname(os.path.realpath(__file__))# pylint: enable=line-too-longBOTTLENECK_TENSOR_NAME = 'pool_3/_reshape:0'BOTTLENECK_TENSOR_SIZE = 2048MODEL_INPUT_WIDTH = 299MODEL_INPUT_HEIGHT = 299MODEL_INPUT_DEPTH = 3JPEG_DATA_TENSOR_NAME = 'DecodeJpeg/contents:0'RESIZED_INPUT_TENSOR_NAME = 'ResizeBilinear:0'MAX_NUM_IMAGES_PER_CLASS = 2 ** 27 - 1 # ~134M

def create_image_lists(image_dir, testing_percentage, validation_percentage): """Builds a list of training images from the file system.
Analyzes the sub folders in the image directory, splits them into stable training, testing, and validation sets, and returns a data structure describing the lists of images for each label and their paths.
Args: image_dir: String path to a folder containing subfolders of images. testing_percentage: Integer percentage of the images to reserve for tests. validation_percentage: Integer percentage of images reserved for validation.
Returns: A dictionary containing an entry for each label subfolder, with images split into training, testing, and validation sets within each label. """ if not gfile.Exists(image_dir): print("Image directory '" + image_dir + "' not found.") return None result = {} sub_dirs = [x[0] for x in gfile.Walk(image_dir)] # The root directory comes first, so skip it. is_root_dir = True for sub_dir in sub_dirs: if is_root_dir: is_root_dir = False continue extensions = ['jpg', 'jpeg', 'JPG', 'JPEG'] file_list = [] dir_name = os.path.basename(sub_dir) if dir_name == image_dir: continue print("Looking for images in '" + dir_name + "'") for extension in extensions: file_glob = os.path.join(image_dir, dir_name, '*.' + extension) file_list.extend(gfile.Glob(file_glob)) if not file_list: print('No files found') continue if len(file_list) < 20: print('WARNING: Folder has less than 20 images, which may cause issues.') elif len(file_list) > MAX_NUM_IMAGES_PER_CLASS: print('WARNING: Folder {} has more than {} images. Some images will ' 'never be selected.'.format(dir_name, MAX_NUM_IMAGES_PER_CLASS)) label_name = re.sub(r'[^a-z0-9]+', ' ', dir_name.lower()) training_images = [] testing_images = [] validation_images = [] for file_name in file_list: base_name = os.path.basename(file_name) # We want to ignore anything after '_nohash_' in the file name when # deciding which set to put an image in, the data set creator has a way of # grouping photos that are close variations of each other. For example # this is used in the plant disease data set to group multiple pictures of # the same leaf. hash_name = re.sub(r'_nohash_.*$', '', file_name) # This looks a bit magical, but we need to decide whether this file should # go into the training, testing, or validation sets, and we want to keep # existing files in the same set even if more files are subsequently # added. # To do that, we need a stable way of deciding based on just the file name # itself, so we do a hash of that and then use that to generate a # probability value that we use to assign it. hash_name_hashed = hashlib.sha1(compat.as_bytes(hash_name)).hexdigest() percentage_hash = ((int(hash_name_hashed, 16) % (MAX_NUM_IMAGES_PER_CLASS + 1)) * (100.0 / MAX_NUM_IMAGES_PER_CLASS)) if percentage_hash < validation_percentage: validation_images.append(base_name) elif percentage_hash < (testing_percentage + validation_percentage): testing_images.append(base_name) else: training_images.append(base_name) result[label_name] = { 'dir': dir_name, 'training': training_images, 'testing': testing_images, 'validation': validation_images, } return result

def get_image_path(image_lists, label_name, index, image_dir, category): """"Returns a path to an image for a label at the given index.
Args: image_lists: Dictionary of training images for each label. label_name: Label string we want to get an image for. index: Int offset of the image we want. This will be moduloed by the available number of images for the label, so it can be arbitrarily large. image_dir: Root folder string of the subfolders containing the training images. category: Name string of set to pull images from - training, testing, or validation.
Returns: File system path string to an image that meets the requested parameters.
""" if label_name not in image_lists: tf.logging.fatal('Label does not exist %s.', label_name) label_lists = image_lists[label_name] if category not in label_lists: tf.logging.fatal('Category does not exist %s.', category) category_list = label_lists[category] if not category_list: tf.logging.fatal('Label %s has no images in the category %s.', label_name, category) mod_index = index % len(category_list) base_name = category_list[mod_index] sub_dir = label_lists['dir'] full_path = os.path.join(image_dir, sub_dir, base_name) return full_path

def get_bottleneck_path(image_lists, label_name, index, bottleneck_dir, category): """"Returns a path to a bottleneck file for a label at the given index.
Args: image_lists: Dictionary of training images for each label. label_name: Label string we want to get an image for. index: Integer offset of the image we want. This will be moduloed by the available number of images for the label, so it can be arbitrarily large. bottleneck_dir: Folder string holding cached files of bottleneck values. category: Name string of set to pull images from - training, testing, or validation.
Returns: File system path string to an image that meets the requested parameters. """ return get_image_path(image_lists, label_name, index, bottleneck_dir, category) + '.txt'

def create_inception_graph(): """"Creates a graph from saved GraphDef file and returns a Graph object.
Returns: Graph holding the trained Inception network, and various tensors we'll be manipulating. """ with tf.Session() as sess: model_filename = os.path.join( FLAGS.model_dir, 'classify_image_graph_def.pb') with gfile.FastGFile(model_filename, 'rb') as f: graph_def = tf.GraphDef() graph_def.ParseFromString(f.read()) bottleneck_tensor, jpeg_data_tensor, resized_input_tensor = ( tf.import_graph_def(graph_def, name='', return_elements=[ BOTTLENECK_TENSOR_NAME, JPEG_DATA_TENSOR_NAME, RESIZED_INPUT_TENSOR_NAME])) return sess.graph, bottleneck_tensor, jpeg_data_tensor, resized_input_tensor

def run_bottleneck_on_image(sess, image_data, image_data_tensor, bottleneck_tensor): """Runs inference on an image to extract the 'bottleneck' summary layer.
Args: sess: Current active TensorFlow Session. image_data: String of raw JPEG data. image_data_tensor: Input data layer in the graph. bottleneck_tensor: Layer before the final softmax.
Returns: Numpy array of bottleneck values. """ bottleneck_values = sess.run( bottleneck_tensor, {image_data_tensor: image_data}) bottleneck_values = np.squeeze(bottleneck_values) return bottleneck_values
def ensure_dir_exists(dir_name): """Makes sure the folder exists on disk.
Args: dir_name: Path string to the folder we want to create. """ if not os.path.exists(dir_name): os.makedirs(dir_name)

def write_list_of_floats_to_file(list_of_floats , file_path): """Writes a given list of floats to a binary file.
Args: list_of_floats: List of floats we want to write to a file. file_path: Path to a file where list of floats will be stored.
"""
s = struct.pack('d' * BOTTLENECK_TENSOR_SIZE, *list_of_floats) with open(file_path, 'wb') as f: f.write(s)

def read_list_of_floats_from_file(file_path): """Reads list of floats from a given file.
Args: file_path: Path to a file where list of floats was stored. Returns: Array of bottleneck values (list of floats).
"""
with open(file_path, 'rb') as f: s = struct.unpack('d' * BOTTLENECK_TENSOR_SIZE, f.read()) return list(s)

bottleneck_path_2_bottleneck_values = {}

def get_or_create_bottleneck(sess, image_lists, label_name, index, image_dir, category, bottleneck_dir, jpeg_data_tensor, bottleneck_tensor): """Retrieves or calculates bottleneck values for an image.
If a cached version of the bottleneck data exists on-disk, return that, otherwise calculate the data and save it to disk for future use.
Args: sess: The current active TensorFlow Session. image_lists: Dictionary of training images for each label. label_name: Label string we want to get an image for. index: Integer offset of the image we want. This will be modulo-ed by the available number of images for the label, so it can be arbitrarily large. image_dir: Root folder string of the subfolders containing the training images. category: Name string of which set to pull images from - training, testing, or validation. bottleneck_dir: Folder string holding cached files of bottleneck values. jpeg_data_tensor: The tensor to feed loaded jpeg data into. bottleneck_tensor: The output tensor for the bottleneck values.
Returns: Numpy array of values produced by the bottleneck layer for the image. """ label_lists = image_lists[label_name] sub_dir = label_lists['dir'] sub_dir_path = os.path.join(bottleneck_dir, sub_dir) ensure_dir_exists(sub_dir_path) bottleneck_path = get_bottleneck_path(image_lists, label_name, index, bottleneck_dir, category) if not os.path.exists(bottleneck_path): print('Creating bottleneck at ' + bottleneck_path) image_path = get_image_path(image_lists, label_name, index, image_dir, category) if not gfile.Exists(image_path): tf.logging.fatal('File does not exist %s', image_path) image_data = gfile.FastGFile(image_path, 'rb').read() bottleneck_values = run_bottleneck_on_image(sess, image_data, jpeg_data_tensor, bottleneck_tensor) bottleneck_string = ','.join(str(x) for x in bottleneck_values) with open(bottleneck_path, 'w') as bottleneck_file: bottleneck_file.write(bottleneck_string)
with open(bottleneck_path, 'r') as bottleneck_file: bottleneck_string = bottleneck_file.read() bottleneck_values = [float(x) for x in bottleneck_string.split(',')] return bottleneck_values

def cache_bottlenecks(sess, image_lists, image_dir, bottleneck_dir, jpeg_data_tensor, bottleneck_tensor): """Ensures all the training, testing, and validation bottlenecks are cached.
Because we're likely to read the same image multiple times (if there are no distortions applied during training) it can speed things up a lot if we calculate the bottleneck layer values once for each image during preprocessing, and then just read those cached values repeatedly during training. Here we go through all the images we've found, calculate those values, and save them off.
Args: sess: The current active TensorFlow Session. image_lists: Dictionary of training images for each label. image_dir: Root folder string of the subfolders containing the training images. bottleneck_dir: Folder string holding cached files of bottleneck values. jpeg_data_tensor: Input tensor for jpeg data from file. bottleneck_tensor: The penultimate output layer of the graph.
Returns: Nothing. """ how_many_bottlenecks = 0 ensure_dir_exists(bottleneck_dir) for label_name, label_lists in image_lists.items(): for category in ['training', 'testing', 'validation']: category_list = label_lists[category] for index, unused_base_name in enumerate(category_list): get_or_create_bottleneck(sess, image_lists, label_name, index, image_dir, category, bottleneck_dir, jpeg_data_tensor, bottleneck_tensor) how_many_bottlenecks += 1 if how_many_bottlenecks % 100 == 0: print(str(how_many_bottlenecks) + ' bottleneck files created.')

def get_random_cached_bottlenecks(sess, image_lists, how_many, category, bottleneck_dir, image_dir, jpeg_data_tensor, bottleneck_tensor): """Retrieves bottleneck values for cached images.
If no distortions are being applied, this function can retrieve the cached bottleneck values directly from disk for images. It picks a random set of images from the specified category.
Args: sess: Current TensorFlow Session. image_lists: Dictionary of training images for each label. how_many: The number of bottleneck values to return. category: Name string of which set to pull from - training, testing, or validation. bottleneck_dir: Folder string holding cached files of bottleneck values. image_dir: Root folder string of the subfolders containing the training images. jpeg_data_tensor: The layer to feed jpeg image data into. bottleneck_tensor: The bottleneck output layer of the CNN graph.
Returns: List of bottleneck arrays and their corresponding ground truths. """ class_count = len(image_lists.keys()) bottlenecks = [] ground_truths = [] for unused_i in range(how_many): label_index = random.randrange(class_count) label_name = list(image_lists.keys())[label_index] image_index = random.randrange(MAX_NUM_IMAGES_PER_CLASS + 1) bottleneck = get_or_create_bottleneck(sess, image_lists, label_name, image_index, image_dir, category, bottleneck_dir, jpeg_data_tensor, bottleneck_tensor) ground_truth = np.zeros(class_count, dtype=np.float32) ground_truth[label_index] = 1.0 bottlenecks.append(bottleneck) ground_truths.append(ground_truth) return bottlenecks, ground_truths

def get_random_distorted_bottlenecks( sess, image_lists, how_many, category, image_dir, input_jpeg_tensor, distorted_image, resized_input_tensor, bottleneck_tensor): """Retrieves bottleneck values for training images, after distortions.
If we're training with distortions like crops, scales, or flips, we have to recalculate the full model for every image, and so we can't use cached bottleneck values. Instead we find random images for the requested category, run them through the distortion graph, and then the full graph to get the bottleneck results for each.
Args: sess: Current TensorFlow Session. image_lists: Dictionary of training images for each label. how_many: The integer number of bottleneck values to return. category: Name string of which set of images to fetch - training, testing, or validation. image_dir: Root folder string of the subfolders containing the training images. input_jpeg_tensor: The input layer we feed the image data to. distorted_image: The output node of the distortion graph. resized_input_tensor: The input node of the recognition graph. bottleneck_tensor: The bottleneck output layer of the CNN graph.
Returns: List of bottleneck arrays and their corresponding ground truths. """ class_count = len(image_lists.keys()) bottlenecks = [] ground_truths = [] for unused_i in range(how_many): label_index = random.randrange(class_count) label_name = list(image_lists.keys())[label_index] image_index = random.randrange(MAX_NUM_IMAGES_PER_CLASS + 1) image_path = get_image_path(image_lists, label_name, image_index, image_dir, category) if not gfile.Exists(image_path): tf.logging.fatal('File does not exist %s', image_path) jpeg_data = gfile.FastGFile(image_path, 'rb').read() # Note that we materialize the distorted_image_data as a numpy array before # sending running inference on the image. This involves 2 memory copies and # might be optimized in other implementations. distorted_image_data = sess.run(distorted_image, {input_jpeg_tensor: jpeg_data}) bottleneck = run_bottleneck_on_image(sess, distorted_image_data, resized_input_tensor, bottleneck_tensor) ground_truth = np.zeros(class_count, dtype=np.float32) ground_truth[label_index] = 1.0 bottlenecks.append(bottleneck) ground_truths.append(ground_truth) return bottlenecks, ground_truths

def should_distort_images(flip_left_right, random_crop, random_scale, random_brightness): """Whether any distortions are enabled, from the input flags.
Args: flip_left_right: Boolean whether to randomly mirror images horizontally. random_crop: Integer percentage setting the total margin used around the crop box. random_scale: Integer percentage of how much to vary the scale by. random_brightness: Integer range to randomly multiply the pixel values by.
Returns: Boolean value indicating whether any distortions should be applied. """ return (flip_left_right or (random_crop != 0) or (random_scale != 0) or (random_brightness != 0))

def add_input_distortions(flip_left_right, random_crop, random_scale, random_brightness): """Creates the operations to apply the specified distortions.
During training it can help to improve the results if we run the images through simple distortions like crops, scales, and flips. These reflect the kind of variations we expect in the real world, and so can help train the model to cope with natural data more effectively. Here we take the supplied parameters and construct a network of operations to apply them to an image.
Cropping ~~~~~~~~
Cropping is done by placing a bounding box at a random position in the full image. The cropping parameter controls the size of that box relative to the input image. If it's zero, then the box is the same size as the input and no cropping is performed. If the value is 50%, then the crop box will be half the width and height of the input. In a diagram it looks like this:
< width > +---------------------+ | | | width - crop% | | < > | | +------+ | | | | | | | | | | | | | | +------+ | | | | | +---------------------+
Scaling ~~~~~~~
Scaling is a lot like cropping, except that the bounding box is always centered and its size varies randomly within the given range. For example if the scale percentage is zero, then the bounding box is the same size as the input and no scaling is applied. If it's 50%, then the bounding box will be in a random range between half the width and height and full size.
Args: flip_left_right: Boolean whether to randomly mirror images horizontally. random_crop: Integer percentage setting the total margin used around the crop box. random_scale: Integer percentage of how much to vary the scale by. random_brightness: Integer range to randomly multiply the pixel values by. graph.
Returns: The jpeg input layer and the distorted result tensor. """
jpeg_data = tf.placeholder(tf.string, name='DistortJPGInput') decoded_image = tf.image.decode_jpeg(jpeg_data, channels=MODEL_INPUT_DEPTH) decoded_image_as_float = tf.cast(decoded_image, dtype=tf.float32) decoded_image_4d = tf.expand_dims(decoded_image_as_float, 0) margin_scale = 1.0 + (random_crop / 100.0) resize_scale = 1.0 + (random_scale / 100.0) margin_scale_value = tf.constant(margin_scale) resize_scale_value = tf.random_uniform(tensor_shape.scalar(), minval=1.0, maxval=resize_scale) scale_value = tf.mul(margin_scale_value, resize_scale_value) precrop_width = tf.mul(scale_value, MODEL_INPUT_WIDTH) precrop_height = tf.mul(scale_value, MODEL_INPUT_HEIGHT) precrop_shape = tf.stack([precrop_height, precrop_width]) precrop_shape_as_int = tf.cast(precrop_shape, dtype=tf.int32) precropped_image = tf.image.resize_bilinear(decoded_image_4d, precrop_shape_as_int) precropped_image_3d = tf.squeeze(precropped_image, squeeze_dims=[0]) cropped_image = tf.random_crop(precropped_image_3d, [MODEL_INPUT_HEIGHT, MODEL_INPUT_WIDTH, MODEL_INPUT_DEPTH]) if flip_left_right: flipped_image = tf.image.random_flip_left_right(cropped_image) else: flipped_image = cropped_image brightness_min = 1.0 - (random_brightness / 100.0) brightness_max = 1.0 + (random_brightness / 100.0) brightness_value = tf.random_uniform(tensor_shape.scalar(), minval=brightness_min, maxval=brightness_max) brightened_image = tf.mul(flipped_image, brightness_value) distort_result = tf.expand_dims(brightened_image, 0, name='DistortResult') return jpeg_data, distort_result

def variable_summaries(var): """Attach a lot of summaries to a Tensor (for TensorBoard visualization).""" with tf.name_scope('summaries'): mean = tf.reduce_mean(var) tf.summary.scalar('mean', mean) with tf.name_scope('stddev'): stddev = tf.sqrt(tf.reduce_mean(tf.square(var - mean))) tf.summary.scalar('stddev', stddev) tf.summary.scalar('max', tf.reduce_max(var)) tf.summary.scalar('min', tf.reduce_min(var)) tf.summary.histogram('histogram', var)

def add_final_training_ops(class_count, final_tensor_name, bottleneck_tensor): """Adds a new softmax and fully-connected layer for training.
We need to retrain the top layer to identify our new classes, so this function adds the right operations to the graph, along with some variables to hold the weights, and then sets up all the gradients for the backward pass.
The set up for the softmax and fully-connected layers is based on: https://tensorflow.org/versions/master/tutorials/mnist/beginners/index.html
Args: class_count: Integer of how many categories of things we're trying to recognize. final_tensor_name: Name string for the new final node that produces results. bottleneck_tensor: The output of the main CNN graph.
Returns: The tensors for the training and cross entropy results, and tensors for the bottleneck input and ground truth input. """ with tf.name_scope('input'): bottleneck_input = tf.placeholder_with_default( bottleneck_tensor, shape=[None, BOTTLENECK_TENSOR_SIZE], name='BottleneckInputPlaceholder')
ground_truth_input = tf.placeholder(tf.float32, [None, class_count], name='GroundTruthInput')
# Organizing the following ops as `final_training_ops` so they're easier # to see in TensorBoard layer_name = 'final_training_ops' with tf.name_scope(layer_name): with tf.name_scope('weights'): layer_weights = tf.Variable(tf.truncated_normal([BOTTLENECK_TENSOR_SIZE, class_count], stddev=0.001), name='final_weights') variable_summaries(layer_weights) with tf.name_scope('biases'): layer_biases = tf.Variable(tf.zeros([class_count]), name='final_biases') variable_summaries(layer_biases) with tf.name_scope('Wx_plus_b'): logits = tf.matmul(bottleneck_input, layer_weights) + layer_biases tf.summary.histogram('pre_activations', logits)
final_tensor = tf.nn.softmax(logits, name=final_tensor_name) tf.summary.histogram('activations', final_tensor)
with tf.name_scope('cross_entropy'): cross_entropy = tf.nn.softmax_cross_entropy_with_logits( logits=logits, labels=ground_truth_input) with tf.name_scope('total'): cross_entropy_mean = tf.reduce_mean(cross_entropy) tf.summary.scalar('cross_entropy', cross_entropy_mean)
with tf.name_scope('train'): train_step = tf.train.GradientDescentOptimizer(FLAGS.learning_rate).minimize( cross_entropy_mean)
return (train_step, cross_entropy_mean, bottleneck_input, ground_truth_input, final_tensor)

def add_evaluation_step(result_tensor, ground_truth_tensor): """Inserts the operations we need to evaluate the accuracy of our results.
Args: result_tensor: The new final node that produces results. ground_truth_tensor: The node we feed ground truth data into.
Returns: Nothing. """ with tf.name_scope('accuracy'): with tf.name_scope('correct_prediction'): correct_prediction = tf.equal(tf.argmax(result_tensor, 1), \ tf.argmax(ground_truth_tensor, 1)) with tf.name_scope('accuracy'): evaluation_step = tf.reduce_mean(tf.cast(correct_prediction, tf.float32)) tf.summary.scalar('accuracy', evaluation_step) return evaluation_step

def main(_): # Setup the directory we'll write summaries to for TensorBoard if tf.gfile.Exists(FLAGS.summaries_dir): tf.gfile.DeleteRecursively(FLAGS.summaries_dir) tf.gfile.MakeDirs(FLAGS.summaries_dir)
graph, bottleneck_tensor, jpeg_data_tensor, resized_image_tensor = ( create_inception_graph())
# Look at the folder structure, and create lists of all the images. image_lists = create_image_lists(FLAGS.image_dir, FLAGS.testing_percentage, FLAGS.validation_percentage) class_count = len(image_lists.keys()) if class_count == 0: print('No valid folders of images found at ' + FLAGS.image_dir) return -1 if class_count == 1: print('Only one valid folder of images found at ' + FLAGS.image_dir + ' - multiple classes are needed for classification.') return -1
# See if the command-line flags mean we're applying any distortions. do_distort_images = should_distort_images( FLAGS.flip_left_right, FLAGS.random_crop, FLAGS.random_scale, FLAGS.random_brightness) sess = tf.Session()
if do_distort_images: # We will be applying distortions, so setup the operations we'll need. distorted_jpeg_data_tensor, distorted_image_tensor = add_input_distortions( FLAGS.flip_left_right, FLAGS.random_crop, FLAGS.random_scale, FLAGS.random_brightness) else: # We'll make sure we've calculated the 'bottleneck' image summaries and # cached them on disk. cache_bottlenecks(sess, image_lists, FLAGS.image_dir, FLAGS.bottleneck_dir, jpeg_data_tensor, bottleneck_tensor)
# Add the new layer that we'll be training. (train_step, cross_entropy, bottleneck_input, ground_truth_input, final_tensor) = add_final_training_ops(len(image_lists.keys()), FLAGS.final_tensor_name, bottleneck_tensor)
# Create the operations we need to evaluate the accuracy of our new layer. evaluation_step = add_evaluation_step(final_tensor, ground_truth_input)
# Merge all the summaries and write them out to /tmp/retrain_logs (by default) merged = tf.summary.merge_all() train_writer = tf.summary.FileWriter(FLAGS.summaries_dir + '/train', sess.graph) validation_writer = tf.summary.FileWriter(FLAGS.summaries_dir + '/validation')
# Set up all our weights to their initial default values. init = tf.global_variables_initializer() sess.run(init)
# Run the training for as many cycles as requested on the command line. for i in range(FLAGS.how_many_training_steps): # Get a batch of input bottleneck values, either calculated fresh every time # with distortions applied, or from the cache stored on disk. if do_distort_images: train_bottlenecks, train_ground_truth = get_random_distorted_bottlenecks( sess, image_lists, FLAGS.train_batch_size, 'training', FLAGS.image_dir, distorted_jpeg_data_tensor, distorted_image_tensor, resized_image_tensor, bottleneck_tensor) else: train_bottlenecks, train_ground_truth = get_random_cached_bottlenecks( sess, image_lists, FLAGS.train_batch_size, 'training', FLAGS.bottleneck_dir, FLAGS.image_dir, jpeg_data_tensor, bottleneck_tensor) # Feed the bottlenecks and ground truth into the graph, and run a training # step. Capture training summaries for TensorBoard with the `merged` op. train_summary, _ = sess.run([merged, train_step], feed_dict={bottleneck_input: train_bottlenecks, ground_truth_input: train_ground_truth}) train_writer.add_summary(train_summary, i)
# Every so often, print out how well the graph is training. is_last_step = (i + 1 == FLAGS.how_many_training_steps) if (i % FLAGS.eval_step_interval) == 0 or is_last_step: train_accuracy, cross_entropy_value = sess.run( [evaluation_step, cross_entropy], feed_dict={bottleneck_input: train_bottlenecks, ground_truth_input: train_ground_truth}) print('%s: Step %d: Train accuracy = %.1f%%' % (datetime.now(), i, train_accuracy * 100)) print('%s: Step %d: Cross entropy = %f' % (datetime.now(), i, cross_entropy_value)) validation_bottlenecks, validation_ground_truth = ( get_random_cached_bottlenecks( sess, image_lists, FLAGS.validation_batch_size, 'validation', FLAGS.bottleneck_dir, FLAGS.image_dir, jpeg_data_tensor, bottleneck_tensor)) # Run a validation step and capture training summaries for TensorBoard # with the `merged` op. validation_summary, validation_accuracy = sess.run( [merged, evaluation_step], feed_dict={bottleneck_input: validation_bottlenecks, ground_truth_input: validation_ground_truth}) validation_writer.add_summary(validation_summary, i) print('%s: Step %d: Validation accuracy = %.1f%%' % (datetime.now(), i, validation_accuracy * 100))
# We've completed all our training, so run a final test evaluation on # some new images we haven't used before. test_bottlenecks, test_ground_truth = get_random_cached_bottlenecks( sess, image_lists, FLAGS.test_batch_size, 'testing', FLAGS.bottleneck_dir, FLAGS.image_dir, jpeg_data_tensor, bottleneck_tensor) test_accuracy = sess.run( evaluation_step, feed_dict={bottleneck_input: test_bottlenecks, ground_truth_input: test_ground_truth}) print('Final test accuracy = %.1f%%' % (test_accuracy * 100))
# Write out the trained graph and labels with the weights stored as constants. output_graph_def = graph_util.convert_variables_to_constants( sess, graph.as_graph_def(), [FLAGS.final_tensor_name]) with gfile.FastGFile(FLAGS.output_graph, 'wb') as f: f.write(output_graph_def.SerializeToString()) with gfile.FastGFile(FLAGS.output_labels, 'w') as f: f.write('\n'.join(image_lists.keys()) + '\n')

if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument( '--image_dir', type=str, default='', help='Path to folders of labeled images.' ) parser.add_argument( '--output_graph', type=str, default='/tmp/output_graph.pb', help='Where to save the trained graph.' ) parser.add_argument( '--output_labels', type=str, default='/tmp/output_labels.txt', help='Where to save the trained graph\'s labels.' ) parser.add_argument( '--summaries_dir', type=str, default='/tmp/retrain_logs', help='Where to save summary logs for TensorBoard.' ) parser.add_argument( '--how_many_training_steps', type=int, default=4000, help='How many training steps to run before ending.' ) parser.add_argument( '--learning_rate', type=float, default=0.01, help='How large a learning rate to use when training.' ) parser.add_argument( '--testing_percentage', type=int, default=10, help='What percentage of images to use as a test set.' ) parser.add_argument( '--validation_percentage', type=int, default=10, help='What percentage of images to use as a validation set.' ) parser.add_argument( '--eval_step_interval', type=int, default=10, help='How often to evaluate the training results.' ) parser.add_argument( '--train_batch_size', type=int, default=100, help='How many images to train on at a time.' ) parser.add_argument( '--test_batch_size', type=int, default=500, help="""\ How many images to test on at a time. This test set is only used infrequently to verify the overall accuracy of the model.\ """ ) parser.add_argument( '--validation_batch_size', type=int, default=100, help="""\ How many images to use in an evaluation batch. This validation set is used much more often than the test set, and is an early indicator of how accurate the model is during training.\ """ ) parser.add_argument( '--model_dir', type=str, default='/tmp/imagenet', help="""\ Path to classify_image_graph_def.pb, imagenet_synset_to_human_label_map.txt, and imagenet_2012_challenge_label_map_proto.pbtxt.\ """ ) parser.add_argument( '--bottleneck_dir', type=str, default='/tmp/bottleneck', help='Path to cache bottleneck layer values as files.' ) parser.add_argument( '--final_tensor_name', type=str, default='final_result', help="""\ The name of the output classification layer in the retrained graph.\ """ ) parser.add_argument( '--flip_left_right', default=False, help="""\ Whether to randomly flip half of the training images horizontally.\ """, action='store_true' ) parser.add_argument( '--random_crop', type=int, default=0, help="""\ A percentage determining how much of a margin to randomly crop off the training images.\ """ ) parser.add_argument( '--random_scale', type=int, default=0, help="""\ A percentage determining how much to randomly scale up the size of the training images by.\ """ ) parser.add_argument( '--random_brightness', type=int, default=0, help="""\ A percentage determining how much to randomly multiply the training image input pixels up or down by.\ """ ) FLAGS, unparsed = parser.parse_known_args() tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)