Spark-2.4 Deep Learning Pipelines (Keras)Image Claasifer
(原文链接)-这是Spark2018 Submit 的一个演讲Demo, 针对Keras图片分类和使用Spark做分类的方法做了讲解，供学习使用。
keras_dlp_image_classifier(Python)
Import Notebook
Part 1: Exploring and Classifying Images with Pretrained Models
We will use Keras with TensorFlow as the backend, and download VGG16 from Keras.
VGG16
from keras.preprocessing import image
from keras.applications.vgg16 import preprocess_input, decode_predictions, VGG16
import numpy as np
import os
vgg16Model = VGG16(weights='imagenet')
Function to predict category
def predict_images(images, m):
for i in images:
print ('processing image:', i)
img = image.load_img(i, target_size=(224, 224))
#convert to numpy array for Keras image formate processing
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)
preds = m.predict(x)
# decode the results into a list of tuples (class, description, probability
print('Predicted:', decode_predictions(preds, top=3)[0], '\n')
Classify African and Indian Elephants
Load and predict images using VGG16 pretrained model
elephants_img_paths = ["/dbfs/brooke/spark-summit-sf/elephants/" + path for path in os.listdir("/dbfs/brooke/spark-summit-sf/elephants/")]
predict_images(elephants_img_paths, vgg16Model)
processing image: /dbfs/brooke/spark-summit-sf/elephants/african_elephant_1.jpg Downloading data from https://s3.amazonaws.com/deep-learning-models/image-models/imagenet_class_index.json
...
processing image: /dbfs/brooke/spark-summit-sf/elephants/indian_elephant_1.jpeg
Predicted: [('n01871265', 'tusker', 0.63030255), ('n02504013', 'Indian_elephant', 0.3172723), ('n02504458', 'African_elephant', 0.052417696)]
processing image: /dbfs/brooke/spark-summit-sf/elephants/indian_elephant_3.jpeg
Predicted: [('n03980874', 'poncho', 0.27748922), ('n02504013', 'Indian_elephant', 0.15591854), ('n03884397', 'panpipe', 0.118131705)]
hotdog_img_paths = ["/dbfs/brooke/spark-summit-sf/hotdog/" + path for path in os.listdir("/dbfs/brooke/spark-summit-sf/hotdog/")]
predict_images(hotdog_img_paths, vgg16Model)
processing image: /dbfs/brooke/spark-summit-sf/hotdog/hamburger_1.jpeg
Predicted: [('n07697313', 'cheeseburger', 0.98272187), ('n07693725', 'bagel', 0.0061851414), ('n07613480', 'trifle', 0.005066536)]
processing image: /dbfs/brooke/spark-summit-sf/hotdog/hotdog_1.jpeg
Predicted: [('n07697537', 'hotdog', 0.9997451), ('n07697313', 'cheeseburger', 8.424303e-05), ('n07615774', 'ice_lolly', 4.0920193e-05)] ...
DeepImagePredictor
Let's make these predictions in parallel on our Spark cluster!
from pyspark.ml.image import ImageSchema
from sparkdl.image import imageIO
from sparkdl import DeepImagePredictor
nerds_df = ImageSchema.readImages("brooke/spark-summit-sf/nerds/")
predictor = DeepImagePredictor(inputCol="image", outputCol="predicted_labels", modelName="VGG16", decodePredictions=True, topK=5)
predictions_df = predictor.transform(nerds_df).cache()
predictions_df.count()
INFO:tensorflow:Froze 32 variables. Converted 32 variables to const ops. INFO:tensorflow:Froze 0 variables. Converted 0 variables to const ops. Out[5]: 8
display(predictions_df)
| image | predicted_labels |
|---|---|
 |
[{"class":"n03045698","description":"cloak","probability":0.39211166},{"class":"n03787032","description":"mortarboard","probability":0.091029376},{"class":"n03404251","description":"fur_coat","probability":0.08471853},{"class":"n04371774","description":"swing","probability":0.056981083},{"class":"n04370456","description":"sweatshirt","probability":0.028172707}] |
 |
[{"class":"n04350905","description":"suit","probability":0.41138414},{"class":"n02916936","description":"bulletproof_vest","probability":0.10152984},{"class":"n03763968","description":"military_uniform","probability":0.09318812},{"class":"n04591157","description":"Windsor_tie","probability":0.07702819},{"class":"n02669723","description":"academic_gown","probability":0.03608404}] |
 |
[{"class":"n03630383","description":"lab_coat","probability":0.23492548},{"class":"n04591157","description":"Windsor_tie","probability":0.09487544},{"class":"n03838899","description":"oboe","probability":0.049194943},{"class":"n04350905","description":"suit","probability":0.043708242},{"class":"n03832673","description":"notebook","probability":0.041520175}] |
| [{"class":"n04350905","description":"suit","probability":0.25813994},{"class":"n01440764","description":"tench","probability":0.03799466},{"class":"n03838899","description":"oboe","probability":0.03496751},{"class":"n02883205","description":"bow_tie","probability":0.033893984},{"class":"n03394916","description":"French_horn","probability":0.03332546}] | |
 |
[{"class":"n03595614","description":"jersey","probability":0.3530513},{"class":"n04370456","description":"sweatshirt","probability":0.13232166},{"class":"n03942813","description":"ping-pong_ball","probability":0.097091846},{"class":"n03141823","description":"crutch","probability":0.018438742},{"class":"n04270147","description":"spatula","probability":0.017245641}] |
 |
[{"class":"n03000247","description":"chain_mail","probability":0.14295407},{"class":"n02672831","description":"accordion","probability":0.10376813},{"class":"n02787622","description":"banjo","probability":0.069579415},{"class":"n02804610","description":"bassoon","probability":0.061210092},{"class":"n03838899","description":"oboe","probability":0.058611386}] |
 |
[{"class":"n03630383","description":"lab_coat","probability":0.6628539},{"class":"n04317175","description":"stethoscope","probability":0.12459004},{"class":"n04370456","description":"sweatshirt","probability":0.038363792},{"class":"n04039381","description":"racket","probability":0.0132558},{"class":"n03595614","description":"jersey","probability":0.008680802}] |
 |
[{"class":"n04350905","description":"suit","probability":0.17892571},{"class":"n02883205","description":"bow_tie","probability":0.0947369},{"class":"n04591157","description":"Windsor_tie","probability":0.08621924},{"class":"n04162706","description":"seat_belt","probability":0.07562429},{"class":"n03630383","description":"lab_coat","probability":0.06052583}] |
Show image preview
Detected data types for which enhanced rendering is supported. For details, see the Databricks Guide.
Let's change the model
inception = DeepImagePredictor(inputCol="image", outputCol="predicted_labels", modelName="InceptionV3", decodePredictions=True, topK=5)
inception_df = inception.transform(nerds_df).cache()
inception_df.count()
Downloading data from https://github.com/fchollet/deep-learning-models/releases/download/v0.5/inception_v3_weights_tf_dim_ordering_tf_kernels.h5
0us/step INFO:tensorflow:Froze 378 variables. Converted 378 variables to const ops. INFO:tensorflow:Froze 0 variables. Converted 0 variables to const ops. Out[7]: 8
display(inception_df)
| image | predicted_labels |
|---|---|
 |
[{"class":"n04350905","description":"suit","probability":0.14165702},{"class":"n04479046","description":"trench_coat","probability":0.11945703},{"class":"n03404251","description":"fur_coat","probability":0.027757034},{"class":"n04370456","description":"sweatshirt","probability":0.024130477},{"class":"n02837789","description":"bikini","probability":0.021668304}] |
 |
[{"class":"n02916936","description":"bulletproof_vest","probability":0.81811774},{"class":"n03787032","description":"mortarboard","probability":0.036147255},{"class":"n03763968","description":"military_uniform","probability":0.02605114},{"class":"n02669723","description":"academic_gown","probability":0.01372046},{"class":"n04350905","description":"suit","probability":0.008794671}] |
 |
[{"class":"n04479046","description":"trench_coat","probability":0.054543473},{"class":"n03838899","description":"oboe","probability":0.042113375},{"class":"n03630383","description":"lab_coat","probability":0.0317195},{"class":"n02787622","description":"banjo","probability":0.029045274},{"class":"n02804610","description":"bassoon","probability":0.026370155}] |
 |
[{"class":"n04350905","description":"suit","probability":0.7987358},{"class":"n02883205","description":"bow_tie","probability":0.053425536},{"class":"n04591157","description":"Windsor_tie","probability":0.011151478},{"class":"n02992529","description":"cellular_telephone","probability":0.0053684525},{"class":"n03763968","description":"military_uniform","probability":0.0039382246}] |
 |
[{"class":"n03595614","description":"jersey","probability":0.15574361},{"class":"n03942813","description":"ping-pong_ball","probability":0.05970348},{"class":"n04370456","description":"sweatshirt","probability":0.048369024},{"class":"n02804610","description":"bassoon","probability":0.034532476},{"class":"n03838899","description":"oboe","probability":0.03400313}] |
 |
[{"class":"n03763968","description":"military_uniform","probability":0.117212564},{"class":"n04350905","description":"suit","probability":0.035018962},{"class":"n02787622","description":"banjo","probability":0.033046678},{"class":"n04584207","description":"wig","probability":0.032433487},{"class":"n04317175","description":"stethoscope","probability":0.028688557}] |
 |
[{"class":"n03630383","description":"lab_coat","probability":0.36856785},{"class":"n04317175","description":"stethoscope","probability":0.037452906},{"class":"n03832673","description":"notebook","probability":0.03503557},{"class":"n04350905","description":"suit","probability":0.028838113},{"class":"n03787032","description":"mortarboard","probability":0.020943912}] |
 |
[{"class":"n03763968","description":"military_uniform","probability":0.9151895},{"class":"n03787032","description":"mortarboard","probability":0.012946242},{"class":"n04350905","description":"suit","probability":0.007009363},{"class":"n02669723","description":"academic_gown","probability":0.0068341326},{"class":"n02883205","description":"bow_tie","probability":0.0064667594}] |
Show image preview
Detected data types for which enhanced rendering is supported. For details, see the Databricks Guide.
Part 2: Transfer Learning with Deep Learning Pipelines (DLP)
Deep Learning Pipelines provides utilities to perform transfer learning on images, which is one of the fastest (code and run-time-wise) ways to start using deep learning. Using Deep Learning Pipelines, it can be done in just several lines of code.
The idea behind transfer learning is to take knowledge from one model doing some task, and transfer it to build another model doing a similar task.
from pyspark.ml.image import ImageSchema
from pyspark.sql.functions import lit
from sparkdl.image import imageIO
img_dir = 'dbfs:/brooke/spark-summit-sf'
cats_df = ImageSchema.readImages(img_dir + "/cats").withColumn("label", lit(1))
dogs_df = ImageSchema.readImages(img_dir + "/dogs").withColumn("label", lit(0))
cats_train, cats_test = cats_df.randomSplit([.8, .2], seed=42)
dogs_train, dogs_test = dogs_df.randomSplit([.8, .2], seed=42)
train_df = cats_train.unionAll(dogs_train).cache()
test_df = cats_test.unionAll(dogs_test).cache()
display(train_df.select("image", "label"))
| image | label |
|---|---|
 |
1 |
 |
1 |
 |
1 |
 |
1 |
 |
1 |
 |
0 |
 |
0 |
 |
0 |
 |
0 |
 |
0 |
Show image preview
Detected data types for which enhanced rendering is supported. For details, see the Databricks Guide.
Build the MLlib Pipeline
Use DeepImageFeaturizer and LogisticRegression
from pyspark.ml.classification import LogisticRegression
from pyspark.ml import Pipeline
from sparkdl import DeepImageFeaturizer
featurizer = DeepImageFeaturizer(inputCol="image", outputCol="features", modelName="InceptionV3")
lr = LogisticRegression(maxIter=20, regParam=0.05, elasticNetParam=0.3, labelCol="label")
p = Pipeline(stages=[featurizer, lr])
p_model = p.fit(train_df)
Evaluate the Accuracy
from pyspark.ml.evaluation import MulticlassClassificationEvaluator
pred_df = p_model.transform(test_df).cache()
evaluator = MulticlassClassificationEvaluator(metricName="accuracy")
print("Test set accuracy = " + str(evaluator.evaluate(pred_df.select("prediction", "label"))*100) + "%")
Test set accuracy = 100.0%
display(pred_df.select("image", "label", "probability"))
| image | label | probability |
|---|---|---|
 |
1 | [1,2,[],[0.07983768538504338,0.9201623146149567]] |
 |
1 | [1,2,[],[0.0735124824751803,0.9264875175248197]] |
 |
1 | [1,2,[],[0.0688419453818859,0.9311580546181142]] |
 |
0 | [1,2,[],[0.9475188514834973,0.0524811485165027]] |
 |
0 | [1,2,[],[0.9026450467442289,0.0973549532557711]] |
 |
0 | [1,2,[],[0.7731177886783923,0.2268822113216076]] |
Show image preview
Detected data types for which enhanced rendering is supported. For details, see the Databricks Guide.