Notebook

SETI CNN using TF and Binary DS¶

In [ ]:

import requests
import json
#import ibmseti
import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline
import tensorflow as tf
import pickle
import time
#!sudo pip install sklearn
import os
from sklearn.metrics import confusion_matrix
from sklearn import metrics

Set your team folder¶

In [ ]:

### SET YOUR TEAM NAME HERE! Use this folder to save intermediate results
team_name = 'Saeed_team'
mydatafolder = os.path.join( os.environ['PWD'], team_name )  #Change my_data_folder to your team name
if os.path.exists(mydatafolder) is False:
    os.makedirs(mydatafolder)
print mydatafolder

Import dataset reader¶

The following cell will load a python code to read the SETI dataset.

In [ ]:

!wget --output-document SETI.zip  https://ibm.box.com/shared/static/jhqdhcblhua5dx2t7ixwm88okitjrl6l.zip
!unzip -o SETI.zip
import SETI

Download data¶

In [ ]:

ds_directory = mydatafolder + '/SETI/SETI_ds_64x128/'

In [ ]:

print os.popen("ls -lrt "+ ds_directory).read() # to verify

Load data SETI¶

In [ ]:

#from tensorflow.examples.tutorials.mnist import input_data
#dataset = input_data.read_data_sets("MNIST_data/", one_hot=True)
dataset = SETI.read_data_sets(ds_directory, one_hot=True, validation_size=0)
dataset.train.images.shape

Network Parameters¶

In [ ]:

# Parameters
decay_rate=0.96
decay_steps=1000
learning_rate = 0.005
training_epochs = 200
batch_size = 50
display_step = 100

#check point directory
chk_directory = mydatafolder+'/save/'
checkpoint_path = chk_directory+'model.ckpt'


n_classes = 4 # number of possible classifications for the problem
dropout = 0.50 # Dropout, probability to keep units

height = 64 # height of the image in pixels 
width = 128 # width of the image in pixels 
n_input = width * height # number of pixels in one image 

Inputs¶

In [ ]:

x  = tf.placeholder(tf.float32, shape=[None, n_input])
y_ = tf.placeholder(tf.float32, shape=[None, n_classes])

In [ ]:

x_image = tf.reshape(x, [-1,height,width,1]) 
x_image

Convolutional Layer 1¶

In [ ]:

W_conv1 = tf.Variable(tf.truncated_normal([5, 5, 1, 32], stddev=0.1))
b_conv1 = tf.Variable(tf.constant(0.1, shape=[32])) # need 32 biases for 32 outputs
convolve1 = tf.nn.conv2d(x_image, W_conv1, strides=[1, 1, 1, 1], padding='SAME') + b_conv1
h_conv1 = tf.nn.relu(convolve1)
conv1 = tf.nn.max_pool(h_conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME') #max_pool_2x2
conv1

Convolutional Layer 2¶

In [ ]:

W_conv2 = tf.Variable(tf.truncated_normal([5, 5, 32, 64], stddev=0.1))
b_conv2 = tf.Variable(tf.constant(0.1, shape=[64])) #need 64 biases for 64 outputs
convolve2= tf.nn.conv2d(conv1, W_conv2, strides=[1, 1, 1, 1], padding='SAME')+ b_conv2
h_conv2 = tf.nn.relu(convolve2)
conv2 = tf.nn.max_pool(h_conv2, ksize=[1, 2, 2, 1], strides=[1, 4, 4, 1], padding='SAME') #max_pool_2x2
conv2

Convolutional Layer 3¶

W_conv3 = tf.Variable(tf.truncated_normal([5, 5, 64, 128], stddev=0.1)) b_conv3 = tf.Variable(tf.constant(0.1, shape=[128])) #need 64 biases for 64 outputs convolve3= tf.nn.conv2d(conv2, W_conv3, strides=[1, 1, 1, 1], padding='SAME')+ b_conv3 h_conv3 = tf.nn.relu(convolve3) conv3 = tf.nn.max_pool(h_conv3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME') #max_pool_2x2 conv3

Convolutional Layer 4¶

W_conv4 = tf.Variable(tf.truncated_normal([5, 5, 128, 256], stddev=0.1)) b_conv4 = tf.Variable(tf.constant(0.1, shape=[256])) #need 64 biases for 64 outputs convolve4= tf.nn.conv2d(conv3, W_conv4, strides=[1, 1, 1, 1], padding='SAME')+ b_conv4 h_conv4 = tf.nn.relu(convolve4) conv4 = tf.nn.max_pool(h_conv4, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME') #max_pool_2x2

Fully Connected Layer 1¶

In [ ]:

input_layer = conv2
dim = input_layer.get_shape().as_list()
dim

In [ ]:

dims= dim[1]*dim[2]*dim[3]
nodes1 = 1024
prv_layer_matrix = tf.reshape(input_layer, [-1, dims])
W_fc1 = tf.Variable(tf.truncated_normal([dims, nodes1], stddev=0.1))
b_fc1 = tf.Variable(tf.constant(0.1, shape=[nodes1])) # need 1024 biases for 1024 outputs
h_fcl1  = tf.matmul(prv_layer_matrix, W_fc1) + b_fc1
fc_layer1 = tf.nn.relu(h_fcl1) # ???
fc_layer1

Droupout 1¶

In [ ]:

keep_prob = tf.placeholder(tf.float32)
layer_drop1 = tf.nn.dropout(fc_layer1, keep_prob)

Fully Connected Layer 2¶

nodes2 = 256 W_fc2 = tf.Variable(tf.truncated_normal([layer_drop1.get_shape().as_list()[1], nodes2], stddev=0.1)) b_fc2 = tf.Variable(tf.constant(0.1, shape=[nodes2])) h_fcl2 = tf.matmul(layer_drop1, W_fc2) + b_fc2 fc_layer2 = tf.nn.relu(h_fcl2) # ??? fc_layer2

Droupout 2¶

layer_drop2 = tf.nn.dropout(fc_layer2, keep_prob)

Readout Layer¶

In [ ]:

W_fc = tf.Variable(tf.truncated_normal([nodes1, n_classes], stddev=0.1)) #1024 neurons
b_fc = tf.Variable(tf.constant(0.1, shape=[n_classes])) # 10 possibilities for classes [0,1,2,3]
fc = tf.matmul(layer_drop1, W_fc) + b_fc
y_CNN= tf.nn.softmax(fc)

Loss function¶

In [ ]:

cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=y_CNN, labels=y_))

Training¶

In [ ]:

# Create a variable to track the global step.
global_step = tf.Variable(0, trainable=False)

# create learning_decay
lr = tf.train.exponential_decay( learning_rate,
                                 global_step,
                                 decay_steps,
                                 decay_rate, staircase=True )

In [ ]:

# Use the optimizer to apply the gradients that minimize the loss
# (and also increment the global step counter) as a single training step.
optimizer = tf.train.GradientDescentOptimizer(lr)

train_op = optimizer.minimize(cross_entropy, global_step=global_step)
#train_op = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cross_entropy)

Evaluation¶

In [ ]:

correct_prediction = tf.equal(tf.argmax(y_CNN,1), tf.argmax(y_,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

Create checkpoint directory¶

In [ ]:

directory = os.path.dirname(chk_directory)
try:
    os.stat(directory)
    ckpt = tf.train.get_checkpoint_state(chk_directory)
    print ckpt
except:
    os.mkdir(directory) 

Training¶

In [ ]:

# Initializing the variables
init = tf.global_variables_initializer()

In [ ]:

loss_values = []
with tf.Session() as sess:

    
    X_test = dataset.test.images
    y_test = dataset.test.labels
    sess.run(init)
    saver = tf.train.Saver(tf.global_variables())
    
    # load previously trained model if appilcable
    ckpt = tf.train.get_checkpoint_state(chk_directory)
    if ckpt:
        print "loading model: ",ckpt.model_checkpoint_path
        #saver.restore(sess, ckpt.model_checkpoint_path)
    
    
    #step = 0
    num_examples = dataset.train.num_examples
    # Training cycle
    for epoch in range(training_epochs):
        avg_loss = 0.
        avg_accuracy = 0.
        #dataset.shuffle_data()
        total_batch = int(num_examples / batch_size)

        # Loop over all batches
        start = time.time()
        for step in range(total_batch):
            x_batch, y_batch = dataset.train.next_batch(batch_size,shuffle=True)
            train_op.run(feed_dict={x: x_batch, y_: y_batch, keep_prob: dropout})
            loss, acc = sess.run([cross_entropy, accuracy], feed_dict={x: x_batch,y_: y_batch,keep_prob: 1.})
            
            avg_loss += loss / total_batch
            avg_accuracy += acc / total_batch
            if step % display_step == 1000:

                
                # Calculate batch loss and accuracy
                loss, acc = sess.run([cross_entropy, accuracy], feed_dict={x: x_batch,y_: y_batch,keep_prob: 1.})
                #train_accuracy = accuracy.eval(feed_dict={x:x_batch, y_: y_batch,  keep_prob: 0.5})

                test_accuracy = sess.run(accuracy, feed_dict={x: X_test[0:100], y_: y_test[0:100], keep_prob: 1.})

                print("Iter " + str(step) + \
                    ", Minibatch Loss= " +  "{:.6f}".format(loss) +  \
                    ", Training Accuracy= " + "{:.5f}".format(acc)  + \
                    ", Test Accuracy= " + "{:.5f}".format(test_accuracy) )
        
        # save model every 1 epochs
        if epoch >= 0 and epoch % 1 == 0:
            # Save model
            #print ("model saved to {}".format(checkpoint_path))
            #saver.save(sess, checkpoint_path, global_step = epoch)
            end = time.time()
            plr = sess.run(lr)
            loss_values.append(avg_loss)
            #print(sess.run(tf.train.global_step()))
            print "Epoch:", '%04d' % (epoch+1) , ", Epoch time=" , "{:.5f}".format(end - start) , ", lr=", "{:.9f}".format(plr), ", cost=", "{:.9f}".format(avg_loss) ,", Acc=", "{:.9f}".format(avg_accuracy)

    print("Optimization Finished!")
    print ("model saved to {}".format(checkpoint_path))
    saver.save(sess, checkpoint_path, global_step = (epoch+1)*step)

    
    
    # Calculate accuracy for test images
    #print("Testing Accuracy:", sess.run(accuracy, feed_dict={x: X_test[0:30], y_: y_test[0:30], keep_prob: 1.}))
        
    # Find the labels of test set
    y_pred_lb = sess.run(tf.argmax(y_CNN,1), feed_dict={x: X_test[0:100], y_: y_test[0:100], keep_prob: 1.})
    y_pred = sess.run(y_CNN, feed_dict={x: X_test[0:100], y_: y_test[0:100], keep_prob: 1.})
    
    # lets save kernels
    kernels_l1 = sess.run(tf.reshape(tf.transpose(W_conv1, perm=[2, 3, 0, 1]),[32,-1]))
    kernels_l2 = sess.run(tf.reshape(tf.transpose(W_conv2, perm=[2, 3, 0, 1]),[32*64,-1]))

In [ ]:

%matplotlib inline
import numpy as np
import matplotlib.pyplot as plt
plt.plot([np.mean(loss_values[i:i+5]) for i in range(len(loss_values))])
plt.show()

Evaluation¶

Accuracy is depend on the number of epoch that you set in partametrs part.

In [ ]:

y_ = np.argmax(y_test[0:100],1) # ground truth
print metrics.classification_report(y_true= y_, y_pred= y_pred_lb)
print metrics.confusion_matrix(y_true= y_, y_pred= y_pred_lb)
print("Classification accuracy: %0.6f" % metrics.accuracy_score(y_true= y_, y_pred= y_pred_lb) )
print("Log Loss: %0.6f" % metrics.log_loss(y_true= y_, y_pred= y_pred, labels=range(4)) )

Generate CSV file for Scoreboard¶

Here's an example of what the CSV file should look like for submission to the scoreboard. Although, in this case, we only have 4 classes instead of 7.

NOTE: This uses the test set created in Step_5c, which only contain the BASIC4 test data set. The code challenge and hackathon will be based on the Primary Data Set which contains 7 signal classes, and different test set.¶

In [ ]:

my_output_results = mydatafolder + '/' + 'DL_scores.csv'
with open(my_output_results, 'w') as csvfile:
    np.savetxt(my_output_results, y_pred, delimiter=",")

In [ ]:

print os.popen("ls -lrt "+ mydatafolder).read() # to verify

Viz¶

In [ ]:

!wget --output-document utils1.py http://deeplearning.net/tutorial/code/utils.py
import utils1
from utils1 import tile_raster_images

In [ ]:

#from utils import tile_raster_images
import matplotlib.pyplot as plt
from PIL import Image
%matplotlib inline
image = Image.fromarray(tile_raster_images(kernels_l1, img_shape=(5, 5) ,tile_shape=(4, 8), tile_spacing=(1, 1)))
### Plot image
plt.rcParams['figure.figsize'] = (18.0, 18.0)
imgplot = plt.imshow(image)
imgplot.set_cmap('gray')  

In [ ]:

image = Image.fromarray(tile_raster_images(kernels_l2, img_shape=(5, 5) ,tile_shape=(4, 12), tile_spacing=(1, 1)))
### Plot image
plt.rcParams['figure.figsize'] = (18.0, 18.0)
imgplot = plt.imshow(image)
imgplot.set_cmap('gray')  

In [ ]:

import numpy as np
plt.rcParams['figure.figsize'] = (5.0, 5.0)
sampleimage1 = X_test[3]
plt.imshow(np.reshape(sampleimage1,[64,128]), cmap="gray")

In [ ]:

# Launch the graph
with tf.Session() as sess:
    sess.run(init)
    saver = tf.train.Saver(tf.all_variables())
    
    # load previously trained model if appilcable
    ckpt = tf.train.get_checkpoint_state(chk_directory)
    if ckpt:
        print "loading model: ",ckpt.model_checkpoint_path
        saver.restore(sess, ckpt.model_checkpoint_path)
    ActivatedUnits1 = sess.run(convolve1,feed_dict={x:np.reshape(sampleimage1,[1,64*128],order='F'),keep_prob:1.0})
    plt.figure(1, figsize=(20,20))
    n_columns = 3
    n_rows = 3
    for i in range(9):
        plt.subplot(n_rows, n_columns, i+1)
        plt.title('Filter ' + str(i))
        plt.imshow(ActivatedUnits1[0,:,:,i], interpolation="nearest", cmap="gray")

Authors:

Saeed Aghabozorgi

Saeed Aghabozorgi, PhD is Sr. Data Scientist in IBM with a track record of developing enterprise level applications that substantially increases clients’ ability to turn data into actionable knowledge. He is a researcher in data mining field and expert in developing advanced analytic methods like machine learning and statistical modelling on large datasets.

In [ ]: