The following notebook represent one of the solution for the Statoil challange hosted on kaggle in which we had to binary classify the data taken from satelites.
There are two highlight of my proposed solution:
- Training data augmentation
- CNN classifier
I used 12 fold augmentation. The augmentation is done in the folowing fashion-
- Rotation by +ve degree
- Rotation by -ve degree
- Horizonation translation to right
- Horizonation translation to left
- Vertical translation downward
- Vertical translation upward
- SE translation
- NW translation
- NE translation
- SW translation
- Vertical flip
- Horizontal flip
- Horizontal flip
- Zoom in
- Zoom out
The score without image augmentation was 0.299 (2556th rank). After augmentation the score was 0.1571(400th rank). Couple of tweaks and optimazation may yet be needed to generate better result.
import numpy as np
import pandas as pd
from sklearn.model_selection import train_test_split
from os.path import join as opj
from matplotlib import pyplot as plt
from mpl_toolkits.mplot3d import Axes3D
import pylab
plt.rcParams['figure.figsize'] = 20, 14
%matplotlib inline
import mathtrain = pd.read_json("./statoil-iceberg/train.json")train[train['inc_angle']=='na'] = train[train['inc_angle']!='na']['inc_angle'].mean()
train['inc_angle'] = train['inc_angle'].apply(lambda x: math.radians(x))def standardise_vector(vector):
'''standardise vector'''
standardised_vector = (np.array(vector) - np.mean(vector)) / np.std(vector)
return standardised_vector.tolist()train['band_1'] = train['band_1'].apply(standardise_vector)
train['band_2'] = train['band_2'].apply(standardise_vector)def find_missing_data(series, shape):
'''function which return the count and the index of mismatched data'''
count = 0
missing_list = []
for i,x in enumerate(series):
if np.shape(series.iloc[i]) != shape:
missing_list.append(i)
count += 1
return missing_list, countmissing_list1, count1 = find_missing_data(train.band_1, (5625,))
print("count: ", count1)count: 133
missing_list2, count2 = find_missing_data(train.band_2, (5625,))
print("count: ", count1)
missing_list1 == missing_list2count: 133
True
train =train.drop(train.index[missing_list1])train.head(5)
<style scoped>
.dataframe tbody tr th:only-of-type {
vertical-align: middle;
}
</style>
.dataframe tbody tr th {
vertical-align: top;
}
.dataframe thead th {
text-align: right;
}
| band_1 | band_2 | id | inc_angle | is_iceberg | |
|---|---|---|---|---|---|
| 0 | [0.01182174564467684, 0.27378275943768166, -0.... | [1.1573585860406173, 0.15631457838356574, -0.4... | dfd5f913 | 0.766617 | 0.0 |
| 1 | [0.42137323598577087, -0.43078366534450846, -0... | [-2.0950953014724543, -0.8948057535299927, -0.... | e25388fd | 0.665951 | 0.0 |
| 2 | [-0.6969623073924855, -0.6969793944501068, -0.... | [-0.014649839717716016, 0.35157292381049343, 1... | 58b2aaa0 | 0.790388 | 1.0 |
| 3 | [0.2946591706447792, 0.04985985071106006, -0.3... | [0.7023040788844376, 0.8569610477684707, 1.005... | 4cfc3a18 | 0.764988 | 0.0 |
| 4 | [-0.31533024179271146, 0.9175181439939705, 0.9... | [-0.693769690678262, -2.0346940986353044, -2.0... | 271f93f4 | 0.621784 | 0.0 |
band_1 = np.array([np.array(band).astype(np.float32).reshape(75, 75) for band in train["band_1"]])
band_2 = np.array([np.array(band).astype(np.float32).reshape(75, 75) for band in train["band_2"]])labels = train.is_iceberg.as_matrix()
angles = train.inc_angle.as_matrix()# randomly choosing the train and validation indices
train_indices = np.random.choice(len(labels), round(len(labels)*0.75), replace=False)
validation_indices = np.array(list(set(range(len(labels))) - set(train_indices)))# extract train set
band_1_train = band_1[train_indices]
band_2_train = band_2[train_indices]
angles_train = angles[train_indices]
labels_train = labels[train_indices]
# extract validation set
band_1_validation = band_1[validation_indices]
band_2_validation = band_2[validation_indices]
angles_validation = angles[validation_indices]
labels_validation = labels[validation_indices]
# # extract test set
# band_1_test = band_1_test
# band_2_test = band_2_test
# angles_test = test_data.inc_angle.as_matrix()
# iD = test_data.id.as_matrix()# Converting the data to floating point
band_1_train = band_1_train.astype(np.float32)
band_1_validation = band_1_validation.astype(np.float32)
# band_1_test = band_1_test.astype(np.float32)
band_2_train = band_2_train.astype(np.float32)
band_2_validation = band_2_validation.astype(np.float32)
# band_2_test = band_2_test.astype(np.float32)
angles_train = angles_train.astype(np.float32)
angles_validation = angles_validation.astype(np.float32)
# angles_test = angles_test.astype(np.float32)
labels_train = labels_train.astype(np.float32)
labels_validation = labels_validation.astype(np.float32)
# iD = iD.astype(np.str)from agument import *
plt.rcParams['figure.figsize'] = (20.0, 14.0)
image = band_1_train[3].copy()
plt.subplot(3, 5, 1)
plt.title("Original Image")
plt.imshow(image)
plt.subplot(3, 5, 2)
generated_image = rotate_image(image,40)
plt.title("Rotation by +ve degree")
plt.imshow(generated_image)
plt.subplot(3, 5, 3)
generated_image = rotate_image(image,-40)
plt.title("Rotation by -ve degree")
plt.imshow(generated_image)
plt.subplot(3, 5, 4)
generated_image = translate_horizontal(image,10)
plt.title("Horizonation translation to right")
plt.imshow(generated_image)
plt.subplot(3, 5, 5)
generated_image = translate_horizontal(image,-10)
plt.title("Horizonation translation to left")
plt.imshow(generated_image)
plt.subplot(3, 5, 6)
generated_image = translate_vertical(image,10)
plt.title("Vertical translation downward")
plt.imshow(generated_image)
plt.subplot(3, 5, 7)
generated_image = translate_vertical(image,-10)
plt.title("Vertical translation upward")
plt.imshow(generated_image)
plt.subplot(3, 5, 8)
generated_image = translate_positive_diagonal(image,10)
plt.title("SE translation")
plt.imshow(generated_image)
plt.subplot(3, 5, 9)
generated_image = translate_positive_diagonal(image,-10)
plt.title("NW translation")
plt.imshow(generated_image)
plt.subplot(3, 5, 10)
generated_image = translate_negative_diagonal(image,10)
plt.title("NE translation")
plt.imshow(generated_image)
plt.subplot(3, 5, 11)
generated_image = translate_negative_diagonal(image,-10)
plt.title("SW translation")
plt.imshow(generated_image)
plt.subplot(3, 5, 12)
generated_image = flip(image,0)
plt.title("Vertical flip")
plt.imshow(generated_image)
plt.subplot(3, 5, 13)
generated_image = flip(image,1)
plt.title("Horizontal flip")
plt.imshow(generated_image)
plt.subplot(3, 5, 14)
generated_image = zoom(image,10)
plt.title("Zoom in")
plt.imshow(generated_image)
plt.subplot(3, 5, 15)
generated_image = zoom(image,-10)
plt.title("Zoom out")
plt.imshow(generated_image)
plt.show()
def augment_data(band1, band2, angles, labels):
'''a function to augment band1 and band2 image'''
# list to store the generated data
band1_generated = []
band2_generated = []
angles_generated = []
labels_generated = []
# iterate through each point in train set
for i in range(labels.shape[0]):
# rotate by positive degree
angle = np.random.randint(5,20)
band1_generated.append(rotate_image(band1[i],angle))
band2_generated.append(rotate_image(band2[i],angle))
angles_generated.append(angles[i])
labels_generated.append(labels[i])
# rotate by negative degree
angle = np.random.randint(5,20)
band1_generated.append(rotate_image(band1[i],-angle))
band2_generated.append(rotate_image(band2[i],-angle))
angles_generated.append(angles[i])
labels_generated.append(labels[i])
# positive horizontal shift
shift = np.random.randint(3,7)
band1_generated.append(translate_horizontal(band1[i],+shift))
band2_generated.append(translate_horizontal(band2[i],+shift))
angles_generated.append(angles[i])
labels_generated.append(labels[i])
# negative horizontal shift
shift = np.random.randint(3,7)
band1_generated.append(translate_horizontal(band1[i],-shift))
band2_generated.append(translate_horizontal(band2[i],-shift))
angles_generated.append(angles[i])
labels_generated.append(labels[i])
# positive vertical shift
shift = np.random.randint(0,7)
band1_generated.append(translate_vertical(band1[i],+shift))
band2_generated.append(translate_vertical(band2[i],+shift))
angles_generated.append(angles[i])
labels_generated.append(labels[i])
# negative vertical shift
shift = np.random.randint(3,7)
band1_generated.append(translate_vertical(band1[i],-shift))
band2_generated.append(translate_vertical(band2[i],-shift))
angles_generated.append(angles[i])
labels_generated.append(labels[i])
# translate along positive diagonal in positive direction
shift = np.random.randint(3,7)
band1_generated.append(translate_positive_diagonal(band1[i],+shift))
band2_generated.append(translate_positive_diagonal(band2[i],+shift))
angles_generated.append(angles[i])
labels_generated.append(labels[i])
# translate along positive diagonal in negative direction
shift = np.random.randint(3,7)
band1_generated.append(translate_positive_diagonal(band1[i],-shift))
band2_generated.append(translate_positive_diagonal(band2[i],-shift))
angles_generated.append(angles[i])
labels_generated.append(labels[i])
# translate along negative diagonal in positive direction
shift = np.random.randint(3,7)
band1_generated.append(translate_negative_diagonal(band1[i],+shift))
band2_generated.append(translate_negative_diagonal(band2[i],+shift))
angles_generated.append(angles[i])
labels_generated.append(labels[i])
# translate along negative diagonal in negative direction
shift = np.random.randint(3,7)
band1_generated.append(translate_negative_diagonal(band1[i],-shift))
band2_generated.append(translate_negative_diagonal(band2[i],-shift))
angles_generated.append(angles[i])
labels_generated.append(labels[i])
# vertical flip
band1_generated.append(flip(band1[i],0))
band2_generated.append(flip(band2[i],0))
angles_generated.append(angles[i])
labels_generated.append(labels[i])
# horizontal flip
band1_generated.append(flip(band1[i],1))
band2_generated.append(flip(band2[i],1))
angles_generated.append(angles[i])
labels_generated.append(labels[i])
# zoom in image
zoom_shift = np.random.randint(2,5)
band1_generated.append(zoom(band1[i],zoom_shift))
band2_generated.append(zoom(band2[i],zoom_shift))
angles_generated.append(angles[i])
labels_generated.append(labels[i])
# zoom out image
zoom_shift = np.random.randint(2,5)
band1_generated.append(zoom(band1[i],-zoom_shift))
band2_generated.append(zoom(band2[i],-zoom_shift))
angles_generated.append(angles[i])
labels_generated.append(labels[i])
# convert the generated data into numpy array
band1_generated = np.array(band1_generated)
band2_generated = np.array(band2_generated)
angles_generated = np.array(angles_generated)
labels_generated = np.array(labels_generated)
# concatenate the generated data to original train set
band1_augmented = np.concatenate((band1, band1_generated),axis=0)
band2_augmented = np.concatenate((band2, band2_generated),axis=0)
angles_augmented = np.concatenate((angles, angles_generated),axis=0)
labels_augmented = np.concatenate((labels, labels_generated),axis=0)
return band1_augmented, band2_augmented, angles_augmented, labels_augmentedband_1_train, band_2_train, angles_train, labels_train = \
augment_data(band_1_train, band_2_train, angles_train, labels_train)
print("Shape of band_1_train:",band_1_train.shape)
print("Shape of band_2_train:",band_2_train.shape)
print("Shape of angles_train:",angles_train.shape)
print("Shape of labels_train:",labels_train.shape)Shape of band_1_train: (16545, 75, 75)
Shape of band_2_train: (16545, 75, 75)
Shape of angles_train: (16545,)
Shape of labels_train: (16545,)
image_train = np.concatenate([band_1_train[:, :, :, np.newaxis],
band_2_train[:, :, :, np.newaxis],
((band_1_train+band_2_train)/2)[:, :, :, np.newaxis]],
axis=-1)image_validation = np.concatenate([band_1_validation[:, :, :, np.newaxis],
band_2_validation[:, :, :, np.newaxis],
((band_1_validation+band_2_validation)/2)[:, :, :, np.newaxis]],
axis=-1)print("Shape of image_train:",image_train.shape)
print("Shape of image_validation:",image_validation.shape)Shape of image_train: (16545, 75, 75, 3)
Shape of image_validation: (368, 75, 75, 3)
import plotly.offline as py
import plotly.graph_objs as go
py.init_notebook_mode(connected=True)
def plotmy3d(c, name):
data = [go.Surface(z=c)]
layout = go.Layout(
title=name,
autosize=False,
width=700,
height=700,
margin=dict(
l=65,
r=50,
b=65,
t=90
)
)
fig = go.Figure(data=data, layout=layout)
py.iplot(fig)
plotmy3d(band_1_train[17,:,:], 'Ship!!!')del(band_1_train, band_1_validation, band_2_train, band_2_validation)from matplotlib import pyplot
from keras.preprocessing.image import ImageDataGenerator
from keras.models import Sequential
from keras.layers import Conv2D, MaxPooling2D, Dense, Dropout, Input, Flatten, Activation
from keras.layers import GlobalMaxPooling2D
from keras.layers.normalization import BatchNormalization
from keras.layers.merge import Concatenate
from keras.models import Model
from keras import initializers
from keras.optimizers import Adam
from keras.callbacks import ModelCheckpoint, Callback, EarlyStopping/usr/local/lib/python3.5/dist-packages/h5py/__init__.py:36: FutureWarning:
Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.
Using TensorFlow backend.
def getModel():
#Building the model
gmodel=Sequential()
#Conv Layer 1
gmodel.add(Conv2D(64, kernel_size=(3, 3),activation='relu', input_shape=(75, 75, 3)))
gmodel.add(MaxPooling2D(pool_size=(3, 3), strides=(2, 2)))
gmodel.add(Dropout(0.2))
#Conv Layer 2
gmodel.add(Conv2D(128, kernel_size=(3, 3), activation='relu' ))
gmodel.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
gmodel.add(Dropout(0.2))
# #Conv Layer 3.1
# gmodel.add(Conv2D(256, kernel_size=(3, 3), activation='relu'))
# gmodel.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
# gmodel.add(Dropout(0.2))
#Conv Layer 3.2
gmodel.add(Conv2D(128, kernel_size=(3, 3), activation='relu'))
gmodel.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
gmodel.add(Dropout(0.2))
#Conv Layer 4
gmodel.add(Conv2D(64, kernel_size=(3, 3), activation='relu'))
gmodel.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
gmodel.add(Dropout(0.2))
#Flatten the data for upcoming dense layers
gmodel.add(Flatten())
#Dense Layers
gmodel.add(Dense(512))
gmodel.add(Activation('relu'))
gmodel.add(Dropout(0.2))
#Dense Layer 2
gmodel.add(Dense(256))
gmodel.add(Activation('relu'))
gmodel.add(Dropout(0.2))
#Sigmoid Layer
gmodel.add(Dense(1))
gmodel.add(Activation('sigmoid'))
mypotim=Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)
gmodel.compile(loss='binary_crossentropy',
optimizer=mypotim,
metrics=['accuracy'])
gmodel.summary()
return gmodelX_train_cv, _, y_train_cv,_ = train_test_split(image_train, labels_train, random_state=1, train_size=0.90)
_, X_valid,_, y_valid = train_test_split(image_validation, labels_validation, random_state=1, train_size=0.10)/usr/local/lib/python3.5/dist-packages/sklearn/model_selection/_split.py:2026: FutureWarning:
From version 0.21, test_size will always complement train_size unless both are specified.
del (image_train, image_validation, labels_train, labels_validation)import os
gmodel=getModel()
gmodel.fit(X_train_cv, y_train_cv,
batch_size=24,
epochs=8,
verbose=1,
validation_data=(X_valid, y_valid))_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_1 (Conv2D) (None, 73, 73, 64) 1792
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 36, 36, 64) 0
_________________________________________________________________
dropout_1 (Dropout) (None, 36, 36, 64) 0
_________________________________________________________________
conv2d_2 (Conv2D) (None, 34, 34, 128) 73856
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 17, 17, 128) 0
_________________________________________________________________
dropout_2 (Dropout) (None, 17, 17, 128) 0
_________________________________________________________________
conv2d_3 (Conv2D) (None, 15, 15, 128) 147584
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 7, 7, 128) 0
_________________________________________________________________
dropout_3 (Dropout) (None, 7, 7, 128) 0
_________________________________________________________________
conv2d_4 (Conv2D) (None, 5, 5, 64) 73792
_________________________________________________________________
max_pooling2d_4 (MaxPooling2 (None, 2, 2, 64) 0
_________________________________________________________________
dropout_4 (Dropout) (None, 2, 2, 64) 0
_________________________________________________________________
flatten_1 (Flatten) (None, 256) 0
_________________________________________________________________
dense_1 (Dense) (None, 512) 131584
_________________________________________________________________
activation_1 (Activation) (None, 512) 0
_________________________________________________________________
dropout_5 (Dropout) (None, 512) 0
_________________________________________________________________
dense_2 (Dense) (None, 256) 131328
_________________________________________________________________
activation_2 (Activation) (None, 256) 0
_________________________________________________________________
dropout_6 (Dropout) (None, 256) 0
_________________________________________________________________
dense_3 (Dense) (None, 1) 257
_________________________________________________________________
activation_3 (Activation) (None, 1) 0
=================================================================
Total params: 560,193
Trainable params: 560,193
Non-trainable params: 0
_________________________________________________________________
Train on 14890 samples, validate on 332 samples
Epoch 1/8
14890/14890 [>>>>>>>>>>>>>>>>>>>>>>>>>>>] - ETA: 8:20 - loss: 0.2981 - acc: 0.8541