RegressionProject/train_phone_finder.py at master · Binbose/RegressionProject · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
import keras
from keras import Model
from keras.layers import Conv2D, Input, Dense,Flatten,LeakyReLU, BatchNormalization
from keras.callbacks import TensorBoard, ModelCheckpoint, EarlyStopping
from keras import backend as K
import tensorflow as tf

import os
import argparse

from image_helper import *
from MyDataGenerator import DataGenerator


###########################################################################################
############################### READ COMMANDLINE ##########################################
###########################################################################################

parser = argparse.ArgumentParser(description='Augments the data and trains the model on this data.')
parser.add_argument('path', metavar='p', help='Path to the train images')
parser.add_argument('use_pretrained', metavar='u', help='Boolean: Whether to use pretrained model and fine tune it, or to train model from scratch', defeault=True)
args = parser.parse_args()

path = args.path
use_pretrained = args.path


###########################################################################################
############################### LOAD DATA #################################################
###########################################################################################

lines = [line.rstrip('\n') for line in open(path + "labels.txt")]
labels = list()
for line in lines:
    labels.append(line.split(" "))


data, data_label = load_data(path, labels)

split = int(data.shape[0]*0.8)
train_data = data[:split]
validation_data = data[split:]


train_label = data_label[:split]
validation_label = data_label[split:]


###########################################################################################
############################### AUGMENT DATA ##############################################
###########################################################################################

if not os.path.exists("./augmented_data/"):
    os.makedirs("./augmented_data/")

    augmented_train_data, augmented_train_label = augment_all(train_data, train_label, n_translations=0, n_scales=0)
    save_augmented_data("./augmented_data/", augmented_train_data, augmented_train_label, "train")

    augmented_validation_data, augmented_validation_label = augment_all(validation_data, validation_label, n_translations=0, n_scales=0)
    save_augmented_data("./augmented_data/", augmented_validation_data, augmented_validation_label, "val")
else:
    augmented_train_data, augmented_train_label = load_augmented_data("./augmented_data/", "train")
    augmented_validation_data, augmented_validation_label = load_augmented_data("./augmented_data/", "val")
    augmented_test_data, augmented_test_label = load_augmented_data("./augmented_data/", "test")

# In case dataset wouldbe bigger and doesnt fit in memory, this is for using a data generator:
# partition_dict, labels_dict = get_generator_dicts("./augmented_data/")

# Note: Data is not normalized (only scaled in range [0,1]), as the dataset is to small, so that calculating mean and stdv
# would be biased and not estimate real mean and stdv of the distribution correctly, thereby leading to overfitting, as
# experiments also confirmed


###########################################################################################
############################### DOWNSAMPLING ##############################################
###########################################################################################

resize_factor = 0.6
n_sample_train = augmented_train_data.shape[0]
n_sample_validation = augmented_train_validation.shape[0]
n_rows_new, n_cols_new, n_channel = imresize(augmented_train_data[0], resize_factor).shape

augmented_train_data_resized = np.zeros((n_sample_train, n_rows_new, n_cols_new, 3))
augmented_validation_data_resized = np.zeros((n_sample_validation, n_cols_new, 3))

for i, im in enumerate(augmented_train_data):
    augmented_train_data_resized[i] = imresize(im, resize_factor)/255.0


for i, im in enumerate(augmented_validation_data):
    augmented_validation_data_resized[i] = imresize(im, resize_factor)/255.0

augmented_train_data = augmented_train_data_resized
augmented_validation_data = augmented_validation_data_resized


###########################################################################################
############################### TRAIN MODEL ###############################################
###########################################################################################

# Define some callbacks for early stopping and checkpoints, as well as tensorboard log files for analysis of the model
tbCallBack = TensorBoard(log_dir='./logs', histogram_freq=2, batch_size=32, write_graph=True, write_grads=True, write_images=True)
cpCallBack = ModelCheckpoint('./checkpoints/model.hdf5', monitor='val_loss', verbose=0, save_best_only=True, save_weights_only=False, mode='auto', period=1)
esCallBack = EarlyStopping(monitor='val_loss', min_delta=0, patience=6, verbose=0, mode='auto')

# Either load pretrained model to fine tune it, or learn it from scratch
if use_pretrained:
    model = load_model("./pretrained_model/model.hdf5")
else:
    input_img = Input(shape=(augmented_train_data[0].shape[0], augmented_train_data[0].shape[1], 3))

    # Note: No Batch Normalization, as it seemed to cause overfitting. Same reason as before.
    #(Alternative explanation: As some images look quite similiar, the data could get artificially stretched to much)
    x = Conv2D(8, (3, 3), padding='same')(input_img)
    x = LeakyReLU(alpha=0.3)(x)
    x = Conv2D(8, (3, 3), padding='same')(x)
    x = LeakyReLU(alpha=0.3)(x)
    x = Conv2D(1, (1, 1), padding='same')(x)
    x = LeakyReLU(alpha=0.3)(x)
    x = Flatten()(x)

    # Set bias to false in this layer. This prevents network from learning local minima: "always predict mean (x=y=0.5)"
    # by learning all weights to become 0 except bias weight to 0.5
    output = Dense(2, use_bias=False)(x)

    model = Model(inputs=input_img, outputs=output)
    model.compile(optimizer="adam", loss='mean_squared_error')


# fit model either with data generator, if data doesnt fit in memory,
# or just normally (which is done here, as the dataset is small):
#params = {'dim': (augmented_train_data[0].shape[0],augmented_train_data[0].shape[1]),
#          'batch_size': 32,
#          'n_channels': 3,
#          'shuffle': True}
#training_generator = DataGenerator("./augmented_data/", partition_dict['train'], labels_dict, **params)
#validation_generator = DataGenerator("./augmented_data/", partition_dict['val'], labels_dict, **params)

#history = model.fit_generator(generator=training_generator, validation_data=[augmented_validation_data, augmented_validation_label], epochs=30, callbacks=[cpCallBack, esCallBack])

history = model.fit(augmented_train_data, augmented_train_label, validation_data=[augmented_validation_data, augmented_validation_label] ,epochs=500, batch_size=32, callbacks=[tbCallBack, cpCallBack, esCallBack])