def read_attributes(path):
print("Reading execution attributes from ", path)
attr = ExecutionAttribute()
lines = [line.rstrip('\n') for line in open(path)]
attr.seq = int(lines[0].split("=", 1)[1])
attr.img_width = int(lines[1].split("=", 1)[1])
attr.img_height = int(lines[2].split("=", 1)[1])
attr.path = lines[3].split("=", 1)[1]
attr.summ_basename = lines[4].split("=", 1)[1]
attr.epochs = int(lines[5].split("=", 1)[1])
attr.batch_size = int(lines[6].split("=", 1)[1])
attr.train_data_dir = lines[7].split("=", 1)[1]
attr.validation_data_dir = lines[8].split("=", 1)[1]
attr.test_data_dir = lines[9].split("=", 1)[1]
attr.steps_train = int(lines[10].split("=", 1)[1])
attr.steps_valid = int(lines[11].split("=", 1)[1])
attr.steps_test = int(lines[12].split("=", 1)[1])
attr.architecture = lines[13].split("=", 1)[1]
attr.curr_basename = lines[14].split("=", 1)[1]
return attr
SUMMARY_PATH = "/mnt/data/results"
# SUMMARY_PATH="c:/temp/results"
# SUMMARY_PATH="/tmp/results"
NETWORK_FORMAT = "Unimodal"
IMAGE_FORMAT = "2D"
SUMMARY_BASEPATH = create_results_dir(SUMMARY_PATH, NETWORK_FORMAT, IMAGE_FORMAT)
# how many times to execute the training/validation/test cycle
CYCLES = 20
#
# Execution Attributes
attr = ExecutionAttribute()
# dimensions of our images.
attr.img_width, attr.img_height = 96, 96
# network parameters
# attr.path='C:/Users/hp/Downloads/cars_train'
# attr.path='/home/amenegotto/dataset/2d/sem_pre_proc_mini/
attr.path = '/mnt/data/image/2d/com_pre_proc/'
attr.summ_basename = get_base_name(SUMMARY_BASEPATH)
attr.s3_path = NETWORK_FORMAT + '/' + IMAGE_FORMAT
attr.epochs = 100
attr.batch_size = 128
attr.set_dir_names()
if K.image_data_format() == 'channels_first':
input_s = (3, attr.img_width, attr.img_height)
else:
input_s = (attr.img_width, attr.img_height, 3)
attr.path = '/mnt/data/image/2d/sem_pre_proc'
attr.set_dir_names()
attr.batch_size = 128 # try 4, 8, 16, 32, 64, 128, 256 dependent on CPU/GPU memory capacity (powers of 2 values).
attr.epochs = 500
# how many times to execute the training/validation/test cycle
CYCLES = 1
for i in range(0, CYCLES):
# create the base pre-trained model
base_model = InceptionV3(weights='imagenet', include_top=False)
# dimensions of our images.
# Inception input size
attr.img_width, attr.img_height = 299, 299
# add a global spatial average pooling layer
x = base_model.output
x = GlobalAveragePooling2D()(x)
# let's add a fully-connected layer
x = Dense(1024, activation='relu')(x)
drop = Dropout(0.20)(x)
# and a logistic layer -- we have 2 classes
predictions = Dense(2, activation='softmax')(drop)
# this is the model we will train
attr.model = Model(inputs=base_model.input, outputs=predictions)
# first: train only the top layers (which were randomly initialized)
attr = ExecutionAttribute()
attr.architecture = 'vgg19'
attr.csv_path = 'csv/clinical_data.csv'
attr.s3_path = NETWORK_FORMAT + '/' + IMAGE_FORMAT
attr.numpy_path = '/mnt/data/image/2d/numpy/' + IMG_TYPE
# attr.numpy_path = '/home/amenegotto/dataset/2d/numpy/' + IMG_TYPE
attr.path = '/mnt/data/image/2d/' + IMG_TYPE
results_path = create_results_dir(SUMMARY_BASEPATH, 'fine-tuning',
attr.architecture)
attr.summ_basename = get_base_name(results_path)
attr.set_dir_names()
attr.batch_size = 128
attr.epochs = 500
attr.img_width = 224
attr.img_height = 224
input_attributes_s = (20, )
# how many times to execute the training/validation/test cycle
CYCLES = 1
for i in range(0, CYCLES):
#Load the VGG model
vgg_conv = VGG19(weights='imagenet',
include_top=False,
input_shape=(attr.img_width, attr.img_height, 3))
# Freeze the layers except the last 4 layers
from keras.optimizers import RMSprop, Adam
from ExecutionAttributes import ExecutionAttribute
import cv2
import os
import pandas as pd
import numpy as np
from sklearn.metrics import classification_report, confusion_matrix, cohen_kappa_score, roc_auc_score, roc_curve
from keras.preprocessing.image import load_img, img_to_array
os.environ["PATH"] += os.pathsep + 'C:/Program Files (x86)/Graphviz2.38/bin'
# Execution Attributes
attr = ExecutionAttribute()
# dimensions of our images.
attr.img_width, attr.img_height = 32, 32
# network parameters
attr.path = 'C:/Users/hp/Downloads/cars_train'
attr.epochs = 200
attr.batch_size = 8
attr.set_dir_names()
if K.image_data_format() == 'channels_first':
input_s = (3, attr.img_width, attr.img_height)
else:
input_s = (attr.img_width, attr.img_height, 3)
def load_data(filepath):
files = pd.read_csv(filepath)
# Execution Attributes
INITIAL_EPOCH=1
attr = ExecutionAttribute()
attr.architecture = 'Xception'
results_path = create_results_dir('/mnt/data', 'fine-tuning', attr.architecture)
attr.summ_basename = get_base_name(results_path)
attr.path='/mnt/data/image/2d/com_pre_proc'
attr.set_dir_names()
attr.batch_size = 64 # try 4, 8, 16, 32, 64, 128, 256 dependent on CPU/GPU memory capacity (powers of 2 values).
attr.epochs = 15
# hyper parameters for model
nb_classes = 2 # number of classes
based_model_last_block_layer_number = 126 # value is based on based model selected.
attr.img_width, attr.img_height = 299, 299 # change based on the shape/structure of your images
learn_rate = 1e-4 # sgd learning rate
momentum = .9 # sgd momentum to avoid local minimum
transformation_ratio = .05 # how aggressive will be the data augmentation/transformation
if K.image_data_format() == 'channels_first':
input_s = (1, attr.img_width, attr.img_height)
else:
input_s = (attr.img_width, attr.img_height, 1)
# define model
#attr.model = load_model(attr.summ_basename + '-mid-ckweights.h5')
callbacks = [EarlyStopping(monitor='val_loss', patience=5, mode='min', restore_best_weights=True),
ModelCheckpoint(attr.summ_basename + "-ckweights.h5", mode='min', verbose=1, monitor='val_loss', save_best_only=True)]