In [1]:
%matplotlib inline
import numpy as np
import pandas as pd
import matplotlib
import seaborn as sns
import matplotlib.pyplot as plt
import os
from os import getcwd
from os import listdir
from os import mkdir
from os.path import isfile, join, isdir
import skimage
from skimage import measure
from skimage import io
from PIL import Image
from sklearn.model_selection import train_test_split
from sklearn.metrics import log_loss
from sklearn.preprocessing import LabelEncoder
from skimage.transform import resize
# from keras.models import Sequential
# from keras.layers import Dropout, Flatten, Convolution2D, MaxPooling2D, ZeroPadding2D, Dense, Activation
# from keras.optimizers import RMSprop, Adam
# from keras.callbacks import EarlyStopping
# from keras.utils import np_utils
# from keras import backend as K
In [2]:
import tensorflow as tf
In [3]:
import tflearn
from tflearn.data_utils import shuffle
from tflearn.layers.core import input_data, dropout, fully_connected
from tflearn.layers.conv import conv_2d, max_pool_2d
from tflearn.layers.estimator import regression
from tflearn.data_preprocessing import ImagePreprocessing
from tflearn.data_augmentation import ImageAugmentation
from tflearn.metrics import Accuracy
In [4]:
def get_paths(foldNames):
paths = dict.fromkeys(foldNames)
for idx,g in enumerate(foldNames):
fileNames = [f for f in listdir(join(trainPath,g)) if isfile(join(trainPath,g, f))]
for i,f in enumerate(fileNames):
fileNames[i] = join(trainPath,g,f)
paths[g] = fileNames
return paths
In [5]:
ROWS = 90 #90 720
COLS = 160 #160 1280
CHANNELS = 3
trainPath = '../train'
testPath = '../test_stg1'
fish_classes = [f for f in listdir(trainPath) if isdir(join(trainPath, f))]
groupData = pd.DataFrame ({'group': fish_classes})
fish_paths = get_paths(fish_classes)
numFish = len(fish_classes)
In [6]:
for idx,fish in enumerate(fish_classes):
groupData.ix[idx,'num files'] = int(len(fish_paths[fish]))
files = []
Y_cat = []
for fish in fish_classes:
fish_files = fish_paths[fish]
files.extend(fish_files)
y_fish = np.tile(fish, len(fish_files))
Y_cat.extend(y_fish)
Y_cat = np.array(Y_cat)
In [ ]:
# One Hot Encoding Labels
# Transform the categorical array Y_all into matrix of the same height,
# but with a boolean column for each category.
Y_all = LabelEncoder().fit_transform(Y_cat)
# Y_all = np_utils.to_categorical(Y_all)
Y_all = tflearn.data_utils.to_categorical(Y_all,len(fish_classes))
# test_size: between 0 and 1. proportion of the dataset to include in the test split
# random_state: Pseudo-random number generator state used for random sampling. How to shoose this?
# stratify: this is ensuring that the split datasets are balanced, i.e. contains the same
# percentage of classes
X_train, X_valid, Y_train, Y_valid = train_test_split(X_all, Y_all,
test_size=0.2, random_state=23,
stratify=Y_all)
In [ ]:
def dnn_test1():
#needed to run this tensorflow operation in order to build the network and subsequently
#create the model, multiple times. Rebuilding without resetting the tf.Graph object produces
#errors. Could also get around this issue by restarting kernel, but that's annoying.
#python with statement: it will use this session for the code following the with statement
# and then automatically close the session
with tf.Graph().as_default():
# n_nodes_hl1 = 500
# n_nodes_hl2 = 500
# n_nodes_hl3 = 500
# n_classes = 10
# batch_size = 100
# normalisation of images
img_prep = ImagePreprocessing()
img_prep.add_featurewise_zero_center()
img_prep.add_featurewise_stdnorm()
# Create extra synthetic training data by flipping & rotating images
img_aug = ImageAugmentation()
img_aug.add_random_flip_leftright()
img_aug.add_random_rotation(max_angle=25.)
#specific a specific device or gpu
# with tf.device("/gpu:1"):
#input layer with shape of data specified. In this case, dimensions of our images,
#rows X cols X rgb array. The initial 'None' is for an unknown dimension reflecting the
#"number of samples that are processed in a batch"
network = input_data(shape=[None, ROWS, COLS, 3],
data_preprocessing=img_prep,
data_augmentation=img_aug)
# 1: Convolution layer with 32 filters, each 3x3x3
conv_1 = conv_2d(network, 32, 3, activation='relu', name='conv_1')
# 2: Max pooling layer
network = max_pool_2d(conv_1, 2)
# 3: Convolution layer with 64 filters
conv_2 = conv_2d(network, 64, 3, activation='relu', name='conv_2')
# 4: Convolution layer with 64 filters
conv_3 = conv_2d(conv_2, 64, 3, activation='relu', name='conv_3')
# 5: Max pooling layer
network = max_pool_2d(conv_3, 2)
# 6: Fully-connected 512 node layer
network = fully_connected(network, 512, activation='relu')
# 7: Dropout layer to combat overfitting
network = dropout(network, 0.5)
#output latyer
network = fully_connected(network, 8, activation='softmax')
# Configure how the network will be trained
acc = Accuracy(name="Accuracy")
network = regression(network, optimizer='adam',
loss='categorical_crossentropy',
learning_rate=0.0005, metric=acc)
return tflearn.DNN(network)
In [ ]:
# Define
model_tf = dnn_test1()
# Start training (apply gradient descent algorithm). Will want to specify multiple epochs
# typically unless just testing
model_tf.fit(X_train, Y_train, n_epoch=1,validation_set=(X_valid,Y_valid),
show_metric=True, batch_size=15)
In [7]:
""" Finetuning Example. Using weights from model trained in
convnet_cifar10.py to retrain network for a new task (your own dataset).
All weights are restored except last layer (softmax) that will be retrained
to match the new task (finetuning).
"""
from __future__ import division, print_function, absolute_import
import tflearn
from tflearn.layers.core import input_data, dropout, fully_connected
from tflearn.layers.conv import conv_2d, max_pool_2d
from tflearn.layers.estimator import regression
# Data loading
# Note: You input here any dataset you would like to finetune
X = X_train
Y = Y_train
num_classes = numFish
# Redefinition of convnet_cifar10 network
network = input_data(shape=[None, 32, 32, 3])
network = conv_2d(network, 32, 3, activation='relu')
network = max_pool_2d(network, 2)
network = dropout(network, 0.75)
network = conv_2d(network, 64, 3, activation='relu')
network = conv_2d(network, 64, 3, activation='relu')
network = max_pool_2d(network, 2)
network = dropout(network, 0.5)
network = fully_connected(network, 512, activation='relu')
network = dropout(network, 0.5)
# Finetuning Softmax layer (Setting restore=False to not restore its weights)
softmax = fully_connected(network, num_classes, activation='softmax', restore=False)
regression = regression(softmax, optimizer='adam',
loss='categorical_crossentropy',
learning_rate=0.001)
model = tflearn.DNN(regression, checkpoint_path='model_finetuning',
max_checkpoints=3, tensorboard_verbose=0)
# Load pre-existing model, restoring all weights, except softmax layer ones
model.load('cifar10_cnn')
In [ ]:
# Start finetuning
model.fit(X, Y, n_epoch=1, validation_set=0.1, shuffle=True,
show_metric=True, batch_size=4, snapshot_step=200,
snapshot_epoch=False, run_id='model_finetuning')
model.save('model_finetuning')
In [ ]:
BUILD_HDF5_DATASET = False
IMAGE_SIZE = 128
VALIDATION_SPLIT = True
output_path = join(rawdataPath, 'fish_dataset_{}x{}.h5'.format(IMAGE_SIZE, IMAGE_SIZE))
input_path = join(rawdataPath, 'train')
if BUILD_HDF5_DATASET:
# Build a HDF5 dataset (only required once)
from tflearn.data_utils import build_hdf5_image_dataset
build_hdf5_image_dataset(target_path =input_path,
image_shape =(IMAGE_SIZE, IMAGE_SIZE),
mode ='folder',
output_path =output_path,
categorical_labels =True,
normalize =True)
In [ ]:
%%time
# Load HDF5 dataset
import h5py
h5f = h5py.File(output_path, 'r')
X_all = h5f['X'][()]
Y_all = h5f['Y'][()]
# Split into
if VALIDATION_SPLIT:
X_train, X_valid, Y_train, Y_valid = train_test_split(X_all, Y_all,
test_size =0.2,
random_state =23,
stratify =Y_all)