def Load(loc, wordEmb, emb_size=50):
res = []
table = str.maketrans(dict.fromkeys(string.punctuation))
f = open(loc, 'r')
exceptions = 0
while True:
line = f.readline()
if not line:
break
if len(line) < 1:
continue
# Remove Punctuations
# Should I handle with 'd -> ed???
line = line.replace("'d", "ed").translate(table).lower().rstrip()
line = line.split(" ")
for elem in line:
if elem == "":
continue
try:
res.append(wordEmb[elem])
except:
# logger.PrintDebug("Word not found! %s"%elem,col="r")
res.append(np.zeros(emb_size))
exceptions += 1
f.close()
r = np.array(res)
logger.PrintDebug("Excepted word count : %d" % exceptions, col="r")
logger.PrintDebug("Loaded word size : %s" % (str(r.shape)), col="g")
return r
def _glove6B(sz=50, logDiv=100000):
d = dict()
if sz not in [50, 100, 200, 300]:
logger.PrintDebug(
"Not supported word embedding size, returning empty dict")
return d
path = "data/glove6B/glove.6B." + str(sz) + "d.txt"
logger.PrintDebug("Reading embedded word file from" + path)
f = open(path, 'r', encoding="utf-8")
while True:
line = f.readline()
if not line: break
line = line.replace("\n", "").split(" ")
if len(line) != sz + 1:
break
l = np.zeros(sz)
for i in range(sz):
l[i] = float(line[i + 1])
d[line[0]] = l
# Print log
if logDiv > 0 and len(d) % logDiv == 0:
logger.PrintDebug(" Data Count = " + str(len(d)))
f.close()
logger.PrintDebug("Loaded " + str(len(d)) + " embedded word data")
return d
def LoadRawData(loc):
logger.PrintDebug("Loading data from " + loc)
with gzip.open(loc, 'rb') as f:
raw = f.read()
if int.from_bytes(raw[:4], byteorder='big') == 2051:
# This is image file
nImages = int.from_bytes(raw[4:8], byteorder='big')
rows = int.from_bytes(raw[8:12], byteorder='big')
cols = int.from_bytes(raw[12:16], byteorder='big')
res = np.zeros((nImages, 1, rows, cols))
for imageIdx in range(nImages):
for rowIdx in range(rows):
for colIdx in range(cols):
res[imageIdx, 0, rowIdx,
colIdx] = int(raw[16 + imageIdx * rows * cols +
rowIdx * cols + colIdx]) / 255.0
if imageIdx % 10000 == 9999 and imageIdx + 1 != nImages:
logger.PrintDebug(" " + str(imageIdx + 1) +
" Images loaded",
end='\r')
logger.PrintDebug("Total " + str(nImages) + " Images loaded")
elif int.from_bytes(raw[:4], byteorder='big') == 2049:
# This is label file
nImages = int.from_bytes(raw[4:8], byteorder='big')
res = np.zeros((nImages), dtype=np.int32)
for imageIdx in range(nImages):
res[imageIdx] = int(raw[8 + imageIdx])
logger.PrintDebug("Total " + str(nImages) + " Labels loaded")
else:
logger.PrintDebug("File is not a MNIST image or label file.",
col="r")
return -1
return res
def DownloadGlove6B():
# Create directory if not exists
if not os.path.exists("./data/"):
os.mkdir("./data/")
logger.PrintDebug("Downloading glove6B...")
p = "./data/glove6B.zip"
request.urlretrieve(" http://nlp.stanford.edu/data/glove.6B.zip", p)
logger.PrintDebug("Unpacking glove6B...")
with zipfile.ZipFile(p, "r") as zr:
zr.extractall("data/glove6B/")
os.remove(p)
def LoadData(loc, rawloc):
if not os.path.exists(loc + ".npy"):
res = LoadRawData(rawloc)
np.save(loc, res)
else:
res = np.load(loc + ".npy")
logger.PrintDebug("Loaded " + loc + ".npy")
return res
def glove6B(sz=50, logDiv=100000):
loc = "data/glove6B/glove6B%s" % str(sz) + ".dump"
if not os.path.exists(loc):
res = _glove6B(sz, logDiv)
pickle.dump(res, open(loc, 'wb'))
else:
res = pickle.load(open(loc, 'rb'))
logger.PrintDebug("Loaded " + loc)
return res
def Forward(self, input_image, input_word):
for i in range(len(self.cnn)):
input_image = self.cnn[i].Forward(input_image)
if i == 10:
input_image_c4 = input_image
logger.PrintDebug("CNN " + str(i) + " " + str(input_image.shape))
input_image = (input_image_c4, input_image)
for i in range(len(self.concat)):
input_image = self.concat[i].Forward(input_image)
logger.PrintDebug("Concat " + str(i) + " " +
str(input_image.shape))
for i in range(len(self.emb)):
input_word = self.emb[i].Forward(input_word)
logger.PrintDebug("Embedding " + str(i) + " " +
str(input_word.shape))
input_linear = (input_image, input_word)
for i in range(len(self.linear)):
input_linear = self.linear[i].Forward(input_linear)
logger.PrintDebug("Linear " + str(i) + " " +
str(input_linear.shape))
return input_linear
def Download():
# Delete existing file
logger.PrintDebug("Clearing Folder...")
top = ("./data/")
for root, dirs, files in os.walk(top, topdown=False):
for name in files:
os.remove(os.path.join(root, name))
for name in dirs:
os.rmdir(os.path.join(root, name))
# Create directory if not exists
if not os.path.exists("./data/"):
os.mkdir("./data/")
# Download Flickr8k
logger.PrintDebug("Downloading Flickr8k Text...")
p = "./data/Flickr8k_text.zip"
request.urlretrieve(
"https://github.com/jbrownlee/Datasets/releases/download/Flickr8k/Flickr8k_text.zip",
p)
logger.PrintDebug("Unpacking Flickr8k Text...")
with zipfile.ZipFile(p, "r") as zr:
zr.extractall("data/Flickr8k/")
os.remove(p)
logger.PrintDebug("Downloading Flickr8k Image...")
p = "./data/Flickr8k_Dataset.zip"
request.urlretrieve(
"https://github.com/jbrownlee/Datasets/releases/download/Flickr8k/Flickr8k_Dataset.zip",
p)
logger.PrintDebug("Unpacking Flickr8k Image...")
with zipfile.ZipFile(p, "r") as zr:
zr.extractall("data/Flickr8k/")
os.remove(p)
DownloadGlove6B()
def Flickr8k(mode="test", logDiv=2000, imageSize=[64, 64]):
data = list()
dataL = list()
# Load image list
imageList = list()
path = "data/Flickr8k/Flickr_8k." + mode + "Images.txt"
logger.PrintDebug("Loading " + mode + " image list from " + path)
f = open(path, 'r', encoding="utf-8")
while True:
line = f.readline().replace("\n", "")
if not line: break
imageList.append(line)
logger.PrintDebug("Total of " + str(len(imageList)) + " " + mode +
" image list loaded")
# Load token
tokens = dict()
path = "data/Flickr8k/Flickr8k.token.txt"
logger.PrintDebug("Loading tokens from " + path)
f = open(path, 'r', encoding="utf-8")
while True:
line = f.readline().replace("\n", "")
if not line: break
line = line.split("\t")
if line[0][:-2] not in tokens: # Add token
tokens[line[0][:-2]] = [line[1]]
else:
tokens[line[0][:-2]].append(line[1])
logger.PrintDebug(str(len(tokens)) + " image tokens are loaded")
# Load images and tokens
logger.PrintDebug("Loading " + mode + " images from " + path)
for ind, val in enumerate(imageList):
# Load image
data.append(
np.asarray(
Image.open('data/Flickr8k/Flicker8k_Dataset/' + val).resize(
(imageSize[0], imageSize[1]))))
# Load tag
dataL.append(tokens[val])
# Print log
if logDiv > 0 and (ind + 1) % logDiv == 0:
logger.PrintDebug(" Data Count = " + str(ind + 1))
logger.PrintDebug("Total of " + str(len(data)) + " " + mode +
" images loaded")
return data, dataL
continue
try:
res.append(wordEmb[elem])
except:
# logger.PrintDebug("Word not found! %s"%elem,col="r")
res.append(np.zeros(emb_size))
exceptions += 1
f.close()
r = np.array(res)
logger.PrintDebug("Excepted word count : %d" % exceptions, col="r")
logger.PrintDebug("Loaded word size : %s" % (str(r.shape)), col="g")
return r
if __name__ == "__main__":
logger.PrintDebug("Simple RNN Network Training", col='b')
# Parse args
args = args.parse()
# Numpy options
np.set_printoptions(precision=3)
np.set_printoptions(suppress=True)
# np.set_printoptions(threshold=sys.maxsize)
# Download data if not exists
locData = "./data_tinyshakespeare/"
url = "http://raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt"
locFile = "input.txt"
if not os.path.exists(locData):
os.mkdir(locData)
if not os.path.exists(locData + locFile):
return -1
return res
def LoadData(loc, rawloc):
if not os.path.exists(loc + ".npy"):
res = LoadRawData(rawloc)
np.save(loc, res)
else:
res = np.load(loc + ".npy")
logger.PrintDebug("Loaded " + loc + ".npy")
return res
if __name__ == "__main__":
logger.PrintDebug("Simple CNN Network Training", col='b')
# Parse args
args = args.parse()
# Numpy options
np.set_printoptions(precision=3)
np.set_printoptions(suppress=True)
# np.set_printoptions(threshold=sys.maxsize)
# Download data if not exists
locData = "./data_mnist/"
url = "http://yann.lecun.com/exdb/mnist/"
locTrainLabel = "train-labels-idx1-ubyte.gz"
locTrainData = "train-images-idx3-ubyte.gz"
locTestLabel = "t10k-labels-idx1-ubyte.gz"
locTestData = "t10k-images-idx3-ubyte.gz"
import args
import model
import load
import image
from log import logger
import numpy as np
from model import Model
if __name__ == "__main__":
logger.PrintDebug("Initiated")
# parse args
args = args.parse()
# Numpy set precision
np.set_printoptions(precision=3)
# Download Data if argument is checked or falied to verify
if args.download or not load.Verify():
load.Download()
dataTrain = []; dataTrainLabel = []
# load data to memory
# dataTrain, dataTrainLabel = load.Flickr8k(mode = "train")
dataTest, dataTestLabel = load.Flickr8k(mode = "test")
# Normalize images
for i in range(len(dataTest)):
dataTest[i] = image.Normalize(dataTest[i])
for i in range(len(dataTrain)):
dataTrain[i] = image.Normalize(dataTrain[i])
logger.PrintDebug("Image Normalized")
def Verify():
logger.PrintDebug("Verifing Data...")
logger.PrintDebug("Data Verified!")
return True
