In [6]:
import sys
sys.path.append('..')
import numpy as np
import tensorflow as tf
from keras.callbacks import *
from keras.models import *
from keras.layers import *
from keras.optimizers import *
from keras.utils.np_utils import to_categorical
from keras.utils import plot_model
import keras.backend as K
from PIL import Image
from matplotlib.pyplot import imshow
%matplotlib inline
import random
from LineFilterLayer import LineFilterLayer
modelPath = 'model/BasicStrategyTwo.h5'
In [2]:
print(K.image_data_format())
# expected output: channels_last
In [3]:
def dotsAndBoxesToCategorical(inputData):
inp = np.copy(inputData)
inp[inp == 255] = 1 # Line - comes first so that target data only has two categories
inp[inp == 65] = 2 # Box A
inp[inp == 150] = 3 # Box B
inp[inp == 215] = 4 # Dot
cat = to_categorical(inp)
newShape = inp.shape + (cat.shape[-1],)
return cat.reshape(newShape)
In [4]:
firstTryDataset = np.load('basicStrategy5x4.npz')
x_train = firstTryDataset['x_train']
y_train = firstTryDataset['y_train']
x_train_cat = dotsAndBoxesToCategorical(x_train)
y_train_cat = y_train[:,LineFilterLayer.lineFilterMatrix(x_train.shape[-1], x_train.shape[-2], True)]
np.set_printoptions(precision=2)
print("original data:")
print(x_train[0])
print(y_train[0])
print(x_train.shape)
print(y_train.shape)
print("\nnormalized data:")
print(np.transpose(x_train_cat[0]))
print(np.transpose(y_train_cat[0]))
print(x_train_cat.shape)
print(y_train_cat.shape)
In [9]:
# LineFilterLayer has to be set before loading the model
LineFilterLayer.imgWidth = x_train.shape[-1]
LineFilterLayer.imgHeight = x_train.shape[-2]
model = load_model(modelPath, custom_objects={'LineFilterLayer':LineFilterLayer})
#model.save(modelPath)
In [5]:
img_input = Input(shape=(None,None,5,))
kernelSize = (5,5)
x = Conv2D(64, kernelSize, padding='same', activation='relu')(img_input)
x = BatchNormalization()(x)
x = Conv2D(64, kernelSize, padding='same', activation='relu')(x)
x = BatchNormalization()(x)
x = Conv2D(64, kernelSize, padding='same', activation='relu')(x)
x = BatchNormalization()(x)
x = Conv2D(64, kernelSize, padding='same', activation='relu')(x)
x = BatchNormalization()(x)
x = Conv2D(64, kernelSize, padding='same', activation='relu')(x)
x = BatchNormalization()(x)
x = Conv2D(64, kernelSize, padding='same', activation='relu')(x)
x = BatchNormalization()(x)
x = Conv2D(1, kernelSize, padding='same')(x)
x = LineFilterLayer(x_train.shape[-1], x_train.shape[-2])(x)
x = Activation('sigmoid', name='sigmoid')(x)
model = Model(inputs=img_input, outputs=x)
model.compile(optimizer='adam', loss='binary_crossentropy')
model.summary()
In [7]:
# Training
callbacks = []
checkpoint = ModelCheckpoint(filepath='model/BasicStrategyTwoCheckpoint.h5', save_weights_only=False)
callbacks.append(checkpoint)
progbar = ProgbarLogger()
callbacks.append(progbar)
tensorboard = TensorBoard(log_dir='model/log2', write_grads=True, write_graph=True, write_images=True, histogram_freq=1)
callbacks.append(tensorboard)
model.fit(x_train_cat, y_train_cat, epochs=10, batch_size=64, callbacks=callbacks, validation_split=0.001)
model.save(modelPath)
In [42]:
example = random.randrange(x_train.shape[0])
print("example: "+str(example))
input_data = x_train[example:example+1]
input_data_cat = x_train_cat[example:example+1]
prediction = model.predict(input_data_cat)
# print input data
input_data_print = x_train[example,:,:]
input_data_print = input_data_print.astype(np.uint8)
print("input: ")
print(input_data_print)
# generate greyscale image data from input data
target_imgdata = x_train[example,:,:]
target_imgdata = target_imgdata.astype(np.uint8)
# print prediction
prediction_data_print = prediction[0,:,:,1] * 100
prediction_data_print = prediction_data_print.astype(np.uint8)
print("prediction: ")
print(prediction_data_print)
# generate greyscale image data from prediction data
prediction_imgdata = prediction[0] * 255
prediction_imgdata = prediction_imgdata.astype(np.uint8)
# merge image data in color channels
tmp = np.zeros((prediction[0].shape[0], prediction[0].shape[1]), dtype=np.uint8)
merged_imgdata = np.stack([target_imgdata, prediction_imgdata[:,:,1], tmp], axis=2)
#create image
img = Image.fromarray(merged_imgdata, 'RGB')
img = img.resize(size=(img.size[0]*10, img.size[1]*10))
img
Out[42]:
In [ ]: