Changeset - 006c6f1aab13
[Not reviewed]
default
0 3 1
Laman - 6 years ago 2019-05-11 16:29:29

Euclidean distance as a loss function
4 files changed with 39 insertions and 13 deletions:
0 comments (0 inline, 0 general)
exp/kerokero/k_util.py
Show inline comments
 
new file 100644
 
import math
 

	
 
import keras.backend as K
 

	
 
def singleUnorderedLoss(yTrue,yPred):
 
	d1=sum(math.sqrt(min((yTrue[i*2]-yPred[j*2])**2+(yTrue[i*2+1]-yPred[j*2+1])**2 for j in range(4))) for i in range(4))
 
	d2=sum(math.sqrt(min((yTrue[i*2]-yPred[j*2])**2+(yTrue[i*2+1]-yPred[j*2+1])**2 for i in range(4))) for j in range(4))
 
	return (d1+d2)/2
 

	
 

	
 
def averageDistance(yTrue,yPred):
 
	squares=K.square(yTrue-yPred)
 
	distances=K.sqrt(K.sum(squares,-1))
 
	return K.mean(distances,-1)
exp/kerokero/prepare_data.py
Show inline comments
 
@@ -32,25 +32,25 @@ class Sample:
 
		m=getIdentity()
 
		t1=getTranslation(-center.x,-center.y)
 
		proj=getProjection()
 
		rot=getRotation()
 
		mir=getMirroring()
 
		for mi in [t1,mir,proj,rot]:
 
			m=np.matmul(mi,m)
 
		m=np.matmul(self._computeCrop(m),m)
 
		img=cv.warpPerspective(self.img,m,(self.SIDE,self.SIDE))
 
		img=np.float32(img)/128-1
 
		grid=Corners(c.transform(m) for c in self.grid)
 
		grid=list(map(lambda p: 2*p/self.SIDE-EPoint(1,1), grid))
 
		return (img,grid,list(itertools.chain.from_iterable(grid)))
 
		return (img,grid)
 

	
 
	def _getCenter(self):
 
		(a,b,c,d)=self.grid
 
		p=Line.fromPoints(a,c)
 
		q=Line.fromPoints(b,d)
 
		return p.intersect(q)
 

	
 
	def _computeCrop(self,m):
 
		grid=Corners(c.transform(m) for c in self.grid)
 
		(x1,y1,x2,y2)=computeBoundingBox(grid)
 
		(wg,hg)=(x2-x1,y2-y1)
 
		(left,top,right,bottom)=[random.uniform(0.05,0.2) for i in range(4)]
 
@@ -79,25 +79,25 @@ def traverseDirs(root):
 

	
 
def harvestDir(path):
 
	annotations=DataFile(os.path.join(path,"annotations.json.gz"))
 
	imgFilter=lambda f: f.is_file() and re.match(r".*\.(jpg|jpeg|png|gif)$", f.name.lower())
 
	files=sorted(filter(imgFilter,os.scandir(path)),key=lambda f: f.name)
 
	boards=annotations["."]
 
	for f in files:
 
		img=cv.imread(f.path)
 
		img=cv.cvtColor(img,cv.COLOR_BGR2GRAY)
 
		for b in boards:
 
			sample=Sample(img,b.grid)
 
			# sample.show()
 
			(transformedImg,transformedGrid,label)=sample.transform()
 
			(transformedImg,label)=sample.transform()
 
			# Sample(np.uint8((transformedImg+1)*128),map(lambda c: (c+EPoint(1,1))*Sample.SIDE/2,transformedGrid)).show()
 
			yield (transformedImg,label)
 

	
 

	
 
def loadDataset(root):
 
	testRatio=0.1
 
	trainRatio=1-testRatio
 
	images=[]
 
	labels=[]
 
	for d in traverseDirs(root):
 
		for (img,label) in harvestDir(d):
 
			images.append(img)
exp/kerokero/test.py
Show inline comments
 
import argparse
 
import logging as log
 

	
 
import numpy as np
 
from keras.models import load_model
 
import keras.losses
 
import keras.metrics
 

	
 
from prepare_data import loadDataset,Sample
 
from analyzer.epoint import EPoint
 
from analyzer.corners import Corners
 
from k_util import averageDistance
 
import config as cfg
 

	
 
keras.losses.averageDistance=averageDistance
 
keras.metrics.averageDistance=averageDistance
 

	
 
parser=argparse.ArgumentParser()
 
parser.add_argument("model")
 
parser.add_argument("data")
 
args=parser.parse_args()
 

	
 
model=load_model(args.model)
 
model.summary()
 

	
 
log.info("loading data...")
 
with np.load(args.data) as data:
 
	testImages=data["testImages"]
 
	testLabels=data["testLabels"]
 
log.info("done")
 

	
 
log.info(model.evaluate(testImages.reshape((-1,224,224,1)),testLabels))
 
log.info(model.evaluate(testImages.reshape((-1,224,224,1)),testLabels.reshape((-1,4,2))))
 

	
 
for img in testImages:
 
	label=model.predict(np.reshape(img,(1,224,224,1)))
 
	print(label)
 
	points=[]
 
	for i in range(4):
 
		points.append(EPoint((label[0][i*2]+1)*112,(label[0][i*2+1]+1)*112))
 
		points.append(EPoint((label[0][i][0]+1)*112,(label[0][i][1]+1)*112))
 
	corners=Corners(points)
 
	sample=Sample(np.uint8((img+1)*128),corners)
 
	sample.show()
exp/kerokero/train.py
Show inline comments
 
import os
 
from time import time
 
import argparse
 
import logging as log
 

	
 
import numpy as np
 
from keras.layers import Conv2D,Dropout,Dense,Flatten,MaxPooling2D,BatchNormalization,GlobalAveragePooling2D
 
from keras.layers import Conv2D,Dropout,Dense,Flatten,MaxPooling2D,BatchNormalization,GlobalAveragePooling2D,Reshape
 
from keras.models import Sequential,load_model
 
from keras.callbacks import TensorBoard,ModelCheckpoint
 
import keras.losses
 
import keras.metrics
 

	
 
import config as cfg
 
from k_util import averageDistance
 

	
 
keras.losses.averageDistance=averageDistance
 
keras.metrics.averageDistance=averageDistance
 

	
 
parser=argparse.ArgumentParser()
 
parser.add_argument("data")
 
parser.add_argument("--load_model")
 
parser.add_argument("--save_model",default="/tmp/gogo-{0:03}.h5")
 
parser.add_argument("--epochs",type=int,default=100)
 
parser.add_argument("--initial_epoch",type=int,default=0)
 
parser.add_argument("--log_dir",default="/tmp/tflogs")
 
args=parser.parse_args()
 

	
 

	
 
def createFullyConnected():
 
@@ -32,62 +38,63 @@ def createFullyConnected():
 
	model.compile(
 
		optimizer='adam',
 
		loss='mse',
 
		metrics=['mae','accuracy']
 
	)
 
	return model
 

	
 
def createCNN():
 
	model=Sequential()
 

	
 
	model.add(BatchNormalization(input_shape=(224,224,1)))
 

	
 
	model.add(Conv2D(24,(5,5),border_mode="same",init="he_normal",activation="relu",input_shape=(224,224,1),dim_ordering="tf"))
 
	model.add(MaxPooling2D(pool_size=(2,2),strides=(2,2),border_mode="valid"))
 
	model.add(Conv2D(24,(5,5),padding="same",kernel_initializer="he_normal",activation="relu",input_shape=(224,224,1),data_format="channels_last"))
 
	model.add(MaxPooling2D(pool_size=(2,2),strides=(2,2),padding="valid"))
 

	
 
	model.add(Conv2D(36,(5,5),activation="relu"))
 
	model.add(MaxPooling2D(pool_size=(2,2),strides=(2,2),border_mode="valid"))
 
	model.add(MaxPooling2D(pool_size=(2,2),strides=(2,2),padding="valid"))
 

	
 
	model.add(Conv2D(48,(5,5),activation="relu"))
 
	model.add(MaxPooling2D(pool_size=(2,2),strides=(2,2),border_mode="valid"))
 
	model.add(MaxPooling2D(pool_size=(2,2),strides=(2,2),padding="valid"))
 

	
 
	model.add(Conv2D(64,(3,3),activation="relu"))
 
	model.add(MaxPooling2D(pool_size=(2,2),strides=(2,2),border_mode="valid"))
 
	model.add(MaxPooling2D(pool_size=(2,2),strides=(2,2),padding="valid"))
 

	
 
	model.add(Conv2D(64,(3,3),activation="relu"))
 

	
 
	model.add(GlobalAveragePooling2D())
 

	
 
	model.add(Dense(500,activation="relu"))
 
	model.add(Dense(90,activation="relu"))
 
	model.add(Dense(8))
 
	model.add(Reshape((4,2)))
 

	
 
	model.compile(optimizer="rmsprop",loss="mse",metrics=["mae","accuracy"])
 
	model.compile(optimizer="rmsprop",loss=averageDistance,metrics=["mae","accuracy"])
 
	return model
 

	
 

	
 
model=createCNN()
 
if args.load_model:
 
	model=load_model(args.load_model)
 

	
 
log.info("loading data...")
 
with np.load(args.data) as data:
 
	trainImages=data["trainImages"]
 
	trainLabels=data["trainLabels"]
 
	testImages=data["testImages"]
 
	testLabels=data["testLabels"]
 
log.info("done")
 

	
 
tensorboard=TensorBoard(log_dir=os.path.join(args.log_dir,"{}".format(time())))
 
checkpoint=ModelCheckpoint(args.save_model,monitor="val_loss",period=10)
 

	
 
model.fit(
 
	trainImages.reshape((-1,224,224,1)),
 
	trainLabels,
 
	trainLabels.reshape((-1,4,2)),
 
	epochs=args.epochs,
 
	initial_epoch=args.initial_epoch,
 
	batch_size=20,
 
	validation_split=0.2,
 
	callbacks=[tensorboard,checkpoint]
 
)
 

	
 
log.info(model.evaluate(testImages.reshape((-1,224,224,1)),testLabels))
 
log.info(model.evaluate(testImages.reshape((-1,224,224,1)),testLabels.reshape((-1,4,2))))
0 comments (0 inline, 0 general)