diff --git a/exp/kerokero/prepare_data.py b/exp/kerokero/prepare_data.py --- a/exp/kerokero/prepare_data.py +++ b/exp/kerokero/prepare_data.py @@ -20,7 +20,7 @@ class Sample: SIDE=224 def __init__(self,img,grid): - """:param img: a greyscale image as a 2D np.uint8 + """:param img: an image as a 3D np.uint8, channels-last :param grid: iterable of 4 EPoints, ie. Corners""" self.img=img self.grid=grid @@ -38,7 +38,7 @@ class Sample: 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)/255 + img=cv.cvtColor(img,cv.COLOR_BGR2RGB) 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))) @@ -59,7 +59,7 @@ class Sample: return np.matmul(scale,t2) def show(self): - img=cv.cvtColor(self.img,cv.COLOR_GRAY2BGR) + img=np.copy(self.img) for c in self.grid: cv.circle(img,(int(c.x),int(c.y)),3,[0,255,0],-1) show(img) @@ -80,15 +80,15 @@ 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) + files=sorted(filter(imgFilter,os.scandir(path)),key=lambda f: f.name)[::3] 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() - # Sample(np.uint8(transformedImg*255),map(lambda c: (c+EPoint(1,1))*Sample.SIDE/2,transformedGrid)).show() + # Sample(transformedImg,map(lambda c: (c+EPoint(1,1))*Sample.SIDE/2,transformedGrid)).show() yield (transformedImg,label) @@ -108,8 +108,8 @@ def loadDataset(root): labels=[labels[k] for k in keys] m=int(n*trainRatio) return ( - (np.float32(images[:m]),np.float32(labels[:m])), - (np.float32(images[m:]),np.float32(labels[m:])) + (images[:m],np.float32(labels[:m])), + (images[m:],np.float32(labels[m:])) ) diff --git a/exp/kerokero/test.py b/exp/kerokero/test.py --- a/exp/kerokero/test.py +++ b/exp/kerokero/test.py @@ -3,6 +3,7 @@ import logging as log import numpy as np from keras.models import load_model +from keras.applications.inception_v3 import preprocess_input from prepare_data import loadDataset,Sample from analyzer.epoint import EPoint @@ -16,23 +17,22 @@ 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: - trainImages=data["trainImages"] - trainLabels=data["trainLabels"] testImages=data["testImages"] testLabels=data["testLabels"] log.info("done") -log.info(model.evaluate(testImages.reshape((-1,224,224,1)),testLabels)) +log.info(model.evaluate(preprocess_input(testImages).reshape((-1,224,224,3)),testLabels)) for img in testImages: - label=model.predict(np.reshape(img,(1,224,224,1))) + label=model.predict(preprocess_input(np.reshape(img,(1,224,224,3)))) print(label) points=[] for i in range(4): points.append(EPoint((label[0][i*2]+1)*112,(label[0][i*2+1]+1)*112)) corners=Corners(points) - sample=Sample(np.uint8(img*255),corners) + sample=Sample(img,corners) sample.show() diff --git a/exp/kerokero/train.py b/exp/kerokero/train.py --- a/exp/kerokero/train.py +++ b/exp/kerokero/train.py @@ -4,9 +4,11 @@ import argparse import logging as log import numpy as np -from keras.layers import Conv2D,Dropout,Dense,Flatten,MaxPooling2D,BatchNormalization -from keras.models import Sequential,load_model +from keras.layers import Conv2D,Dropout,Dense,Flatten,MaxPooling2D,GlobalAveragePooling2D,BatchNormalization +from keras.models import Sequential,load_model,Model +from keras.optimizers import SGD from keras.callbacks import TensorBoard +from keras.applications.inception_v3 import InceptionV3,preprocess_input import config as cfg import ftp @@ -17,6 +19,7 @@ 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() @@ -72,24 +75,50 @@ def createCNN(): return model -model=createCNN() +def createPretrained(): + base=InceptionV3(weights="imagenet",include_top=False,input_shape=(224,224,3)) + + x=base.output + x=GlobalAveragePooling2D()(x) + x=Dense(1024,activation="relu")(x) + predictions=Dense(8)(x) + + model=Model(inputs=base.input,outputs=predictions) + for layer in base.layers: + layer.trainable=False + + model.compile(optimizer='adam',loss='mse',metrics=['mae','accuracy']) + return model + + if args.load_model: model=load_model(args.load_model) +else: + model=createPretrained() log.info("loading data...") with np.load(args.data) as data: - trainImages=data["trainImages"] + trainImages=preprocess_input(data["trainImages"]) trainLabels=data["trainLabels"] - testImages=data["testImages"] + testImages=preprocess_input(data["testImages"]) testLabels=data["testLabels"] log.info("done") -tensorboard = TensorBoard(log_dir=os.path.join(cfg.thisDir,"../logs","{}".format(time()))) -BIG_STEP=50 +tensorboard = TensorBoard(log_dir=os.path.join(args.log_dir,"{}".format(time()))) + +if not args.load_model: + model.fit(trainImages.reshape((-1,224,224,3)),trainLabels,epochs=10,batch_size=128,validation_split=0.2,callbacks=[tensorboard]) +for layer in model.layers[:249]: + layer.trainable = False +for layer in model.layers[249:]: + layer.trainable = True +model.compile(optimizer=SGD(lr=0.0001,momentum=0.9),loss='mse') + +BIG_STEP=20 for i in range(args.initial_epoch//BIG_STEP,args.epochs//BIG_STEP): - model.fit(trainImages.reshape((-1,224,224,1)),trainLabels,epochs=(i+1)*BIG_STEP,initial_epoch=i*BIG_STEP,batch_size=128,validation_split=0.2,callbacks=[tensorboard]) + model.fit(trainImages.reshape((-1,224,224,3)),trainLabels,epochs=(i+1)*BIG_STEP,initial_epoch=i*BIG_STEP,batch_size=128,validation_split=0.2,callbacks=[tensorboard]) path=args.save_model.format((i+1)*BIG_STEP) log.info("saving model...") model.save(path) # ftp.push(path) -log.info(model.evaluate(testImages.reshape((-1,224,224,1)),testLabels)) +log.info(model.evaluate(testImages.reshape((-1,224,224,3)),testLabels))