traingridmodel_light.py

import numpy as np
from keras import optimizers
from keras.preprocessing.image import ImageDataGenerator
from keras.models import Sequential
from keras.layers import Dense, Conv2D, MaxPooling2D, Dropout, Flatten, Reshape, LeakyReLU, BatchNormalization
from keras import regularizers
from keras.callbacks import ModelCheckpoint

from jugglingdataloader import JugglingDataLoader
from utils import handleTensorflowSession
from losses import grid_loss_with_hands


handleTensorflowSession(memoryLimit=0.6)


def trainModel():
    augFactor = 1
    imageGenerator = ImageDataGenerator(horizontal_flip=True, width_shift_range=0.1*augFactor, height_shift_range=0.1*augFactor, zoom_range=0.15*augFactor) #zoom_range was 0.15
    jugglingDataLoader = JugglingDataLoader(shape=(64,64), gridShape=(15,15), batch_size=8, expressFactor=1.0, imageGenerator=imageGenerator, dataType='SUBMOVAVG')

    valx, valy = jugglingDataLoader.getValidationSet()

    nRows, nCols, nDims = valx.shape[1:]
    w = 8
    l2_conv = 0 #0.00000001

    model = Sequential()
    model.add(Conv2D(w, (3,3), padding='same', kernel_regularizer=regularizers.l2(l2_conv), input_shape=(nRows, nCols, nDims)))
    model.add(LeakyReLU())
    model.add(Conv2D(w, (3,3), padding='same', kernel_regularizer=regularizers.l2(l2_conv)))
    model.add(LeakyReLU())
    model.add(MaxPooling2D())

    model.add(Conv2D(w*2, (3,3), padding='same', kernel_regularizer=regularizers.l2(l2_conv)))
    model.add(LeakyReLU())
    model.add(Conv2D(w*2, (3,3), padding='same', kernel_regularizer=regularizers.l2(l2_conv)))
    model.add(LeakyReLU())
    model.add(MaxPooling2D())

    model.add(Conv2D(w*4, (3,3), padding='same', kernel_regularizer=regularizers.l2(l2_conv)))
    model.add(LeakyReLU())
    model.add(Conv2D(w*4, (3,3), padding='same', kernel_regularizer=regularizers.l2(l2_conv)))
    model.add(LeakyReLU())
    model.add(MaxPooling2D())

    l2_dense = 0 # 0.000001 * 5

    model.add(Flatten())
    # model.add(Dropout(0.5))
    model.add(Dense(w*64, kernel_regularizer=regularizers.l2(l2_dense)))
    model.add(LeakyReLU())
    model.add(Dense(15*15*9, activation='sigmoid'))
    model.add(Reshape((15,15,9)))

    model.summary()

    model.compile(optimizer='Adadelta', loss=grid_loss_with_hands)
    checkpoint = ModelCheckpoint('grid_model_submovavg_64x64_light.h5', verbose=1, save_best_only=True, period=1)

    model.fit_generator(
        jugglingDataLoader,
        workers=4,
        use_multiprocessing=True,
        epochs=300,
        callbacks=[checkpoint],
        validation_data=(valx, valy)
        )


trainModel()