Skip to content
Snippets Groups Projects
Commit cb919de9 authored by maali's avatar maali
Browse files

Update the paths in the training part

parent 2b5ffe1c
No related branches found
No related tags found
No related merge requests found
Hide whitespace changes
Inline Side-by-side
Showing
with 40 additions and 80 deletions
models.py
+ 40
− 80
View file @ cb919de9
......@@ -5,6 +5,7 @@ from tensorflow.keras import backend as K
from classes import modelSet, modelDictVal
import random
from data import *
import argparse
huberDelta = .5
......@@ -237,7 +238,8 @@ def stvNetNew(inputShape=(480, 640, 3), outVectors=True, outClasses=True, modelN
return tf.keras.Model(inputs=xIn, outputs=outputs, name=modelName)
def uNet(inputShape=(480, 640, 3), outVectors=True, outClasses=True, modelName="uNet"): # neural net structure used for image segmentation
def uNet(inputShape=(480, 640, 3), outVectors=True, outClasses=True,
modelName="uNet"): # neural net structure used for image segmentation
xIn = tf.keras.Input(inputShape, dtype=np.dtype('uint8'))
x = tf.keras.layers.Lambda(lambda x: x / 255)(xIn)
......@@ -305,7 +307,7 @@ def uNet(inputShape=(480, 640, 3), outVectors=True, outClasses=True, modelName="
return tf.keras.Model(inputs=[xIn], outputs=outputs, name=modelName)
def trainModel(modelStruct, modelGen, modelClass='cat', batchSize=2, optimizer=tf.keras.optimizers.Adam,
def trainModel_Fruits(modelStruct, modelGen, modelClass='cat', batchSize=2, optimizer=tf.keras.optimizers.Adam,
learning_rate=0.01, losses=None, metrics=None, saveModel=True, modelName='stvNet_weights',
epochs=1, loss_weights=None, outVectors=False, outClasses=False, dataSplit=True, altLabels=True,
augmentation=True): # train and save model weights
......@@ -321,11 +323,12 @@ def trainModel(modelStruct, modelGen, modelClass='cat', batchSize=2, optimizer=t
trainData, validData = None, None
if dataSplit: # if using datasplit, otherwise all available data is used
trainData, validData = data.getDataSplit(modelClass=modelClass)
trainData, validData = data.getDataSplit_Fruits(modelClass=modelClass)
logger = tf.keras.callbacks.CSVLogger("models/history/" + modelName + "_" + modelClass + "_history.csv",
append=True)
# evalLogger = tf.keras.callbacks.CSVLogger("models/history/" + modelName + "_" + modelClass + "_eval_history.csv", append = True)
evalLogger = tf.keras.callbacks.CSVLogger("models/history/" + modelName + "_" + modelClass + "_eval_history.csv",
append=True)
history, valHistory = [], []
......@@ -336,14 +339,15 @@ def trainModel(modelStruct, modelGen, modelClass='cat', batchSize=2, optimizer=t
print("Epoch {0} of {1}".format(i + 1, epochs))
hist = model.fit(modelGen(modelClass, batchSize, masterList=trainData, out0=outKeys[0], out1=outKeys[1],
altLabels=altLabels, augmentation=augmentation),
steps_per_epoch=math.ceil(len(trainData) / batchSize), max_queue_size=2,
steps_per_epoch=math.ceil(len(trainData)), max_queue_size=2,
callbacks=[logger])
history.append(hist.history)
if dataSplit:
print("Validation:")
print("################ Validation: ############")
valHist = model.evaluate(
modelGen(modelClass, batchSize, masterList=validData, out0=outKeys[0], out1=outKeys[1],
altLabels=altLabels, augmentation=False), steps=math.ceil(len(validData) / batchSize),
modelGen(modelClass, batchSize, val=True, masterList=validData, out0=outKeys[0],
out1=outKeys[1],
altLabels=altLabels, augmentation=False), steps=math.ceil(len(validData)),
max_queue_size=2)
valHistory.append(valHist)
else:
......@@ -353,13 +357,14 @@ def trainModel(modelStruct, modelGen, modelClass='cat', batchSize=2, optimizer=t
print("Epoch {0} of {1}".format(i + 1, epochs))
hist = model.fit(
modelGen(modelClass, batchSize, masterList=trainData, altLabels=altLabels, augmentation=augmentation),
steps_per_epoch=math.ceil(len(trainData) / batchSize), max_queue_size=2, callbacks=[logger])
steps_per_epoch=math.ceil(len(trainData)), max_queue_size=2, callbacks=[logger])
history.append(hist.history)
if dataSplit:
print("Validation:")
print("########## Validation: ############2")
valHist = model.evaluate(
modelGen(modelClass, batchSize, masterList=validData, altLabels=altLabels, augmentation=False),
steps=math.ceil(len(validData) / batchSize), max_queue_size=2)
modelGen(modelClass, batchSize, val=True, masterList=validData, altLabels=altLabels,
augmentation=False),
steps=math.ceil(len(validData)), max_queue_size=2)
valHistory.append(valHist)
historyLog = {"struct": modelStruct.__name__,
......@@ -391,11 +396,10 @@ def trainModel(modelStruct, modelGen, modelClass='cat', batchSize=2, optimizer=t
return model
def trainModel_Fruits(modelStruct, modelGen, modelClass='cat', batchSize=2, optimizer=tf.keras.optimizers.Adam,
learning_rate=0.01, losses=None, metrics=None, saveModel=True, modelName='stvNet_weights',
epochs=1, loss_weights=None, outVectors=False, outClasses=False, dataSplit=True, altLabels=True,
augmentation=True): # train and save model weights
def trainModel(modelStruct, modelGen, modelClass='cat', batchSize=2, optimizer=tf.keras.optimizers.Adam,
learning_rate=0.01, losses=None, metrics=None, saveModel=True, modelName='stvNet_weights',
epochs=1, loss_weights=None, outVectors=False, outClasses=False, dataSplit=True, altLabels=True,
augmentation=True): # train and save model weights
if metrics is None:
metrics = ['accuracy']
if not (outVectors or outClasses):
......@@ -427,7 +431,7 @@ def trainModel_Fruits(modelStruct, modelGen, modelClass='cat', batchSize=2, opti
callbacks=[logger])
history.append(hist.history)
if dataSplit:
print("Validation:")
print("############# Validation: ##############")
valHist = model.evaluate(
modelGen(modelClass, batchSize, masterList=validData, out0=outKeys[0], out1=outKeys[1],
altLabels=altLabels, augmentation=False), steps=math.ceil(len(validData) / batchSize),
......@@ -443,7 +447,7 @@ def trainModel_Fruits(modelStruct, modelGen, modelClass='cat', batchSize=2, opti
steps_per_epoch=math.ceil(len(trainData) / batchSize), max_queue_size=2, callbacks=[logger])
history.append(hist.history)
if dataSplit:
print("Validation:")
print("############ Validation: ##########2")
valHist = model.evaluate(
modelGen(modelClass, batchSize, masterList=validData, altLabels=altLabels, augmentation=False),
steps=math.ceil(len(validData) / batchSize), max_queue_size=2)
......@@ -482,7 +486,7 @@ def trainModels(modelSets, shutDown=False):
for modelSet in modelSets:
print("Training {0}".format(modelSet.name))
model = modelsDict[modelSet.name]
trainModel(model.structure, model.generator, modelClass=modelSet.modelClass, epochs=model.epochs,
trainModel_Fruits(model.structure, model.generator, modelClass=modelSet.modelClass, epochs=model.epochs,
losses=model.losses, modelName=modelSet.name, outClasses=model.outClasses,
outVectors=model.outVectors, learning_rate=model.lr, metrics=model.metrics,
altLabels=model.altLabels, augmentation=model.augmentation)
......@@ -501,7 +505,7 @@ def evaluateModel(modelStruct, modelName, evalGen, modelClass='cat', outVectors=
metrics = ['accuracy']
model = tf.keras.models.load_model(
os.path.dirname(os.path.realpath(__file__)) + '/models/' + modelName + '_' + modelClass)
model.evaluate(evalGen(modelClass, batchSize), steps=samples // batchSize)
model.evaluate(evalGen(modelClass, batchSize), steps=samples, val=True)
def evaluateModels(modelSets, batchSize=2, dataSplit=True):
......@@ -516,13 +520,13 @@ def evaluateModels(modelSets, batchSize=2, dataSplit=True):
model.evaluate(
modelEnt.generator(modelSet.modelClass, batchSize=batchSize, masterList=validData, out0=outKeys[0],
out1=outKeys[1], altLabels=modelEnt.altLabels, augmentation=False),
steps=math.ceil(len(validData) / batchSize), max_queue_size=2)
steps=math.ceil(len(validData)), max_queue_size=2)
else:
raise Exception("Probably shouldn't be here ever..")
else:
model.evaluate(modelEnt.generator(modelSet.modelClass, batchSize=batchSize, masterList=validData,
model.evaluate(modelEnt.generator(modelSet.modelClass, val=True, batchSize=batchSize, masterList=validData,
altLabels=modelEnt.altLabels, augmentation=False),
steps=math.ceil(len(validData) / batchSize), max_queue_size=2)
steps=math.ceil(len(validData)), max_queue_size=2)
def trainModelClassGen(modelStruct, modelName, losses, modelClass='cat', batchSize=2,
......@@ -666,70 +670,26 @@ class generatorClass: # simulates generator behaviour, unused
output[self.outVecName] = np.array(yCoordBatch)
if self.outClasses:
output[self.outClassName] = np.array(yClassBatch)
return (False, np.array(xBatch), output)
return False, np.array(xBatch), output
modelsDict = {
'uNet_classes': modelDictVal(uNet, data.classTrainingGenerator, tf.keras.losses.BinaryCrossentropy(), False, True,
epochs=20, lr=0.001, augmentation=False),
'uNet_coords': modelDictVal(uNet, data.coordsTrainingGenerator, tf.keras.losses.Huber(), True, False, epochs=5,
lr=0.001, metrics=['mae', 'mse']),
'uNet_coords_smooth': modelDictVal(uNet, data.coordsTrainingGenerator, smoothL1, True, False, epochs=3, lr=0.0001,
metrics=['mae', 'mse']),
'stvNet': modelDictVal(stvNet, data.combinedTrainingGenerator,
{'coordsOut': tf.keras.losses.Huber(), 'classOut': tf.keras.losses.BinaryCrossentropy()},
True, True, epochs=5, lr=0.00005,
metrics={'coordsOut': ['mae', 'mse'], "classOut": ['accuracy']}),
'stvNet_coords_slow_learner': modelDictVal(stvNet, data.coordsTrainingGenerator, tf.keras.losses.Huber(), True,
False, epochs=40, lr=0.00001, metrics=['mae', 'mse'],
outVecName='coordsOut'),
'stvNetAltLabels': modelDictVal(stvNet, data.combinedTrainingGenerator, {'coordsOut': tf.keras.losses.Huber(),
'classOut': tf.keras.losses.BinaryCrossentropy()},
True, True, epochs=10, lr=0.001,
metrics={'coordsOut': ['mae', 'mse'], "classOut": ['accuracy']}, altLabels=True,
augmentation=True),
'stvNetNormLabels': modelDictVal(stvNet, data.combinedTrainingGenerator, {'coordsOut': tf.keras.losses.Huber(),
'classOut': tf.keras.losses.BinaryCrossentropy()},
True, True, epochs=10, lr=0.001,
metrics={'coordsOut': ['mae', 'mse'], "classOut": ['accuracy']}, altLabels=False,
augmentation=True),
'stvNet_coords': modelDictVal(stvNet, data.coordsTrainingGenerator, tf.keras.losses.Huber(), True, False, epochs=20,
lr=0.001, metrics=['mae', 'mse'], altLabels=False, augmentation=True),
'stvNet_coords_altLabels': modelDictVal(stvNet, data.coordsTrainingGenerator, tf.keras.losses.Huber(), True, False,
epochs=20, lr=0.001, metrics=['mae', 'mse'], altLabels=True,
augmentation=True),
'stvNet_coords_altLabels_noAug': modelDictVal(stvNet, data.coordsTrainingGenerator, tf.keras.losses.Huber(), True,
False, epochs=20, lr=0.001, metrics=['mae', 'mse'], altLabels=True,
augmentation=False),
'stvNet_coords_noAug': modelDictVal(stvNet, data.coordsTrainingGenerator, tf.keras.losses.Huber(), True, False,
epochs=20, lr=0.001, metrics=['mae', 'mse'], altLabels=False,
augmentation=False),
'stvNet_classes': modelDictVal(stvNet, data.classTrainingGenerator, tf.keras.losses.BinaryCrossentropy(), False,
True, epochs=10, lr=0.001, altLabels=False, augmentation=True),
'stvNet_classes_noAug': modelDictVal(stvNet, data.classTrainingGenerator, tf.keras.losses.BinaryCrossentropy(),
False, True, epochs=10, lr=0.001, altLabels=False, augmentation=False),
'stvNet_new_coords_alt': modelDictVal(stvNetNew, data.coordsTrainingGenerator, tf.keras.losses.Huber(), True, False,
epochs=20, lr=0.001, metrics=['mae', 'mse'], altLabels=True,
augmentation=False),
epochs=50, lr=0.0001, augmentation=True),
'stvNet_new_coords': modelDictVal(stvNetNew, data.coordsTrainingGenerator, tf.keras.losses.Huber(), True, False,
epochs=20, lr=0.001, metrics=['mae', 'mse'], altLabels=False, augmentation=False),
'stvNet_new_coords_alt_aug': modelDictVal(stvNetNew, data.coordsTrainingGenerator, tf.keras.losses.Huber(), True,
False, epochs=20, lr=0.001, metrics=['mae', 'mse'], altLabels=True,
augmentation=True),
'stvNet_new_coords_aug': modelDictVal(stvNetNew, data.coordsTrainingGenerator, tf.keras.losses.Huber(), True, False,
epochs=20, lr=0.001, metrics=['mae', 'mse'], altLabels=False,
augmentation=True),
'stvNet_new_classes': modelDictVal(stvNetNew, data.classTrainingGenerator, tf.keras.losses.BinaryCrossentropy(),
False, True, epochs=20, lr=0.001, augmentation=False),
'stvNet_new_combined': modelDictVal(stvNetNew, data.combinedTrainingGenerator,
{'coordsOut': tf.keras.losses.Huber(),
'classOut': tf.keras.losses.BinaryCrossentropy()}, True, True, epochs=20,
lr=0.001, metrics={'coordsOut': ['mae', 'mse'], "classOut": ['accuracy']},
augmentation=False),
epochs=50, lr=0.0001, metrics=['mae', 'mse'], altLabels=False, augmentation=False)
}
if __name__ == "__main__":
class_name = 'pear'
ap = argparse.ArgumentParser()
ap.add_argument("-cls_name", "--class_name", type=str,
default='kiwi1',
help="[kiwi1, pear2, banana1, orange, peach1]")
args = vars(ap.parse_args())
class_name = args["class_name"]
modelSets = [modelSet('uNet_classes', class_name), modelSet('stvNet_new_coords', class_name)]
trainModels(modelSets)
evaluateModels(modelSets)
......
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment