def __init__(self):
goods_dataset = GoodsDataset("dataset-181018.list",
"dataset-181018.labels",
settings.IMAGE_SIZE, settings.train_batch,
settings.valid_batch, settings.multiply,
settings.valid_percentage)
train_set = goods_dataset.get_train_dataset()
valid_set = goods_dataset.get_valid_dataset()
input_tensor = keras.layers.Input(shape=(IMAGE_SIZE[0], IMAGE_SIZE[1],
3))
base_model = InceptionV3(weights='imagenet',
include_top=False,
pooling='avg',
input_tensor=input_tensor)
output_layer_number = 248
intermediate_layer_model = keras.Model(
inputs=base_model.input,
outputs=base_model.layers[output_layer_number].output)
def _intermediate_processing(images, labels):
images = intermediate_layer_model.predict(images, steps=77)
return images, labels
self.train_set = train_set.map(
_intermediate_processing) #, num_parallel_calls=8)
self.valid_set = valid_set.map(
_intermediate_processing) #, num_parallel_calls=8)
predictions = keras.layers.Dense(settings.num_classes,
activation='softmax')(main_model.output)
new_model = keras.Model(inputs=main_model.input, outputs=predictions)
new_model.compile(optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy', top_6])
session = keras.backend.get_session()
print(new_model.summary())
for i in range(0, len(base_model.layers)):
new_model.layers[i + 248].set_weights(base_model.layers[i].get_weights())
print("new model inputs")
for node in new_model.inputs:
print(node.op.name)
print("new model outputs")
for node in new_model.outputs:
print(node.op.name)
dataset = GoodsDataset("dataset-181018.list", "dataset-181018.labels",
(IMAGE_SIZE[0], IMAGE_SIZE[1]), 32, 32, 5, 0.1)
results = new_model.evaluate(dataset.get_valid_dataset(), steps=77)
print(results)
new_model.save("output/inception_{0}.hdf5".format(model_name))
metrics=['accuracy', top_6])
callbacks = [
keras.callbacks.ModelCheckpoint(
"./checkpoints/FINE_TUNE_MODEL_4_DIRECT_inceptionv3-181018-{epoch:02d}-{acc:.3f}-{val_acc:.3f}[{val_top_6:.3f}].hdf5",
save_best_only=True,
monitor='val_top_6',
mode='max'),
keras.callbacks.TensorBoard(
log_dir='./tensorboard-incv4',
write_images=True,
)
]
goods_dataset = GoodsDataset("dataset-181018.list", "dataset-181018.labels",
settings.IMAGE_SIZE, settings.train_batch,
settings.valid_batch, settings.multiply,
settings.valid_percentage)
train_dataset = goods_dataset.get_train_dataset()
valid_dataset = goods_dataset.get_valid_dataset()
results = model.evaluate(
goods_dataset.get_images_for_label(94).batch(16).repeat(), steps=6)
print(results)
model.fit(
train_dataset.prefetch(2).repeat(),
callbacks=callbacks,
epochs=30,
steps_per_epoch=1157,
validation_data=valid_dataset.repeat(),
validation_steps=77,
ax2.grid(color='g', linestyle='-', linewidth=0.2)
ymaxval = max(results['valid_top6'])
ymin = 0.9 if ymaxval > 0.95 else (0.8 if ymaxval > 0.85 else 0.6)
ax2.set_ylim(ymin, 1.0)
#plt.show()
outfile = '_plot_[{}].png'.format(net_model_name)
plt.savefig(outfile)
#------------
# dataset
from dataset_factory import GoodsDataset
#from dataset_factory_imgaug import GoodsDatasetImgaug as GoodsDataset
goods_dataset = GoodsDataset(settings.dataset_list, settings.labels_list,
settings.IMAGE_SIZE, settings.train_batch,
settings.valid_batch, settings.multiply,
settings.valid_percentage)
train_dataset = goods_dataset.get_train_dataset()
valid_dataset = goods_dataset.get_valid_dataset()
num_epochs = 500
epochs_checkpoint = 20 # interval for saving checkpoints and pb-file
train_steps_per_epoch = 724 #1157
valid_steps_per_epoch = 78 #77
train_dataset = train_dataset.repeat()
valid_dataset = valid_dataset.repeat()
"""
def model_function(next_element):
x, y = next_element
logits, end_points = inception.inception_v3(
flags.DEFINE_string(
'data_dir', '/tmp/mnist/data',
'Directory where mnist data will be downloaded'
' if the data is not already there')
flags.DEFINE_string('model_dir', '/tmp/mnist/model',
'Directory where all models are saved')
flags.DEFINE_integer('batch_size', 100, 'Batch size.')
flags.DEFINE_integer('num_epochs', 1, 'Num of batches to train (epochs).')
flags.DEFINE_float('learning_rate', 0.001, 'Learning Rate')
FLAGS = flags.FLAGS
#----------
from dataset_factory import GoodsDataset
goods_dataset = GoodsDataset("dataset-181018.list", "dataset-181018.labels",
settings.IMAGE_SIZE, settings.train_batch,
settings.valid_batch, settings.multiply,
settings.valid_percentage)
train_dataset = goods_dataset.get_train_dataset()
valid_dataset = goods_dataset.get_valid_dataset()
def train_data():
#data = dataset.train(FLAGS.data_dir)
#data = data.cache()
#data = data.batch(FLAGS.batch_size)
#return data
return train_dataset.prefetch(2).repeat()
def eval_data():
#data = dataset.test(FLAGS.data_dir)
callbacks = [
# Interrupt training if `val_loss` stops improving for over 2 epochs
# keras.callbacks.EarlyStopping(patience=2, monitor='val_loss'),
# Write TensorBoard logs to `./logs` directory
keras.callbacks.ModelCheckpoint(
"./checkpoints/se_recognizer131018.{epoch:02d}-{val_loss:.2f}.hdf5"),
keras.callbacks.TensorBoard(
log_dir='./tensorboard',
write_images=True,
)
]
#goods_dataset = GoodsDataset("dataset.list", OUTPUT_FOLDER + "/" + OUTPUT_MODEL_NAME + ".txt",
# IMAGE_SIZE, 20, 20, 5, 0.1)
goods_dataset = GoodsDataset("dataset-181018.list", "dataset-181018.labels",
(IMAGE_SIZE[0], IMAGE_SIZE[1]),
settings.train_batch, settings.valid_batch,
settings.multiply, settings.valid_percentage)
input_tensor = keras.layers.Input(shape=(IMAGE_SIZE[0], IMAGE_SIZE[1], 3))
# input_shape=(IMAGE_SIZE[0], IMAGE_SIZE[1], 3)
model = InceptionResNetV2(include_top=True,
weights=None,
classes=goods_dataset.classes_count,
input_tensor=input_tensor)
# tf.train.RMSPropOptimizer(0.001),
optimizer = keras.optimizers.RMSprop(lr=0.005)
model.compile(optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy'])
import tensorflow as tf
from tensorflow import keras
from tensorflow.python.framework import graph_io
from dataset_factory import GoodsDataset
# tf.enable_eager_execution()
def top_6(y_true, y_pred):
k = 6
return tf.keras.metrics.top_k_categorical_accuracy(y_true, y_pred, k)
goods_dataset = GoodsDataset("dataset-181018.list", "dataset-181018.labels",
(299, 299), 32, 32, 5, 0.1)
base_model = keras.models.load_model(
"./output/inpcetionv3_top60_181018-02-0.876-0.584[0.881]_rnd_adagrad.hdf5",
custom_objects={'top_6': top_6})
base_model.compile(optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy', top_6])
results = base_model.evaluate(
goods_dataset.get_images_for_label(94).batch(16).repeat(), steps=6)
print(results)
# for i, (img, lbl) in enumerate(goods_dataset.get_ambroziya().batch(32).repeat()):
# l = tf.argmax(lbl)
# r = tf.math.equal(l, tf.constant(45, dtype=tf.int64))