def fit_model(model, train_data, test_data, weights_dir, input_shape):
try:
if os.path.exists(weights_dir):
model.load_weights(weights_dir)
print('Load weights')
train_generator = sequence_generator(train_data, BatchSize,
input_shape, N_CLASSES)
test_generator = sequence_generator(test_data, BatchSize, input_shape,
N_CLASSES)
print('Start fitting model')
checkpointer = keras.callbacks.ModelCheckpoint(weights_dir,
save_best_only=True,
save_weights_only=True)
adam = Adam()
model.compile(loss='categorical_crossentropy',
optimizer=adam,
metrics=['accuracy'])
model.fit_generator(train_generator,
steps_per_epoch=300,
epochs=200,
validation_data=test_generator,
validation_steps=100,
verbose=2,
callbacks=[checkpointer])
except KeyboardInterrupt:
print('Training time:')
def fit_model(model,
train_data,
test_data,
weights_dir,
input_shape,
optical_flow=False,
project_dir='./'):
try:
# using sequence or image_from_sequnece generator
if optical_flow:
train_generator = sequence_generator(train_data, BatchSize,
input_shape, N_CLASSES)
test_generator = sequence_generator(test_data, BatchSize,
input_shape, N_CLASSES)
else:
train_generator = image_from_sequence_generator(
train_data, BatchSize, input_shape, N_CLASSES)
test_generator = image_from_sequence_generator(
test_data, BatchSize, input_shape, N_CLASSES)
# frames_dir = '/home/changan/ActionRecognition/data/flow_images'
# train_generator = image_generator(train_data, frames_dir, BatchSize, input_shape, N_CLASSES, mean_sub=False,
# normalization=True, random_crop=True, horizontal_flip=True)
# test_generator = image_generator(test_data, frames_dir, BatchSize, input_shape, N_CLASSES, mean_sub=False,
# normalization=True, random_crop=True, horizontal_flip=True)
sgd = SGD(lr=0.001, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(optimizer=sgd,
loss='categorical_crossentropy',
metrics=['accuracy'])
print(model.summary())
print('Start fitting model')
while True:
checkpointer = keras.callbacks.ModelCheckpoint(
weights_dir, save_best_only=True, save_weights_only=True)
earlystopping = keras.callbacks.EarlyStopping(monitor='val_loss',
min_delta=0.001,
patience=20,
verbose=2,
mode='auto')
tensorboard = keras.callbacks.TensorBoard(log_dir='./logs/try',
histogram_freq=0,
write_graph=True,
write_images=True)
model.fit_generator(
train_generator,
steps_per_epoch=200,
epochs=2000,
validation_data=test_generator,
validation_steps=100,
verbose=2,
callbacks=[checkpointer, tensorboard, earlystopping])
data_dir = os.path.join(project_dir, 'data')
list_dir = os.path.join(data_dir, 'ucfTrainTestlist')
UCF_dir = os.path.join(data_dir, 'UCF-101')
regenerate_data(data_dir, list_dir, UCF_dir)
except KeyboardInterrupt:
print('Training is interrupted')
from utils.UCF_utils import sequence_generator, get_data_list
import os
import numpy as np
from train_CNN import LockedIterator
cwd = os.getcwd()
data_dir = os.path.join(cwd, 'data')
list_dir = os.path.join(data_dir, 'ucfTrainTestlist')
video_dir = os.path.join(data_dir, 'OF_data')
input_shape = (216, 216, 18)
N_CLASSES = 101
train_data, test_data, class_index = get_data_list(list_dir, video_dir)
c = LockedIterator(sequence_generator(train_data, 32, input_shape, N_CLASSES))
i = next(c)[0]
print("i shape = " + str(i.shape))
print(i)