def train_input_fn():
train_files = inputs.get_tfrecord_paths(FLAGS.train_path)
return input_fn.train_input_fn(
tfrecord_fpaths=train_files,
batch_size=FLAGS.train_batch_size,
shuffle_buffer_size=FLAGS.shuffle_buffer_size,
)
def test_train_input_fn_single(self):
features, labels = train_input_fn(
['test_data/input_fn_test.tfrecord'],
1, # num_features
1, # buffer_size
1, # batch_size
1, # num_epochs
)
session = tf.Session()
result_features, result_labels = session.run([features, labels])
length = result_features[LENGTH]
self.assertEqual(8, length)
feature = result_features[FEATURE]
self.assertListEqual([1., 0., 1., 2., 1., 2., 2., 1.],
list(feature.flatten()))
self.assertListEqual([0, 0, 0, 1, 0, 2, 0, 3, 0, 4, 0, 5, 0, 6],
list(result_labels.indices.flatten()))
self.assertListEqual([1, 0, 1, 2, 1, 2, 1], list(result_labels.values))
self.assertListEqual([1, 7], list(result_labels.dense_shape))
def main(argv):
param_namespace = parse(argv[1:])
params = create_params(param_namespace)
estimator = create_estimator(params)
tf.logging.set_verbosity(tf.logging.INFO)
estimator.train(input_fn=lambda: train_input_fn(
param_namespace.filenames, param_namespace.num_features,
param_namespace.buffer_size, param_namespace.batch_size,
param_namespace.num_epochs),
steps=param_namespace.steps)
predictions = estimator.predict(input_fn=eval_input_fn)
for key, value in enumerate(predictions):
print('------------------key:', key)
for row in value[SOFTMAX]:
row = list(row)
print('max index: %1d' % row.index(max(row)))
def test_train_input_fn_batch(self):
features, labels = train_input_fn(
['test_data/input_fn_test.tfrecord'],
1, # num_features
5, # buffer_size
2, # batch_size
1, # num_epochs
)
session = tf.Session()
result_features, result_labels = session.run([features, labels])
length = result_features[LENGTH]
self.assertEqual(2, len(length))
feature = result_features[FEATURE]
self.assertEqual(2, len(feature))
self.assertEqual(max(length), len(feature[0]))
self.assertEqual(max(length), len(feature[1]))
self.assertEqual(len(result_labels.values), len(result_labels.indices))
self.assertEqual(2, len(result_labels.dense_shape))
#layers['prob'] = tf.nn.softmax(layers['fc1000'])
with tf.name_scope('fc256'):
layers['fc256'] = fully_connected(layers['fc1000'], tf.get_variable('weight_256',shape=[1000,256],initializer=tf.truncated_normal_initializer(stddev=0.1)),tf.get_variable('biases_256',[256],initializer=tf.constant_initializer(0.0)))
return layers['fc256']
def resnet50(input):
#weigths and biases for tensorflow
net = np.load('./resnet50.npy',encoding='latin1').item()
weights = {}
for name in net.keys():
weights[name] = {}
for i in net[name].keys():
if name == 'fc1000':
weights[name][i] = tf.Variable(tf.constant(net[name][i]), dtype='float32' , name=name + '_' + i, trainable=True)
else:
weights[name][i] = tf.Variable(tf.constant(net[name][i]), dtype='float32' , name=name + '_' + i, trainable=False)
return model(input, weights)
if __name__ == "__main__":
from input_fn import train_input_fn
x = train_input_fn('../dataset/train.csv')
[y0,y1] = x.__next__()
y0 = y0.astype(np.float32)
tmp = resnet50(y0)
init = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init)
z = sess.run(tmp)
print(z)
coco_imgs.append((img_name, word_anno))
return coco_imgs
if __name__ == '__main__':
# python train.py --model_dir ./save/
params = Params('./params.json')
tf.reset_default_graph()
tf.logging.set_verbosity(tf.logging.INFO)
# Define the model
config = tf.estimator.RunConfig(tf_random_seed=230,
model_dir=args.model_dir,
save_summary_steps=50,
save_checkpoints_steps=10000)
estimator = tf.estimator.Estimator(model_fn, params=params, config=config)
# Read real image data
coco_imgs = read_coco(args.coco_path)
# Train the model
# # Evaluate the model on the test set
train_spec = tf.estimator.TrainSpec(
input_fn=lambda: train_input_fn(params, coco_imgs))
eval_spec = tf.estimator.EvalSpec(input_fn=lambda: val_input_fn(params),
throttle_secs=100)
# tf.estimator.train_and_evaluate(estimator, train_spec, eval_spec)
estimator.train(input_fn=lambda: train_input_fn(params, coco_imgs))