Transferring the inlet piping to the TensorFlow score
I'm a noob to TF, so it's easy on me.
I need to train a simple CNN from a bunch of images in a shortcut directory. After looking around a lot, I cooked up this code that prepares the TF input pipeline and I was able to print an array of images.
image_list, label_list = load_dataset()
imagesq = ops.convert_to_tensor(image_list, dtype=dtypes.string)
labelsq = ops.convert_to_tensor(label_list, dtype=dtypes.int32)
# Makes an input queue
input_q = tf.train.slice_input_producer([imagesq, labelsq],
shuffle=True)
file_content = tf.read_file(input_q[0])
train_image = tf.image.decode_png(file_content,channels=3)
train_label = input_q[1]
train_image.set_shape([120,120,3])
# collect batches of images before processing
train_image_batch, train_label_batch = tf.train.batch(
[train_image, train_label],
batch_size=5
# ,num_threads=1
)
with tf.Session() as sess:
# initialize the variables
sess.run(tf.global_variables_initializer())
# initialize the queue threads to start to shovel data
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
# print "from the train set:"
for i in range(len(image_list)):
print sess.run(train_image_batch)
# sess.run(train_image)
# sess.run(train_label)
# classifier.fit(input_fn=lambda: (train_image, train_label),
# steps=100,
# monitors=[logging_hook])
# stop our queue threads and properly close the session
coord.request_stop()
coord.join(threads)
sess.close()
But looking at the MNIST example given in the TF docs, I can see that they use cnn_model_fn as well as the Estimator class .
I have defined my own cnn_model_fn and would like to combine the two. Please help me how to move on. This code doesn't work
classifier = learn.Estimator(model_fn=cnn_model_fn, model_dir='./test_model')
classifier.fit(input_fn=lambda: (train_image, train_label),
steps=100,
monitors=[logging_hook])
The pipeline seems to only fill up when the session is started, otherwise its empty and it gives the InputError 'Input graph and Layer graph' are not the same '
Please help me.
source to share
I'm new to tensorflow, so take this with a grain of salt.
AFAICT, when you call any of the APIs tf
that create "tensors" or "operations", they are created in a context named Graph
.
Also, I believe that on startup Estimator
it creates a new empty one Graph
for every startup. It fills Graph
in by launching model_fn
and input_fn
which should call APIs tf
that add "tensors" and "operations" in the context of this new one Graph
.
The return values from model_fn
and input_fn
just provide references so the parts can be connected correctly - Graph
already contains them.
However, in this example, the input operations were already created before it was Estimator
created Graph
, and therefore their associated operations were added to the implicit default Graph
(one is created automatically, I believe). So when Estimator
creating a new one and filling the model with with model_fn
, the input and the model will be on two different graphs.
To fix this, you need to change input_fn
. Don't just wrap the pair (image, labels)
in lambda
, but rather wrap the entire input construct in a function so that when Estimator
triggered input_fn
as a side effect of all API calls, all operation inputs and tensors will be created in the correct context Graph
.
source to share