Develop Reference

Single Image Inference in Tensorflow [Python]

I have already converted a pre-trained .ckpt file to .pb file freezing the model and saving the weighs as well. What I am trying to do now is to make a simple inference using that .pb file and extract and save output image. The model is a (Fully Convolutional Network for Semantic Segmentation) downloaded from here : https://github.com/MarvinTeichmann/KittiSeg . So far I have managed to, load the image, set the default tf graph and import the graph defined by the model on that, read the input and the output tensors and run the session (error here).
import tensorflow as tf
import os
import numpy as np
from tensorflow.python.platform import gfile
from PIL import Image
# Read the image & get statstics
img=Image.open('/path-to-image/demoImage.png')
img.show()
width, height = img.size
print(width)
print(height)
#Plot the image
#image.show()
with tf.Graph().as_default() as graph:
with tf.Session() as sess:
# Load the graph in graph_def
print("load graph")
# We load the protobuf file from the disk and parse it to retrive the unserialized graph_drf
with gfile.FastGFile("/path-to-FCN-model/FCN8.pb",'rb') as f:
#Set default graph as current graph
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
#sess.graph.as_default() #new line
# Import a graph_def into the current default Graph
tf.import_graph_def(graph_def, name='')
# Print the name of operations in the session
#for op in sess.graph.get_operations():
#print "Operation Name :",op.name # Operation name
#print "Tensor Stats :",str(op.values()) # Tensor name
# INFERENCE Here
l_input = graph.get_tensor_by_name('Placeholder:0')
l_output = graph.get_tensor_by_name('save/Assign_38:0')
print "l_input", l_input
print "l_output", l_output
print
print
# Acceptable feed values include Python scalars, strings, lists, numpy ndarrays, or TensorHandles.
result = sess.run(l_output, feed_dict={l_input : img})
print(results)
print("Inference done")
# Info
# First Tensor name : Placeholder:0
# Last tensor name : save/Assign_38:0"
Can the error come from the format of the image (e.g should I convert .png to another format?). Is it another fundamental error?

I managed to fix the error, below is the working script to inference a single image on Fully Convolutional Networks (for whoever is interesting in an alternative segmentation algorithm from SEGNET) . This model use billinear interpolation for scaling rather than an un-pooling layer. Anyway, because the model is available to download in a .chkpt format, you must first freeze the model and save it as a .pb file. Later on, you must pass the network from TF optimizer to set Dropout probabilities to 1. Afterwards, set the correct input and output tensor name in this script and the inference works correctly, extracting the segmented image.
import tensorflow as tf # Default graph is initialized when the library is imported
import os
from tensorflow.python.platform import gfile
from PIL import Image
import numpy as np
import scipy
from scipy import misc
import matplotlib.pyplot as plt
import cv2
with tf.Graph().as_default() as graph: # Set default graph as graph
with tf.Session() as sess:
# Load the graph in graph_def
print("load graph")
# We load the protobuf file from the disk and parse it to retrive the unserialized graph_drf
with gfile.FastGFile("/path-to-protobuf/FCN8_Freezed.pb",'rb') as f:
print("Load Image...")
# Read the image & get statstics
image = scipy.misc.imread('/Path-To-Image/uu_000010.png')
image = image.astype(float)
Input_image_shape=image.shape
height,width,channels = Input_image_shape
print("Plot image...")
#scipy.misc.imshow(image)
# Set FCN graph to the default graph
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
sess.graph.as_default()
# Import a graph_def into the current default Graph (In this case, the weights are (typically) embedded in the graph)
tf.import_graph_def(
graph_def,
input_map=None,
return_elements=None,
name="",
op_dict=None,
producer_op_list=None
)
# Print the name of operations in the session
for op in graph.get_operations():
print "Operation Name :",op.name # Operation name
print "Tensor Stats :",str(op.values()) # Tensor name
# INFERENCE Here
l_input = graph.get_tensor_by_name('Inputs/fifo_queue_Dequeue:0') # Input Tensor
l_output = graph.get_tensor_by_name('upscore32/conv2d_transpose:0') # Output Tensor
print "Shape of input : ", tf.shape(l_input)
#initialize_all_variables
tf.global_variables_initializer()
# Run Kitty model on single image
Session_out = sess.run( l_output, feed_dict = {l_input : image}

Have you already looked at the demo.py. There is shown at line 141 how they modify the input of the graph:
# Create placeholder for input
image_pl = tf.placeholder(tf.float32)
image = tf.expand_dims(image_pl, 0)
# build Tensorflow graph using the model from logdir
prediction = core.build_inference_graph(hypes, modules,
image=image)
And at line 164 how the image is opened:
image = scp.misc.imread(input_image)
Which is fed directly to image_pl. The only point is that core.build_inference_graph is a TensorVision call.
Note, it would be interesting to provide the exact error message as as input as well.

How do I convert a numpy Array to a data type that tensorflow can classify?

I am writing a Python program to detect the state of a chess board and I am using a sliding window to detect the position of each piece. My main program detects the chessboard within an image and passes its cropped picture to the my_sliding_window method. This is supposed to use Tensorflow to detect a piece in the sliding window. From this tutorial I saw that pictures are read like this:
image_data = tf.gfile.FastGFile('picture.jpg', 'rb').read()
But I don't want to read it from file as I already have the picture in a numpy array. How do make my numpy array in such a way that it can be classified by Tensorflow?
Thank you.
Code:
import tensorflow as tf, sys
import cv2
image_path = sys.argv[1]
img = cv2.imread('picture.jpg')
image_data = tf.convert_to_tensor(img)
print type(image_data) # this returns <class 'tensorflow.python.framework.ops.Tensor'>
# This is what is used in the tutorial I mentioned above
image_data2 = tf.gfile.FastGFile(image_path, 'rb').read()
print type(image_data2) # this returns <type 'str'>
# Loads label file, strips off carriage return
label_lines = [line.rstrip() for line
in tf.gfile.GFile("retrained_labels.txt")]
# Unpersists graph from file
with tf.gfile.FastGFile("retrained_graph.pb", 'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
_ = tf.import_graph_def(graph_def, name='')
with tf.Session() as sess:
# Feed the image_data as input to the graph and get first prediction
softmax_tensor = sess.graph.get_tensor_by_name('final_result:0')
predictions = sess.run(softmax_tensor, \
{'DecodeJpeg/contents:0': image_data})
# Sort to show labels of first prediction in order of confidence
top_k = predictions[0].argsort()[-len(predictions[0]):][::-1]
for node_id in top_k:
human_string = label_lines[node_id]
score = predictions[0][node_id]
print('%s (score = %.5f)' % (human_string, score))

You could use tf.convert_to_tensor() to convert your numpy array into a TensorFlow tensor:
This function converts Python objects of various types to Tensor objects. It accepts Tensor objects, numpy arrays, Python lists, and Python scalars.
Update
Ok, so what you're trying to do is to feed the numpy array image_data, with dimensions [123, 82] to the placeholder DecodeJpeg/contents:0. However that placeholder was defined with shape=() meaning it only accepts 0D tensors as input (see tensor shapes), hence throwing you an error.
What the original code does is to read an image as a dimensionless string with:
image_data = tf.gfile.FastGFile(image_path, 'rb').read()
which is then fed to the DecodeJpeg/contents:0 placeholder in:
predictions = sess.run(softmax_tensor, {'DecodeJpeg/contents:0': image_data})
The easiest way to proceed and try to run your images through the pretrained graph would be to use the same tf.gfile.FastGFile() call for loading the images.

I solved the problem by using this one-liner:
image_data = cv2.imencode('.jpg', cv_image)[1].tostring()

Tensorflow Label_Image for PNG, GIF etc

I was playing around with Tensorflow for image classification. I used the image_retraining/retrain.py to retrain the inception library with new categories and used it to classify images using label_image.py from https://github.com/llSourcell/tensorflow_image_classifier/blob/master/src/label_image.py as below:
import tensorflow as tf
import sys
# change this as you see fit
image_path = sys.argv[1]
# Read in the image_data
image_data = tf.gfile.FastGFile(image_path, 'rb').read()
# Loads label file, strips off carriage return
label_lines = [line.rstrip() for line
in tf.gfile.GFile("/root/tf_files/output_labels.txt")]
# Unpersists graph from file
with tf.gfile.FastGFile("/root/tf_files/output_graph.pb", 'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
_ = tf.import_graph_def(graph_def, name='')
with tf.Session() as sess:
# Feed the image_data as input to the graph and get first prediction
softmax_tensor = sess.graph.get_tensor_by_name('final_result:0')
#predictions = sess.run(softmax_tensor,{'DecodeJpeg/contents:0': image_data})
predictions = sess.run(softmax_tensor,{'DecodePng/contents:0': image_data})
# Sort to show labels of first prediction in order of confidence
top_k = predictions[0].argsort()[-len(predictions[0]):][::-1]
for node_id in top_k:
human_string = label_lines[node_id]
score = predictions[0][node_id]
print('%s (score = %.5f)' % (human_string, score))
I noticed two issues. When I retrain with new categories, it only trains JPG images. I am a noob in machine learning so not sure whether this is a limitation or is it possible to train other extension images like PNG, GIF?
Another one is when classifying the images the input is again only for JPG. I tried to change DecodeJpeg to DecodePng in label_image.py above but couldn't work. Another way I tried was to convert other formats into JPG before passing them in for classification like:
im = Image.open('/root/Desktop/200_s.gif').convert('RGB')
im.save('/root/Desktop/test.jpg', "JPEG")
image_path1 = '/root/Desktop/test.jpg'
Is there any other way to do this? Does Tensorflow have functions to handle other image formats other than JPG?
I tried the following by feeding in parsed image as compared to JPEG as suggested by #mrry
import tensorflow as tf
import sys
import numpy as np
from PIL import Image
# change this as you see fit
image_path = sys.argv[1]
# Read in the image_data
image_data = tf.gfile.FastGFile(image_path, 'rb').read()
image = Image.open(image_path)
image_array = np.array(image)[:,:,0:3] # Select RGB channels only.
# Loads label file, strips off carriage return
label_lines = [line.rstrip() for line
in tf.gfile.GFile("/root/tf_files/output_labels.txt")]
# Unpersists graph from file
with tf.gfile.FastGFile("/root/tf_files/output_graph.pb", 'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
_ = tf.import_graph_def(graph_def, name='')
with tf.Session() as sess:
# Feed the image_data as input to the graph and get first prediction
softmax_tensor = sess.graph.get_tensor_by_name('final_result:0')
predictions = sess.run(softmax_tensor,{'DecodeJpeg:0': image_array})
# Sort to show labels of first prediction in order of confidence
top_k = predictions[0].argsort()[-len(predictions[0]):][::-1]
for node_id in top_k:
human_string = label_lines[node_id]
score = predictions[0][node_id]
print('%s (score = %.5f)' % (human_string, score))
It works for JPEG images but when I use PNG or GIF it throws
Traceback (most recent call last):
File "label_image.py", line 17, in <module>
image_array = np.array(image)[:,:,0:3] # Select RGB channels only.
IndexError: too many indices for array

The model can only train on (and evaluate) JPEG images, because the GraphDef that you've saved in /root/tf_files/output_graph.pb only contains a tf.image.decode_jpeg() op, and uses the output of that op for making predictions. There are at least a couple of options for using other image formats:
Feed in parsed images rather than JPEG data. In the current program, you feed in a JPEG-encoded image as a string value for the tensor "DecodeJpeg/contents:0". Instead, you can feed in a 3-D array of decoded image data for the tensor "DecodeJpeg:0" (which represents the output of the tf.image.decode_jpeg() op), and you can use NumPy, PIL, or some other Python library to create this array.
Remap the image input in tf.import_graph_def(). The tf.import_graph_def() function enables you to connect two different graphs together by remapping individual tensor values. For example, you could do something like the following to add a new image-processing op to the existing graph:
image_string_input = tf.placeholder(tf.string)
image_decoded = tf.image.decode_png(image_string_input)
# Unpersists graph from file
with tf.gfile.FastGFile("/root/tf_files/output_graph.pb", 'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
softmax_tensor, = tf.import_graph_def(
graph_def,
input_map={"DecodeJpeg:0": image_decoded},
return_operations=["final_result:0"])
with tf.Session() as sess:
# Feed the image_data as input to the graph and get first prediction
predictions = sess.run(softmax_tensor, {image_string_input: image_data})
# ...

You should have a look at the tf.image package. It's got good functions to decode / encode JPEGs, GIFs and PNGs.

Following #mrry's suggestion to feed in parsed image, converted the image data into array and convert into RGB as stated below in the code. Now I am able to feed in JPG,PNG and GIF.
import tensorflow as tf
import sys
import numpy as np
from PIL import Image
# change this as you see fit
image_path = sys.argv[1]
# Read in the image_data
image_data = tf.gfile.FastGFile(image_path, 'rb').read()
image = Image.open(image_path)
image_array = image.convert('RGB')
# Loads label file, strips off carriage return
label_lines = [line.rstrip() for line
in tf.gfile.GFile("/root/tf_files/output_labels.txt")]
# Unpersists graph from file
with tf.gfile.FastGFile("/root/tf_files/output_graph.pb", 'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
_ = tf.import_graph_def(graph_def, name='')
with tf.Session() as sess:
# Feed the image_data as input to the graph and get first prediction
softmax_tensor = sess.graph.get_tensor_by_name('final_result:0')
predictions = sess.run(softmax_tensor,{'DecodeJpeg:0': image_array})
# Sort to show labels of first prediction in order of confidence
top_k = predictions[0].argsort()[-len(predictions[0]):][::-1]
for node_id in top_k:
human_string = label_lines[node_id]
score = predictions[0][node_id]
print('%s (score = %.5f)' % (human_string, score))

Tensorflow Inception_Resnet_V2 Classify Image

I trained an inception_resnet_v2 model for the flowers images following the README at https://github.com/tensorflow/models/tree/master/slim
I got my graph.pbtxt file out of this after training with which I converted to a graph.pb file with the following code:
import tensorflow as tf
from google.protobuf import text_format
def convert_pbtxt_to_graphdef(filename):
"""Returns a `tf.GraphDef` proto representing the data in the given pbtxt file.
Args:
filename: The name of a file containing a GraphDef pbtxt (text-formatted
`tf.GraphDef` protocol buffer data).
Returns:
A `tf.GraphDef` protocol buffer.
"""
with tf.gfile.FastGFile(filename, 'r') as f:
graph_def = tf.GraphDef()
file_content = f.read()
# Merges the human-readable string in `file_content` into `graph_def`.
text_format.Merge(file_content, graph_def)
return graph_def
with tf.gfile.FastGFile('/foo/bar/workspace/results/graph.pb', 'wb') as f:
f.write(convert_pbtxt_to_graphdef('/foo/bar/workspace/results/graph.pbtxt'))
After getting this file I tried feeding the trained model a random image using tensorflow's classify_image.py found here: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/models/image/imagenet/classify_image.py
using my .pb, .pbtxt, and my labels file, however, I get the following error:
Traceback (most recent call last):
File "classify_image.py", line 212, in <module>
tf.app.run()
File "/usr/local/lib/python2.7/site-packages/tensorflow/python/platform/app.py", line 30, in run
sys.exit(main(sys.argv[:1] + flags_passthrough))
File "classify_image.py", line 208, in main
run_inference_on_image(image)
File "classify_image.py", line 170, in run_inference_on_image
softmax_tensor = sess.graph.get_tensor_by_name('softmax:0')
File "/usr/local/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 2615, in get_tensor_by_name
return self.as_graph_element(name, allow_tensor=True, allow_operation=False)
File "/usr/local/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 2466, in as_graph_element
return self._as_graph_element_locked(obj, allow_tensor, allow_operation)
File "/usr/local/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 2508, in _as_graph_element_locked
"graph." % (repr(name), repr(op_name)))
KeyError: "The name 'softmax:0' refers to a Tensor which does not exist. The operation, 'softmax', does not exist in the graph."

The problem with slim, in fact tensorflow/models, is that the framework and so the produced models don't really fit the prediction use case:
TF-slim is a new lightweight high-level API of TensorFlow
(tensorflow.contrib.slim) for defining, training and evaluating
complex models.
(Source https://github.com/tensorflow/models/tree/master/slim)
The main problem at prediction time is, that it only works well with the checkpoint files created by the Saver class. When using checkpoint files the assign_from_checkpoint_fn() method can be used to initialize all the variables with the trained parameters contained in the checkpoint. On the other hand, in the situation when having the GraphDef file *.pb only, you kinda lost. There is a nice trick though.
The key idea is to inject a tf.placeholder variable for the input image(s) into the computation graph after you saved your trained model as a checkpoint. The following script (convert_checkpoint_to_pb.py) reads a checkpoint, inserts a placeholder, converts the graph variables to constants and dumps it to a *.pb file.
import tensorflow as tf
from tensorflow.contrib import slim
from nets import inception
from tensorflow.python.framework.graph_util import convert_variables_to_constants
from tensorflow.python.tools.optimize_for_inference_lib import optimize_for_inference
from preprocessing import inception_preprocessing
checkpoints_dir = '/path/to/your/checkpoint_dir/'
OUTPUT_PB_FILENAME = 'minimal_graph.proto'
NUM_CLASSES = 2
# We need default size of image for a particular network.
# The network was trained on images of that size -- so we
# resize input image later in the code.
image_size = inception.inception_resnet_v2.default_image_size
with tf.Graph().as_default():
# Inject placeholder into the graph
input_image_t = tf.placeholder(tf.string, name='input_image')
image = tf.image.decode_jpeg(input_image_t, channels=3)
# Resize the input image, preserving the aspect ratio
# and make a central crop of the resulted image.
# The crop will be of the size of the default image size of
# the network.
# I use the "preprocess_for_eval()" method instead of "inception_preprocessing()"
# because the latter crops all images to the center by 85% at
# prediction time (training=False).
processed_image = inception_preprocessing.preprocess_for_eval(image,
image_size,
image_size, central_fraction=None)
# Networks accept images in batches.
# The first dimension usually represents the batch size.
# In our case the batch size is one.
processed_images = tf.expand_dims(processed_image, 0)
# Load the inception network structure
with slim.arg_scope(inception.inception_resnet_v2_arg_scope()):
logits, _ = inception.inception_resnet_v2(processed_images,
num_classes=NUM_CLASSES,
is_training=False)
# Apply softmax function to the logits (output of the last layer of the network)
probabilities = tf.nn.softmax(logits)
model_path = tf.train.latest_checkpoint(checkpoints_dir)
# Get the function that initializes the network structure (its variables) with
# the trained values contained in the checkpoint
init_fn = slim.assign_from_checkpoint_fn(
model_path,
slim.get_model_variables())
with tf.Session() as sess:
# Now call the initialization function within the session
init_fn(sess)
# Convert variables to constants and make sure the placeholder input_image is included
# in the graph as well as the other neccesary tensors.
constant_graph = convert_variables_to_constants(sess, sess.graph_def, ["input_image", "DecodeJpeg",
"InceptionResnetV2/Logits/Predictions"])
# Define the input and output layer properly
optimized_constant_graph = optimize_for_inference(constant_graph, ["eval_image"],
["InceptionResnetV2/Logits/Predictions"],
tf.string.as_datatype_enum)
# Write the production ready graph to file.
tf.train.write_graph(optimized_constant_graph, '.', OUTPUT_PB_FILENAME, as_text=False)
(The models/slim code must be in your python path to execute this code)
To predict new images with the converted model (now present as a *.pb file) use the code from file minimal_predict.py:
import tensorflow as tf
import urllib2
def create_graph(model_file):
"""Creates a graph from saved GraphDef file and returns a saver."""
# Creates graph from saved graph_def.pb.
with tf.gfile.FastGFile(model_file, 'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
_ = tf.import_graph_def(graph_def, name='')
model_file = "/your/path/to/minimal_graph.proto"
url = ("http://pictureparadise.net/funny-babies/funny-babies02/funny-babies-053.jpg")
# Open specified url and load image as a string
image_string = urllib2.urlopen(url).read()
with tf.Graph().as_default():
with tf.Session() as new_sess:
create_graph(model_file)
softmax = new_sess.graph.get_tensor_by_name("InceptionResnetV2/Logits/Predictions:0")
# Loading the injected placeholder
input_placeholder = new_sess.graph.get_tensor_by_name("input_image:0")
probabilities = new_sess.run(softmax, {input_placeholder: image_string})
print probabilities
To use these scripts simply run python
python convert_checkpoint_to_pb.py
python minimal_predicty.py
while having tensorflow and tensorflow/models/slim in your PYTHONPATH.

In the convert_checkpoint_to_pb.pyprovided by #Maximilian,["eval_image"] in
optimized_constant_graph = optimize_for_inference(constant_graph, ["eval_image"],["InceptionResnetV2/Logits/Predictions"],tf.string.as_datatype_enum)
should be replaced by ["input_image"] else you will get an error stating "The following input nodes were not found: {'eval_image'}\n".

Tensorflow docker container predictions won't work outside of container

I went through the Tensorflow for poets Tutorial and then classified my own images, this is all done in the TF docker container provided. The model has a validation accuracy in the mid to low 90's. There is a separate file that makes predictions for new images(below).
I copied the files 'retrained_labels_corn.txt' and 'retrained_graph_corn.pd' and the file holding the code seen below to a directory(and changed the file paths) to see if I could make predictions while not in the docker container. I made sure I give it a valid image path as the system arg but it always predicts the image as one class with a probability above 97%. When I do the same thing while in the docker container everything works fine. I even tried pointing the labeling file to the exact same files that docker container uses and I am getting the same result of it always predicting one class with a high degree of certainty.
What did I do wrong?
import tensorflow as tf, sys
image_path = sys.argv[1]
# Read in the image_data
image_data = tf.gfile.FastGFile(image_path, 'rb').read()
# Loads label file, strips off carriage return
label_lines = [line.rstrip() for line
in tf.gfile.GFile("/tf_files/retrained_labels_corn.txt")]
# Unpersists graph from file
with tf.gfile.FastGFile("/tf_files/retrained_graph_corn.pb", 'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
_ = tf.import_graph_def(graph_def, name='')
with tf.Session() as sess:
# Feed the image_data as input to the graph and get first prediction
softmax_tensor = sess.graph.get_tensor_by_name('final_result:0')
predictions = sess.run(softmax_tensor, \
{'DecodeJpeg/contents:0': image_data})
# Sort to show labels of first prediction in order of confidence
top_k = predictions[0].argsort()[-len(predictions[0]):][::-1]
for node_id in top_k:
human_string = label_lines[node_id]
score = predictions[0][node_id]
print('%s (score = %.5f)' % (human_string, score))
I am on Ubuntu version 16.04 and TF 0.10