Tensorflow : Android demo accuracy - tensorflow

Question

What related GitHub issues or Stack Overflow threads have you found by searching the web for your problem?
I searched #1269 #504
Environment info
Mac OS for build and Android version 5 to run .apk demo.
If possible, provide a minimal reproducible example (We usually don't have time to read hundreds of lines of your code)
I followed the steps mentioned in #1269 and could able to run the example successfully, but the accuracy of the result is very low and often wrong. I have trained my systems on 25 different daily used products like soap, soup, noodles, etc.
Where as when i run the same example using following script it give me very high accuracy (approx. 90-95%)
import sys
import tensorflow as tf
// 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("/tf_files/retrained_labels.txt")]
// Unpersists graph from file
with tf.gfile.FastGFile("/tf_files/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))
The only difference I see here is that the model file used in the Android demo is stripped because it does not support DecodeJpeg, whereas in the above code its the actually generated unstripped model. Is there any specific reason or somewhere I am wrong here?
I also tried using optimize_for_inference
but unfortunately, it fails with following error:
[milinddeore#P028: ~/tf/tensorflow ] bazel-bin/tensorflow/python/tools/optimize_for_inference --input=/Users/milinddeore/tf_files_nm/retrained_graph.pb --output=/Users/milinddeore/tf/tensorflow/tensorflow/examples/android/assets/tf_ul_stripped_graph.pb --input_names=DecodeJpeg/content —-output_names=final_result
Traceback (most recent call last):
File "/Users/milinddeore/tf/tensorflow/bazel-bin/tensorflow/python/tools/optimize_for_inference.runfiles/org_tensorflow/tensorflow/python/tools/optimize_for_inference.py", line 141, in <module>
app.run(main=main, argv=[sys.argv[0]] + unparsed)
File "/Users/milinddeore/tf/tensorflow/bazel-bin/tensorflow/python/tools/optimize_for_inference.runfiles/org_tensorflow/tensorflow/python/platform/app.py", line 44, in run
_sys.exit(main(_sys.argv[:1] + flags_passthrough))
File "/Users/milinddeore/tf/tensorflow/bazel-bin/tensorflow/python/tools/optimize_for_inference.runfiles/org_tensorflow/tensorflow/python/tools/optimize_for_inference.py", line 90, in main
FLAGS.output_names.split(","), FLAGS.placeholder_type_enum)
File "/Users/milinddeore/tf/tensorflow/bazel-bin/tensorflow/python/tools/optimize_for_inference.runfiles/org_tensorflow/tensorflow/python/tools/optimize_for_inference_lib.py", line 91, in optimize_for_inference
placeholder_type_enum)
File "/Users/milinddeore/tf/tensorflow/bazel-bin/tensorflow/python/tools/optimize_for_inference.runfiles/org_tensorflow/tensorflow/python/tools/strip_unused_lib.py", line 71, in strip_unused
output_node_names)
File "/Users/milinddeore/tf/tensorflow/bazel-bin/tensorflow/python/tools/optimize_for_inference.runfiles/org_tensorflow/tensorflow/python/framework/graph_util_impl.py", line 141, in extract_sub_graph
assert d in name_to_node_map, "%s is not in graph" % d
AssertionError: is not in graph
I suspect that this problem is due to the android not being parse DecodeJpeg, but please correct me if i am wrong.
What other attempted solutions have you tried?
Yes, I the above script and it gives me quite high accuracy result.

Answer 1

Well, the reason for bad accuracy is following:
I ran this example code on Lenovo Vibe K5 mobile (This has SanpDragon 415), this wasn't compiled for hexagon DSP, even through the DSP on 415 is very old as compared to 835 (Hexagon DSP 682), in fact i am not quite sure if the Hexagon SDK will work with 415 or not, i haven't tried this though. This means the example was running on CPU to first detect motion and later classify them and hence the poor performance.
Slow FPS, will capture the images very slowly and hence moving objects will be really difficult.
So if you have bad image, there are very strong chances that the prediction will also be bad.
Camera capture and classification is taking long time, due to latency its not quite real-time.