Object detection net(eg:faster_rcnn_resnet101) not worked with deconv_visualization

[jidebingfeng]

Object detection net(eg:faster_rcnn_resnet101) not worked with deconv_visualization. With activation_visualization works well. Error:

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-4-9aba9f0fced7> in <module>()
      3                                      layers=layers,
      4                                      path_logdir=os.path.join("Log","Inception5"),
----> 5                                      path_outdir=os.path.join("Output","Inception5"))

/notebooks/workspace/github/tf_cnnvis/tf_cnnvis/tf_cnnvis.py in deconv_visualization(sess_graph_path, value_feed_dict, input_tensor, layers, path_logdir, path_outdir)
    408 def deconv_visualization(sess_graph_path, value_feed_dict, input_tensor = None,  layers = 'r', path_logdir = './Log', path_outdir = "./Output"):
    409     is_success = _get_visualization(sess_graph_path, value_feed_dict, input_tensor = input_tensor, layers = layers, method = "deconv",
--> 410         path_logdir = path_logdir, path_outdir = path_outdir)
    411     return is_success
    412 

/notebooks/workspace/github/tf_cnnvis/tf_cnnvis/tf_cnnvis.py in _get_visualization(sess_graph_path, value_feed_dict, input_tensor, layers, path_logdir, path_outdir, method)
    167                 elif layer != None and layer.lower() in dict_layer.keys():
    168                     layer_type = dict_layer[layer.lower()]
--> 169                     is_success = _visualization_by_layer_type(g, value_feed_dict, input_tensor, layer_type, method, path_logdir, path_outdir)
    170                 else:
    171                     print("Skipping %s . %s is not valid layer name or layer type" % (layer, layer))

/notebooks/workspace/github/tf_cnnvis/tf_cnnvis/tf_cnnvis.py in _visualization_by_layer_type(graph, value_feed_dict, input_tensor, layer_type, method, path_logdir, path_outdir)
    225 
    226     for layer in layers:
--> 227         is_success = _visualization_by_layer_name(graph, value_feed_dict, input_tensor, layer, method, path_logdir, path_outdir)
    228     return is_success
    229 

/notebooks/workspace/github/tf_cnnvis/tf_cnnvis/tf_cnnvis.py in _visualization_by_layer_name(graph, value_feed_dict, input_tensor, layer_name, method, path_logdir, path_outdir)
    289         elif method == "deconv":
    290             # deconvolution
--> 291             results = _deconvolution(graph, sess, op_tensor, X, feed_dict)
    292         elif method == "deepdream":
    293             # deepdream

/notebooks/workspace/github/tf_cnnvis/tf_cnnvis/tf_cnnvis.py in _deconvolution(graph, sess, op_tensor, X, feed_dict)
    335                         c += 1
    336                 if c > 0:
--> 337                     out.extend(sess.run(reconstruct[:c], feed_dict = feed_dict))
    338     return out
    339 def _deepdream(graph, sess, op_tensor, X, feed_dict, layer, path_outdir, path_logdir):

/usr/local/lib/python3.5/dist-packages/tensorflow/python/client/session.py in run(self, fetches, feed_dict, options, run_metadata)
    898     try:
    899       result = self._run(None, fetches, feed_dict, options_ptr,
--> 900                          run_metadata_ptr)
    901       if run_metadata:
    902         proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)

/usr/local/lib/python3.5/dist-packages/tensorflow/python/client/session.py in _run(self, handle, fetches, feed_dict, options, run_metadata)
   1118     # Create a fetch handler to take care of the structure of fetches.
   1119     fetch_handler = _FetchHandler(
-> 1120         self._graph, fetches, feed_dict_tensor, feed_handles=feed_handles)
   1121 
   1122     # Run request and get response.

/usr/local/lib/python3.5/dist-packages/tensorflow/python/client/session.py in __init__(self, graph, fetches, feeds, feed_handles)
    425     """
    426     with graph.as_default():
--> 427       self._fetch_mapper = _FetchMapper.for_fetch(fetches)
    428     self._fetches = []
    429     self._targets = []

/usr/local/lib/python3.5/dist-packages/tensorflow/python/client/session.py in for_fetch(fetch)
    243     elif isinstance(fetch, (list, tuple)):
    244       # NOTE(touts): This is also the code path for namedtuples.
--> 245       return _ListFetchMapper(fetch)
    246     elif isinstance(fetch, dict):
    247       return _DictFetchMapper(fetch)

/usr/local/lib/python3.5/dist-packages/tensorflow/python/client/session.py in __init__(self, fetches)
    350     """
    351     self._fetch_type = type(fetches)
--> 352     self._mappers = [_FetchMapper.for_fetch(fetch) for fetch in fetches]
    353     self._unique_fetches, self._value_indices = _uniquify_fetches(self._mappers)
    354 

/usr/local/lib/python3.5/dist-packages/tensorflow/python/client/session.py in <listcomp>(.0)
    350     """
    351     self._fetch_type = type(fetches)
--> 352     self._mappers = [_FetchMapper.for_fetch(fetch) for fetch in fetches]
    353     self._unique_fetches, self._value_indices = _uniquify_fetches(self._mappers)
    354 

/usr/local/lib/python3.5/dist-packages/tensorflow/python/client/session.py in for_fetch(fetch)
    240     if fetch is None:
    241       raise TypeError('Fetch argument %r has invalid type %r' % (fetch,
--> 242                                                                  type(fetch)))
    243     elif isinstance(fetch, (list, tuple)):
    244       # NOTE(touts): This is also the code path for namedtuples.

TypeError: Fetch argument None has invalid type <class 'NoneType'>

Addition:

Variable reconstruct (defined at line 327 in tf_cnnvis.py ) is [None, None, None, None, None, None, None, None]. TensorFlow version is 1.9. When I use TensorFlow 1.4, I got same Error as #26

[wildpig22]

What is your input_sensor set to?

Not familiar with faster rcnn, but with mobilenet ssd (from tensorflow object detection model zoo), I got the same error when using image_tensor as input, switching to FeatureExtractor/MobilenetV2/MobilenetV2/input will work.

You can try skipping the pre-processing blocks and set input_tensor to the input of feature extractor block(or whatever the actual input is). You will need to manually do the pre-processing though.

[aggpankaj2]

@wildpig22

Object detection model is not working with deconv_visualization. With activation_visualization works well. import argparse import sys import os import numpy as np import tensorflow as tf from Save_load_graph import load_graph from scipy.misc import imread, imresize import argparse import cv2 import time from tensorflow.examples.tutorials.mnist import input_data from tf_cnnvis import * def evalFrozenGraphModel(FLAGS): graph = load_graph(FLAGS.model_dir + 'frozen_inference_graph.pb' X = graph.get_tensor_by_name('image_tensor:0') y_predict = graph.get_tensor_by_name("detection_classes:0") im = np.expand_dims(imresize(imread(os.path.join("sample_images", "Lenna.png")), (224, 224)), axis=0) with tf.Session(graph=graph) as sess:

is_success = activation_visualization(sess_graph_path=graph, value_feed_dict=feed_dict,

        #                                       layers=layers, path_logdir=os.path.join("Log", "2.9"),
        #                                       path_outdir=os.path.join("Output", "2.9"))
        is_success = deconv_visualization(sess_graph_path=graph, value_feed_dict=feed_dict,
                                                                               input_tensor=None, layers=layers,
                                                                           path_logdir=os.path.join("Log", "1"),
                                                                               path_outdir=os.path.join("Output", "1"))

if name == 'main': parser = argparse.ArgumentParser() parser.add_argument('--model_dir', type=str, default='/resnet101_correcteddata/inference_resnet_1class_1batch/',

Error outer_forward_ctxt = forward_ctxt.outer_context AttributeError: 'NoneType' object has no attribute 'outer_context'

Tried so many things given here https://github.com/InFoCusp/tf_cnnvis/issues/12 and https://github.com/InFoCusp/tf_cnnvis/issues/26 but still getting error.

[jidebingfeng]

@wildpig22 Thanks very much! It works! And I read the source code, find that some tensor in pre-processing is not trainable.

[jidebingfeng]

@aggpankaj2 this is my code. It works.

PATH_TO_CKPT = 'data/tf1.9.5819/frozen_inference_graph.pb'

detection_graph = tf.Graph()
with detection_graph.as_default():
    od_graph_def = tf.GraphDef()
    with tf.gfile.GFile(PATH_TO_CKPT, 'rb') as fid:
        serialized_graph = fid.read()
        od_graph_def.ParseFromString(serialized_graph)
        tf.import_graph_def(od_graph_def, name='')

image_tensor = detection_graph.get_tensor_by_name('image_tensor:0')
input_tensor_placehodler = detection_graph.get_tensor_by_name('FirstStageFeatureExtractor/resnet_v1_101/resnet_v1_101/Pad:0')
image = mpimg.imread('data/1.jpg')
image_np_expanded = np.expand_dims(image, axis=0)

layers = ['FirstStageFeatureExtractor/resnet_v1_101/resnet_v1_101/block1/unit_1/bottleneck_v1/conv1/Conv2D',]

is_success = deconv_visualization(sess_graph_path='data/tf1.9.5819/model.ckpt.meta', value_feed_dict={image_tensor: image_np_expanded},
                                  input_tensor=input_tensor_placehodler,
                                  layers=layers, path_logdir=os.path.join("Log", "Inception5"),
                                  path_outdir=os.path.join("Log", "Inception5"))

Addtion TensorFlow versoin is 1.9.

[jidebingfeng]

@wildpig22 Did you have any idea about use GPU? I have two GPUs, but still take large time. GPU-Util is 0%, Memory-Usage is 7857MiB / 8114MiB Logs:

Reconstruction Completed for FirstStageFeatureExtractor/resnet_v1_101/resnet_v1_101/conv1/Conv2D layer. Time taken = 7.082620 s
Reconstruction Completed for FirstStageFeatureExtractor/resnet_v1_101/resnet_v1_101/block3/unit_1/bottleneck_v1/conv1/Conv2D layer. Time taken = 123.632443 s
Reconstruction Completed for FirstStageFeatureExtractor/resnet_v1_101/resnet_v1_101/block3/unit_9/bottleneck_v1/conv2/Conv2D layer. Time taken = 111.493410 s
Reconstruction Completed for FirstStageFeatureExtractor/resnet_v1_101/resnet_v1_101/block3/unit_17/bottleneck_v1/conv3/Conv2D layer. Time taken = 602.894801 s

[wildpig22]

@jidebingfeng I haven't figured that out either, sometimes I even have to set clear_devices=True to work...

[jidebingfeng]

@wildpig22 In fact, it works well with the code os.environ["CUDA_VISIBLE_DEVICES"] = "0". Some layer is close to input_tensor or it has few GPU operation, so GPU-Util is 0%.

[aggpankaj2]

@jidebingfeng Thank you very much. It work. Had you tried for deconv visualization for mobilenetv2 ?

[jidebingfeng]

@aggpankaj2 No. I just tried for faster_rcnn_resnet101. I think the deconv_visualization function will work well, if you make sure the layers between input_tensor and the deconv layer are trainable.

[jidebingfeng]

@aggpankaj2 You can judge the layer is trainable by it's dtype.That's the way how I find out the problem of faster_rcnn_resnet101 . Some code:

for op in detection_graph.get_operations():
    t = op.type.lower()
    if t == 'maxpool' or t == 'relu' or t == 'conv2d':
        for input in op.inputs:
            print(input.name, input.dtype)

[aggpankaj2]

@jidebingfeng thanks for your reply. Yes i tried same as per your suggestion but not working

[wildpig22]

@aggpankaj2 do you mean mobilenetV2 or mobilenetV2+SSD?

Try this code for mobilenetV2+SSD (tensorflow 1.9.0)

import os
import sys
import time
import copy
import h5py
import numpy as np

from tf_cnnvis import *

import tensorflow as tf
from scipy.misc import imread, imresize

with tf.gfile.FastGFile('frozen.pb', 'rb') as f:
    graph_def = tf.GraphDef()
    graph_def.ParseFromString(f.read())
t_input = tf.placeholder(np.float32, name='FeatureExtractor/MobilenetV2/MobilenetV2/input') 
tf.import_graph_def(graph_def, {'FeatureExtractor/MobilenetV2/MobilenetV2/input':t_input})

im = np.expand_dims(imresize(imread(os.path.join("./images", "image.jpg")), (300, 300)), axis = 0)
mean = 2/255.0
dev = 1.0
im_processed = im * mean - dev

layers = ["r", "p", "c"]

is_success = deconv_visualization(sess_graph_path = tf.get_default_graph(), value_feed_dict = {t_input : im_processed}, 
                                      layers=layers,
                                      path_logdir=os.path.join("./Log","ssd"), 
                                      path_outdir=os.path.join("./Output","ssd"))

[aggpankaj2]

@wildpig22 thanks for your code, but not working.

[wildpig22]

@aggpankaj2 which model are you using, can you provide a link?

Also, what is the error?