Comments (22)
villanuevab
commented on July 3, 2024
1
@Lisandro79 thanks for your work on this. May I ask how you decided on
["bbox/trimming/bbox:0",
"probability/score:0",
"probability/class_idx:0"]
as the output nodes? These are from the interpretation_graph, and I am trying to understand the difference between this and the forward_graph
from squeezedet.
andreapiso
commented on July 3, 2024
1
Because the operation is not linked by makefile bazel when you compile for
mobile. You could link the operation manually and pass 1 as keep_prob and
it would work.
β¦On Sun, 1 Apr 2018 at 12:11 AM, hoonkai ***@***.***> wrote:
We did optimization for inference and removed dropout (this step is
necessary if you want to run SqueezeDet on some mobile devices)
@Lisandro79 <https://github.com/Lisandro79> @BichenWuUCB
<https://github.com/BichenWuUCB> Can I ask why dropout needs to be
removed? Isn't the dropout layer trivial during inference?
β
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
<#35 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AN_wJqW_ZtWwNK7R0j6u9xCE8ETudlajks5tj6rEgaJpZM4NNDsN>
.
from squeezedet.
keymanchen1215
commented on July 3, 2024
1
Hi all
When I load the original model.ckpt-8700, I check the node name in the graph, it has "image_input".
Then I double check it again after I use the below code, the "image_input" node disappear.
output_graph_def = graph_util.convert_variables_to_constants(sess, input_graph_def, output_node_names.split(",") )
But the "batch/fifo_queue" node mentioned by @venuktan venuktan exists.
The issue should be located in the below codeοΌ
self.image_input, self.input_mask, self.box_delta_input, \ self.box_input, self.labels = tf.train.batch( self.FIFOQueue.dequeue(), batch_size=mc.BATCH_SIZE, capacity=mc.QUEUE_CAPACITY)
Any one know why?
from squeezedet.
Lisandro79
commented on July 3, 2024
In the end I solved this by changing the way I was saving the model to disk.
with tf.Graph().as_default():
mc = kitti_squeezeDetPlus_config()
mc.BATCH_SIZE = 1
mc.LOAD_PRETRAINED_MODEL = False
model = SqueezeDetPlus(mc, FLAGS.gpu)
saver = tf.train.Saver(model.model_params, reshape=True)
graph = tf.get_default_graph()
input_graph_def = graph.as_graph_def()
output_node_names = "bbox/trimming/bbox,probability/class_idx,probability/score"
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
init = tf.initialize_all_variables()
sess.run(init)
tf.train.start_queue_runners(sess=sess)
# Restores from checkpoint
ckpts = set()
ckpt = tf.train.get_checkpoint_state(FLAGS.checkpoint_path)
ckpts.add(ckpt.model_checkpoint_path)
print ('Loading {}...'.format(ckpt.model_checkpoint_path))
saver.restore(sess, ckpt.model_checkpoint_path)
# Use a built-in TF helper to export variables to constants
output_graph_def = graph_util.convert_variables_to_constants(sess,
input_graph_def, output_node_names.split(",") )
# Serialize and dump the output graph to the filesystem
output_graph = "/data/squeezeDet_TF011/logs/test_freeze/test.pb"
with tf.gfile.GFile(output_graph, "wb") as f:
f.write(output_graph_def.SerializeToString())
print(("%d ops in the final graph." % len(output_graph_def.node)))
from squeezedet.
mistiansen
commented on July 3, 2024
Can you offer some advice for getting SqueezeDet to work with your dataset?
Did you just change make your labels in KITTI format and change the number of classes in config?
Thank you
from squeezedet.
Lisandro79
commented on July 3, 2024
Yes, that's all you need to train your own dataset
Cheers
from squeezedet.
mistiansen
commented on July 3, 2024
Thank you. Do you know if changing the image width and height will work? I see the parameters in the config file. I am hoping to work with very large images.
from squeezedet.
Lisandro79
commented on July 3, 2024
Yes, it works perfectly well with other image resolutions. You just need to adjust the parameters for the anchors appropriately
from squeezedet.
mistiansen
commented on July 3, 2024
It seems one would need to make changes to the eval tool (the KITTI-eval tool in dataset/kitti-eval/cpp/evaluate_object.cpp) to handle new detection classes.
Was this necessary for you? Were you able to evaluate accuracy on your custom dataset?
Thanks again
from squeezedet.
Lisandro79
commented on July 3, 2024
No, I do not remember making any modifications to "evaluate_object.cpp". And, yes, I am able to evaluate accuracy on my custom dataset. We did modified the list of classes in the configuration files inside "config" folder
from squeezedet.
mistiansen
commented on July 3, 2024
Thanks. Do you know how to resume training with a model.ckpt file? Or model.ckpt.data file? I tried putting this in pretrained_model_path but it doesn't work. It looks like it wants a .pkl file.
from squeezedet.
Lisandro79
commented on July 3, 2024
Your question is too broad and off topic. There are several tutorials online that show how to do what you are asking. I recommend that you follow them and if you still have difficulties after that, then post a minimal example with a concrete error so that other people can help you.
from squeezedet.
mistiansen
commented on July 3, 2024
OK. I just didn't know if you were able to stop and resume training using this tool and model.ckpt files.
Did you get this to work with SqueezeDet+ on your own data? It seems the rule for grid spacing as 1/16 of image width, height doesn't work there.
from squeezedet.
venuktan
commented on July 3, 2024
I am trying to do inference with the exported .pb and I am getting the following error
Caused by op u'fire7/expand3x3/kernels/read', defined at:
File "inference.py", line 103, in <module>
tf.app.run(main=Inference())
File "inference.py", line 28, in __init__
self.get_results(FLAGS.input_image)
File "inference.py", line 55, in get_results
model = SqueezeDet(mc, FLAGS.gpu)
File "/home/ubuntu/squeezeDet/src/nets/squeezeDet.py", line 24, in __init__
self._add_forward_graph()
File "/home/ubuntu/squeezeDet/src/nets/squeezeDet.py", line 63, in _add_forward_graph
'fire7', fire6, s1x1=48, e1x1=192, e3x3=192, freeze=False)
File "/home/ubuntu/squeezeDet/src/nets/squeezeDet.py", line 104, in _fire_layer
padding='SAME', stddev=stddev, freeze=freeze)
File "/home/ubuntu/squeezeDet/src/nn_skeleton.py", line 533, in _conv_layer
wd=mc.WEIGHT_DECAY, initializer=kernel_init, trainable=(not freeze))
File "/home/ubuntu/squeezeDet/src/nn_skeleton.py", line 66, in _variable_with_weight_decay
var = _variable_on_device(name, shape, initializer, trainable)
File "/home/ubuntu/squeezeDet/src/nn_skeleton.py", line 48, in _variable_on_device
name, shape, initializer=initializer, dtype=dtype, trainable=trainable)
File "/home/ubuntu/.conda/envs/py27sq/lib/python2.7/site-packages/tensorflow/python/ops/variable_scope.py", line 988, in get_variable
custom_getter=custom_getter)
File "/home/ubuntu/.conda/envs/py27sq/lib/python2.7/site-packages/tensorflow/python/ops/variable_scope.py", line 890, in get_variable
custom_getter=custom_getter)
File "/home/ubuntu/.conda/envs/py27sq/lib/python2.7/site-packages/tensorflow/python/ops/variable_scope.py", line 348, in get_variable
validate_shape=validate_shape)
File "/home/ubuntu/.conda/envs/py27sq/lib/python2.7/site-packages/tensorflow/python/ops/variable_scope.py", line 333, in _true_getter
caching_device=caching_device, validate_shape=validate_shape)
File "/home/ubuntu/.conda/envs/py27sq/lib/python2.7/site-packages/tensorflow/python/ops/variable_scope.py", line 684, in _get_single_variable
validate_shape=validate_shape)
File "/home/ubuntu/.conda/envs/py27sq/lib/python2.7/site-packages/tensorflow/python/ops/variables.py", line 226, in __init__
expected_shape=expected_shape)
File "/home/ubuntu/.conda/envs/py27sq/lib/python2.7/site-packages/tensorflow/python/ops/variables.py", line 344, in _init_from_args
self._snapshot = array_ops.identity(self._variable, name="read")
File "/home/ubuntu/.conda/envs/py27sq/lib/python2.7/site-packages/tensorflow/python/ops/gen_array_ops.py", line 1490, in identity
result = _op_def_lib.apply_op("Identity", input=input, name=name)
File "/home/ubuntu/.conda/envs/py27sq/lib/python2.7/site-packages/tensorflow/python/framework/op_def_library.py", line 763, in apply_op
op_def=op_def)
File "/home/ubuntu/.conda/envs/py27sq/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 2395, in create_op
original_op=self._default_original_op, op_def=op_def)
File "/home/ubuntu/.conda/envs/py27sq/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 1264, in __init__
self._traceback = _extract_stack()
FailedPreconditionError (see above for traceback): Attempting to use uninitialized value fire7/expand3x3/kernels
[[Node: fire7/expand3x3/kernels/read = Identity[T=DT_FLOAT, _class=["loc:@fire7/expand3x3/kernels"], _device="/job:localhost/replica:0/task:0/gpu:0"](fire7/expand3x3/kernels)]]
[[Node: probability/score/_7 = _Recv[client_terminated=false, recv_device="/job:localhost/replica:0/task:0/cpu:0", send_device="/job:localhost/replica:0/task:0/gpu:0", send_device_incarnation=1, tensor_name="edge_534_probability/score", tensor_type=DT_FLOAT, _device="/job:localhost/replica:0/task:0/cpu:0"]()]]
I have exported my protobuf this way as instructed above
with tf.gfile.GFile(out_model, "wb") as f:
f.write(output_graph_def.SerializeToString())
from squeezedet.
venuktan
commented on July 3, 2024
@Lisandro79 @mistiansen @BichenWuUCB after you export to a pb whats the input (for image) and output nodes ?
I am using this
det_boxes, det_probs, det_class = sess.run(
["import_5/bbox/trimming/bbox:0", "import_5/probability/score:0", "import_5/probability/class_idx:0"],
feed_dict={"import_5/batch/fifo_queue:0": [input_image]})
I am getting an error
TypeError: Cannot interpret feed_dict key as Tensor: The name 'import_5/batch/fifo_queue:0' refers to a Tensor which does not exist. The operation, 'import_5/batch/fifo_queue', does not exist in the graph.
from squeezedet.
Lisandro79
commented on July 3, 2024
Hi @venuktan,
I think you have the wrong name of the input layer. Here are some snippets of code of what we used. Please note that the input layer is called "Image_Input:0" (not "fifo_queue")
# Test with one image
IMAGE_WIDTH = 720
IMAGE_HEIGHT = 405
BGR_MEANS = np.array([[[103.939, 116.779, 123.68]]])
im_name = './test/test3.jpg'
im = cv2.imread(im_name)
im = im.astype(np.float32, copy=False)
im = cv2.resize(im, (IMAGE_WIDTH, IMAGE_HEIGHT))
input_image = im - BGR_MEANS
cv2.imshow("Image", im)
cv2.imwrite('test.jpg', im.astype(np.int16, copy=False))
cv2.destroyAllWindows()
x = graph.get_tensor_by_name('image_input:0')
bboxes = graph.get_tensor_by_name('bbox/trimming/bbox:0')
scores = graph.get_tensor_by_name('probability/score:0')
clss = graph.get_tensor_by_name('probability/class_idx:0')
keep_prob = graph.get_tensor_by_name('keep_prob:0')
# We launch a Session
with tf.Session(graph=graph) as sess:
det_probs = sess.run(scores, {'image_input:0': [im], 'keep_prob:0': 0.5})
print det_probs
# Unpersists graph from file
with tf.gfile.FastGFile(args.frozen_model_filename, 'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
_ = tf.import_graph_def(graph_def, name='')
with tf.Session() as sess:
x = sess.graph.get_tensor_by_name('image_input:0')
bboxes = sess.graph.get_tensor_by_name('bbox/trimming/bbox:0')
scores = sess.graph.get_tensor_by_name('probability/score:0')
clss = sess.graph.get_tensor_by_name('probability/class_idx:0')
keep_prob = sess.graph.get_tensor_by_name('keep_prob:0')
det_probs = sess.run(scores, {'image_input:0': [input_image], 'keep_prob:0': 0.5})
det_boxes, det_probs, det_class = sess.run([bboxes, scores, clss],
feed_dict={x: [input_image], keep_prob: 1.0})
print det_probs
Hope this helps
from squeezedet.
Lisandro79
commented on July 3, 2024
Hi @mistiansen
If you have the correct numbers for the anchors it should work. What do you mean by "It seems the rule for grid spacing as 1/16 of image width, height doesn't work there"?
You should post a minimal code example with an error so that everybody can understand your problem. It is very difficult to understand the problem from your message.
Cheers
from squeezedet.
Lisandro79
commented on July 3, 2024
I did not "decide" anything, those are the names of the output nodes in the graph.
If you want to check the difference between the graphs, why don't you visualize those graphs and compare the nodes?
Also, @BichenWuUCB provides useful info about this in #73. We did optimization for inference and removed dropout (this step is necessary if you want to run SqueezeDet on some mobile devices)
After freezing for production the speed of inference on a Titan X is ~0.017s for images of 720x400 pixels, which is excellent for our task (great job @BichenWuUCB).
from squeezedet.
villanuevab
commented on July 3, 2024
Thank you, you answered my question. I just wanted to clarify, as stated in #73, which ops and subgraphs were needed for inference.
Thank you and @BichenWuUCB for your helpful responses and great work!
from squeezedet.
hoonkai
commented on July 3, 2024
We did optimization for inference and removed dropout (this step is necessary if you want to run SqueezeDet on some mobile devices)
@Lisandro79 @BichenWuUCB Can I ask why dropout needs to be removed? Isn't the dropout layer trivial during inference?
from squeezedet.
manoj652
commented on July 3, 2024
how to freeze the graph
from squeezedet.
pribadihcr
commented on July 3, 2024
In the end I solved this by changing the way I was saving the model to disk.
with tf.Graph().as_default(): mc = kitti_squeezeDetPlus_config() mc.BATCH_SIZE = 1 mc.LOAD_PRETRAINED_MODEL = False model = SqueezeDetPlus(mc, FLAGS.gpu) saver = tf.train.Saver(model.model_params, reshape=True) graph = tf.get_default_graph() input_graph_def = graph.as_graph_def() output_node_names = "bbox/trimming/bbox,probability/class_idx,probability/score" with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess: init = tf.initialize_all_variables() sess.run(init) tf.train.start_queue_runners(sess=sess) # Restores from checkpoint ckpts = set() ckpt = tf.train.get_checkpoint_state(FLAGS.checkpoint_path) ckpts.add(ckpt.model_checkpoint_path) print ('Loading {}...'.format(ckpt.model_checkpoint_path)) saver.restore(sess, ckpt.model_checkpoint_path) # Use a built-in TF helper to export variables to constants output_graph_def = graph_util.convert_variables_to_constants(sess, input_graph_def, output_node_names.split(",") ) # Serialize and dump the output graph to the filesystem output_graph = "/data/squeezeDet_TF011/logs/test_freeze/test.pb" with tf.gfile.GFile(output_graph, "wb") as f: f.write(output_graph_def.SerializeToString()) print(("%d ops in the final graph." % len(output_graph_def.node)))
@Lisandro79 when I visualize the graph of .pb file using netron, the input node is disappear. ony has queueDequeuemanyV2 or batch node before conv1/convolution node. How to include the input_node?. thanks
from squeezedet.
Related Issues (20)
- Gpu occupancy rate
- where is base_model_config.py?
- Will random initialization parameters have no precision? HOT 1
- Fine-tune SqueezeDet from sparse labels
- How to do hard negative mining HOT 1
- Publish frozen model? HOT 3
- Problem converting to TFLite HOT 3
- low GPU usage
- 8-bit weights
- Deploying squeezeDet on mobile HOT 3
- How to convert checkpoint of squeezedet to frozen graph for tflite conversion?! HOT 1
- Image resolution problem
- How to run demo.py using train.py checkpoint model HOT 1
- Train with different size and Inference with different size.
- Fine tuning with the model
- Train error and Eval error
- Using negative samples for training.
- print weights per layer during training
- Performance issue in src/eval.py (by P3) HOT 1
- The loss plateaus after 100 Epoch
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
π Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. πππ
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google β€οΈ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from squeezedet.