jaspereb / fasterrcnntutorial Goto Github PK

Hey, @jaspereb can you help me with this?

I want to compile the TensorFlow Graph to Movidius Graph. I have used Model Zoo's ssd_mobilenet_v1_coco model to train it on my own dataset.

Then I ran

python object_detection/export_inference_graph.py \
				--input_type=image_tensor \
				--pipeline_config_path=/home/redtwo/nsir/ssd_mobilenet_v1_coco.config \
                                --trained_checkpoint_prefix=/home/redtwo/nsir/train/model.ckpt-3362 \
				--output_directory=/home/redtwo/nsir/output

which generates me frozen_interference_graph.pb & saved_model/saved_model.pb

Now to convert this saved model into Movidius graph. There are commands given

Export GraphDef file

python3 ../tensorflow/tensorflow/python/tools/freeze_graph.py \
		--input_graph=inception_v3.pb \
		--input_binary=true \
		--input_checkpoint=inception_v3.ckpt \
		--output_graph=inception_v3_frozen.pb \
		--output_node_name=InceptionV3/Predictions/Reshape_1

Freeze model for inference

python3 ../tensorflow/tensorflow/python/tools/freeze_graph.py \
		--input_graph=inception_v3.pb \
		--input_binary=true \
		--input_checkpoint=inception_v3.ckpt \
		--output_graph=inception_v3_frozen.pb \
		--output_node_name=InceptionV3/Predictions/Reshape_1

which can finally be feed to NCS Intel Movidius SDK

mvNCCompile -s 12 inception_v3_frozen.pb -in=input -on=InceptionV3/Predictions/Reshape_1

All of this is given at Intel Movidius Website here: https://movidius.github.io/ncsdk/tf_modelzoo.html

My model was already trained i.e. output/frozen_inference_graph. Why do I again freeze it using /slim/export_inference_graph.py or it's the output/saved_model/saved_model.py that will go as input to slim/export_inference_graph.py??

All I want is output_node_name=Inceptionv3/Predictions/Reshape_1. How to get this output_name_name directory structure & anything inside it? I don't know what all it contains

what output node should I use for model zoo's ssd_mobilenet_v1_coco model(trained on my own custom dataset)

python freeze_graph.py \
                 --input_graph=/path/to/graph.pbtxt \
                 --input_checkpoint=/path/to/model.ckpt-22480 \
                 --input_binary=false \
                 --output_graph=/path/to/frozen_graph.pb \
                 --output_node_names="the nodes that you want to output e.g. InceptionV3/Predictions/Reshape_1 for Inception V3 "

Things I understand & don't understand:
input_checkpoint: ✓ [check points that were created during training]
output_graph: ✓ [path to output frozen graph]
out_node_names: X

I don't understand out_node_names parameter & what should inside this considering its ssd_mobilnet not inception_v3

System information

• What is the top-level directory of the model you are using:
• Have I written custom code (as opposed to using a stock example script provided in TensorFlow):
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Linux Ubuntu 16.04
• TensorFlow installed from (source or binary): TensorFlow installed with pip
• TensorFlow version (use command below): 1.13.1
• Bazel version (if compiling from source):
• CUDA/cuDNN version: V10.1.168/7.*
• GPU model and memory: 2080Ti 11Gb
• Exact command to reproduce:

Hi Jasper,

On the faster r-cnn model, the training loss starts at 4. on the same dataset, why is it starting from 7 on ssd model?

Hi jaspereb,
I have a problem with ssd. My images are 600x600 size but with resizing in the config file 300x300. Is there no possibility to work 600x600 in this case? Do my training images have to be 300x300?

Hi Jasper,

I want to apply the mask r-cnn. Are there changes in the implementation steps? For example, while training dataset is being created, I need to train with object boundaries rather than being rectangular. Which tool can I use for this?

Hi jaspereb,

I have trained the faster-rcnn model on size of 600x600 images and the result is quite good in 600x600 size. But when I test on a larger size image of 1024x1024, the success is very low. What would be the reason?

You stated that "FasterRcnn does not like objects smaller than about 30*30 pixels". Is this a fact that can be altered somehow?

For example I would like to be able to detect object smaller than 30x30. If I provide my model with an extra smaller anchor scale that goes below 30x30 this would be enough (I mean even after the image resize application that my objects are smaller than 30x30)? Or should I change the receptive field of my model also (which is hardcoded for each model I think)

Firstly, a very useful video. I applied my work as described on the video. What I need to change for smaller objects. How many images have you trained on your dataset, how many positive objects should be selected (1000, 2000, etc)?

I want to use ssd method with vgg16 model, but it is not available in the models. Do not I have a chance to do this in Tensorflow?

Hi @jaspereb will it be possible to try this tutorial on window OS for my own custom dataset ?
Help will be much grateful..