How to visualize the network structure

[JiYuanFeng]

Hi, Is there any easy way to visualize the network? like Netscope for caffe?
I can use the net_drawer in caffe2, but found it;' so hard to read the network?

[taoari]

The following is a code snippet to visualize:

def get_model(cfg_file, weights_file):
    merge_cfg_from_file(cfg_file)
    cfg.TRAIN.WEIGHTS = '' # NOTE: do not download pretrained model weights
    cfg.TEST.WEIGHTS = weights_file
    cfg.NUM_GPUS = 1
    assert_and_infer_cfg()
    model = infer_engine.initialize_model_from_cfg()
    return model

cfg_file = '{}/configs/12_2017_baselines/e2e_mask_rcnn_R-101-FPN_2x.yaml'.format(DETECTRON_ROOT)
weights_file = 'https://s3-us-west-2.amazonaws.com/detectron/' + \
    '35861858/12_2017_baselines/e2e_mask_rcnn_R-101-FPN_2x.yaml.02_32_51.SgT4y1cO/' + \
    'output/train/coco_2014_train:coco_2014_valminusminival/generalized_rcnn/model_final.pkl'
model = get_model(cfg_file, weights_file)

from caffe2.python import net_drawer
g = net_drawer.GetPydotGraph(model, rankdir="TB")
g.write_dot(model.Proto().name + '.dot')

Then you can use a dot viewer (e.g. xdot) to visualize.

[JiYuanFeng]

Thank you taoari !
I have tested the code in my caffe2 enviorment, And I meet another problem,
File "/home/yuanfeng/Project/Detectron/lib/modeling/ResNet.py", line 95, in add_ResNet_convX_body p = model.AffineChannel(p, 'res_conv1_bn', dim=64, inplace=True) File "/home/yuanfeng/Project/Detectron/lib/modeling/detector.py", line 102, in AffineChannel return self.net.AffineChannel([blob_in, scale, bias], blob_in) File "/home/yuanfeng/Project/caffe2-master/build/caffe2/python/core.py", line 2046, in __getattr__ ",".join(workspace.C.nearby_opnames(op_type)) + ']' AttributeError: Method AffineChannel is not a registered operator. Did you mean: []

i think the original caffe2 don't register the AffinaChannel op,which is belonged to the detectron.
So should i added this op to the caffe2?

[taoari]

@JiYuanFeng You will need the imports in https://github.com/facebookresearch/Detectron/blob/master/tools/infer_simple.py

[JiYuanFeng]

@taoari Thank you for your help！
the problem has solved

[gadcam]

@taoari @JiYuanFeng Maybe you could do a pull request with your script so everyone could benefit from it? :)

[JiYuanFeng]

@gadcam i will do it ~

[liujing1995]

nice

[PacteraOliver]

@taoari

Hi, thank you for your code.

I use part of the code and make it as a tool.
I just start a pull request at #508

[drcege]

@PacteraOliver @taoari
Hi, could you really visualize the graph?

I tried GetPydotGraph and GetPydotGraphMinimal, but it hangs too long like this.

[PacteraOliver]

Yes, the whole program need to load the model (pkl file) and the config file.
The output file is a .dot file. If you open with xdot, the sample is:

The sample part of the graph is like that:

It does not take long time.

[WuZhuoran]

@drcege

Hi, if you like you could try the code in pull request #508

Could you provide more details with the model you want to visualize? You did you gpu to load model right?

[drcege]

@WuZhuoran I tried your pull request, and it works well. One question: the visualization is for model definition, why we must provide a weights_file? Can we get rid of it?

Your code could draw an inference net. However, my problem is that I want to visualize the training net of e2e_faster_rcnn_R-50-FPN_1x.yaml (I have changed it to one gpu). I saved the dot file just after the default proto dump:

Detectron/detectron/utils/train.py

Lines 164 to 169 in dc76799

	nu.broadcast_parameters(model)
	workspace.CreateNet(model.net)
	logger.info('Outputs saved to: {:s}'.format(os.path.abspath(output_dir)))
	dump_proto_files(model, output_dir)

    g = net_drawer.GetPydotGraph(model.net, rankdir='TB')
    g.write('train_net.dot')

I also tried GetPydotGraphMinimal. But even with the minimal dot file, it takes more than 2 hours for xdot to pop up the graph window! Do you have any experience in visualizing training nets?

[drcege]

I think maybe this is because there are too many gradient/lr related nodes. Actually, my purpose is to include the external input nodes (e.g., labels). The gradient nodes are unnecessary.

[tonbing]

I tried to use the same process to get visualize the network but I am getting the following error? Anyone knows how to resolve that?

Traceback (most recent call last):
File "visual.py", line 18, in
model = get_model(cfg_file, weights_file)
File "visual.py", line 12, in get_model
model = infer_engine.initialize_model_from_cfg(1)
File "/detectron/detectron/core/test_engine.py", line 330, in initialize_model_from_cfg
model, weights_file, gpu_id=gpu_id,
File "/detectron/detectron/utils/net.py", line 61, in initialize_gpu_from_weights_file
with open(weights_file, 'r') as f:
TypeError: coercing to Unicode: need string or buffer, int found

[gadcam]

@tonbing Looks like weights_file is an int in this portion of the code. I would first double check that the content of this variable is correct.

[tonbing]

Hey @gadcam thanks for the reply, I figured the problem out. Above code had some issue, that's why it was giving me the above error. But the code for #508 request worked just fine with some tweaks according to the requirement of my model. I believe the code should have worked also, but it did not.

[jiec-msft]

I use anaconda with python2.7.
I installed

• graphviz-2.40.1-h21bd128_1.tar.bz2 from https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main/linux-64/graphviz-2.40.1-h21bd128_1.tar.bz2
• pydot-1.2.3-py27h55c791f_0.tar.bz2 from https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main/linux-64/pydot-1.2.3-py27h55c791f_0.tar.bz2
  and here is my code, it works fine.

#!/usr/bin/env python
# -×- coding: utf-8 -*-
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals

from collections import defaultdict
import argparse
import cv2  # NOQA (Must import before importing caffe2 due to bug in cv2)
import glob
import logging
import os
import sys
reload(sys) 
sys.setdefaultencoding('utf8') # 设置默认编码格式为'utf-8'

import time


from caffe2.python import workspace

from detectron.core.config import assert_and_infer_cfg
from detectron.core.config import cfg
from detectron.core.config import merge_cfg_from_file
from detectron.utils.io import cache_url
from detectron.utils.logging import setup_logging
from detectron.utils.timer import Timer
import detectron.core.test_engine as infer_engine
import detectron.datasets.dummy_datasets as dummy_datasets
import detectron.utils.c2 as c2_utils
import detectron.utils.vis as vis_utils

c2_utils.import_detectron_ops()

# OpenCL may be enabled by default in OpenCV3; disable it because it's not
# thread safe and causes unwanted GPU memory allocations.
cv2.ocl.setUseOpenCL(False)

def get_model(cfg_file, weights_file):
    merge_cfg_from_file(cfg_file)
    cfg.TRAIN.WEIGHTS = '' # NOTE: do not download pretrained model weights
    cfg.TEST.WEIGHTS = weights_file
    cfg.NUM_GPUS = 1
    assert_and_infer_cfg()
    model = infer_engine.initialize_model_from_cfg(cfg.TEST.WEIGHTS)
    return model

# cfg_file = '{}/configs/12_2017_baselines/e2e_mask_rcnn_R-101-FPN_2x.yaml'.format(DETECTRON_ROOT)
cfg_file = '/home/ceej/DetectronRepo/detectron/configs/12_2017_baselines/defects_X-101.yaml'
# weights_file = 'https://s3-us-west-2.amazonaws.com/detectron/' + \
#     '35861858/12_2017_baselines/e2e_mask_rcnn_R-101-FPN_2x.yaml.02_32_51.SgT4y1cO/' + \
#     'output/train/coco_2014_train:coco_2014_valminusminival/generalized_rcnn/model_final.pkl'
weights_file = '/home/ceej/aluminium_surface_defects_detection/m.pkl'
model = get_model(cfg_file, weights_file)

from caffe2.python import net_drawer
g = net_drawer.GetPydotGraph(model, rankdir="TB")
print(model.Proto().name + '.dot')
g.write_dot(model.Proto().name + '.dot')

[jiec-msft]

By the way, I installed xdot on my ubuntu 16.04 lts with

sudo apt-get update
sudo apt-get install xdot

[jiec-msft]

And the result can be like this:

https://note.youdao.com/yws/api/personal/file/WEB86b1c61048404045fa8a514b209dfe06?method=getImage&version=22336&cstk=vij4EFEr

The url listed above doesn't support firefox