distant-viewing / dvt Goto Github PK
View Code? Open in Web Editor NEWDistant Viewing Toolkit for the Analysis of Visual Culture
Home Page: https://distantviewing.org
License: GNU General Public License v2.0
Distant Viewing Toolkit for the Analysis of Visual Culture
Home Page: https://distantviewing.org
License: GNU General Public License v2.0
Describe the bug
DiffAnnotator and CutAggregator do not work with black and white films. I apologize if this is from lack of user knowledge, rather than a bug/missing feature.
To Reproduce
Steps to reproduce the behavior:
Expected behavior
Shots to be detected. This works perfectly fine with color film, but not at all for any black and white movie to which I have access.
Desktop (please complete the following information):
Describe the bug
When running the included sample Colab notebook I hit a AttributeError: 'Node' object has no attribute 'output_masks'
while attempting to initialize the VideoCsvPipeline
!
To Reproduce
from dvt.pipeline.csv import VideoCsvPipeline
from os.path import join
import pandas as pd
from IPython.display import Image
VideoCsvPipeline(finput="video-clip.mp4", dirout="dvt-output-csv").run()
(As a sidenote, the notebook passes include_images=True
to the VideoCsvPipeline
constructor but it seems that kwarg has been removed?)
Expected behavior
I'm not sure what should happen yet! I'm just kicking the tires and seeing what follows :)
Screenshots
NA
Desktop (please complete the following information):
Smartphone (please complete the following information):
NA
Additional context
I'm happy to look into this if that's helpful!
It shows the file you have requested does not exist.
Basically because HSV is not derived from a color awareness model like CIE-Lab is but in comparison to CIE-Lab defines more meaningful axis I wanted to build a custom annotator that turns the input RGB (it is RGB, right? Not openCVs BGR?) into CIE-LCH (Luminance, Chroma, Hue), a color awareness appropriate version of HSV. (Instead of a histogramm I calculate the median value for each channel, but that is not important here). openCV, unfortunately, does not provide a conversion to CIE-LCH, but scikit-image does. In principle everything works out, BUT for a 110 minute movie such as Il Divo with this custom annotator dvt requires 7h30m!!! Considering, that the demo project, on my computer, requires just 30-45 minutes to annotate everything it can (the color annotator and the dominant colors annotator included), I wondered if I got something completely wrong.
My annotator looks like this:
class LchAnnotator(FrameAnnotator):
name = 'lch'
def annotate(self, batch):
img = batch.get_batch()
fnames = batch.get_frame_names()
luminance = list()
chroma = list()
hue = list()
for i in img:
img_lab = rgb2lab(i)
img_lch = lab2lch(img_lab)
lch = img_lch.reshape(-1, 3)
luminance.append(median(lch[:, 0]))
chroma.append(median(lch[:, 1]))
hue.append(median(lch[:, 2]))
output = {'frame': fnames,
'luminance': luminance,
'chroma': chroma,
'hue': hue}
return output
Hi!
One of the annotators I would like to add to dvt is based on the optical flow between subsequent frames (with cv2's calcOpticalFlowFarneback). Yet, I'm not sure how to do this nicely, because for each batch it wouldn't be possible to calculate the optical flow for the last frame in the batch.
Additionally, in our project we segment a video into shots first, before analysing it, and then extract optical flow only for frames within the same shot, as the optical flow between two frames from different cuts is not very meaningful. I guess this can be solved during aggregation, where the frame before the cut is ignored (if necessary), but it would result in a decent amount of wasted computation for videos with many cuts..
Did you have any thoughts on what the 'right' way would be to deal with these kind of annotators within the dvt pipeline? I took a look at the DiffAnnotator, but that appears to be very 'batch oriented', which isn't always ideal I would think.
This issue is part of my review in openjournals/joss-reviews#1800.
In various forms of documentation, I find the consistent claim that dvt
is licensed under GPL v2. I found this in the following places, but there may be more.
The license that is actually included in the source code is the GPL v3: https://github.com/distant-viewing/dvt/blob/master/LICENSE.
Regarding the shield in the README, this could update automatically with the information available on PyPI if you were to use the following image URL (documentation and options here):
https://img.shields.io/pypi/l/dvt
Note that this reflects the information in the setup.py rather than the information in the actual license.
The version on PyPI is 1.0.0 still although the github repo shows 1.0.1. If the most recent commits (PR 36 merge) are considered good, maybe even do a 1.0.2 release?
Disclaimer: I'm not using dvt
but someone on discuss.python.org
was asking about an install problem and I noticed they were getting dvt
from github instead of PyPI (still using pip
but with a github URL). I was hoping to suggest to them they could just pip install dvt
.
Thank you.
This issue is part of my review in openjournals/joss-reviews#1800. However, in this issue I'm only providing suggestions for minor improvements to the documentation. I consider all of these suggestions optional; the review criteria can also be met without adopting them.
dvt
in an Anaconda environment. I tried installing the package in a regular "bare" virtualenv instead and this worked as documented, too. So you could add a clarifying note that Anaconda is not strictly required. A regular virtualenv, or in fact any Python 3.7 environment with pip
and a compiler toolchain to back it, should work..py
file, assuming that I would encounter a way to invoke the code from the command line later on. Instead, I found that the third code block already contained example output, so I was supposed to enter the code into an interactive Python console instead (this worked fine in my "bare" virtualenv, too). It would be helpful to mention this before introducing the example code. Another option is to include the >>>
prompt in the example displays, following the pydoc/doctest convention. You already did this in the DataExtraction API documentation. Doing both might be best.diff
, demo_anno
and demo_agg
examples. In diff
and demo_agg
the differences were all less than 1 and I'm generally not so impressed by differences in floating point numbers, but I found the differences in the demo_anno
output quite apparent because I don't expect different machines to disagree about pixel values in a video file. Has the example video somehow been edited after the tutorials were written? Published example output here, my output below.dextra.get_data()['demo_anno'].head()
frame red green blue
0 0 6 1 0
1 1 6 1 0
2 2 6 1 0
3 3 6 1 0
4 4 4 2 0
python3 -m http.server –directory dvt-output-data
which has an N-dash unicode character instead of two regular dashes before the directory
option. This will fail if users try to copy-paste the command. I don't know Sphinx very well, but I suspect this can be fixed by convincing Sphinx that the following line is code:Line 19 in c833dc8
VisualInput
abstract base class, documented here and code referenced below, the next_batch
method strongly reminds me of the special __next__
method that is used in Python's iterator convention (documented here). Adopting this convention would permit more declarative for batch in my_video
loops with my_video
an instance of VisualInput
.Lines 14 to 55 in c833dc8
dvt.utils.sub_image
documentation, the notion of a three-dimensional image deserves a bit more explanation. What is the third dimension? Color channels?dvt.annotate.cielab.CIElabAnnotator.num_buckets
documentation is very confusing because the documented default bucket size of 16 does not meet the documented requirement of being divisible by 256. Is it the other way round (should the bucket size be a divisor of 256, i.e., one of the first seven powers of 2) or does the requirement perhaps apply only when the bucket size exceeds some particular threshold?dvt.annotate.diff.DiffAnnotator.size
, I presume the intention is “(Size of one side of the square) (used for down sampling [sic] the image)”, but I first read it as “(Size of one side of (the square used for down sampling [sic] the image))”, in which case it is ambiguous whether this square represents the target image size after resampling or a resampling window size (e.g. for a convolution). Perhaps it would be clearer to write “Side length of the square that the image will be downsampled to for the pixel-by-pixel comparison”.ImgAnnotator
.dvt.annotate.obj.ObjectAnnotator
has multiple occurrences of the word “face” that should probably read “object”.This issue is part of my review in openjournals/joss-reviews#1800.
One of the requirements on the paper is to describe how your package compares to other commonly-used packages. Currently, the second paragraph of the summary mentions the existence of numerous standalone algorithm implementations but “limited options” for building end-to-end pipelines. It may be valuable to name these limited options explicitly and discuss their relative advantages and disadvantages.
Generally speaking, the documentation is very well-written. As the software also focusses on layperson users, some parts of the documentation should be expanded.
I'm currently reviewing the JOSS submission of dvt
(openjournals/joss-reviews#1800). One of the review criteria (per JOSS documentation) is that "the full list of paper authors seems appropriate and complete".
While I trust that the author list is correct, I am currently unable to objectively confirm this from the information available to me on GitHub:
Could you please clarify why @nolauren is included in the author list and why @Nanne is not? Please let me emphasize that I trust that there is a good reason; I'm just asking because this reason is currently invisible to me and I need it in order to confirm one of the review criteria. Thanks in advance.
I used pip to install dvt and got a deprecation warning:
DEPRECATION: dvt is being installed using the legacy 'setup.py install' method, because it does not have a 'pyproject.toml' and the 'wheel' package is not installed. pip 23.1 will enforce this behaviour change. A possible replacement is to enable the '--use-pep517' option. Discussion can be found at pypa/pip#8559
The pyproject.toml approach is newer to me, see https://peps.python.org/pep-0621/ or some background and documentation.
I also know from looking into this from my other projects, you may able to use pdm to import your existing setup.py and create a pyproject file
This issue is part of my review in openjournals/joss-reviews#1800.
One of the review requirements is that there is a documented way to verify that the software works correctly. While there is a test suite, I was not able to find any instructions on how to run it. From searching and trial and error, it appears that these are the steps needed in order to run all tests with coverage and without errors after cloning the repository, although I still got a lot of warnings:
pip install .[tests,optional]
cd tests
pytest --cov=dvt
If installing from PyPI instead of cloning, substitute dvt
for .
.
Please include these instructions (or similar) in the Readme. It is also worth mentioning in the Readme that running all tests takes quite long; on my machine, it took nearly 15 minutes.
To simulate to different usage scenarios, I tried to install the software on two plattforms:
Hi there!
Great work on this, this is excellent.
Reporting an error that I did manage to fix on my own but thought you should be aware of.
Heads up: I'm running dvt as a git submodule which I then install into my virtual environment (which i suspect could be linked to the problem).
I was trying to run dvt.AnnoShotBreaks()
, and dvt attempted to download this file : https://github.com/distant-viewing/dvt/releases/download/0.0.1/dvt_detect_shots.pt
(I'm assuming a ML model?)
I was getting the following error:
Traceback (most recent call last):
File "/Users/jacob/Documents/Git Repos/plozevet-archive/Scripts/DVT-Tests/breakpoint-test.py", line 23, in <module>
process(params)
File "/Users/jacob/Documents/Git Repos/plozevet-archive/Scripts/DVT-Tests/breakpoint-test.py", line 19, in process
anno_breaks = dvt.AnnoShotBreaks()
File "/Users/jacob/Documents/Git Repos/plozevet-archive/venv/lib/python3.10/site-packages/dvt/shots.py", line 26, in __init__
model_path = _download_file("dvt_detect_shots.pt")
File "/Users/jacob/Documents/Git Repos/plozevet-archive/venv/lib/python3.10/site-packages/dvt/utils.py", line 135, in _download_file
download_url_to_file(url, cached_file, hash_prefix, progress=True)
File "/Users/jacob/Documents/Git Repos/plozevet-archive/venv/lib/python3.10/site-packages/torch/hub.py", line 625, in download_url_to_file
f = tempfile.NamedTemporaryFile(delete=False, dir=dst_dir)
File "/usr/local/Cellar/[email protected]/3.10.12/Frameworks/Python.framework/Versions/3.10/lib/python3.10/tempfile.py", line 559, in NamedTemporaryFile
file = _io.open(dir, mode, buffering=buffering,
File "/usr/local/Cellar/[email protected]/3.10.12/Frameworks/Python.framework/Versions/3.10/lib/python3.10/tempfile.py", line 556, in opener
fd, name = _mkstemp_inner(dir, prefix, suffix, flags, output_type)
File "/usr/local/Cellar/[email protected]/3.10.12/Frameworks/Python.framework/Versions/3.10/lib/python3.10/tempfile.py", line 256, in _mkstemp_inner
fd = _os.open(file, flags, 0o600)
FileNotFoundError: [Errno 2] No such file or directory: '/Users/jacob/.cache/torch/hub/checkpoints/tmps3k5bm1g'
In the end I went and manually created the path /torch/hub/checkpoints/
. Sounds like a problem that could be easily resolved with a Path(dir).mkdir(parents=True)
? Or maybe the issue is my submodule/venv setup.
Anyway, all working my end and having fun with the toolkit!
Very much appreciate the colab notebook, as I find the documentation a bit hard to parse, given my low skill level. With regard to working with video it'd be good to have code snippets demonstrating the other annotators on video as in examples 9 - 12 using images.
Hiya,
Just trying to follow the minimal example. I've got a new python 3.7 environment (made with conda), I pip install dvt
then give python3 -m dvt video-viz target.mp4
a whirl:
ModuleNotFoundError: No module named 'tensorflow.keras.layers.experimental.preprocessing'
and then
Keras requires TensorFlow 2.2 or higher.
But if I installed tf2.2, I get dvt 0.3.3 requires tensorflow==1.15.2, but you have tensorflow 2.2.0 which is incompatible.
so I'm assuming that I've got something borked on my machine.
When following the tutorial and after launching the local webserver with the provided mp4 file, further navigation becomes impossible. The arrow icons are inactive as well as the scroll bar.
Checked with Firefox 71, Opera 60.0.3255.109, and Chrome 79.
I'm trying to run the face annotation code example in the readme and getting errors. I think I have everything installed correctly. I've tried in Google Colab, jupyter, and python console, so I don't think it's my environment.
Here's the error from running dvt.AnnoFaces()
in a jupyter notebook:
Downloading: "https://github.com/distant-viewing/dvt/releases/download/0.0.1/dvt_detect_face.pt" to /Users/rkoeser/.cache/torch/hub/checkpoints/dvt_detect_face.pt
---------------------------------------------------------------------------
FileNotFoundError Traceback (most recent call last)
Cell In[3], line 1
----> 1 anno_face = dvt.AnnoFaces()
File ~/workarea/env/dvt/lib/python3.9/site-packages/dvt/face.py:32, in AnnoFaces.__init__(self, model_path_mtcnn, model_path_vggface)
29 def __init__(self, model_path_mtcnn=None, model_path_vggface=None):
31 if not model_path_mtcnn:
---> 32 model_path_mtcnn = _download_file("dvt_detect_face.pt")
34 if not model_path_vggface:
35 model_path_vggface = _download_file("dvt_face_embed.pt")
File ~/workarea/env/dvt/lib/python3.9/site-packages/dvt/utils.py:131, in _download_file(url, basename)
129 r = HASH_REGEX.search(filename) # r is Optional[Match[str]]
130 hash_prefix = r.group(1) if r else None
--> 131 download_url_to_file(url, cached_file, hash_prefix, progress=True)
133 return cached_file
File ~/workarea/env/dvt/lib/python3.9/site-packages/torch/hub.py:611, in download_url_to_file(url, dst, hash_prefix, progress)
609 dst = os.path.expanduser(dst)
610 dst_dir = os.path.dirname(dst)
--> 611 f = tempfile.NamedTemporaryFile(delete=False, dir=dst_dir)
613 try:
614 if hash_prefix is not None:
File ~/.pyenv/versions/3.9.13/lib/python3.9/tempfile.py:545, in NamedTemporaryFile(mode, buffering, encoding, newline, suffix, prefix, dir, delete, errors)
542 if _os.name == 'nt' and delete:
543 flags |= _os.O_TEMPORARY
--> 545 (fd, name) = _mkstemp_inner(dir, prefix, suffix, flags, output_type)
546 try:
547 file = _io.open(fd, mode, buffering=buffering,
548 newline=newline, encoding=encoding, errors=errors)
File ~/.pyenv/versions/3.9.13/lib/python3.9/tempfile.py:255, in _mkstemp_inner(dir, pre, suf, flags, output_type)
253 _sys.audit("tempfile.mkstemp", file)
254 try:
--> 255 fd = _os.open(file, flags, 0o600)
256 except FileExistsError:
257 continue # try again
FileNotFoundError: [Errno 2] No such file or directory: '/Users/rkoeser/.cache/torch/hub/checkpoints/tmp3fu4p_3n'
I can't tell where it's going wrong, but hopefully this is meaningful to you. I tested manually and the .pt
file does download.
Is there a reason these files are separate downloads? Would it make sense to include them as part of the python package?
Describe the bug
utils._which_frames
does not filter frames as one would expect with respect to the freq
-Attribute. It applies freq
only within a present batch but not across batches for the entire movie. This leads to inconsistent steps in the output data.
To Reproduce
dextra = DataExtraction(FrameInput(input_path="./VID-20191006-WA0014.mp4")) # default bsize=256
dextra.run_annotators([LchAnnotatorRay(freq=10)], max_batch=2) #my custom annotator, using utils._which_frames function
which gives me data for the frames: [0, 10, 20, ..., 240, 250, 256, 266, 276, ..., 500, 510]
you see wrong frame step after switching from batch 1 to batch 2. This makes sense when looking at the implementation of utils._which_frames
:
Line 218 in f4e5f89
Regardless of the starting point of the current batch, the selection of frames always starts with 0 and the first frame in the second batch is frame 256.
Expected behavior
The correct frame sequence in the above example should be [0, 10, 20, ..., 240, 250, 260, 270, ..., 500, 510]
.
The first frame of the second batch should be 260 which equates to the index 4 (not 0).
Desktop (please complete the following information):
Additional context
I developed a fix which for which I would make a pull request if you agree upon my expected behaviour. It works like this:
The freq
minus the rest of bnum
times the bsize
divided by the freq
for bnum
greater than 0 otherwise 0.
if frames is None:
# return list(range(0, batch.bsize, freq))
first_frame = 0
_, rest = divmod(batch.bnum * batch.bsize, freq)
if rest != 0:
first_frame = freq - rest
return list(range(first_frame, batch.bsize, freq))
This works well for me so far.
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google ❤️ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.