I am trying to create (what I thought was) a simple image classification pipeline between s3 and SageMaker.
Images are stored in an s3 bucket with their class labels in their file names currently, e.g.
My-s3-bucket-dir
cat-1.jpg
dog-1.jpg
cat-2.jpg
..
I've been trying to leverage several related example .py scripts, but most seem to be download data sets already in .rec format or containing special manifest or annotation files I don't have.
All I want is to pass the images from s3 to the SageMaker image classification algorithm that's located in the same region, IAM account, etc. I suppose this means I need a .lst file
When I try to manually create the .lst it doesn't seem to like it and it also takes too long doing manual work to be a good practice.
How can I automatically generate the .lst file (or otherwise send the images/classes for training)?
Things I read made it sound like im2rec.py was a solution, but I don't see how. The example I'm working with now is
Image-classification-fulltraining-highlevel.ipynb
but it seems to download the data as .rec,
download('http://data.mxnet.io/data/caltech-256/caltech-256-60-train.rec')
download('http://data.mxnet.io/data/caltech-256/caltech-256-60-val.rec')
which just skips working with the .jpeg files. I found another that converts them to .rec but again it has essentially the .lst already as .json and just converts it.
I have mostly been working in a Python Jupyter notebook within the AWS console (in my browser) but I have also tried using their GUI.
How can I simply and automatically generate the .lst or otherwise get the data/class info into SageMaker without manually creating a .lst file?
Update
It looks like im2py can't be run against s3. You'd have to completely download everything from all s3 buckets into the notebook's storage...
Please note that [...] im2rec.py is running locally,
therefore cannot take input from the S3 bucket. To generate the list
file, you need to download the data and then use the im2rec tool. - AWS SageMaker Team
There are 3 options to provide annotated data to the Image Classification algo: (1) packing labels in recordIO files, (2) storing labels in a JSON manifest file ("augmented manifest" option), (3) storing labels in a list file. All options are documented here: https://docs.aws.amazon.com/sagemaker/latest/dg/image-classification.html.
Augmented Manifest and .lst files option are quick to do since they just require you to create an annotation file with a usually quick for loop for example. RecordIO requires you to use im2rec.py tool, which is a little more work.
Using .lst files is another option that is reasonably easy: you just need to create annotation them with a quick for loop, like this:
# assuming train_index, train_class, train_pics store the pic index, class and path
with open('train.lst', 'a') as file:
for index, cl, pic in zip(train_index, train_class, train_pics):
file.write(str(index) + '\t' + str(cl) + '\t' + pic + '\n')
Related
We have recently made changes to how we connect to ADLS from Databricks which have removed mount points that were previously established within the environment. We are using databricks to find points in polygons, as laid out in the databricks blog here: https://databricks.com/blog/2019/12/05/processing-geospatial-data-at-scale-with-databricks.html
Previously, a chunk of code read in a GeoJSON file from ADLS into the notebook and then projected it to the cluster(s):
nights = gpd.read_file("/dbfs/mnt/X/X/GeoSpatial/Hex_Nights_400Buffer.geojson")
a_nights = sc.broadcast(nights)
However, the new changes that have been made have removed the mount point and we are now reading files in using the string:
"wasbs://Z#Y.blob.core.windows.net/X/Personnel/*.csv"
This works fine for CSV and Parquet files, but will not load a GeoJSON! When we try this, we get an error saying "File not found". We have checked and the file is still within ADLS.
We then tried to copy the file temporarily to "dbfs" which was the only way we had managed to read files previously, as follows:
dbutils.fs.cp("wasbs://Z#Y.blob.core.windows.net/X/GeoSpatial/Nights_new.geojson", "/dbfs/tmp/temp_nights")
nights = gpd.read_file(filename="/dbfs/tmp/temp_nights")
dbutils.fs.rm("/dbfs/tmp/temp_nights")
a_nights = sc.broadcast(nights)
This works fine on the first use within the code, but then a second GeoJSON run immediately after (which we tried to write to temp_days) fails at the gpd.read_file stage, saying file not found! We have checked with dbutils.fs.ls() and can see the file in the temp location.
So some questions for you kind folks:
Why were we previously having to use "/dbfs/" when reading in GeoJSON but not csv files, pre-changes to our environment?
What is the correct way to read in GeoJSON files into databricks without a mount point set?
Why does our process fail upon trying to read the second created temp GeoJSON file?
Thanks in advance for any assistance - very new to Databricks...!
Pandas uses the local file API for accessing files, and you accessed files on DBFS via /dbfs that provides that local file API. In your specific case, the problem is that even if you use dbutils.fs.cp, you didn't specify that you want to copy file locally, and it's by default was copied onto DBFS with path /dbfs/tmp/temp_nights (actually it's dbfs:/dbfs/tmp/temp_nights), and as result local file API doesn't see it - you will need to use /dbfs/dbfs/tmp/temp_nights instead, or copy file into /tmp/temp_nights.
But the better way would be to copy file locally - you just need to specify that destination is local - that's done with file:// prefix, like this:
dbutils.fs.cp("wasbs://Z#Y.blob.core.windows.net/...Nights_new.geojson",
"file:///tmp/temp_nights")
and then read file from /tmp/temp_nights:
nights = gpd.read_file(filename="/tmp/temp_nights")
I’m building out a pipeline that should execute and train fairly frequently. I’m following this: https://learn.microsoft.com/en-us/azure/machine-learning/service/how-to-create-your-first-pipeline
Anyways, I’ve got a stream analytics job dumping telemetry into .json files on blob storage (soon to be adls gen2). Anyways, I want to find all .json files and use all of those files to train with. I could possibly use just new .json files as well (interesting option honestly).
Currently I just have the store mounted to a data lake and available; and it just iterates the mount for the data files and loads them up.
How can I use data references for this instead?
What does data references do for me that mounting time stamped data does not?
a. From an audit perspective, I have version control, execution time and time stamped read only data. Albeit, doing a replay on this would require additional coding, but is do-able.
As mentioned, the input to the step can be a DataReference to the blob folder.
You can use the default store or add your own store to the workspace.
Then add that as an input. Then when you get a handle to that folder in your train code, just iterate over the folder as you normally would. I wouldnt dynamically add steps for each file, I would just read all the files from your storage in a single step.
ds = ws.get_default_datastore()
blob_input_data = DataReference(
datastore=ds,
data_reference_name="data1",
path_on_datastore="folder1/")
step1 = PythonScriptStep(name="1step",
script_name="train.py",
compute_target=compute,
source_directory='./folder1/',
arguments=['--data-folder', blob_input_data],
runconfig=run_config,
inputs=[blob_input_data],
allow_reuse=False)
Then inside your train.py you access the path as
parser = argparse.ArgumentParser()
parser.add_argument('--data-folder', type=str, dest='data_folder', help='data folder')
args = parser.parse_args()
print('Data folder is at:', args.data_folder)
Regarding benefits, it depends on how you are mounting. For example if you are dynamically mounting in code, then the credentials to mount need to be in your code, whereas a DataReference allows you to register credentials once, and we can use KeyVault to fetch them at runtime. Or, if you are statically making the mount on the machine, you are required to run on that machine all the time, whereas a DataReference can dynamically fetch the credentials from any AMLCompute, and will tear that mount down right after the job is over.
Finally, if you want to train on a regular interval, then its pretty easy to schedule it to run regularly. For example
pub_pipeline = pipeline_run1.publish_pipeline(name="Sample 1",description="Some desc", version="1", continue_on_step_failure=True)
recurrence = ScheduleRecurrence(frequency="Hour", interval=1)
schedule = Schedule.create(workspace=ws, name="Schedule for sample",
pipeline_id=pub_pipeline.id,
experiment_name='Schedule_Run_8',
recurrence=recurrence,
wait_for_provisioning=True,
description="Scheduled Run")
You could pass pointer to folder as an input parameter for the pipeline, and then your step can mount the folder to iterate over the json files.
I am using Google Colabs for my research in machine learning.
I do many variations on a network and run them saving the results.
I have a part of my notebook that used to be separate file (network.py) At the start of a training session I used to save this file in a directory that has the results and logs etc. Now that this part of the code is in the notebook it is easier to edit etc, BUT I do not have a file to copy to the output directory that describes the model. how to i take a section of a google colab notebook and save the raw code as a python file?
Things I have tried:
%%writefile "my_file.py" - is able to write the file however the classes are not available to the runtime.
Is there any way one can append two files in GCS, suppose file one is a full
load and second file is an incremental load. Then what's the way we can append
the two?
Secondly, using gsutil compose will append the two files including the attributes
names as well. So, in the final file I want the data of the two files.
You can append two separate files using compose in the Google Cloud Shell and rename the output file as the first file, like this:
gsutil compose gs://bucket/obj1 [gs://bucket/obj2 ...] gs://bucket/obj1
This command is meant for parallel uploads in which you divide a large object file in smaller objects. They get uploaded to Google Cloud Storage and then you can append them to get the original file. You can find more information on Composite Objects and Parallel Uploads.
I've come up with two possible solutions:
Google Cloud Function solution
The option I would go for is using a Cloud Function. Doing something like the following:
Create an empty bucket like append_bucket.
Upload the first file.
Create a Cloud Function to be triggered by new uploaded files on the
bucket.
Upload the second file.
Read the first and the second file (you will have to download them as string first).
Make the append operation.
Upload the result to the bucket.
Google Dataflow solution
You can also do it with Dataflow for BigQuery (keep in mind it’s still in beta).
Create a BigQuery dataset and table.
Create a Dataflow instance, from the template Cloud Storage Text to BigQuery.
Create a Javascript file with the logic to transform the text.
Upload your files in Json format to the bucket.
Dataflow will read the Json file, execute the Javascript code and append the new data to the BigQuery dataset.
At last, export the BigQuery query result to Cloud Storage.
I'm trying to read the exif data from a .JPG image. I've tried differents solutions found here and there (PIL, piexif, exifread...) and none of them worked for this set of images. It worked for other images taken from another camera but not for this one, all these different methods returning empty dictionaries. It seems that there is no exif data but (I apologies for my newbyness) when I RIGHT-click + properties (I use windows), I do see what is exif data to me : date of creation, etc...
Here is one image :
image.JPG
If another of the thousands of anonymous heroes could help me on this one, I would be very grateful...
Alright so I found a solution which I share now.
The problem is that the libraries that open metadata are not taking all possible configurations for the image file and therefore, they can handle some and some others they cannot. I finally made it using exiftool, an executable that I dowloaded on my windows on this link :
https://sno.phy.queensu.ca/~phil/exiftool/
Then I paste the executable in a folder and I add exiftool.py in that folder, that I got from :
https://github.com/smarnach/pyexiftool/find/master
Then, using this small piece of code (for example):
import exiftool
with exiftool.ExifTool("exiftool.exe") as et:
metadata = et.get_metadata_batch(files)
for d in metadata:
print("{:20.20} {:20.20}".format(d["SourceFile"],
d["File:FileCreateDate"]))
Of course, this is just to show that you indeed can access the metadata, then you can do whatever you want with that. Here is the documentation of the library exiftool : http://smarnach.github.io/pyexiftool/
Cheers, JM