The other day in November I was about to build
an image classification model with SAS tools.
I already started googling for examples, but because
I'm lazy, I gave up and decided to give ChatGPT a chance to shine.
So what I asked first was:
Write me some SAS code for training image classification model
This is what I got as answer:
Certainly! To train an image classification model in SAS, you can use the
deep learning capabilities provided by SAS Viya, which includes a high-level
deep learning API. Below is an example of SAS code for training an image classification model
using a deep learning neural network.
Before you use this code, make sure you have SAS Viya installed and configured on your system.
You also need your dataset of labeled images.
/* Load necessary packages */
%let targetServer = YOUR_TARGET_SERVER;
%let targetPort = YOUR_TARGET_PORT;
%let targetUser = YOUR_USERNAME;
%let targetPassword = YOUR_PASSWORD;
options dlCreateCnnConfig = YES;
/* Connect to the CAS server */
cas mySession sessopts= (caslib='CASUSER',host="&targetServer",port=&targetPort,
user="&targetUser",password="&targetPassword");
/* Load the image data into CAS */
caslib _caslib_ sessopts=(caslib=CASUSER);
data _caslib_.your_image_data;
/* Use the IMAGEFILE function to load image files into a CAS table */
/* Specify your image file directory */
image_bytes = imagefile('/path/to/your/images', '*.jpg', 'Label', 'jpg');
run;
/* Define the deep learning model */
data _caslib_.cnn_model;
retain caslib 'CASUSER';
inputTable = '_caslib_.your_image_data';
modelType = 'CNN';
target = 'Label';
target = { "Label" };
/* Define the convolutional layers */
modelOpt = modelweights('/path/to/pretrained_weights.weights');
numConv = 2;
convLayer = { { 6, 5 }, { 12, 5 } };
convAct = 'RELU';
poolLayer = { { 2 }, { 2 } };
poolAct = 'RELU';
/* Define the fully connected layers */
numFC = 2;
fcLayer = { 100, 50 };
fcAct = 'RELU';
/* Define the output layer */
outputAct = 'SOFTMAX';
n=3;
/* Specify the optimization parameters */
casout = casOut( append=T );
run;
/* Train the deep learning model */
proc deepLearn trainSeed=12345 data=_caslib_.cnn_model;
targetModel=_caslib_.your_trained_model;
savestate state=_caslib_.your_model_state;
casout= { caslib = _caslib_ name = 'cnn_model_training' replace=TRUE };
run;
/* Assess the model's performance */
proc deepLearn assess seed=12345 data=_caslib_.your_image_data;
targetModel=_caslib_.your_trained_model;
casout= { caslib = _caslib_ name = 'cnn_model_assessment' replace=TRUE };
run;
/* Terminate the CAS session */
cas mySession terminate;
/* View the assessment results */
proc print data=_caslib_.cnn_model_assessment;
run;
Great! Nice one with first attempt.
But this was too simple network in my opinion: CNN with only two layers.
I knew that in CAS Deep Learning actions, there are more complex built-in CNN
models for image classification purposes.
So, my next request was:
I know there is built-in Resnet50 in SAS. Rewrite the previous code
Oh yes, ChatGPT knew it too! :
SAS provides a built-in ResNet50 model that you can use for image classification.
You can train a ResNet50 model with your own dataset using the deep learning capabilities in SAS Viya.
Here's how you can modify the previous code to use the ResNet50 model..:
To making the story a bit shorter, I won't include the modified code here.
The given code did most obviously what I needed.
But I wanted to demonstrate the Viya Deep Learning capabilities to my open-source friend.
Could I do the same using Python? I had heard that there is something called 'dlpy' that
enables using Viya deep learning actions in Python.
So, I blindly asked this:
Can you do the same with dlpy?
Answer:
Certainly! You can use SAS DLPy, a Python library for deep learning in SAS,
to train an image classification model using the ResNet50 architecture.
Here is an example of how to do it:
# Import necessary libraries
import dlpy
from dlpy.images import ImageTable
from dlpy.applications import ResNet50
from dlpy.model import Model
from dlpy.layers import Input, Output
from dlpy.utils import display_table
from dlpy import gpuinfo
from dlpy.metrics import accuracy_score
from dlpy.model import Gpu
from swat import CAS
# Initialize a CAS session
s = CAS(hostname='YOUR_TARGET_SERVER', port=YOUR_TARGET_PORT, username='YOUR_USERNAME', password='YOUR_PASSWORD')
# Load the image data into CAS
image_table = ImageTable.load_files(s, path='/path/to/your/images', casout=dict(name='your_image_data', replace=True))
# Define the ResNet50 model
model = ResNet50(s, model_table='ResNet50',
width=224, height=224, random_flip='HV', random_crop='RESIZETHENCROP',
offsets=(103.939, 116.779, 123.68), scale=1, random_mutation='none',
pre_trained_weights=True, pre_trained_weights_file='/path/to/pretrained_weights.weights')
# Specify the training settings
model.print_summary()
model.fit(data=image_table, n_threads=4, max_epochs=100, lr=0.001, gpu=Gpu(devices=[1]), target='Label')
# Evaluate the model on the validation data
model.evaluate(validation_data)
# Display the evaluation results
print(model.valid_res_tbl)
# Disconnect from CAS
s.close()
How cool!
But now it started to interest me if this code really works.
And - it appeared to be almost good for running. The most essential fix was the
name of the Resnet constructor (Resnet50_Caffe instead of Resnet50).
Then, there was some minor fixes and additions like removing unnecessary import-clauses, removing everything that refers into GPU since I didn't have one in my environment, dividing dataset into train-test -partitions, testing and evaluating the model.. I might have got them too if I had played with ChatGPT little bit longer.
BTW, what was my model about?
As Christmas is approaching, I wanted to train a model that classifies pictures of trees: Christmas trees into first category and non-Christmas trees into second.
I used ~600 Christmas tree pictures and ~1000 non-Christmas tree pictures as training material from Open Images Dataset, a rich source of labeled images of all sorts.
Resnet50 performed quite well in predicting these image classes!
Testing with 319 images I got following accuracy figures:
So, only 26 of 319 were misclassified.
To finalize my story, I include a couple of examples of correctly classified and misclassified cases:
Correct classification:
Not correct classification:
I still wonder why the last picture was labeled as Christmas tree in the first place.. Resnet knew it better 😃
... View more
