Implementing ResNet with Base Line Accuracy

 Implementing ResNet with Base Line Accuracy  

ResNet Base Line Accuracies 

https://keras.io/api/applications/ 

 

Model	Size (MB)	Top-1 Accuracy	Top-5 Accuracy	Parameters	Depth	Time (ms) per inference step (CPU)	Time (ms) per inference step (GPU)
ResNet50	98	0.749	0.921	25,636,712	-	58.20	4.55
ResNet101	171	0.764	0.928	44,707,176	-	89.59	5.19

 

Results Summary & Comparison (Reasons and the Implementations described below) 

Model	Size (MB)	Base Line Top-1 Accuracy	Actual Base Line Top 1 Accuracy
ResNet50	98	74.9 %	68.084 %
ResNet101	171	76.4 %	69.976 %

 

ResNet 50 Validation Test Results 

Data Set: ImageNet Validation data set (50,000 images)  

Validation Accuracy (Top1) = 68.084 % with caffe input preprocessing 

 

ResNet 101 Validation Test Results 

Data Set: ImageNet Validation data set (50,000 images) 

Validation Accuracy (Top1) = 69.976 % with caffe input preprocessing 

 

Base Line Setup Details 

https://keras.io/api/applications/#classify-imagenet-classes-with-resnet50 

 

1. Load the images in the format of RGB channel ordering (size = 224, 224, 3) 
2. Convert images in to an array  
3. Expand the image dimension by one  
4. Preprocess Input (on tf, torch and caffe modes) 

 
Mode is one of "caffe", "tf" or "torch". Defaults to "caffe". 

https://www.tensorflow.org/api_docs/python/tf/keras/applications/imagenet_utils/preprocess_input 

 

• caffe: will convert the images from RGB to BGR, then will zero-center each color channel with respect to the ImageNet dataset, without scaling. 

• tf: will scale pixels between -1 and 1, sample-wise. 

• torch: will scale pixels between 0 and 1 and then will normalize each channel with respect to the ImageNet dataset. 

• 
  

     5.Call predict  

 

Tested ResNet101 on all 3 input preprocessing modes 

1. caffe = 71.5% (Top1 Accuracy) with 1000 validation images 
2. tf = 0.3% (Top1 Accuracy) with 1000 validation images 
3. torch = 0.89% (Top1 Accuracy) with 1000 validation images 

Out of all three modes, caffe mode has the highest accuracy 

 

ResNet Baseline Validation Script TensorFlow 1.15 

Install TensorFlow 1.15 

1. Install python 3.6 
2. Create a new PyCharm project 
3. Select python 3.6 interpreter 
4. Create VENV 
5. Pip install tensorflow==1.15 

Python 3.6 

Tensorflow 1.15 

Requirements file link

https://github.com/SaralaSewwandi/ResNet-BaseLine-Validation/blob/main/Tensoflow%201.15%20Libray%20Requirements/requirements.txt

 

 

Error with resnet decode fix in tensorflow 1.15 

resolved AttributeError: 'str' object has no attribute 'decode' 

pip install 'h5py==2.10.0' --force-reinstall 

 

Final TensorFlow Script is written in order to eb compatible with both the TensorFlow versions 2 and 1.5 

 

ResNet Baseline Validation Script TensorFlow 2  

python 3.9  

tensorflow 2.7 

 

Tasks Performed 

1. Implemented the script for TensorFlow 2 & Tensorflow 1.15 
2. Tested for all the preprocessing modes for the base line accuracy (TF, TORCH and CAFFE) 
3. Image size is fine = 224,224,3 
4. Channel ordering and conversion to RGB to BGR confirmed 
5. Input Preprocessing Done According to the API 
6. Inference Validation Script debugged  
7. Tested for all the details mentioned in the Keras site to reproduce the baseline accuruacy 
8. Only thing left to check is whether we are evaluating based on the exact validation data set that they were providing the accuracy on. 
9. We have used the 2012 ImageNet Validation dataset 
10. However, between 2012, 2014 & 2017 there will be changes in labels (which were corrected over time) 
11. So ResNet101 and ResNet50 is to be tested on 2014 ImageNet Validation dataset  

 

ImageNet Class IDs with Labels 

Official Link (here first line has the value background just to avoid the index 0, and assign the indexes from 1 – 1000 for the imagenet class ids) 

https://storage.googleapis.com/download.tensorflow.org/data/ImageNetLabels.txt 

 

https://gist.github.com/yrevar/942d3a0ac09ec9e5eb3a 

https://gist.github.com/yrevar/667fd94b94f1666137f45d1363f60910 

https://github.com/torch/tutorials/blob/master/7_imagenet_classification/synset_words.txt 

 

Other References 

https://www.tensorflow.org/tutorials/images/transfer_learning_with_hub#setup 

https://tfhub.dev/s?module-type=image-classification%252F&tf-version=tf2&q=tf2 

https://tfhub.dev/google/imagenet/resnet_v1_101/classification/5 

 

Final Python Scripts

ResNet Validation Script is developed to support both the TensorFlow 2 and Tensorflow1.15 versions

https://github.com/SaralaSewwandi/ResNet-BaseLine-Validation