Valify

Author

Ahmed Hussein

table of Content

Install for Train Enviroments
Model Information
Dataset
Train
Metric
Inference
Docker Setup
Test API
What Is Next

Install for Train Enviroments

to install you need to create Virtual Environment with conda, activate it then install requirments

conda create -n Test python=3.8.*
conda activate Test
pip install -r requirements.txt

Model Information

the model informations

Architecture

========== Model Summary  ==========
----------------------------------------------------------------
        Layer (type)               Output Shape         Param #
================================================================
            Conv2d-1            [-1, 8, 16, 16]             224
            Conv2d-2             [-1, 16, 8, 8]           1,168
            Conv2d-3             [-1, 32, 4, 4]           4,640
            Linear-4                  [-1, 128]          65,664
            Linear-5                  [-1, 256]          33,024
            Linear-6                   [-1, 29]           7,453
================================================================
Total params: 112,173
Trainable params: 112,173
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.01
Forward/backward pass size (MB): 0.03
Params size (MB): 0.43
Estimated Total Size (MB): 0.47

MACs and FLOPs

you can compute FLOPs (Floating Point Operations) From MACs as one MACs equals roughly two FLOPs

MyNetwork(
  112.17 k, 100.000% Params, 312.48 KMac, 100.000% MACs, 
  (conv1): Conv2d(224, 0.200% Params, 57.34 KMac, 18.351% MACs, 3, 8, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
  (conv2): Conv2d(1.17 k, 1.041% Params, 74.75 KMac, 23.922% MACs, 8, 16, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
  (conv3): Conv2d(4.64 k, 4.136% Params, 74.24 KMac, 23.759% MACs, 16, 32, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
  (fc1): Linear(65.66 k, 58.538% Params, 65.66 KMac, 21.014% MACs, in_features=512, out_features=128, bias=True)
  (fc2): Linear(33.02 k, 29.440% Params, 33.02 KMac, 10.568% MACs, in_features=128, out_features=256, bias=True)
  (fc3): Linear(7.45 k, 6.644% Params, 7.45 KMac, 2.385% MACs, in_features=256, out_features=29, bias=True)
)
========== Model Complex  ==========
Computational complexity:       312.48 KMac
Number of parameters:           112.17 k

Respective Field

the respective field represent the number of pixels the model see in as overall in a CNN for conv layers you can see this link

input_layer: output size = 32; size change relative to original = 1; receptive image size = 1
Conv2d-1: output size = 16; size change relative to original = 2; receptive image size = 3
Conv2d-2: output size = 8; size change relative to original = 4; receptive image size = 7
Conv2d-3: output size = 4; size change relative to original = 8; receptive image size = 15

Dataset

the dataset contain Train and Dev data can be generated from general_functions.py

Hierarchy

Dataset
|
----Train 
    |
    ----29 folder each folder is an arabic letter, each folder contain 2385 image 
|
----Dev 
    |
    ----29 folder each folder is an arabic letter, each folder contain 96 image

Dataset location

you can find the dataset in link

Train

train can be found in train.py

Input Size : 32 width, 32 height and BGR
LR : Start with 1e-3 then end with 0.5*1e-3 through the training
Epochs : 9 epochs
Loss : NLLLoss The negative log likelihood loss, Calculated as −log(y), where y is a prediction corresponding to the true label, after the Softmax Activation
Log : logs can be found in tensorboard you can run the next command or you can see the screenshots

tensorboard --logdir=runs

run the train with

python train.py

Metric

the metric is Precision, Recall, and F1 Score you can generate it from the get_metrics.py you can open Metrics.json

python get_metrics.py

{"ا": {"Precision": 1.0, "Recall": 0.9479166666666666, "F1": 0.9732620320855615}, "ب": {"Precision": 0.8055555555555556, "Recall": 0.90625, "F1": 0.8529411764705882}, "ت": {"Precision": 0.7373737373737373, "Recall": 0.7604166666666666, "F1": 0.7487179487179487}, "ث": {"Precision": 0.7604166666666666, "Recall": 0.7604166666666666, "F1": 0.7604166666666666}, "ج": {"Precision": 0.971830985915493, "Recall": 0.71875, "F1": 0.8263473053892216}, "ح": {"Precision": 0.8640776699029126, "Recall": 0.9270833333333334, "F1": 0.8944723618090452}, "خ": {"Precision": 0.9263157894736842, "Recall": 0.9166666666666666, "F1": 0.9214659685863874}, "د": {"Precision": 0.8942307692307693, "Recall": 0.96875, "F1": 0.93}, "ذ": {"Precision": 0.9368421052631579, "Recall": 0.9270833333333334, "F1": 0.9319371727748692}, "ر": {"Precision": 0.93, "Recall": 0.96875, "F1": 0.9489795918367346}, "ز": {"Precision": 0.9891304347826086, "Recall": 0.9479166666666666, "F1": 0.9680851063829786}, "س": {"Precision": 0.85, "Recall": 0.8854166666666666, "F1": 0.8673469387755102}, "ش": {"Precision": 0.956989247311828, "Recall": 0.9270833333333334, "F1": 0.9417989417989417}, "ص": {"Precision": 0.8105263157894737, "Recall": 0.8020833333333334, "F1": 0.806282722513089}, "ض": {"Precision": 0.946236559139785, "Recall": 0.9166666666666666, "F1": 0.9312169312169313}, "ط": {"Precision": 0.9775280898876404, "Recall": 0.90625, "F1": 0.9405405405405405}, "ظ": {"Precision": 0.9108910891089109, "Recall": 0.9583333333333334, "F1": 0.934010152284264}, "ع": {"Precision": 0.7924528301886793, "Recall": 0.875, "F1": 0.8316831683168316}, "غ": {"Precision": 0.7435897435897436, "Recall": 0.90625, "F1": 0.8169014084507042}, "ف": {"Precision": 0.7560975609756098, "Recall": 0.6458333333333334, "F1": 0.6966292134831461}, "ق": {"Precision": 0.7934782608695652, "Recall": 0.7604166666666666, "F1": 0.7765957446808509}, "ك": {"Precision": 0.9777777777777777, "Recall": 0.9166666666666666, "F1": 0.946236559139785}, "ل": {"Precision": 0.8529411764705882, "Recall": 0.90625, "F1": 0.8787878787878787}, "لا": {"Precision": 0.979381443298969, "Recall": 0.9895833333333334, "F1": 0.9844559585492227}, "م": {"Precision": 0.8736842105263158, "Recall": 0.8645833333333334, "F1": 0.8691099476439791}, "ن": {"Precision": 0.8395061728395061, "Recall": 0.7083333333333334, "F1": 0.768361581920904}, "ه": {"Precision": 0.7889908256880734, "Recall": 0.8958333333333334, "F1": 0.8390243902439025}, "و": {"Precision": 0.9894736842105263, "Recall": 0.9791666666666666, "F1": 0.9842931937172775}, "ي": {"Precision": 0.9139784946236559, "Recall": 0.8854166666666666, "F1": 0.8994708994708994}}

Class	Precision	Recall	F1
ا	1.0	0.9479166666666666	0.9732620320855615
ب	0.8055555555555556	0.90625	0.8529411764705882
ت	0.7373737373737373	0.7604166666666666	0.7487179487179487
ث	0.7604166666666666	0.7604166666666666	0.7604166666666666
ج	0.971830985915493	0.71875	0.8263473053892216
ح	0.8640776699029126	0.9270833333333334	0.8944723618090452
خ	0.9263157894736842	0.9166666666666666	0.9214659685863874
د	0.8942307692307693	0.96875	0.93
ذ	0.9368421052631579	0.9270833333333334	0.9319371727748692
ر	0.93	0.96875	0.9489795918367346
ز	0.9891304347826086	0.9479166666666666	0.9680851063829786
س	0.85	0.8854166666666666	0.8673469387755102
ش	0.956989247311828	0.9270833333333334	0.9417989417989417
ص	0.8105263157894737	0.8020833333333334	0.806282722513089
ض	0.946236559139785	0.9166666666666666	0.9312169312169313
ط	0.9775280898876404	0.90625	0.9405405405405405
ظ	0.9108910891089109	0.9583333333333334	0.934010152284264
ع	0.7924528301886793	0.875	0.8316831683168316
غ	0.7435897435897436	0.90625	0.8169014084507042
ف	0.7560975609756098	0.6458333333333334	0.6966292134831461
ق	0.7934782608695652	0.7604166666666666	0.7765957446808509
ك	0.9777777777777777	0.9166666666666666	0.946236559139785
ل	0.8529411764705882	0.90625	0.8787878787878787
لا	0.979381443298969	0.9895833333333334	0.9844559585492227
م	0.8736842105263158	0.8645833333333334	0.8691099476439791
ن	0.8395061728395061	0.7083333333333334	0.768361581920904
ه	0.7889908256880734	0.8958333333333334	0.8390243902439025
و	0.9894736842105263	0.9791666666666666	0.9842931937172775
ي	0.9139784946236559	0.8854166666666666	0.8994708994708994

Inference

the inference can be found in inference.py, expect dir contain images run inference with

python inference.py <dir-contain-images>

this will print the predicted label

Docker Setup

you can setup the docker from zero from Dockerfile run command

sudo docker build -t app .

or load it after downloading the app.tar

sudo docker load < app.tar

run a container with

sudo docker run -d --name mycontainer2 -p 80:80 app

Test API

you can test the API with test_api.py or see the screenshot

What Is Next

add more testing for the API input format and inference
add CI/CD
Train bigger model or change the input size for better results
add more augmentations
use Pretrained network and adapt on this data

ahmedhessuin / valify Goto Github PK

valify's Introduction