ahmedhessuin / fashon_detection

Fashon_detection

table of Content

1. Instalation
2. Data
3. Data Handling
4. Model
5. Model Details
6. the Model Accuracy and Metric
7. Discussion

Instalation

first thing you need to do is install the requriments

conda create -n Fashon_Classification python=3.8 
conda activate Fashon_Classification
conda install -c anaconda cudatoolkit=11.2
pip install -r requirments.txt

after that the env will be ready to run

Data

• the data used is fashion_mnist dataset provided by tensorflow
• the data contain 60k train images and 10k dev images, for train 6k image for each class and 1k in the dev
• the images are Gray lvl with (28x28)
• the data contain labels as sprase for 10 different types of cloths

'T-shirt/top', 'Trouser', 'Pullover', 'Dress', 'Coat','Sandal', 'Shirt', 'Sneaker', 'Bag', 'Ankle boot'

with annotation like

0, 1, 2, 3, 4, 5, 6, 7, 8, 9

this is for the data section

Data Handling

1. download the training data with

    fashion_mnist = tf.keras.datasets.fashion_mnist
    (train_images, train_labels),(_,_) = fashion_mnist.load_data()

2. change the image channels to be (28,28,1), to use CNN layers

train_images=np.expand_dims(train_images,-1)

3. doing the same to the test data

fashion_mnist = tf.keras.datasets.fashion_mnist
(_,_),(test_images, test_labels)=fashion_mnist.load_data()
test_images=np.expand_dims(test_images,-1)

• All functions used to load the data in the data_script.py

Model

because of the simplicity of the training the model archtecture is so simple, and the method of building the archtecture was Sequential Starting with Simple model ARC

    model = tf.keras.Sequential([
        tf.keras.layers.Input(shape=(28,28,1),name="input_layer"),
        tf.keras.layers.Conv2D(filters=20,kernel_size=3,strides=1,padding='valid',activation='relu',name="conv1"),
        tf.keras.layers.Conv2D(filters=40,kernel_size=3,strides=1,padding='valid',activation='relu',name="conv2"),
        tf.keras.layers.Conv2D(filters=80,kernel_size=3,strides=1,padding='valid',activation='relu',name="conv3"),
        tf.keras.layers.Flatten(name='Flatten'),
        tf.keras.layers.Dense(units=100,activation='relu',name='dense1'),
        tf.keras.layers.Dense(number_of_classes,activation='softmax',name='dense_output_layer')
    ])

this model is so simple contain only 3 conv layers followed by 1 dense then another dense layer for the output

Models Details

• in this section we will talk about the models used and the different between each model
• the FLOPs for each model
• the MACs for each model
• all results can be found in model.py script

First Model

The Model ARCH

    model = tf.keras.Sequential([
        tf.keras.layers.Input(shape=(28,28,1),name="input_layer"),
        tf.keras.layers.Conv2D(filters=20,kernel_size=3,strides=1,padding='valid',activation='relu',name="conv1"),
        tf.keras.layers.Conv2D(filters=40,kernel_size=3,strides=1,padding='valid',activation='relu',name="conv2"),
        tf.keras.layers.Conv2D(filters=80,kernel_size=3,strides=1,padding='valid',activation='relu',name="conv3"),
        tf.keras.layers.Flatten(name='Flatten'),
        tf.keras.layers.Dense(units=100,activation='relu',name='dense1'),
        tf.keras.layers.Dense(number_of_classes,activation='softmax',name='dense_output_layer')
    ])

• the CONV layers are Kernal = 3x3 with Stride = 1 with Activation = Relu
• taking input 28x28x1 indicate that it takes gray image
• Number of Parameters

    # Total params: 3,909,430
    # Trainable params: 3,909,430
    # Non-trainable params: 0

The Model FLOPs & MACs & Respective Field

• the FLOPs
• the equation from
• link
• aslo confirmed that i am doing it right with from keras_flops import get_flops

    # conv1    256880
    # conv2   8317440
    # conv3  27917120 # most layer CONV 3 because the max number of input features in this layer
    # dense1  7744100
    # dense_output_layer 2010
    # total number of FLOPs 44 237 550

• the MACs

    # conv1      180
    # conv2     7200
    # conv3    28800
    # dense1 3872000 # MOST MACs
    # dense_output_layer 1000
    # total number of MACs : 3 909 180

• the Respective Field
• 

[['conv1', 3], ['conv2', 5], ['conv3', 7]]

Second Model

The Model ARCH

    model = tf.keras.Sequential([
        tf.keras.layers.Input(shape=(28,28,1),name="input_layer"),
        tf.keras.layers.Conv2D(filters=20,kernel_size=3,strides=1,padding='valid',activation='relu',name="conv1"),
        tf.keras.layers.Conv2D(filters=40,kernel_size=3,strides=2,padding='valid',activation='relu',name="conv2"),
        tf.keras.layers.Conv2D(filters=80,kernel_size=3,strides=2,padding='valid',activation='relu',name="conv3"),
        tf.keras.layers.Flatten(name='Flatten'),
        tf.keras.layers.Dense(units=100,activation='relu',name='dense1'),
        tf.keras.layers.Dense(number_of_classes,activation='softmax',name='dense_output_layer')
    ])

• the CONV layers are Kernal = 3x3 with Stride = 2 with Activation = Relu except first CONV use Stride = 1
• taking input 28x28x1 indicate that it takes gray image
• Number of Parameters

    # Total params: 237,430
    # Trainable params: 237,430
    # Non-trainable params: 0

• note that the number of parameters is reduced because the size of the input layer for the First Dense is reduced thanks to the Stride = 2

The Model FLOPs & MACs & Respective Field

• the FLOPs
• the equation from
• aslo confirmed that i am doing it right with from keras_flops import get_flops
• link

    # conv1   256880
    # conv2  2079360
    # conv3  1442000 # most layer CONV 3 because the max number of input features in this layer
    # dense1  400100
    # dense_output_layer 2010
    # total number of FLOPs 4 180 350

• note that the number of FLOPs is reduced because the Stride = 2 effect on the input size for each layer
• the MACs

    # conv1     180
    # conv2    7200
    # conv3   28800
    # dense1 200000 # MOST MACs
    # dense_output_layer 1000
    # total number of MACs : 237 180

• note the number of MACs in the most heavy layer dense1 is reduced because the input size is reduced thanks to the Stride = 2 effect
• the Respective Field
• 

[['conv1', 3], ['conv2', 7], ['conv3', 15]]

• note the recptive field change in conv2 and conv3 because of the Stride = 2

the Model Accuracy and Metric

• Mainly used SparseCategoricalCrossentropy as a loss
• Mainly used Precision and Recall for Accuracy metric and HIT RATE (over all right answers / total number of data)
  • choosed this metric because they indicate very well the behaviour of a Classification Model
  • Precision: how many right answers i get for each class if the model made prediction of this class
  • Recall : how many right answer i get for each class from all the Predicted Data
  • HIT RATE : how good the model over all with all classes

Results of the both Models

Model Name	T-shirt/top	Trouser	Pullover	Dress	Coat	Sandal	Shirt	Sneaker	Bag	Ankle boot	Acc
Model_1(stride_1)	PRECISION:0.798 & RECALL:0.866	PRECISION:0.981 & RECALL:0.973	PRECISION:0.851 & RECALL:0.804	PRECISION:0.923 & RECALL:0.854	PRECISION:0.781 & RECALL:0.867	PRECISION:0.982 & RECALL:0.957	PRECISION:0.703 & RECALL:0.67	PRECISION:0.921 & RECALL:0.973	PRECISION:0.968 & RECALL:0.963	PRECISION:0.971 & RECALL:0.941	0.8868%
Model_2(stride_2)	PRECISION:0.855 & RECALL:0.836	PRECISION:0.981 & RECALL:0.976	PRECISION:0.789 & RECALL:0.867	PRECISION:0.879 & RECALL:0.907	PRECISION:0.850 & RECALL:0.75	PRECISION:0.968 & RECALL:0.97	PRECISION:0.715 & RECALL:0.699	PRECISION:0.949 & RECALL:0.952	PRECISION:0.949 & RECAL:0.978	PRECISION:0.957 & RECALL:0.963	0.8898%
---	---	---	---	---	---	---	---	---	---	---	---

• You can get the result from train.py as it save the Chekpoints while training and Test after training done, saving Logs in log folder

Discussion

in this we will discuss

Approch

• downloading the data and see the number of label for each class

• Some Data Analysis

  • knowing the image size
  • knowing the color domain
  • looking about the data distribution for each class
  • looking for outlier

• building the model

  • looking about the data and how complex is the target
  • based on the target complex the model is decided to be kinda right for example, not using resnet 152 for 28x28x1 images
  • because complex model with simple data can lead to a bad result
    • first train with small and avg model
    • increase the model Parameters and experiment it
    • decrease the model parameters and experiment it
    • after some experiments you will find the best model for this task
  • try good optimizer and train with different Learning Rate ( as it can be critical )
  • compile the model and see it's FLOPs & MACs and Parameters
  • see if there is a way to reduce the FLOPs & MACs
  • this is why i choosed model 2 instead of model 1 because model 2 achieve same result and has min FLOPs
  • can try more smaller model or only double layer model with 2 Dense layers (this won't achieve the task requirments for (Respective Field)

• Loss

  • deciding the loss is a critical decision as based on your decision the model can change a lot
  • see the type of loss that meet the requriment of the target
  • using the loss for Classification (CrossEntropy) and decide the activation function of the last layer (softmax)

• Train

  • write a call backs with tensorboard to trace the model training and save checkpoint every 1 epoch
  • checking do i need to augment the data, in my case i didn't need
  • start training with reasonable batchsize and number of epochs

• Test

  • testing the model with the Metric i choosed ( Precision and Recall ) with the default metric (Acc)
  • saving the result in text file to be found later

How to reduce number of FLOPs

this can be done by

• more conv layers with stride 2
  • as we can see in model with Stride 2 the FLOPs got reduced
• Pooling
  • that will reduce the number of operations because the input size is reduced
• Separable Convolutions
  • Spatially Separable Convolutionsdoing like Inception Net breaking the 3x3 to (1x3) (3x1)
  • Depthwise Separable Convolutions Depthwise convolution followed by Pointwise convolution
• Reduce model Size

How to increase the Respective Field

this can be done by

• stacking more layer
  • as we stack more layers the Respective Field increase ( this can be consuming )
• Pooling
  • Pooling or increasing the stride of the CONV increase the Respective FIeld as we saw in the model with Stride 2
• Dilated Conv
  • as it's using wilder kernal size with the same Parameters