Please enable JavaScript.
Coggle requires JavaScript to display documents.
Model - Coggle Diagram
Model
Model
ResNet
U-Net
Inception
GoogleNet
MobileNet (SSD)
Accuracyต่ำ แต่Latencyต่ำ
CNN
FeatureExtraction
In This Case : Transfer Learning (Similar GoogleNet)
Base Model( Cut Head)
Becase DogBody similar CatBody
-Fix Weight ,Bias -Weight (ImageNet)
-Output -ReLU
-Dense Layer -ReLU to Flatten before Dense Layer
Head Custom
-Fix Base Model
-Custom Weight ,Bias - Activation (Sigmoid)
-Prediciton Layer
Training =Base Model+Head Custom
Fine Tuning
Training at Attract (Not Concreate)
Pre-Processing Data
Choose Model
IMG Size (Manual)
Import IMG
Make RGB to Small Scale
Resize for Saving Memory
Random IMG
Get The First IMG
Save ,Donot Train The same IMG again
Check Accuracy >= 90% Good
Tool
Pre-Processing
TensorFlow (ImageDataGenerator)
OpenCV
Feature Extraction
This Case : TranferLeaning
Classification
Binary = True/False
Sigmoid
Catagorical
Cat / Dog or Others
Softmax
Training
Optimizer
MobileNet
RMSprop
Adam
Loss
Binary Classification
Binary_CrossEntropy
Catagorical Classification
Catagorical_Crossentropy
Metric
Accuracy
Epoch