Emotion detection using transfer learning model. #38
Comments
|
if yes then please assign me. |
|
Did you find a method for more accuracy? @keshav340 |
|
yes, I hypertuned his model architecture and got better result |
|
How is hypertuned a model done? can you share the example :-) |
I ran the original code and was only in the mid to high 80's for accuracy, altered the code slightly and now have the model at 95.87% accuracy. What results do you currently have @melikeey, you can change some of the layer coding where the image is being manipulated, you might notice that in the original code he has his layers as 32/64/128 and he follows this with another 128 layer, simply changing that layer to 256 and leaving kernel size the same should get you a small boost in accuracy, |
|
Oh yeah, there was a lot of overfit, my point was more that the OP has submitted their finished version, but we don't know if it is the optimal version unless we try changing some of the settings. I have been trying some slight variations on his model to try and maximise accuracy/validation but it's a spare time kind of thing so maybe trying once or twice per week in google colab to see if I can raise the accuracy from 89% to +95% whilst keeping the validation in a similar range. |
model = Sequential() model.add(Conv2D(128, kernel_size=(3, 3), activation='relu')) model.add(Conv2D(256, kernel_size=(3, 3), activation='relu')) model.add(Flatten()) |
|
Which should (val)test accuracy or training accuracy be considered? As you said, while the training accuracy was 85% for the original code, the test(val) accuracy did not exceed 63%. 5 times, I trained this code. Now I have started to train the model you gave. @Alexander-K-byte |
Which should (val)test accuracy or training accuracy be considered? As you said, while the training accuracy was 85% for the original code, the test(val) accuracy did not exceed 63%. 5 times, I trained this code. Now I have started to train the model you gave. can you also share with me for this model calculation table and accuracy and loss graphs? |
|
Validation accuracy should be noted down. Because the training accuracy
eventually become 2 at infinite epoch.But, the model will get overfitted
after certain epochs,which would be indicated by the validation
accuracy,not the train accuracy.
…On Wed, Apr 13, 2022, 5:30 PM Melike Elif Yıldırım ***@***.***> wrote:
what about the validation test accuracy.Did it get overfit?
Try running this for the layers.
model = Sequential() model.add(Conv2D(64, kernel_size=(3, 3),
activation='relu', input_shape=(48,48,1))) model.add(Conv2D(64,
kernel_size=(3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(128, kernel_size=(3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2))) model.add(Dropout(0.25))
model.add(Conv2D(256, kernel_size=(3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2))) model.add(Dropout(0.25))
model.add(Flatten()) model.add(Dense(1024, activation='relu'))
model.add(Dropout(0.5)) model.add(Dense(7, activation='softmax'))
Should (val)test accuracy or training accuracy be considered? As you said,
while the training accuracy was 85% for the original code, the test(val)
accuracy did not exceed 63%.
Which of val/train accuracy interprets true accuracy. There is some
confusion between these two.
5 times, I trained this code.
I added more layers to my model, I changed the ratios of test and training
data. I even removed the maximum pooling layers (it took too long) and
again I couldn't increase the test(val) accuracy above 63%.
Now I have started to train the model you gave.
—
Reply to this email directly, view it on GitHub
<#38 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AUNXPAD4ZLW3WJKQEN73FDLVE2ZN3ANCNFSM46RH44HA>
.
You are receiving this because you commented.Message ID:
***@***.***>
|
Did you achieve a high rate validation accuracy with this code? I changed model 5 times but I couldn't. my results are around 63%, 42%, 50%, 62%, 57%. I need to solve this :/ |
I change activation relu with sigmoid. It has %42 validation accuracy. |
|
%60-%70 ratio is poor model, %70-%80 is good model intervals. I tried to create a good model at least :D |
The problem with the dataset is that disgust/anger are so alike. Ideas that spring to mind are removing disgust from the equation and placing those files into anger. Ensure the camera you are using for live video capture is better quality than a simple built in webcam on a laptop, also make sure that your face is well lit. Using a Pi cam v1.3 for a few opencv projects and the inbuilt haar cascade had issues picking out my face while wearing glasses if my face was poorly illuminated, I was picked up every 4th or 5th frame whilst wearing glasses, tooe them off and it was able to detect my face a lot easier. Try taking some pictures of the different emotions, turn them in to a video montage with each image on display for 5 to 10 seconds and use this as a video stream through opencv so that the frames it looks at are not constantly changing. I would suggest also looking into transfer learning, but I am guessing you are more time constrained. |
I want to train different transfer learning model on emotion detection so that we get good accuracy and less loss.
The text was updated successfully, but these errors were encountered: