Это видео недоступно.
Сожалеем об этом.
240 - Deep Learning training for age and gender detection
HTML-код
- Опубликовано: 16 авг 2024
- Code generated in the video can be downloaded from here:
github.com/bns...
Train deep learning models to predict age and gender.
Dataset from here: susanqq.github...
This trained model will be later used towards real time emotion detection on Windows and raspberry Pi.
Thank you very much for this training, you are really perfect professor
So nice of you
This is amazing!
Also, I wondered if you would do a video about self-supervised image classification and clustering.
Getting "ResourceExhaustedError" upon executing the exact code on same dataset, anyone else facing this error? If yes, could you please suggest a solution?
Love your videos! Keep on going
Please try to use graph representation for images
Can I get the .h5 model file produced in the process please. Can anyone share a g-drive link of the same if available. Thanks in advanced.
Greetings of the day
I would like to post a question sir.
At the time of loading the any pretrained model (GPU supported environment) I am getting this error
AttributeError: 'str' object has no attribute 'decode'
Please provide me solution.
Again, very nice. I am working on CNN for regression and I was wondering about having 2 outputs. What do you think?? Will it be better to have two separate models?
Designing a model with multiple outputs involves branching out your model to generate multiple outputs. This assumes that you want to use same structure up to certain point and then branch out. If that is the case then you can save some computing time.
I'm running this on my laptop which has an i7 11800, 16GB memory, and an RTX 3060 with 6GB of memory. TensorFlow is configured to use CUDA/cudnn (and is doing so here). However, I absolutely just cannot get it to work at the standard 200x200 pixels with the default CNN configuration here. I've modified it to dynamically resize the UTKFace dataset images to 128x128, and have made the CNN a bit smaller. I (just barely) don't run out of memory now, but the later project "live prediction" shows all kinds of whacky age numbers (I changed the expected image size to 128x128 there, as well). Thoughts? Is there a way to build the model using standard 200x200 images and the CNN as you've configured here? Or am I actually hardware limited? Thanks for the great tutorial!
@Meliodas Hi! I got the same problem I run out of memory. How did you solve the problem ? Also you tried using google colab ?
@@TheXSeby I haven't tried Google collab. I messed around with the CNN configuration until it was happy with my configuration. I kept the image size to 128x128.
@@meliodas4560 how did you change all the images size and what updates you made to the code ? I'm interested because I want to do face and age detection as final year project at uni. Thanks!
@@TheXSeby
if img.shape is not (IMG_WIDTH, IMG_HEIGHT, IMG_DEPTH):
resized = cv2.resize(img, (IMG_WIDTH, IMG_HEIGHT), interpolation = cv2.INTER_AREA)
images.append(np.array(resized))
else:
images.append(np.array(img))
How can I make the code more efficient because I run out of memory on many machines. Any suggestions ?
Sreeni sir, ☕, thank you 😁
You're welcome!
i'm not able to download the dataset file
Hello Sreeni Sir, Thanks for your perfect explanations,
I have a question, is that normal that i'm getting the loss (mse) = 396.01
and mae of 15.45. 🙈, where's the problem ?
I do not see a problem. The values depend on your actual data. I made a quick test using python code and your results are reasonable to me. Run this code in python to understand mse and mae.
import numpy as np
true = [234,285,324,248,423,345]
predicted = [248,310,340,265,403,325]
def mae(y_true, predictions):
y_true, predictions = np.array(y_true), np.array(predictions)
return np.mean(np.abs(y_true - predictions))
print(mae(true, predicted))
#MSE
mse = np.square(np.subtract(true, predicted)).mean()
print(mse)
Sir make it one step further with face recognition.
Thank you
Sir gender is fine but for age each epoch is taking 25-30 mins to run and how many epochs minimum should I run to get a decent accuracy of around 60 percentage for age model
hii can you send me the age epoch file..??pleasee
Dataset is not
accessible.
Please where i can to get the files
The ideology is we know humans are evolving through millions of generation form ape to human so is there any possibility that we can create a dataset(facial bones structure) form start of humanity to current human and create a network which will predict that what will next generations will look like .
Time series forecasting but with human timeline :)
I have very limited understanding of anthropology so I am not the right person to comment on your question. But here is my point of view.... We are training a model for age detection and we know that age can be a challenging thing to predict. This is because the face evolves gradually from baby to youth to adult to old age. At no point can you predict the future to within a few years. In other words, the growth in continuous with incremental changes in facial features. Similarly, human evolution is continuous with incremental changes. There are algorithms that take your current face and predict how you'd look like in your old age. They are based on training a lot of samples representing different ages. How would you predict future human features without having data from future? Evolution depends on ecological features and requires information on future climactic conditions and also cultural conditions. My 2 cents on the topic!
Sir can you do a vedio on face aging with gan project sir