شبکه عصبی با پایتورچ

شرح پروژه

سلام وقت بخیر این یک سوال پیاده سازی یا پایتورچ برای درس شبکه عصبی است که گزارش کار هم می خواهد

Traffic Sign Label Prediction. By utilizing CNN models, you

are going to predict the labels of traffic sign images, the corresponding data is available here(https://www.kaggle.com/datasets/valentynsichkar/traffic-signs-preprocessed) The dataset consists of 34,799 of 32*32 images with certain labels. You should use train.pickle, valid.pickle, test.pickle, label\_names.csv files to solve the problem. By using the pickle library, you should import the necessary files, each file is a dictionary that has 2 keys: features that consists of images and lables that have the corresponding label for each image. Consider the following points when tackling the problem:

• Implement a decent evaluation pipeline and compare all the models that you use to tackle the problem. Some metrics that you can use, are f1\_score, precision, recall, and confusion matrix.
• Try to make a custom CNN model to perform best on the data. Use various techniques to boost your performance.
• Try to solve the problem using Transfer Learning by utilizing various pre-trained models.
• Analyze which signs are usually mispredicted by the model. What could be the reason?
• Check whether using grayscale images affects the performance of the model or not.
• Check whether the data is balanced or not. If not try to solve the problem and achieve better performance. (Hint: you can try up-sampling and down-sampling approaches to tackle this problem. For an up-sampling approach you can consider augmenting or synthesizing images)(Extra point)
• Visualize the output of multiple random images on different layers of your network and try to explain them. (Extra point)
• Visualize different images in the train set and compare them in terms of illumination, aspect ratio, or other aspects. Try to use appropriate transformation to pre-process the images and boost your model’s performance. For, example in terms of illumination, if images differ considerably, you can use Histogram Equalization to solve the issue. (Extra point)

مهارت ها و تخصص های مورد نیاز

درباره کارفرما

کاربر264457

عضویت پنج سال پیش

10 پروژه ثبت شده ،

0 پروژه در حال انجام ،

0 پروژه آماده دریافت پیشنهاد ،

نرخ پذیرش پیشنهاد 50%