beerboaa / color-classification-cnn Goto Github PK

I ran the code but GPU is not been used i have installed tensorflow-gpu==1.14.0 and keras==2.2.4. Anyone noticed the same problem or solved it ?

Was anybody able to find the dataset?

Your implementation use an input size of (224, 224, 3) but the paper says they use (227, 227, 3).

Your implementation use 5 epochs. The paper says that they used 200000 iterations. The training set is composed of 7799 images. Using a batch size of 32 this would be around 820 epochs.

Where I can find the labeled data set of colored cars?

Your implementation uses Batch Normalization but in the paper is said that the network use Local Response Normalization.

thanks for sharing the code.
I couldn't find the image dataset from the original paper and from the reference [2] inside that paper neither. Would you please provide the image dataset as well, thanks!

Validation accuracy reached over 95% by epoch 5 during training. However, I'm only able to reach 54% while inferencing. Please find the code I use for getting model output below -

input = tf.keras.preprocessing.image.load_img("test/3.jpg", target_size=(224,224))
input_arr = keras.preprocessing.image.img_to_array(input)
input_arr = np.array([input_arr])
#input_arr = input_arr.astype('float32')
#input_arr /= 255
t = time.time()
output = model.predict(input_arr)
y_class = output.argmax(axis=-1)
print("time taken = ", (time.time()-t)*1000)
print(output)
print(y_class)

If I don't rescale the image to [0,1], I get an accuracy of 38%. Also, after if I print output - it is always similar to the following -
[[0. 0. 0. 0. 0. 0. 0. 1. 0.]]
Which means probability for whatever class is 1, and for rest all classes is 0.

Run your code but it give error on ln: 136 & 115. No such file or directory: 'train/' ,
kindly tell the technique
i have run color_net_training.py file.