Htydjtk

I have huge car photos.

I want to predict car's "brand-model-body type and production year"

First, I splitted data into train and validation, and I categorized them like this.

Every category has about 1000 train and 900 validation images.

My plan was: I train my keras model with these categories after training, model can predict labels like below:

audi a3 sedan 2008 => %25

audi a3 sedan 2009 => %25

audi a3 sedan 2010 => %25

audi a3 sedan 2011 => %25

And I can tell user that: "This car is Audi A3 Sedan 2008-2011"

My problem is, some of these categories have very similar photos. For example: audi a3 2009 and audi a3 2010 have same body type and there is not much difference between photos (No difference in reality).
Because of that, train accuracy has improved to about 0.9 but validation accuracy hasn't improved above 0.55

When I try some predictions, it usually gives same label, "Ford Focus sedan 2009" :)

Here is my output:

epoch,          acc,              loss,              val_acc,       val_loss

27,    0.7965514530544776, 0.56618134500483, 0.5192149643316993, 1.729015349846447



28,    0.8058803490480816, 0.5408204138258657, 0.5176764522193236, 1.778763979018732



29,    0.8167710489770164, 0.5116128672937693, 0.523258489762041, 1.7806432932022545



30,    0.8256544639818643, 0.4872381848016096, 0.5207534764479939, 1.8059904007678271



31,    0.8355546238309248, 0.4629556378035959, 0.5237253032663666, 1.8191414148756815



32,    0.8424464767701014, 0.4444190686917562, 0.5242512903147193, 1.8496954914466912



33,    0.8508739288802705, 0.422022156655134, 0.5303593149032422, 1.8565427863780883



34,    0.8576819265745635, 0.40545297008116027, 0.5262894901236571, 1.909881308499735

My train code is here:

Image_width, Image_height = 224, 224

num_epoch = 5000

batch_size = 16

learning_rate = 0.0001

model = ResNet50(weights='imagenet', include_top=False, input_shape=(Image_width, Image_height, 3))

fc_neuron_count = 1024

output = model.output

output = GlobalAveragePooling2D()(output)

output = Dense(fc_neuron_count, activation='relu')(output)

predictions = Dense(num_classes, activation='softmax')(output)

model = Model(inputs=model.input, outputs=predictions)



model.compile(optimizer=opt.Adam(lr=learning_rate), loss=losses.categorical_crossentropy,

              metrics=['accuracy'])



history_transfer_learning = model.fit_generator(

    train_generator,

    epochs=num_epoch,

    steps_per_epoch=num_train_samples // batch_size,

    validation_data=validation_generator,

    validation_steps=num_validate_samples // batch_size,

    class_weight='auto',

    callbacks=callbacks_list)

• Am I doing something wrong? How can I achieve this result?




• Should I change validation accuracy calculation, or should I give more photos per category?

I want to say how I solved my problem for anyone who is looking for a similar question.

My categorization was a mistake. I realized later that, I gave the same photos to my model and waited for the different results. For example, I had nearly same photos in Audi A3 Hatchback/5 2009 and Audi A3 Hatchback/5 2010. When the model starts training, first it learns data. Then it predicts and validates output itself. If the output is wrong, it tries a different way to success. But wait a minute, there was no mistake. I gave you the same photos and waited for different results! My categorization failed here.

I categorized my cars by body changes like Audi A3 Hatchback 2008-2013. Except for categories that have the wrong photos, my results are great for now.

Now, we will work on better photos and better year categorization.