How to start kaggling — just make first step

3 min readApr 7, 2019

I guess it is common when someone is fearing to start participating in Kaggle competitions. Someone may think that Kaggle has high entry level — there is a large contention as a lot of highly experienced masters participate in Kaggle competition and newbies without experience will have hard time to compete against them. Someone may think that the participation will take much time and efforts at the begin (as well I was). Someone else may think that only innovative and original solutions will be accepted (such an academic point of view).

In reality neither of above thoughts is completely true. You should start by making the first step right now. Practice is the most valuable reward you will get out of it.

Here is my story. So I decided to make a first step and took part in a playground competition Aerial Cactus Identification which was around here just 2 days (month ago). This competition supports kernel-only submission (you can compile your kernel and submit the output results in csv file).

I forked a kernel Simple CNN and started to explore it. The only change I did is to change the number of epochs from 4 to 6.

m = Sequential()
def tdsNet(m):
    m.add(Conv2D(32, kernel_size=3, activation='relu', input_shape=input_shape))
    m.add(Conv2D(16, kernel_size=3, activation='relu'))
    m.add(Flatten())
    m.add(Dense(units = output_shape, activation='sigmoid'))
tdsNet(m)
# compile adam with decay and use binary_crossentropy for single category dataset
m.compile(optimizer = 'nadam',
          loss = 'binary_crossentropy', 
          metrics = ['accuracy'])

My first commit brought me to 22nd place on leaderboard:

22nd is not bad. Let’s make that a bit better.

My second submission was sadly worse than the first one. Anyway I have got a place right before author of the original kernel:

What is next. You can increase the number of epochs, play with parameters of the model, with data augmentation. Anything. There is a lot of ways to improve the accuracy and you will get a lot of fun along the way.

How to submit prediction from kernel

First we need to commit the kernel i.e. run the kernel from top to bottom. Click button Commit in top menu above the kernel notebook. The committing process will start immedeately. Once the calculations end, you will see the green status Completed