We tolk a lot about new trend writting code with help AI. If you look into it, it will become obviously: AI capable of replacing small parts of modern code in companies.
Today AI much more effective in areas: detecting objects, words bots and computer vision.
On picture not very hard neural network, which based on a series of convolutions and pulls. This particular design names UNet-Segmentation.
- Some useful libraries will help to impact data for training network:
numpy, pandas, matplotlib
df = pd.read_csv('data/train_masks.csv')
train_df = df[:4000]
val_df = df[4000:]
img_name, mask_rle = train_df.iloc[4]
img = cv2.imread('data/train/{}'.format(img_name))
mask = rle_decode(mask_rle)
- Next step to success coding AI: copying achitecture to Python (I usually use Google Colab/Jupyter Notebook). Might help:
keras
conv_1_1 = Conv2D(32, (3, 3), padding='same')(inp)
conv_1_1 = Activation('relu')(conv_1_1)
conv_1_2 = Conv2D(32, (3, 3), padding='same')(conv_1_1)
conv_1_2 = Activation('relu')(conv_1_2)
pool_1 = MaxPooling2D(2)(conv_1_2)
- The last one: model training. Sometimes it takes a little time (for me ~ 7 minutes) for complete all areas
model.fit_generator(keras_generator(train_df, batch_size),
steps_per_epoch=100,
epochs=100,
verbose=1,
callbacks=callbacks,
validation_data=keras_generator(val_df, batch_size),
validation_steps=50,
class_weight=None,
max_queue_size=10,
workers=1,
use_multiprocessing=False,
shuffle=True,
initial_epoch=0)
Top comments (0)