In [1]:
# This Python 3 environment comes with many helpful analytics libraries installed
# It is defined by the kaggle/python docker image: https://github.com/kaggle/docker-python
# For example, here's several helpful packages to load in
import numpy as np # linear algebra
import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)
# Input data files are available in the "../input/" directory.
# For example, running this (by clicking run or pressing Shift+Enter) will list the files in the input directory
import os
print(os.listdir("../input/art-images-drawings-painting-sculpture-engraving"))
# Any results you write to the current directory are saved as output.
In [2]:
from fastai import *
from fastai.vision import *
In [3]:
#Take PATH to be the parent folder of training and valid set.
PATH = "../input/art-images-drawings-painting-sculpture-engraving/art_dataset_cleaned/art_dataset"
In [4]:
torch.cuda.is_available()
Out[4]:
In [5]:
#do_flip=False because we do not want our paintings to be trained flipped or in any other rotation.
tfms = get_transforms(do_flip=False)
#size = the maximum size of our images, use 224 for most cases as told by Jeremy.
#num_workers = 0, the number of CPUs to use. 0 due to lower hardware in Kaggle.
#If training and valid set are already available, direct the function to them via the method below via arguments train and valid.
data = ImageDataBunch.from_folder(PATH, train="training_set", valid="validation_set", ds_tfms=tfms, size=200, num_workers=0)
In [6]:
data.show_batch(rows=3, figsize=(6,6))
In [7]:
data.classes
Out[7]:
In [8]:
#BEFORE THIS STEP, Click on Add Data and Search for Resnet34 from Kaggle. Because Kaggle serves read-only dirs, we cannot download pre-trained weights for Resnet. Now we have to make a dir
#for copying those weights and hence we are making ~/.torch/models folder.
cache_dir = os.path.expanduser(os.path.join('~','.torch'))
if not os.path.exists(cache_dir):
os.makedirs(cache_dir) #first make ~/.torch if not already available.
In [9]:
models_dir = os.path.join(cache_dir,'models')
if not os.path.exists(models_dir):
os.makedirs(models_dir) #then make ~/.torch/models, if not already available.
In [10]:
#Copied resnet34.pth, which are pretrained weights on Resnet34 to our folder into resnet<version>-<sha-hash>.pth
!cp ../input/resnet34/resnet34.pth ~/.torch/models/resnet34-333f7ec4.pth
In [11]:
#MODEL_PATH is declared this way and glued to model_dir attr of cnn_learner.
MODEL_PATH = '/tmp/models'
learn = cnn_learner(data, models.resnet34, metrics=accuracy, model_dir=MODEL_PATH)
In [12]:
#Fitting and checking for the first time.
learn.fit_one_cycle(4)
In [13]:
#Saving the model with ACCURACY = 93.4%
learn.save('stage-1')
In [14]:
#Initiating refit and checking LR
learn.unfreeze()
learn.lr_find()
In [15]:
learn.recorder.plot()
In [16]:
#The lowest loss is at 1e-06 and loss increases from after 1e-04. Refitting by modulating LR
learn.fit_one_cycle(2,max_lr=slice(1e-05,1e-04))
In [17]:
#Saving the model with accuracy 93.6%
learn.save('stage-2')
In [18]:
learn.unfreeze()
learn.lr_find()
In [19]:
learn.recorder.plot()
In [20]:
learn.fit_one_cycle(2,max_lr=slice(1e-06,1e-05))
In [21]:
#Saving model with acc 94.7%
learn.save('stage-3')
In [27]:
#Uploaded a separate dataset for testing each of the above classes.
PRED_PATH = "../input/for-testing-art-images-cnn"
img_icono = open_image(f'{PRED_PATH}/icono.jpg')
img_drawing = open_image(f'{PRED_PATH}/drawing.jpg')
img_engraving = open_image(f'{PRED_PATH}/engraving.png')
img_painting = open_image(f'{PRED_PATH}/painting.jpg')
img_sculpt = open_image(f'{PRED_PATH}/sculpture.jpg')
In [30]:
img_icono
Out[30]:
In [31]:
img_drawing
Out[31]: