Baseline
Contents
Baseline#
This notebook will do a simple pipeline based on fastai with transfer learning concepts using resnet architecture. Typical ideas I want to implement here
Get a working pipeline with the current dataset ( using Fastai)
Do a submission on Kaggle
Do some iterations using simple data augmentations
We will split the train dataset into 80-20 split. The idea is to have an actual measurement of model performance looking at the training data. However, trick for kaggle here is to train the model on entire training dataset before submitting results and reviewing score on the test.
Note
Big Images are not easy to fit on GPU.
Resizing big image to smaller sizes tends to loose certain features
Next ideas
Apply Digit Cleaner concept
Create Clean Dataset / Visualize and review accuracy of digit cleaner
Figure out way to do a bounding box
After that we can apply a few ideas
Apply resnet after digit cleaner ( current data had 0.93 error_rate after 1 epoch)
Do bounding box and break down dataset into individual digits. Then do clustering to 10 categories. Then use labels to calculate sum_digit
Split into individual digit images -> Resize -> Merge (permutations) into single image then train model on new dataset
Imports#
%load_ext autoreload
%autoreload 2
import torch
torch.cuda.empty_cache()
!nvidia-smi
Sat Mar 12 09:21:11 2022
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 470.103.01 Driver Version: 470.103.01 CUDA Version: 11.4 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 Tesla K80 Off | 00000000:00:1E.0 Off | 0 |
| N/A 32C P8 30W / 149W | 3MiB / 11441MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| No running processes found |
+-----------------------------------------------------------------------------+
import fastai
from fastai.vision.all import *
from fastai.vision.all import *
from fastai.vision.widgets import *
from aiking.data.external import * #We need to import this after fastai modules
import warnings
from matplotlib import cm
warnings.filterwarnings("ignore")
path = untar_data("kaggle_competitions::ultra-mnist"); path
(path/"train").ls()
(#28000) [Path('/Landmark2/pdo/aiking/data/ultra-mnist/train/ypuccwrtnt.jpeg'),Path('/Landmark2/pdo/aiking/data/ultra-mnist/train/jsutblirhq.jpeg'),Path('/Landmark2/pdo/aiking/data/ultra-mnist/train/nbekltglqy.jpeg'),Path('/Landmark2/pdo/aiking/data/ultra-mnist/train/ercjirbxrl.jpeg'),Path('/Landmark2/pdo/aiking/data/ultra-mnist/train/cmuvttuftw.jpeg'),Path('/Landmark2/pdo/aiking/data/ultra-mnist/train/hmqywwuolp.jpeg'),Path('/Landmark2/pdo/aiking/data/ultra-mnist/train/iwigdcizvr.jpeg'),Path('/Landmark2/pdo/aiking/data/ultra-mnist/train/nekpgbwcjl.jpeg'),Path('/Landmark2/pdo/aiking/data/ultra-mnist/train/jeuyqrvgrt.jpeg'),Path('/Landmark2/pdo/aiking/data/ultra-mnist/train/qdrqstkfop.jpeg')...]
df_train = pd.read_csv(path/'train_train.csv')
df_train['id'] = df_train['id'] + ".jpeg"
df_train
| id | digit_sum | |
|---|---|---|
| 0 | oyihlzmbal.jpeg | 4 |
| 1 | qghmkvofim.jpeg | 9 |
| 2 | fbrfhdmlzf.jpeg | 1 |
| 3 | kopyiltoku.jpeg | 5 |
| 4 | aozdekkrks.jpeg | 7 |
| ... | ... | ... |
| 22395 | acishewpws.jpeg | 22 |
| 22396 | nmgyqgimgu.jpeg | 22 |
| 22397 | zqgvuxngmf.jpeg | 20 |
| 22398 | qugvixgehi.jpeg | 12 |
| 22399 | yhqkvjsdqr.jpeg | 1 |
22400 rows × 2 columns
dls = ImageDataLoaders.from_df(df_train, path, folder='train_black', valid_pct=0.2, fn_col=0, label_col=1, bs=16, item_tfms=Resize(299)); dls
<fastai.data.core.DataLoaders at 0x65ca2612fd0>
dls.show_batch()
!nvidia-smi
Sat Mar 12 09:36:40 2022
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 470.103.01 Driver Version: 470.103.01 CUDA Version: 11.4 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 Tesla K80 Off | 00000000:00:1E.0 Off | 0 |
| N/A 47C P0 69W / 149W | 557MiB / 11441MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| 0 N/A N/A 5666 C ...da/envs/aiking/bin/python 554MiB |
+-----------------------------------------------------------------------------+
learn = cnn_learner(dls, resnet34, metrics=error_rate); learn
<fastai.learner.Learner at 0x65ca1c3acd0>
learn.fine_tune(4)
| epoch | train_loss | valid_loss | error_rate | time |
|---|
learn