a_deployment – Rahul Saraf

Compatibility Block

Check Platform

Platform & Environment Configuration

Imports

Public Imports

from fastai.vision.all import *
from fastcore.all import *
from functools import partial
import warnings
import panel as pn

pn.extension(design='material')
warnings.filterwarnings("ignore")

Private Imports

from aiking.data.external import *
from aiking.core import aiking_settings
from aiking.integrations.datasette import *
from aiking.dl.widgets import PersistentImageClassifierCleaner
from aiking.dl.panelwidgets import PanelImageClassifierCleaner

Pandemic Safety

Our dataset consist of images labelled as mask or no_mask. We will do following steps

Build a classifier on raw version of data
Review images with highest confusion. Identify images we would, like to keep, relabel or skip

dsname = 'PandemicSafety'
datasette_base_url = "https://datasette.zealmaker.com"

path = data_frm_datasette(dsname, datasette_base_url); path

Path('/mnt/d/rahuketu/programming/AIKING_HOME/data/PandemicSafety')

DataBlocks and DataLoaders

In present example, I will use image.csv to generate items and labels. Advantages are:-

I don’t have to delete any file
I can manage modifications to dataset in a new csv file and upload the version on Datasette for future review
dataloaders has a path argument. To be consistent with fastai dataloader we need to supply path for df execution

We will get list of fnames and labels from dataframe. If we require cleaning we will save any data modifications in new csv which we can then upload to datasette. This way our data stays immutable.

df = get_image_df(path)
get_images_from_df(path, df=df)

(#309) [Path('/mnt/d/rahuketu/programming/AIKING_HOME/data/PandemicSafety/No_Mask/7a8b0909-4f5c-4d53-b1bc-f35a83a022c9.jpg'),Path('/mnt/d/rahuketu/programming/AIKING_HOME/data/PandemicSafety/No_Mask/1f54d9ac-4a0e-42fe-98cb-09938a3104b0.jpg'),Path('/mnt/d/rahuketu/programming/AIKING_HOME/data/PandemicSafety/No_Mask/d08a498f-27cf-4668-a4f1-af32dfd15416.jpg'),Path('/mnt/d/rahuketu/programming/AIKING_HOME/data/PandemicSafety/No_Mask/52601346-132e-4217-ab57-6c324c1e4eee.jpg'),Path('/mnt/d/rahuketu/programming/AIKING_HOME/data/PandemicSafety/No_Mask/78ac55fb-2db3-48e9-bdcc-4caa2a50e3f6.jpg'),Path('/mnt/d/rahuketu/programming/AIKING_HOME/data/PandemicSafety/No_Mask/5eeec12d-174e-4d1d-a261-8dcd99ea9b8b.jpg'),Path('/mnt/d/rahuketu/programming/AIKING_HOME/data/PandemicSafety/No_Mask/f048f38a-3758-428d-8d9f-8462a4e272d0.jpeg'),Path('/mnt/d/rahuketu/programming/AIKING_HOME/data/PandemicSafety/No_Mask/cfbe3224-4ec8-47f6-8ee8-d637572b2787.jpg'),Path('/mnt/d/rahuketu/programming/AIKING_HOME/data/PandemicSafety/No_Mask/b2b14d80-ddf9-4e4a-9b61-afdbabb1cf9d.jpg'),Path('/mnt/d/rahuketu/programming/AIKING_HOME/data/PandemicSafety/No_Mask/d7a7b4e3-4756-4294-88d6-6c23a7ba6741.jpg')...]

dls = ImageDataLoaders.from_lists(path, 
                            fnames=get_images_from_df(path, df=df), 
                            labels=df['label'].values.tolist(), 
                            valid_pct=0.5,
                            item_tfms=[Resize(192, method='squish')])

dls.valid.show_batch(max_n=6)

Model Training

learn = vision_learner(dls, resnet18, metrics=[error_rate, accuracy]); learn

<fastai.learner.Learner at 0x7f26f76976a0>

learn.fine_tune(8)

epoch	train_loss	valid_loss	error_rate	accuracy	time
0	1.429095	3.356303	0.564935	0.435065	00:24

epoch	train_loss	valid_loss	error_rate	accuracy	time
0	0.771222	2.435999	0.532468	0.467532	00:21
1	0.655559	2.050427	0.512987	0.487013	00:19
2	0.518054	1.739411	0.409091	0.590909	00:19
3	0.391933	1.455391	0.344156	0.655844	00:20
4	0.315473	1.216033	0.233766	0.766234	00:19
5	0.265082	0.994206	0.214286	0.785714	00:19
6	0.222884	0.825363	0.214286	0.785714	00:19
7	0.191623	0.693891	0.168831	0.831169	00:18

Classification Interpretation

interp = ClassificationInterpretation.from_learner(learn)
interp.plot_confusion_matrix()

interp.plot_top_losses(k=20)

Where do we see high losses?

When model predicts incorrect class with high confidence.
When model predicts correct class but the confidence low.

Data Cleaning

w = PanelImageClassifierCleaner(learn, max_n=25); w

def get_new_dataset(df, mods, join_col='fn', skip_col='skipped', new_label_col='new_label', old_label_col='label'):
    mods = [{join_col: k.name, skip_col: v == '<Delete>', new_label_col:[None, v] [v !=  '<Delete>'] } for k,v in mods.items()]
    df_mods = pd.DataFrame(mods); df_mods
    df_new = df[:]
    merged_df = df_new.merge(df_mods, on=join_col, how='outer')
    # Assign old label is new_label is NaN.
    merged_df.loc[merged_df[new_label_col].isnull(), new_label_col] = merged_df[merged_df[new_label_col].isnull()][old_label_col]
    # Assign skip_col = False of skip_col is NaN
    merged_df.loc[merged_df[skip_col].isnull(), skip_col] = False
    return merged_df

Model and Cleaned Data Saving

df_new_ds = get_new_dataset(df, mods = w.get_mods())
df_new_ds.to_csv(path/"cleaned_v1.csv", index=False)

learn.export(aiking_path('model')/"pandemic_v1.pkl")