the_picnic_vision_training.py

# Data Science Libraries
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
# %matplotlib inline

# Fastai Library
import fastai
from fastai.vision import *
from fastai.vision.models import *
import torch

# Others
from pathlib import Path
import glob
from PIL import ImageFile
from google.colab import drive
import re

# making some settings.
np.random.seed(42)
ImageFile.LOAD_TRUNCATED_IMAGES = True

# Accessing The Data via Google Drive.
drive.mount('/content/gdrive')

# Make sure to setup the path to yours.
path_to_folder = 'gdrive/My Drive/Dataset/The Picnic Hackathon 2019/'
train_df = pd.read_csv(path_to_folder + "train.tsv",  sep='\t')
# Take look at the structure.
train_df.head()

# We will take off some images from the training process.
bananas = [66, 425, 751, 758, 913, 1517, 1555, 1633, 2175, 2860, 3116, 3473, 4104, 4713, 4862, 4983, 5416, 5431, 5436, 5671, 5690, 6106, 7022, 7141, 7154]
aspags = [599,1300,2595]
potatos = [91, 131, 1184, 2998, 3373, 3397, 3653, 4599, 5718, 5918, 6646]
pudding = [734,1059,1203,1252,1788,2028,2161,2310,2597,2633,3448,3629,4151,4879,4928,5064,5662,5771,5970,5973,7016]
poultry = [200,2147,2337,2606,3003,5602,6932]
salad = [1519,1778,3088,3190,4609,5430,6707,6921,7127]
fresh_bread= [1405,1967,2775,3606]
fresh_herbs= [1419,3737]
pork = [699,982,3451,4408,5146,5644,5936,7057]
minced_meat = [2842,2920,3911,4442,5119,5405,5721,6607,7214]
nectarine = [839,1018,1465,2051,2773,3294,3295,3340,4679,5887,6735,6909]
kiwis = [183,477,501,681,2106,2348,3660,4090,4277,5538,6854]
lunch_deli = [730,1506,3024,3909,4101,5498]
milk = [1779,1854,2784,3797,4849,6910]
eggs = [2403,2729,2755,4946,4949,5200,5248,5908,6346,6665]
fish = [19,57,1685,2428,4992,6137,4748]
onion = [1490,2327,3231,6075,4127,5606,6208]
broccli = [395,1044,1317,1420,11575,2561,2929,3741,3945,4488,4902,5616,5643,5666,7028]
cheese = [1009,1026,4782,4892,5024,5133,5365,6394]
citrus = [183,560,694,2274,6050]
cucumber = [1135,1401,1865,2137,2386,3019,3558,3624,3740,3850,4097,4119,4206,4305,4319,4349,5179,5480,6272,6307]
pinapls = [573,1866,3885,5585,7048,7225]
prepacked_bread = [52,406,520,3535,4540,5219,5432,5847,6705,6882,7085]
bell = [128,276,388,1740,3285,5150,5915,6069,6114,6376,7017]
berries = [1379,1482,1725,2817,2883,3695,5094,5342,5401,5550,5635,6220]

# making them in one big list.
all_ = bananas + aspags + potatos + pudding + poultry + salad + fresh_bread + fresh_herbs + pork + minced_meat + nectarine + kiwis + lunch_deli + milk + eggs + fish + onion+ broccli +cheese+citrus+cucumber+pinapls+prepacked_bread+bell+berries

# searching their index in the dataframe.
serie = train_df['file'].values
index = []
for i, value in enumerate(serie):
    for j in all_:
        a = r"^"+str(j)+"\.."        
        if re.match(a,value):
            index.append(i)
            
# deleting files from the data frame.
train_df_new = train_df.drop(index)

print("The New Number of Images is : {}".format(len(train_df_new)))
print("The Past Number of Images is : {}".format(len(train_df)))

# getting the count of each class
label_counts = train_df_new.label.value_counts()
plt.figure(figsize = (12,6))
sns.barplot(label_counts.index, label_counts.values, alpha = 0.9)
plt.xticks(rotation = 'vertical')
plt.xlabel('Image Labels', fontsize =12)
plt.ylabel('Counts', fontsize = 12)
plt.show()

# Defining The Model, Batch Size, size of the Images and path where the training Images are.
MODEL= densenet161
BATCH = 32
SIZE = 224
pathTrain = Path("gdrive/My Drive/train/")

# Loading The Images 

# defining Transformations for Image Augmentation.
tfms = get_transforms(do_flip=True, flip_vert=True, max_rotate=300)

# Making Objet to hold all the images of training and validation, be sure to normalize them using imagenet stats.
data = ImageDataBunch.from_df(pathTrain, train_df_new, size= SIZE, bs = BATCH, ds_tfms = tfms, valid_pct = 0.0).normalize(imagenet_stats)

# show sample of images.
data.show_batch(rows=3, figsize=(12,12))

# Model Preparation

from sklearn.metrics import f1_score

# Making Function to use it in metrics for the CNN.
def f1_macro_score(y_pred, y_true, tens=True):
    score = f1_score(y_true, np.argmax(y_pred, axis = 1), average = 'macro')
    if tens:
        score= tensor(score)
    else:
        score= score
    return score

# Function to calculate the F1-Score for validation data using Confusion Matrix.
def f1_from_cm(cm):
    TP = np.diag(cm)
    FP = np.sum(cm, axis=0) - TP
    FN = np.sum(cm, axis=1) - TP
    num_classes = 25
    TN = []
    for i in range(num_classes):
      temp = np.delete(cm, i, 0)
      temp = np.delete(temp, i, 1)
      TN.append(sum(sum(temp)))
    precision = TP/(TP+FP)
    recall = TP/(TP+FN)
    F1 = 2 * (precision.mean() * recall.mean()) / (precision.mean() + recall.mean())
    return F1

def performane_indiators(model):
    # Plotting some curves to see the performance.
    model.recorder.plot_losses()
    model.recorder.plot_metrics()
    model.recorder.plot()
    
    # Print confusion matrix, F1 score and top losses.
    result = ClassificationInterpretation.from_learner(model)
    result.plot_top_losses(9, figsize=(15,15) , heatmap = False)
    result.plot_confusion_matrix(figsize=(12,12), dpi = 50)
    print("F1-Score for validation is {}".format(f1_from_cm(result.confusion_matrix())))

# Making The CNN Model Using Our Images And Densenet161 Architeture.
model = cnn_learner(data, MODEL, metrics=[accuracy])

# Model Training

# Fitting the model to data using OneCycle Policy.
model.fit_one_cycle(20)

# show results.
performane_indiators(model)

model.save('stage-one')

# unfreeze the model to train top layers.
model.unfreeze()
# Fitting the unfrozzen model to data using OneCycle Policy.
model.fit_one_cycle(15)

# show result.
performane_indiators(model)

model.save('stage-two')

# Exporting Model

# we will export the model which got the best result on validation set.
# in my case, the best was the stage-one.
model.load('stage-one')
model.export('densenet161_best_so_far.pkl')

print("Saved to path: {}".format(model.path))