Loading
Feedback

Seismic Facies Identification Challenge

[Explainer] Introduction and General Approach Final Pack!

Introduction to this challenge, general approach, my approach, and what I learn from the others

This is my final explainer trying to summarize what I know and learned from this challenge especially from other wonderful participants.

Notebook list:

I'll try to add more notebooks in the comment about my round 1 approach and what new things I learned (post-processing method, etc).

I hope you guys enjoy it and learn something from this explainer.

 

---

 

Thanks to SEAM AI and Aicrowd for organizing this event.

Also, shout out to other contestant's explainer :

This has been a wonderful experience, I'm glad that I spent my time for this challenge. Always checked the forum and leaderboard every time I got back from work.

I learned a lot from this community and was surprised that I can push my score to F1:0.901 Acc:0.941 (Round 1 Unweighted) and F1:0.770 Acc:0.737 (Round 2 weighted). I'm pretty noob at this.



welcome to.gif

ses.gif

using dl.gif

with.gif leo2.gif


Shout-out

First of all, thank you for SEAM AI and Aicrowd for organizing this event.

Also shout out to other contestants explainer (I suggest you read it too) :

I learned a lot from this community and suprised that I can push my score to F1:0.901 Acc:0.941 (Round 1 Unweighted) and F1:0.770 Acc:0.737 (Round 2 weighted).

Also checkout my previous explainer too here :P

ps: for geoscientist, what I meant for processing here is to transform input data provided from this challenge.

Introduction

Please watch :)


In [ ]:
#@title
from IPython.display import HTML,clear_output
from base64 import b64encode
!gdown "https://drive.google.com/uc?id=1PuQU_NZzKYAhXMYLMBU1Ff3VQ02PuOs7"
mp4 = open('render.mp4','rb').read()
clear_output()
data_url = "data:video/mp4;base64," + b64encode(mp4).decode()
HTML("""
<video width=480 controls>
      <source src="%s" type="video/mp4">
</video>
""" % data_url)
Output hidden; open in https://colab.research.google.com to view.

Sorry for the tone difference lol, I only got time to do this while my family is asleep.

First, we download the training data for the challenge :

In [ ]:
!gdown "https://drive.google.com/uc?id=14u7fkARS8WRJUdhvU79kDxg8EKTqg606"
!gdown "https://drive.google.com/uc?id=1--tADAa10l2M1iaSEslGXK-RaBv8UbMf"
Downloading...
From: https://drive.google.com/uc?id=14u7fkARS8WRJUdhvU79kDxg8EKTqg606
To: /content/data_train.npz
1.72GB [00:18, 94.5MB/s]
Downloading...
From: https://drive.google.com/uc?id=1--tADAa10l2M1iaSEslGXK-RaBv8UbMf
To: /content/labels_train.npz
7.16MB [00:00, 63.3MB/s]

Let's install some package that'll make our job easier :

In [ ]:
!pip install segmentation-models-pytorch==0.1.2  # easy to use some famous model architecture. visit https://github.com/qubvel/segmentation_models.pytorch/
!pip install albumentations               # easy image manipulation for data augmentation
Collecting segmentation-models-pytorch==0.1.2
  Downloading https://files.pythonhosted.org/packages/03/36/37b6b0e54a98ff15eb36ce36c9181fdb627b3e789e23fc764f9e5f01dc68/segmentation_models_pytorch-0.1.2-py3-none-any.whl (53kB)
     |████████████████████████████████| 61kB 8.1MB/s 
Requirement already satisfied: torchvision>=0.3.0 in /usr/local/lib/python3.6/dist-packages (from segmentation-models-pytorch==0.1.2) (0.8.1+cu101)
Collecting pretrainedmodels==0.7.4
  Downloading https://files.pythonhosted.org/packages/84/0e/be6a0e58447ac16c938799d49bfb5fb7a80ac35e137547fc6cee2c08c4cf/pretrainedmodels-0.7.4.tar.gz (58kB)
     |████████████████████████████████| 61kB 8.9MB/s 
Collecting timm==0.1.20
  Downloading https://files.pythonhosted.org/packages/89/26/ba294669cc5cc4d09efd1964c8df752dc0955ac26f86bdeec582aed77d1d/timm-0.1.20-py3-none-any.whl (161kB)
     |████████████████████████████████| 163kB 36.5MB/s 
Collecting efficientnet-pytorch==0.6.3
  Downloading https://files.pythonhosted.org/packages/b8/cb/0309a6e3d404862ae4bc017f89645cf150ac94c14c88ef81d215c8e52925/efficientnet_pytorch-0.6.3.tar.gz
Requirement already satisfied: torch==1.7.0 in /usr/local/lib/python3.6/dist-packages (from torchvision>=0.3.0->segmentation-models-pytorch==0.1.2) (1.7.0+cu101)
Requirement already satisfied: numpy in /usr/local/lib/python3.6/dist-packages (from torchvision>=0.3.0->segmentation-models-pytorch==0.1.2) (1.19.5)
Requirement already satisfied: pillow>=4.1.1 in /usr/local/lib/python3.6/dist-packages (from torchvision>=0.3.0->segmentation-models-pytorch==0.1.2) (7.0.0)
Collecting munch
  Downloading https://files.pythonhosted.org/packages/cc/ab/85d8da5c9a45e072301beb37ad7f833cd344e04c817d97e0cc75681d248f/munch-2.5.0-py2.py3-none-any.whl
Requirement already satisfied: tqdm in /usr/local/lib/python3.6/dist-packages (from pretrainedmodels==0.7.4->segmentation-models-pytorch==0.1.2) (4.41.1)
Requirement already satisfied: dataclasses in /usr/local/lib/python3.6/dist-packages (from torch==1.7.0->torchvision>=0.3.0->segmentation-models-pytorch==0.1.2) (0.8)
Requirement already satisfied: future in /usr/local/lib/python3.6/dist-packages (from torch==1.7.0->torchvision>=0.3.0->segmentation-models-pytorch==0.1.2) (0.16.0)
Requirement already satisfied: typing-extensions in /usr/local/lib/python3.6/dist-packages (from torch==1.7.0->torchvision>=0.3.0->segmentation-models-pytorch==0.1.2) (3.7.4.3)
Requirement already satisfied: six in /usr/local/lib/python3.6/dist-packages (from munch->pretrainedmodels==0.7.4->segmentation-models-pytorch==0.1.2) (1.15.0)
Building wheels for collected packages: pretrainedmodels, efficientnet-pytorch
  Building wheel for pretrainedmodels (setup.py) ... done
  Created wheel for pretrainedmodels: filename=pretrainedmodels-0.7.4-cp36-none-any.whl size=60963 sha256=15ecf5d77b10b7173e744ea8e29e484fd320dd124df60adf1fb7da5e67054433
  Stored in directory: /root/.cache/pip/wheels/69/df/63/62583c096289713f22db605aa2334de5b591d59861a02c2ecd
  Building wheel for efficientnet-pytorch (setup.py) ... done
  Created wheel for efficientnet-pytorch: filename=efficientnet_pytorch-0.6.3-cp36-none-any.whl size=12421 sha256=d62b69be61ee765ae545f905231aa4a6576329cedad28cb4234bf47486d6586b
  Stored in directory: /root/.cache/pip/wheels/42/1e/a9/2a578ba9ad04e776e80bf0f70d8a7f4c29ec0718b92d8f6ccd
Successfully built pretrainedmodels efficientnet-pytorch
Installing collected packages: munch, pretrainedmodels, timm, efficientnet-pytorch, segmentation-models-pytorch
Successfully installed efficientnet-pytorch-0.6.3 munch-2.5.0 pretrainedmodels-0.7.4 segmentation-models-pytorch-0.1.2 timm-0.1.20
Requirement already satisfied: albumentations in /usr/local/lib/python3.6/dist-packages (0.1.12)
Requirement already satisfied: numpy>=1.11.1 in /usr/local/lib/python3.6/dist-packages (from albumentations) (1.19.5)
Collecting imgaug<0.2.7,>=0.2.5
  Downloading https://files.pythonhosted.org/packages/ad/2e/748dbb7bb52ec8667098bae9b585f448569ae520031932687761165419a2/imgaug-0.2.6.tar.gz (631kB)
     |████████████████████████████████| 634kB 16.2MB/s 
Requirement already satisfied: opencv-python in /usr/local/lib/python3.6/dist-packages (from albumentations) (4.1.2.30)
Requirement already satisfied: scipy in /usr/local/lib/python3.6/dist-packages (from albumentations) (1.4.1)
Requirement already satisfied: scikit-image>=0.11.0 in /usr/local/lib/python3.6/dist-packages (from imgaug<0.2.7,>=0.2.5->albumentations) (0.16.2)
Requirement already satisfied: six in /usr/local/lib/python3.6/dist-packages (from imgaug<0.2.7,>=0.2.5->albumentations) (1.15.0)
Requirement already satisfied: matplotlib!=3.0.0,>=2.0.0 in /usr/local/lib/python3.6/dist-packages (from scikit-image>=0.11.0->imgaug<0.2.7,>=0.2.5->albumentations) (3.2.2)
Requirement already satisfied: PyWavelets>=0.4.0 in /usr/local/lib/python3.6/dist-packages (from scikit-image>=0.11.0->imgaug<0.2.7,>=0.2.5->albumentations) (1.1.1)
Requirement already satisfied: networkx>=2.0 in /usr/local/lib/python3.6/dist-packages (from scikit-image>=0.11.0->imgaug<0.2.7,>=0.2.5->albumentations) (2.5)
Requirement already satisfied: pillow>=4.3.0 in /usr/local/lib/python3.6/dist-packages (from scikit-image>=0.11.0->imgaug<0.2.7,>=0.2.5->albumentations) (7.0.0)
Requirement already satisfied: imageio>=2.3.0 in /usr/local/lib/python3.6/dist-packages (from scikit-image>=0.11.0->imgaug<0.2.7,>=0.2.5->albumentations) (2.4.1)
Requirement already satisfied: cycler>=0.10 in /usr/local/lib/python3.6/dist-packages (from matplotlib!=3.0.0,>=2.0.0->scikit-image>=0.11.0->imgaug<0.2.7,>=0.2.5->albumentations) (0.10.0)
Requirement already satisfied: python-dateutil>=2.1 in /usr/local/lib/python3.6/dist-packages (from matplotlib!=3.0.0,>=2.0.0->scikit-image>=0.11.0->imgaug<0.2.7,>=0.2.5->albumentations) (2.8.1)
Requirement already satisfied: pyparsing!=2.0.4,!=2.1.2,!=2.1.6,>=2.0.1 in /usr/local/lib/python3.6/dist-packages (from matplotlib!=3.0.0,>=2.0.0->scikit-image>=0.11.0->imgaug<0.2.7,>=0.2.5->albumentations) (2.4.7)
Requirement already satisfied: kiwisolver>=1.0.1 in /usr/local/lib/python3.6/dist-packages (from matplotlib!=3.0.0,>=2.0.0->scikit-image>=0.11.0->imgaug<0.2.7,>=0.2.5->albumentations) (1.3.1)
Requirement already satisfied: decorator>=4.3.0 in /usr/local/lib/python3.6/dist-packages (from networkx>=2.0->scikit-image>=0.11.0->imgaug<0.2.7,>=0.2.5->albumentations) (4.4.2)
Building wheels for collected packages: imgaug
  Building wheel for imgaug (setup.py) ... done
  Created wheel for imgaug: filename=imgaug-0.2.6-cp36-none-any.whl size=654020 sha256=2bd0fd1798120cc3cc19e0c808f3723d34695735e6d8b8883223305a26ebfe2f
  Stored in directory: /root/.cache/pip/wheels/97/ec/48/0d25896c417b715af6236dbcef8f0bed136a1a5e52972fc6d0
Successfully built imgaug
Installing collected packages: imgaug
  Found existing installation: imgaug 0.2.9
    Uninstalling imgaug-0.2.9:
      Successfully uninstalled imgaug-0.2.9
Successfully installed imgaug-0.2.6

Import the packages :

In [ ]:
import numpy as np
import matplotlib.pyplot as plt
import torch
import torch.optim as optim
from torch.utils.data import DataLoader, Dataset
import segmentation_models_pytorch as smp
import albumentations as A
import os
from ipywidgets import IntProgress
from IPython.display import display
import time
import cv2
os.environ['CUDA_LAUNCH_BLOCKING'] = "1"

Load the data and see the shape:

In [ ]:
train_data_full = np.load('data_train.npz', allow_pickle=True, mmap_mode='r')['data']
train_label_full = np.load('labels_train.npz', allow_pickle=True, mmap_mode='r')['labels']
In [ ]:
print('shape            :',train_data_full.shape)
print('min-max amplitude:',train_data_full.min(),'&',train_data_full.max())
shape            : (1006, 782, 590)
min-max amplitude: -5195.5234 & 5151.7188

If you see the EDA by sergeytsimfer, especially the amplitude distribution, you can improve your score by doing quantile to the data or for me gain+rms is what perform the best.

You can also add it to extra channel (so the dimension will be vanilla+processed,size x, size y) but for me the improvement is really small and not worth the extra computation time.

In [ ]:
import math
from scipy.signal.windows import triang
from scipy.signal import convolve2d as conv2
def gain(data,dt,parameters):
    nt,nx = data.shape
    dout = np.zeros(data.shape)
    L = parameters/dt+1
    L = np.floor(L/2)
    h = triang(2*L+1)
    shaped_h  = h.reshape(len(h),1)
    for k in range(nx):
        aux = data[:,k]
        e = aux**2
        shaped_e = e.reshape(len(e),1)
        rms = np.sqrt(conv2(shaped_e,shaped_h,"same"))
        epsi = 1e-10*max(rms)
        op = rms/(rms**2+epsi)
        op = op.reshape(len(op),)
        dout[:,k] = data[:,k]*op
    for k in range(nx):
        aux = dout[:,k]
        amax = np.sqrt(sum(aux**2)/nt)
        dout[:,k] = dout[:,k]/amax
    return dout

Let's test it for one slice :

In [ ]:
test_proc=train_data_full[:,0,:]
dat_proc=gain(test_proc,3e-3,0.8)
In [ ]:
fig, ax = plt.subplots(1, 2, figsize=(10,8))
ax[0].imshow(test_proc,interpolation='none',cmap='seismic')
ax[1].imshow(dat_proc,interpolation='none',cmap='seismic')

ax[0].set_title("Vanilla")
ax[1].set_title("Gain+RMS")
plt.show()

Now you can run these to process all the data, but it'll take some time to finish!

In [ ]:
print('preprocessing the data :')
f = IntProgress(min=0, max=train_data_full.shape[1])
display(f)
for i in range(0,train_data_full.shape[1]):
    #print('reprocess : ',i+1,'of',train_data_full.shape[1])
    train_data_full[:,i,:]=gain(train_data_full[:,i,:],3e-3,0.8)
    f.value += 1

if you don't want to wait, then you can use this instead.

In [ ]:
!gdown "https://drive.google.com/uc?id=1JZ5LZz_f2Vfg9BxuGGBY9LliJQAAHi_H"
train_data_full=np.load('data_train_processed.npz')['data']
Downloading...
From: https://drive.google.com/uc?id=1JZ5LZz_f2Vfg9BxuGGBY9LliJQAAHi_H
To: /content/data_train_processed.npz
1.73GB [00:13, 132MB/s] 

then we rescale the data :

In [ ]:
train_data_full = (train_data_full - train_data_full.min()) / (train_data_full.max() - train_data_full.min())

labelRR.png

And let's see the label distribution :

In [ ]:
fig = plt.figure(figsize=(10,5))
labels    = ["1", "2", "3", "4", "5", "6"]
colors = ['hotpink', 'lightskyblue', 'mediumpurple','cornsilk', 'pink', 'lightgrey']
N, bins, patches = plt.hist(train_label_full.flatten(),6,density=True, edgecolor='gray', linewidth=1)
for i in range(6):
    patches[i].set_facecolor(colors[i])
    patches[i].set_label(labels[i])
plt.gca().axes.xaxis.set_ticklabels([])
plt.title('Full Train Data Dist.')
plt.show()

Now, about the test cube, it's better to pick the part that also got the label-5.

labeEl.png

Don't do this :

labeEnol.png

As you can see, if we pick this area as a test, then no label-5 that'll be represented.

In [ ]:
from matplotlib.patches import Rectangle
fig, ax = plt.subplots(figsize=(5,10))
im = ax.imshow(train_label_full[:,-1,:],interpolation='none',cmap='Pastel1')
rect = plt.Rectangle((420,0),300,1500,facecolor='red',alpha=0.5)
ax.add_patch(rect)
plt.show()

Now let's split the train and test data :

In [ ]:
split_sample = 0.7
test_data = train_data_full[:,:train_data_full.shape[1]-int(train_data_full.shape[1]*split_sample),:]
train_data = train_data_full[:,train_data_full.shape[1]-int(train_data_full.shape[1]*split_sample):,:]
test_label = train_label_full[:,:train_label_full.shape[1]-int(train_label_full.shape[1]*split_sample),:]
train_label = train_label_full[:,train_label_full.shape[1]-int(train_label_full.shape[1]*split_sample):,:]
In [ ]:
print('train cube shape :',train_data.shape)
print('test cube shape :',test_data.shape)
train cube shape : (1006, 547, 590)
test cube shape : (1006, 235, 590)
In [ ]:
fig = plt.figure(figsize=(10,10))
plt.imshow(train_label[:,-1,:],interpolation='none',cmap='Pastel1')
plt.colorbar()
plt.show()

and see the label distribution :

In [ ]:
fig, ax = plt.subplots(1, 2, figsize=(10,5))
labels    = ["1", "2", "3", "4", "5", "6"]
colors = ['hotpink', 'lightskyblue', 'mediumpurple','cornsilk', 'pink', 'lightgrey']
N, bins, patches = ax[0].hist(train_label.flatten(),[1, 2, 3, 4, 5, 6, 7],density=True, edgecolor='gray', linewidth=1)
for i in range(6):
    patches[i].set_facecolor(colors[i])
N2, bins2, patches2 = ax[1].hist(test_label.flatten(),[1, 2, 3, 4, 5, 6, 7],density=True, edgecolor='gray', linewidth=1)
for i in range(6):
    patches2[i].set_facecolor(colors[i])
    patches2[i].set_label(labels[i])
ax[0].get_xaxis().set_visible(False)
ax[1].get_xaxis().set_visible(False)
ax[0].set_title("Training Set Dist.")
ax[1].set_title("Testing Set Dist.")
plt.legend(title="Label")
plt.show()

Yikes, that's still quite an imbalance that we got here.

In [ ]:
print('label 5 count of train data : ',np.count_nonzero(train_label.flatten() == 5))
print('label 5 count of test data : ',np.count_nonzero(test_label.flatten() == 5))
label 5 count of train data :  4190167
label 5 count of test data :  428833

Still it's not empty if we splice it the other way.

Setting up the Model, Hyperparameter, Dataloader, etc.

Let's set up some stuff for training the models.

First, the metric that we will use from Aicrowd.

In [ ]:
from sklearn.metrics import multilabel_confusion_matrix
from sklearn.metrics import f1_score,accuracy_score  #for score metric calculation
def _prf_divide(numerator, denominator, ):
    """Performs division and handles divide-by-zero.
    On zero-division, sets the corresponding result elements equal to
    0 or 1 (according to ``zero_division``). 
    """
    mask = denominator == 0.0
    denominator = denominator.copy()
    denominator[mask] = 1  # avoid infs/nans
    result = numerator / denominator

    return result

def compute_scores(y_true, y_pred, class_weights=[1, 1, 1, 1, 20, 20]):
    """
    Computes the weighted & unweighted f1_score and accuracy
    Using the standard F1-Score and class-wise accuracy computations were quite 
    slow as we were doing a lot of redundant work across all score computations,
    hence we have implemented this from the base principles.
    Please refer to the inline comments.
    """

    # Initial Housekeeping Taks1
    y_true = np.array(y_true).flatten()
    y_pred = np.array(y_pred).flatten()
    class_weights = np.array(class_weights)
    # print(np.max(y_true))
    # print(np.max(y_pred))
    # print(np.min(y_true))
    # print(np.min(y_pred))
    # Computing Multilabel Confusion Matrix
    #print("--------- Computing MCM... ")
    #begin_time = time.time()
    MCM = multilabel_confusion_matrix(y_true, y_pred,labels=[1,2,3,4,5,6])
    #print("MCM computation time  : ", time.time() - begin_time)
    
    """
    Gather True Positives, True Negatives, False Positives, False Negatives
    """
    tp_sum = MCM[:, 1, 1]
    tn_sum = MCM[:, 0, 0]
    fn_sum = MCM[:, 1, 0]
    fp_sum = MCM[:, 0, 1]
    
    #print("--------- Computing per class instances... ")
    per_class_instances = np.bincount(y_true) # Helps keep a track of total number of instances per class
    per_class_instances = per_class_instances[1:] # as the class names in the dataset are NOT zero-indexed
    
    assert class_weights.shape == per_class_instances.shape
    
    #print("--------- Computing precision... ")
    # precision : tp / (tp + fp)
    precision = _prf_divide(
                    tp_sum,
                    (tp_sum + fp_sum)
                )
    #print("--------- Computing recall... ")                        
    # recall : tp / (tp + fn)
    recall = _prf_divide(
                    tp_sum,
                    (tp_sum + fn_sum)
                )

    #print("--------- Computing F1 score... ")
    # f1 : 2 * (recall * precision) / (recall + precision)
    f1_score = _prf_divide(
                    2 * precision * recall,
                    precision + recall
                )
    #print("--------- Computing Accuracy... ")
    # accuracy = tp_sum / instances_per_class
    # NOTE: we are computing the accuracy independently for all the class specific subgroups
    # accuracy = _prf_divide(
    #                 tp_sum,
    #                 per_class_instances
    #             )
    # print(class_weights)
    # print(f1_score)
    f1_score_weighted = np.dot(class_weights, f1_score) / np.sum(class_weights)
    f1_score_unweighted = f1_score.mean()

    # accuracy_weighted = np.dot(class_weights, accuracy) / np.sum(class_weights)
    # accuracy_unweighted = accuracy.mean()

    return f1_score_weighted, f1_score_unweighted#, accuracy_weighted, f1_score_unweighted, accuracy_unweighted

Then the training parameters:

In [ ]:
batch_size = 8      
num_epochs = 40      
num_classes = 6       
learning_rate = 0.00085

Set up the architecture, pretty simple using smp, right?

Also we don't use pretrained weight here. (After some experiment, pretrained weight got worse score).

In [ ]:
model = smp.PSPNet(
    encoder_name="efficientnet-b3",        
    encoder_weights=None,     
    in_channels=1,                  
    classes=num_classes,                    
)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model.to(device)
print("")

Now let's put some random value to test if it works:

In [ ]:
test = torch.rand(1, 1, 320, 320).cuda()
out = model(test)
out.shape
Out[ ]:
torch.Size([1, 6, 320, 320])

And set up the optimizer and loss function:

In [ ]:
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(),lr = learning_rate)

Next, let's make the train dataloader:

In [ ]:
class seisdataset_train(Dataset):
    def __init__(self, x_set, y_set):
        self.x, self.y = x_set, y_set
        self.n_sample = self.x.shape[1]+self.x.shape[2]+self.x.shape[1]
        self.aug = A.Compose([
            A.Resize(p=1, height=640, width=320, interpolation=1)
        ]) 
        self.aug2 = A.Compose([
            #A.RandomSizedCrop(p=1.0, min_max_height=(1006, 1006), height=1006, width=256, w2h_ratio=1.0, interpolation=0),
            #A.GridDistortion(p=0.3, num_steps=6, distort_limit=(-0.2, -0.05), border_mode=1),
            A.ElasticTransform(p=0.2,alpha=100, sigma=8, alpha_affine=0, border_mode=1),
            A.ShiftScaleRotate(p=0.5, shift_limit=(0.0, 0.0), scale_limit=(0.01, 0.25), rotate_limit=(-15, 15), interpolation=0, border_mode=1),
            A.RandomCrop(900, 250, p=0.3),
            A.Resize(p=1, height=640, width=320, interpolation=0)
            
        ])  

    def __len__(self):
        return self.n_sample
    
    def __getitem__(self, index):
        if index < self.x.shape[1]:
          idx = index
          batch_x = self.x[:,idx,:]
          batch_y = self.y[:,idx,:]-1
          augmented = self.aug(image=batch_x, mask=batch_y)
        elif self.x.shape[1] <= index < (self.x.shape[1]+self.x.shape[2]):
          idx = index-self.x.shape[1]
          batch_x = self.x[:,:,idx]
          batch_y = self.y[:,:,idx]-1
          augmented = self.aug2(image=batch_x, mask=batch_y)
        elif index >= self.x.shape[1]+self.x.shape[2]:
          idx = index-self.x.shape[1]-self.x.shape[2]
          batch_x = self.x[:,idx,:]
          batch_y = self.y[:,idx,:]-1
          augmented = self.aug2(image=batch_x, mask=batch_y)          
        image, mask = augmented['image'], augmented['mask']
        return image[None,:,:], mask

Wait, what are you doing with albumentation there?

We try to augment some new data.

Well augmentation is.. making new data from available dataset by adding image manipulation algorithm like:

  • flip
  • rotation
  • scale
  • elastic transform
  • adding noise, etc

Just like those Indonesia & India TV drama, where they only got like 5 sec footage but they add multiple effect to the footage and merge it so they get 1 minute unnecessary dramatic moment.