NN/Image Classifier: Shape Detection¶

• NN requires more data
• Called deep learning as they become deeper as the number of hidden layers increase, like a funnel

In [ ]:

from tensorflow.keras.layers import Dense, Conv2D,  MaxPool2D, Flatten
import tensorflow as tf
import pandas as pd

from tensorflow.keras.layers import Dense, Conv2D, MaxPool2D, Flatten import tensorflow as tf import pandas as pd

For unzipping Kaggle datasets

In [ ]:

import zipfile

def unzip_file(filename):
  zip_ref = zipfile.ZipFile(filename, 'r')
  zip_ref.extractall()
  zip_ref.close()

import zipfile def unzip_file(filename): zip_ref = zipfile.ZipFile(filename, 'r') zip_ref.extractall() zip_ref.close()

In [ ]:

unzip_file("archive.zip")

unzip_file("archive.zip")

Data Preprocessing¶

In [ ]:

import matplotlib.pyplot as pplt
pplt.imshow(pplt.imread("./shapes/square/101.png"))
# Print to check

import matplotlib.pyplot as pplt pplt.imshow(pplt.imread("./shapes/square/101.png")) # Print to check

Out[ ]:

<matplotlib.image.AxesImage at 0x7f49e94d1d50>

In [ ]:

from tensorflow.keras.preprocessing.image import ImageDataGenerator as idg

from tensorflow.keras.preprocessing.image import ImageDataGenerator as idg

Incase of lack to image data use ImageDataGenerator to generate images, using previous images

• Shering the images
• Moving around and adding extra blank pixels
• Rotating the images
• Brightness etc

Data Augmentation¶

Runs along with the new model, generating data on the fly, cant be visualized or seen on the explorer

In [ ]:

# 1/255 for rescale as its better to get everything into 0 to 1 range - RGB Value
# Smaller number 
train_data_gen = idg(rescale=1/255, rotation_range=37, width_shift_range=0.10, height_shift_range=0.21, validation_split=0.2)

# Reshape every file to target size, better to keep it at 256 or close to 300
# Use it for training and not for validation
train_data = train_data_gen.flow_from_directory("./shapes", target_size=(256, 256), subset="training") 
validation_data = train_data_gen.flow_from_directory("./shapes", target_size=(256, 256), subset="validation")

# 1/255 for rescale as its better to get everything into 0 to 1 range - RGB Value # Smaller number train_data_gen = idg(rescale=1/255, rotation_range=37, width_shift_range=0.10, height_shift_range=0.21, validation_split=0.2) # Reshape every file to target size, better to keep it at 256 or close to 300 # Use it for training and not for validation train_data = train_data_gen.flow_from_directory("./shapes", target_size=(256, 256), subset="training") validation_data = train_data_gen.flow_from_directory("./shapes", target_size=(256, 256), subset="validation")

Found 11968 images belonging to 4 classes.
Found 2990 images belonging to 4 classes.

Remove some samples from the train set for testing

In [ ]:

!mv shapes/triangle/13.png ./test_data
!mv shapes/triangle/130.png ./test_data
!mv shapes/triangle/1300.png ./test_data

!mv shapes/triangle/13.png ./test_data !mv shapes/triangle/130.png ./test_data !mv shapes/triangle/1300.png ./test_data

Modelling¶

Trail 01 - low accuracy¶

In [ ]:

model_1 = tf.keras.Sequential([
    Flatten(),
    Dense(10, activation="relu"),
    Dense(10, activation="relu"),
    Dense(4, activation="softmax")
])

model_1.compile(loss=tf.keras.losses.CategoricalCrossentropy(),  # Loss function that we want
                optimizer=tf.keras.optimizers.Adam(),  # Similar to Gradient Descent
                metrics=["accuracy"]) # Metrics to measure

history_1 = model_1.fit(train_data, epochs=5, validation_data=validation_data)

model_1 = tf.keras.Sequential([ Flatten(), Dense(10, activation="relu"), Dense(10, activation="relu"), Dense(4, activation="softmax") ]) model_1.compile(loss=tf.keras.losses.CategoricalCrossentropy(), # Loss function that we want optimizer=tf.keras.optimizers.Adam(), # Similar to Gradient Descent metrics=["accuracy"]) # Metrics to measure history_1 = model_1.fit(train_data, epochs=5, validation_data=validation_data)

Epoch 1/5
374/374 [==============================] - 243s 644ms/step - loss: 3.0598 - accuracy: 0.2516 - val_loss: 1.3864 - val_accuracy: 0.2485
Epoch 2/5
374/374 [==============================] - 237s 634ms/step - loss: 1.3864 - accuracy: 0.2489 - val_loss: 1.3863 - val_accuracy: 0.2515
Epoch 3/5
374/374 [==============================] - 239s 639ms/step - loss: 1.3864 - accuracy: 0.2515 - val_loss: 1.3863 - val_accuracy: 0.2515
Epoch 4/5
374/374 [==============================] - 240s 642ms/step - loss: 1.3863 - accuracy: 0.2495 - val_loss: 1.3863 - val_accuracy: 0.2515
Epoch 5/5
374/374 [==============================] - 240s 641ms/step - loss: 1.3864 - accuracy: 0.2506 - val_loss: 1.3863 - val_accuracy: 0.2515

In [ ]:

pd.DataFrame(history_1.history).to_csv("model_1_history")

pd.DataFrame(history_1.history).to_csv("model_1_history")

Plotting the loss metrics curve¶

In [ ]:

pd.DataFrame(history_1.history).plot()

pd.DataFrame(history_1.history).plot()

Out[ ]:

<Axes: >

Trial 02 - More layers, higher accuracy¶

In [ ]:

model2 = tf.keras.Sequential([
    Conv2D(10, 3, activation="relu"),
    Conv2D(10, 3, activation="relu"),
    MaxPool2D(2),
    Conv2D(10, 3, activation="relu"),
    Conv2D(10, 3, activation="relu"),
    MaxPool2D(2),
    Flatten(),
    Dense(10, activation="relu"),
    Dense(4, activation="softmax")
])

model2.compile(loss=tf.keras.losses.CategoricalCrossentropy(),
              optimizer=tf.keras.optimizers.Adam(),
              metrics=["accuracy"])

history2 = model2.fit(train_data,
                      epochs=5,
                      validation_data=validation_data)

model2 = tf.keras.Sequential([ Conv2D(10, 3, activation="relu"), Conv2D(10, 3, activation="relu"), MaxPool2D(2), Conv2D(10, 3, activation="relu"), Conv2D(10, 3, activation="relu"), MaxPool2D(2), Flatten(), Dense(10, activation="relu"), Dense(4, activation="softmax") ]) model2.compile(loss=tf.keras.losses.CategoricalCrossentropy(), optimizer=tf.keras.optimizers.Adam(), metrics=["accuracy"]) history2 = model2.fit(train_data, epochs=5, validation_data=validation_data)

Epoch 1/5
374/374 [==============================] - 279s 730ms/step - loss: 0.5839 - accuracy: 0.7314 - val_loss: 0.1579 - val_accuracy: 0.9458
Epoch 2/5
374/374 [==============================] - 245s 655ms/step - loss: 0.1455 - accuracy: 0.9516 - val_loss: 0.0669 - val_accuracy: 0.9816
Epoch 3/5
374/374 [==============================] - 247s 662ms/step - loss: 0.1000 - accuracy: 0.9713 - val_loss: 0.0464 - val_accuracy: 0.9853
Epoch 4/5
374/374 [==============================] - 281s 753ms/step - loss: 0.0322 - accuracy: 0.9901 - val_loss: 0.0136 - val_accuracy: 0.9960
Epoch 5/5
374/374 [==============================] - 247s 662ms/step - loss: 0.0319 - accuracy: 0.9913 - val_loss: 0.0180 - val_accuracy: 0.9960

Plotting the loss metrics curve¶

In [ ]:

pd.DataFrame(history2.history).plot()

pd.DataFrame(history2.history).plot()

Out[ ]:

<Axes: >

both losses and accuraccies should converge independently, if they diverge it means it's an overfit

• Underfit: Low accuraccy & High loss

Preprocesss image into numbers¶

In [ ]:

def preproc_img(path):
  img = tf.io.read_file(path) #decode_img reads only bin
  img = tf.io.decode_image(img) # Converts image to 3D matrix of numbers
  if img.shape[-1] == 1:
    img = tf.image.grayscale_to_rgb(img) # Convert to RGB cause file is gs but model trained on RGB
  img = tf.image.resize(img, [256, 256])/255. # Resize and rescale
  return img

def preproc_img(path): img = tf.io.read_file(path) #decode_img reads only bin img = tf.io.decode_image(img) # Converts image to 3D matrix of numbers if img.shape[-1] == 1: img = tf.image.grayscale_to_rgb(img) # Convert to RGB cause file is gs but model trained on RGB img = tf.image.resize(img, [256, 256])/255. # Resize and rescale return img

In [ ]:

import numpy as np

img = preproc_img("./test_data/1000.png")
print(img.shape)
model_output = model2.predict(tf.expand_dims(img, 0))
print(model_output)
np.argmax(model_output)

import numpy as np img = preproc_img("./test_data/1000.png") print(img.shape) model_output = model2.predict(tf.expand_dims(img, 0)) print(model_output) np.argmax(model_output)

(256, 256, 3)
1/1 [==============================] - 0s 19ms/step
[[9.9999928e-01 6.8465027e-07 1.3748462e-18 1.8264706e-23]]

Out[ ]:

0

0 as the output means that the output is predicted to belong to the 0th category, which in the above case is a circle