In [ ]:
Copied!
%load_ext autoreload
%autoreload 2
%load_ext autoreload
%autoreload 2
The autoreload extension is already loaded. To reload it, use: %reload_ext autoreload
Fine Tuning¶
Transfer Learning Part 2¶
Previous notebook was about using Pretrained models as feature extractors. We transfer the knowledge learned by these models on a similar task and use them directly for our downstream model. This was Transfer Learning - Feature Extraction.
- Used in Low data Volume scenario, or when the pretrained model task is very very similar to our task. E.g. Pretrained model task = "Classifying 1000 types of foods from images", Downstream task = "Classifying 10 fast food dishes".
In this notebook, we will use the pretrained models and fine-tune them (their weights) on our task. This is Transfer Learning - Fine Tuning
- Used in Medium to High volume scenario
- The initial layers of the pretrained models have learned the lower lever features which will still be important as is
- So instead we will adjust the top few layers by gradually unfreezing them
Some imports¶
In [ ]:
Copied!
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from src.image import ClassicImageDataDirectory, ImageDataset
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from src.image import ClassicImageDataDirectory, ImageDataset
Setup all data¶
In [ ]:
Copied!
DATA_DIR = {
'1_percent': '../data/10_food_classes_1_percent',
'10_percent': '../data/10_food_classes_10_percent',
'100_percent': '../data/10_food_classes_all_data',
}
IMAGE_DIM = (224, 224)
IMAGE_DATA_DIRECTORY = {data_pct: ClassicImageDataDirectory(data_dir, IMAGE_DIM, np.uint8) for data_pct, data_dir in DATA_DIR.items()}
DATA_DIR = {
'1_percent': '../data/10_food_classes_1_percent',
'10_percent': '../data/10_food_classes_10_percent',
'100_percent': '../data/10_food_classes_all_data',
}
IMAGE_DIM = (224, 224)
IMAGE_DATA_DIRECTORY = {data_pct: ClassicImageDataDirectory(data_dir, IMAGE_DIM, np.uint8) for data_pct, data_dir in DATA_DIR.items()}
10 Food Classes - Working with less data (10 percent)¶
Load the data¶
In [ ]:
Copied!
data_pct = '10_percent'
data_dir = DATA_DIR[data_pct]
imgdir = IMAGE_DATA_DIRECTORY[data_pct]
imgdir
data_pct = '10_percent'
data_dir = DATA_DIR[data_pct]
imgdir = IMAGE_DATA_DIRECTORY[data_pct]
imgdir
Out[ ]:
<src.image.datadirectory.ClassicImageDataDirectory at 0x15461736048>
Explore the data¶
In [ ]:
Copied!
imgdir.class_names
imgdir.class_names
Out[ ]:
('chicken_curry',
'chicken_wings',
'fried_rice',
'grilled_salmon',
'hamburger',
'ice_cream',
'pizza',
'ramen',
'steak',
'sushi')
In [ ]:
Copied!
imgdir.labelcountdf
imgdir.labelcountdf
Out[ ]:
| label | name | count_train | count_test | |
|---|---|---|---|---|
| 0 | 0 | chicken_curry | 75 | 250 |
| 1 | 1 | chicken_wings | 75 | 250 |
| 2 | 2 | fried_rice | 75 | 250 |
| 3 | 3 | grilled_salmon | 75 | 250 |
| 4 | 4 | hamburger | 75 | 250 |
| 5 | 5 | ice_cream | 75 | 250 |
| 6 | 6 | pizza | 75 | 250 |
| 7 | 7 | ramen | 75 | 250 |
| 8 | 8 | steak | 75 | 250 |
| 9 | 9 | sushi | 75 | 250 |
In [ ]:
Copied!
imgdir.plot_labelcounts();
imgdir.plot_labelcounts();
View Example Images¶
In [ ]:
Copied!
batch = imgdir.load(32)
imgen = next(batch)
imgen.view_random_images(class_names='all');
batch = imgdir.load(32)
imgen = next(batch)
imgen.view_random_images(class_names='all');
Imports¶
In [ ]:
Copied!
import tensorflow as tf
from tensorflow.keras import layers, losses, optimizers
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from src.evaluate import KerasMetrics
from src.visualize import plot_confusion_matrix, plot_keras_model, plot_learning_curve
from sklearn import metrics as skmetrics
import tensorflow as tf
from tensorflow.keras import layers, losses, optimizers
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from src.evaluate import KerasMetrics
from src.visualize import plot_confusion_matrix, plot_keras_model, plot_learning_curve
from sklearn import metrics as skmetrics
EfficientNetB0 - Creating a model using the Keras Functional API¶
In [ ]:
Copied!
from tensorflow.keras.applications import EfficientNetB0, ResNet50V2
from tensorflow.keras.applications import EfficientNetB0, ResNet50V2
Set up initial params¶
In [ ]:
Copied!
DATA_PCT = '10_percent'
CLASSIFICATION_TYPE = 'categorical'
imgdir = IMAGE_DATA_DIRECTORY[DATA_PCT]
IMAGE_DIM = (224, 224)
NUM_CHANNELS = 3
INPUT_SHAPE = (*IMAGE_DIM, NUM_CHANNELS)
N_CLASSES = imgdir.n_classes
BATCH_SIZE = 32
VALIDATION_SPLIT = 0.2
SEED = 42
PRETRAINED_MODEL = 'EfficientNetB0'
DATA_PCT = '10_percent'
CLASSIFICATION_TYPE = 'categorical'
imgdir = IMAGE_DATA_DIRECTORY[DATA_PCT]
IMAGE_DIM = (224, 224)
NUM_CHANNELS = 3
INPUT_SHAPE = (*IMAGE_DIM, NUM_CHANNELS)
N_CLASSES = imgdir.n_classes
BATCH_SIZE = 32
VALIDATION_SPLIT = 0.2
SEED = 42
PRETRAINED_MODEL = 'EfficientNetB0'
Preprocess the data¶
In [ ]:
Copied!
train_datagen = ImageDataGenerator(validation_split=VALIDATION_SPLIT)
test_datagen = ImageDataGenerator()
train_datagen = ImageDataGenerator(validation_split=VALIDATION_SPLIT)
test_datagen = ImageDataGenerator()
Setup the train and test directories¶
In [ ]:
Copied!
train_dir = imgdir.train['dir']
test_dir = imgdir.test['dir']
train_dir = imgdir.train['dir']
test_dir = imgdir.test['dir']
Import the data from directories¶
In [ ]:
Copied!
train_data = train_datagen.flow_from_directory(directory=train_dir, target_size=IMAGE_DIM, class_mode=CLASSIFICATION_TYPE,
batch_size=BATCH_SIZE, subset='training', seed=SEED)
validation_data = train_datagen.flow_from_directory(directory=train_dir, target_size=IMAGE_DIM, class_mode=CLASSIFICATION_TYPE,
batch_size=BATCH_SIZE, subset='validation', seed=SEED)
test_data = test_datagen.flow_from_directory(directory=test_dir, target_size=IMAGE_DIM, class_mode=CLASSIFICATION_TYPE,
batch_size=BATCH_SIZE, shuffle=False)
train_data = train_datagen.flow_from_directory(directory=train_dir, target_size=IMAGE_DIM, class_mode=CLASSIFICATION_TYPE,
batch_size=BATCH_SIZE, subset='training', seed=SEED)
validation_data = train_datagen.flow_from_directory(directory=train_dir, target_size=IMAGE_DIM, class_mode=CLASSIFICATION_TYPE,
batch_size=BATCH_SIZE, subset='validation', seed=SEED)
test_data = test_datagen.flow_from_directory(directory=test_dir, target_size=IMAGE_DIM, class_mode=CLASSIFICATION_TYPE,
batch_size=BATCH_SIZE, shuffle=False)
Found 600 images belonging to 10 classes. Found 150 images belonging to 10 classes. Found 2500 images belonging to 10 classes.
Create the model (using Keras Functional API)¶
In [ ]:
Copied!
# Create the pretrained model (This will download the model)
pretrained_model = EfficientNetB0(include_top=False)
# Make it non trainable
pretrained_model.trainable = False
# input layer
inputs = layers.Input(shape=INPUT_SHAPE, name='input_layer')
print(f'Input shape: {inputs.shape}')
# Scale the input (if model = ResNet50V2) The EfficientNetB0 model in keras.application doesn't require scaling
# inputs = tf.keras.layers.experimental.preprocessing.Rescaling(1/255.)(inputs)
# Extract features from pretrained model
features = pretrained_model(inputs)
print(f'Shape of features extracted from pretrained model: {features.shape}')
# Add a GlobalAveragePool2D layer
pooled = layers.GlobalAvgPool2D(name='global_avg_pooling_layer')(features)
print(f'Shape after global pooling features: {pooled.shape}')
# Now add the final output layer
outputs = layers.Dense(N_CLASSES, activation='softmax', name='output_layer')(pooled)
# Now combine the inputs and outputs into a model
model = tf.keras.Model(inputs=inputs, outputs=outputs)
# Compile the model
model.compile(loss=losses.categorical_crossentropy, optimizer=optimizers.Adam(), metrics=[KerasMetrics.f1, 'accuracy'])
# Summary
model.summary()
# Create the pretrained model (This will download the model)
pretrained_model = EfficientNetB0(include_top=False)
# Make it non trainable
pretrained_model.trainable = False
# input layer
inputs = layers.Input(shape=INPUT_SHAPE, name='input_layer')
print(f'Input shape: {inputs.shape}')
# Scale the input (if model = ResNet50V2) The EfficientNetB0 model in keras.application doesn't require scaling
# inputs = tf.keras.layers.experimental.preprocessing.Rescaling(1/255.)(inputs)
# Extract features from pretrained model
features = pretrained_model(inputs)
print(f'Shape of features extracted from pretrained model: {features.shape}')
# Add a GlobalAveragePool2D layer
pooled = layers.GlobalAvgPool2D(name='global_avg_pooling_layer')(features)
print(f'Shape after global pooling features: {pooled.shape}')
# Now add the final output layer
outputs = layers.Dense(N_CLASSES, activation='softmax', name='output_layer')(pooled)
# Now combine the inputs and outputs into a model
model = tf.keras.Model(inputs=inputs, outputs=outputs)
# Compile the model
model.compile(loss=losses.categorical_crossentropy, optimizer=optimizers.Adam(), metrics=[KerasMetrics.f1, 'accuracy'])
# Summary
model.summary()
Input shape: (None, 224, 224, 3) Shape of features extracted from pretrained model: (None, 7, 7, 1280) Shape after global pooling features: (None, 1280) Model: "model_1" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= input_layer (InputLayer) [(None, 224, 224, 3)] 0 _________________________________________________________________ efficientnetb0 (Functional) (None, None, None, 1280) 4049571 _________________________________________________________________ global_avg_pooling_layer (Gl (None, 1280) 0 _________________________________________________________________ output_layer (Dense) (None, 10) 12810 ================================================================= Total params: 4,062,381 Trainable params: 12,810 Non-trainable params: 4,049,571 _________________________________________________________________
The output shape of EfficientNetB0 is (None, None, None, 1280), which implies that the output shape is dependent on the input shape. However there maybe a recommended input shape described by the authors of this model.
Fit the model¶
In [ ]:
Copied!
history = model.fit(train_data, steps_per_epoch=len(train_data), validation_data=validation_data, validation_steps=len(validation_data), epochs=10)
history = model.fit(train_data, steps_per_epoch=len(train_data), validation_data=validation_data, validation_steps=len(validation_data), epochs=10)
Epoch 1/10 19/19 [==============================] - 14s 512ms/step - loss: 2.2085 - f1: 0.0000e+00 - accuracy: 0.2123 - val_loss: 1.6883 - val_f1: 0.0121 - val_accuracy: 0.5867 Epoch 2/10 19/19 [==============================] - 6s 337ms/step - loss: 1.3832 - f1: 0.1118 - accuracy: 0.7037 - val_loss: 1.2918 - val_f1: 0.3371 - val_accuracy: 0.6933 Epoch 3/10 19/19 [==============================] - 8s 437ms/step - loss: 0.9837 - f1: 0.5548 - accuracy: 0.7723 - val_loss: 1.1034 - val_f1: 0.5064 - val_accuracy: 0.7333 Epoch 4/10 19/19 [==============================] - 7s 352ms/step - loss: 0.7523 - f1: 0.7338 - accuracy: 0.8296 - val_loss: 1.0094 - val_f1: 0.6135 - val_accuracy: 0.7533 Epoch 5/10 19/19 [==============================] - 8s 425ms/step - loss: 0.6182 - f1: 0.7978 - accuracy: 0.8699 - val_loss: 0.9417 - val_f1: 0.6539 - val_accuracy: 0.7533 Epoch 6/10 19/19 [==============================] - 7s 358ms/step - loss: 0.5287 - f1: 0.8374 - accuracy: 0.8984 - val_loss: 0.9044 - val_f1: 0.6878 - val_accuracy: 0.7467 Epoch 7/10 19/19 [==============================] - 7s 354ms/step - loss: 0.4788 - f1: 0.8490 - accuracy: 0.9104 - val_loss: 0.8780 - val_f1: 0.6889 - val_accuracy: 0.7467 Epoch 8/10 19/19 [==============================] - 7s 366ms/step - loss: 0.4796 - f1: 0.8540 - accuracy: 0.9002 - val_loss: 0.8525 - val_f1: 0.7075 - val_accuracy: 0.7600 Epoch 9/10 19/19 [==============================] - 7s 358ms/step - loss: 0.3983 - f1: 0.8986 - accuracy: 0.9309 - val_loss: 0.8310 - val_f1: 0.7295 - val_accuracy: 0.7733 Epoch 10/10 19/19 [==============================] - 7s 350ms/step - loss: 0.3710 - f1: 0.8997 - accuracy: 0.9159 - val_loss: 0.8190 - val_f1: 0.7399 - val_accuracy: 0.7667
Learning Curve¶
In [ ]:
Copied!
plot_learning_curve(model, extra_metric='f1');
plot_learning_curve(model, extra_metric='f1');
Maybe a bit overfitted, but we got so much better performance with just 10% of the data!
Prediction Evaluation¶
In [ ]:
Copied!
y_test_pred_probs = model.predict(test_data)
y_test_preds = y_test_pred_probs.argmax(axis=1)
y_test_pred_probs = model.predict(test_data)
y_test_preds = y_test_pred_probs.argmax(axis=1)
Classifcation Report¶
In [ ]:
Copied!
print(skmetrics.classification_report(test_data.labels, y_test_preds, target_names=test_data.class_indices))
print(skmetrics.classification_report(test_data.labels, y_test_preds, target_names=test_data.class_indices))
precision recall f1-score support
chicken_curry 0.81 0.70 0.75 250
chicken_wings 0.89 0.81 0.85 250
fried_rice 0.83 0.92 0.87 250
grilled_salmon 0.71 0.74 0.73 250
hamburger 0.87 0.92 0.89 250
ice_cream 0.94 0.82 0.88 250
pizza 0.95 0.96 0.96 250
ramen 0.88 0.89 0.88 250
steak 0.68 0.80 0.74 250
sushi 0.87 0.84 0.85 250
accuracy 0.84 2500
macro avg 0.84 0.84 0.84 2500
weighted avg 0.84 0.84 0.84 2500
Confusion Matrix¶
In [ ]:
Copied!
plot_confusion_matrix(test_data.labels, y_test_preds, classes=test_data.class_indices, figsize=(14, 14), text_size=8);
plot_confusion_matrix(test_data.labels, y_test_preds, classes=test_data.class_indices, figsize=(14, 14), text_size=8);
Layers of the Pretrained Model¶
In [ ]:
Copied!
for layer_number, layer in enumerate(pretrained_model.layers):
print(layer_number, layer.name)
for layer_number, layer in enumerate(pretrained_model.layers):
print(layer_number, layer.name)
0 input_2 1 rescaling_1 2 normalization_1 3 stem_conv_pad 4 stem_conv 5 stem_bn 6 stem_activation 7 block1a_dwconv 8 block1a_bn 9 block1a_activation 10 block1a_se_squeeze 11 block1a_se_reshape 12 block1a_se_reduce 13 block1a_se_expand 14 block1a_se_excite 15 block1a_project_conv 16 block1a_project_bn 17 block2a_expand_conv 18 block2a_expand_bn 19 block2a_expand_activation 20 block2a_dwconv_pad 21 block2a_dwconv 22 block2a_bn 23 block2a_activation 24 block2a_se_squeeze 25 block2a_se_reshape 26 block2a_se_reduce 27 block2a_se_expand 28 block2a_se_excite 29 block2a_project_conv 30 block2a_project_bn 31 block2b_expand_conv 32 block2b_expand_bn 33 block2b_expand_activation 34 block2b_dwconv 35 block2b_bn 36 block2b_activation 37 block2b_se_squeeze 38 block2b_se_reshape 39 block2b_se_reduce 40 block2b_se_expand 41 block2b_se_excite 42 block2b_project_conv 43 block2b_project_bn 44 block2b_drop 45 block2b_add 46 block3a_expand_conv 47 block3a_expand_bn 48 block3a_expand_activation 49 block3a_dwconv_pad 50 block3a_dwconv 51 block3a_bn 52 block3a_activation 53 block3a_se_squeeze 54 block3a_se_reshape 55 block3a_se_reduce 56 block3a_se_expand 57 block3a_se_excite 58 block3a_project_conv 59 block3a_project_bn 60 block3b_expand_conv 61 block3b_expand_bn 62 block3b_expand_activation 63 block3b_dwconv 64 block3b_bn 65 block3b_activation 66 block3b_se_squeeze 67 block3b_se_reshape 68 block3b_se_reduce 69 block3b_se_expand 70 block3b_se_excite 71 block3b_project_conv 72 block3b_project_bn 73 block3b_drop 74 block3b_add 75 block4a_expand_conv 76 block4a_expand_bn 77 block4a_expand_activation 78 block4a_dwconv_pad 79 block4a_dwconv 80 block4a_bn 81 block4a_activation 82 block4a_se_squeeze 83 block4a_se_reshape 84 block4a_se_reduce 85 block4a_se_expand 86 block4a_se_excite 87 block4a_project_conv 88 block4a_project_bn 89 block4b_expand_conv 90 block4b_expand_bn 91 block4b_expand_activation 92 block4b_dwconv 93 block4b_bn 94 block4b_activation 95 block4b_se_squeeze 96 block4b_se_reshape 97 block4b_se_reduce 98 block4b_se_expand 99 block4b_se_excite 100 block4b_project_conv 101 block4b_project_bn 102 block4b_drop 103 block4b_add 104 block4c_expand_conv 105 block4c_expand_bn 106 block4c_expand_activation 107 block4c_dwconv 108 block4c_bn 109 block4c_activation 110 block4c_se_squeeze 111 block4c_se_reshape 112 block4c_se_reduce 113 block4c_se_expand 114 block4c_se_excite 115 block4c_project_conv 116 block4c_project_bn 117 block4c_drop 118 block4c_add 119 block5a_expand_conv 120 block5a_expand_bn 121 block5a_expand_activation 122 block5a_dwconv 123 block5a_bn 124 block5a_activation 125 block5a_se_squeeze 126 block5a_se_reshape 127 block5a_se_reduce 128 block5a_se_expand 129 block5a_se_excite 130 block5a_project_conv 131 block5a_project_bn 132 block5b_expand_conv 133 block5b_expand_bn 134 block5b_expand_activation 135 block5b_dwconv 136 block5b_bn 137 block5b_activation 138 block5b_se_squeeze 139 block5b_se_reshape 140 block5b_se_reduce 141 block5b_se_expand 142 block5b_se_excite 143 block5b_project_conv 144 block5b_project_bn 145 block5b_drop 146 block5b_add 147 block5c_expand_conv 148 block5c_expand_bn 149 block5c_expand_activation 150 block5c_dwconv 151 block5c_bn 152 block5c_activation 153 block5c_se_squeeze 154 block5c_se_reshape 155 block5c_se_reduce 156 block5c_se_expand 157 block5c_se_excite 158 block5c_project_conv 159 block5c_project_bn 160 block5c_drop 161 block5c_add 162 block6a_expand_conv 163 block6a_expand_bn 164 block6a_expand_activation 165 block6a_dwconv_pad 166 block6a_dwconv 167 block6a_bn 168 block6a_activation 169 block6a_se_squeeze 170 block6a_se_reshape 171 block6a_se_reduce 172 block6a_se_expand 173 block6a_se_excite 174 block6a_project_conv 175 block6a_project_bn 176 block6b_expand_conv 177 block6b_expand_bn 178 block6b_expand_activation 179 block6b_dwconv 180 block6b_bn 181 block6b_activation 182 block6b_se_squeeze 183 block6b_se_reshape 184 block6b_se_reduce 185 block6b_se_expand 186 block6b_se_excite 187 block6b_project_conv 188 block6b_project_bn 189 block6b_drop 190 block6b_add 191 block6c_expand_conv 192 block6c_expand_bn 193 block6c_expand_activation 194 block6c_dwconv 195 block6c_bn 196 block6c_activation 197 block6c_se_squeeze 198 block6c_se_reshape 199 block6c_se_reduce 200 block6c_se_expand 201 block6c_se_excite 202 block6c_project_conv 203 block6c_project_bn 204 block6c_drop 205 block6c_add 206 block6d_expand_conv 207 block6d_expand_bn 208 block6d_expand_activation 209 block6d_dwconv 210 block6d_bn 211 block6d_activation 212 block6d_se_squeeze 213 block6d_se_reshape 214 block6d_se_reduce 215 block6d_se_expand 216 block6d_se_excite 217 block6d_project_conv 218 block6d_project_bn 219 block6d_drop 220 block6d_add 221 block7a_expand_conv 222 block7a_expand_bn 223 block7a_expand_activation 224 block7a_dwconv 225 block7a_bn 226 block7a_activation 227 block7a_se_squeeze 228 block7a_se_reshape 229 block7a_se_reduce 230 block7a_se_expand 231 block7a_se_excite 232 block7a_project_conv 233 block7a_project_bn 234 top_conv 235 top_bn 236 top_activation
In [ ]:
Copied!
pretrained_model.summary()
pretrained_model.summary()
Model: "efficientnetb0"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_2 (InputLayer) [(None, None, None, 0
__________________________________________________________________________________________________
rescaling_1 (Rescaling) (None, None, None, 3 0 input_2[0][0]
__________________________________________________________________________________________________
normalization_1 (Normalization) (None, None, None, 3 7 rescaling_1[0][0]
__________________________________________________________________________________________________
stem_conv_pad (ZeroPadding2D) (None, None, None, 3 0 normalization_1[0][0]
__________________________________________________________________________________________________
stem_conv (Conv2D) (None, None, None, 3 864 stem_conv_pad[0][0]
__________________________________________________________________________________________________
stem_bn (BatchNormalization) (None, None, None, 3 128 stem_conv[0][0]
__________________________________________________________________________________________________
stem_activation (Activation) (None, None, None, 3 0 stem_bn[0][0]
__________________________________________________________________________________________________
block1a_dwconv (DepthwiseConv2D (None, None, None, 3 288 stem_activation[0][0]
__________________________________________________________________________________________________
block1a_bn (BatchNormalization) (None, None, None, 3 128 block1a_dwconv[0][0]
__________________________________________________________________________________________________
block1a_activation (Activation) (None, None, None, 3 0 block1a_bn[0][0]
__________________________________________________________________________________________________
block1a_se_squeeze (GlobalAvera (None, 32) 0 block1a_activation[0][0]
__________________________________________________________________________________________________
block1a_se_reshape (Reshape) (None, 1, 1, 32) 0 block1a_se_squeeze[0][0]
__________________________________________________________________________________________________
block1a_se_reduce (Conv2D) (None, 1, 1, 8) 264 block1a_se_reshape[0][0]
__________________________________________________________________________________________________
block1a_se_expand (Conv2D) (None, 1, 1, 32) 288 block1a_se_reduce[0][0]
__________________________________________________________________________________________________
block1a_se_excite (Multiply) (None, None, None, 3 0 block1a_activation[0][0]
block1a_se_expand[0][0]
__________________________________________________________________________________________________
block1a_project_conv (Conv2D) (None, None, None, 1 512 block1a_se_excite[0][0]
__________________________________________________________________________________________________
block1a_project_bn (BatchNormal (None, None, None, 1 64 block1a_project_conv[0][0]
__________________________________________________________________________________________________
block2a_expand_conv (Conv2D) (None, None, None, 9 1536 block1a_project_bn[0][0]
__________________________________________________________________________________________________
block2a_expand_bn (BatchNormali (None, None, None, 9 384 block2a_expand_conv[0][0]
__________________________________________________________________________________________________
block2a_expand_activation (Acti (None, None, None, 9 0 block2a_expand_bn[0][0]
__________________________________________________________________________________________________
block2a_dwconv_pad (ZeroPadding (None, None, None, 9 0 block2a_expand_activation[0][0]
__________________________________________________________________________________________________
block2a_dwconv (DepthwiseConv2D (None, None, None, 9 864 block2a_dwconv_pad[0][0]
__________________________________________________________________________________________________
block2a_bn (BatchNormalization) (None, None, None, 9 384 block2a_dwconv[0][0]
__________________________________________________________________________________________________
block2a_activation (Activation) (None, None, None, 9 0 block2a_bn[0][0]
__________________________________________________________________________________________________
block2a_se_squeeze (GlobalAvera (None, 96) 0 block2a_activation[0][0]
__________________________________________________________________________________________________
block2a_se_reshape (Reshape) (None, 1, 1, 96) 0 block2a_se_squeeze[0][0]
__________________________________________________________________________________________________
block2a_se_reduce (Conv2D) (None, 1, 1, 4) 388 block2a_se_reshape[0][0]
__________________________________________________________________________________________________
block2a_se_expand (Conv2D) (None, 1, 1, 96) 480 block2a_se_reduce[0][0]
__________________________________________________________________________________________________
block2a_se_excite (Multiply) (None, None, None, 9 0 block2a_activation[0][0]
block2a_se_expand[0][0]
__________________________________________________________________________________________________
block2a_project_conv (Conv2D) (None, None, None, 2 2304 block2a_se_excite[0][0]
__________________________________________________________________________________________________
block2a_project_bn (BatchNormal (None, None, None, 2 96 block2a_project_conv[0][0]
__________________________________________________________________________________________________
block2b_expand_conv (Conv2D) (None, None, None, 1 3456 block2a_project_bn[0][0]
__________________________________________________________________________________________________
block2b_expand_bn (BatchNormali (None, None, None, 1 576 block2b_expand_conv[0][0]
__________________________________________________________________________________________________
block2b_expand_activation (Acti (None, None, None, 1 0 block2b_expand_bn[0][0]
__________________________________________________________________________________________________
block2b_dwconv (DepthwiseConv2D (None, None, None, 1 1296 block2b_expand_activation[0][0]
__________________________________________________________________________________________________
block2b_bn (BatchNormalization) (None, None, None, 1 576 block2b_dwconv[0][0]
__________________________________________________________________________________________________
block2b_activation (Activation) (None, None, None, 1 0 block2b_bn[0][0]
__________________________________________________________________________________________________
block2b_se_squeeze (GlobalAvera (None, 144) 0 block2b_activation[0][0]
__________________________________________________________________________________________________
block2b_se_reshape (Reshape) (None, 1, 1, 144) 0 block2b_se_squeeze[0][0]
__________________________________________________________________________________________________
block2b_se_reduce (Conv2D) (None, 1, 1, 6) 870 block2b_se_reshape[0][0]
__________________________________________________________________________________________________
block2b_se_expand (Conv2D) (None, 1, 1, 144) 1008 block2b_se_reduce[0][0]
__________________________________________________________________________________________________
block2b_se_excite (Multiply) (None, None, None, 1 0 block2b_activation[0][0]
block2b_se_expand[0][0]
__________________________________________________________________________________________________
block2b_project_conv (Conv2D) (None, None, None, 2 3456 block2b_se_excite[0][0]
__________________________________________________________________________________________________
block2b_project_bn (BatchNormal (None, None, None, 2 96 block2b_project_conv[0][0]
__________________________________________________________________________________________________
block2b_drop (Dropout) (None, None, None, 2 0 block2b_project_bn[0][0]
__________________________________________________________________________________________________
block2b_add (Add) (None, None, None, 2 0 block2b_drop[0][0]
block2a_project_bn[0][0]
__________________________________________________________________________________________________
block3a_expand_conv (Conv2D) (None, None, None, 1 3456 block2b_add[0][0]
__________________________________________________________________________________________________
block3a_expand_bn (BatchNormali (None, None, None, 1 576 block3a_expand_conv[0][0]
__________________________________________________________________________________________________
block3a_expand_activation (Acti (None, None, None, 1 0 block3a_expand_bn[0][0]
__________________________________________________________________________________________________
block3a_dwconv_pad (ZeroPadding (None, None, None, 1 0 block3a_expand_activation[0][0]
__________________________________________________________________________________________________
block3a_dwconv (DepthwiseConv2D (None, None, None, 1 3600 block3a_dwconv_pad[0][0]
__________________________________________________________________________________________________
block3a_bn (BatchNormalization) (None, None, None, 1 576 block3a_dwconv[0][0]
__________________________________________________________________________________________________
block3a_activation (Activation) (None, None, None, 1 0 block3a_bn[0][0]
__________________________________________________________________________________________________
block3a_se_squeeze (GlobalAvera (None, 144) 0 block3a_activation[0][0]
__________________________________________________________________________________________________
block3a_se_reshape (Reshape) (None, 1, 1, 144) 0 block3a_se_squeeze[0][0]
__________________________________________________________________________________________________
block3a_se_reduce (Conv2D) (None, 1, 1, 6) 870 block3a_se_reshape[0][0]
__________________________________________________________________________________________________
block3a_se_expand (Conv2D) (None, 1, 1, 144) 1008 block3a_se_reduce[0][0]
__________________________________________________________________________________________________
block3a_se_excite (Multiply) (None, None, None, 1 0 block3a_activation[0][0]
block3a_se_expand[0][0]
__________________________________________________________________________________________________
block3a_project_conv (Conv2D) (None, None, None, 4 5760 block3a_se_excite[0][0]
__________________________________________________________________________________________________
block3a_project_bn (BatchNormal (None, None, None, 4 160 block3a_project_conv[0][0]
__________________________________________________________________________________________________
block3b_expand_conv (Conv2D) (None, None, None, 2 9600 block3a_project_bn[0][0]
__________________________________________________________________________________________________
block3b_expand_bn (BatchNormali (None, None, None, 2 960 block3b_expand_conv[0][0]
__________________________________________________________________________________________________
block3b_expand_activation (Acti (None, None, None, 2 0 block3b_expand_bn[0][0]
__________________________________________________________________________________________________
block3b_dwconv (DepthwiseConv2D (None, None, None, 2 6000 block3b_expand_activation[0][0]
__________________________________________________________________________________________________
block3b_bn (BatchNormalization) (None, None, None, 2 960 block3b_dwconv[0][0]
__________________________________________________________________________________________________
block3b_activation (Activation) (None, None, None, 2 0 block3b_bn[0][0]
__________________________________________________________________________________________________
block3b_se_squeeze (GlobalAvera (None, 240) 0 block3b_activation[0][0]
__________________________________________________________________________________________________
block3b_se_reshape (Reshape) (None, 1, 1, 240) 0 block3b_se_squeeze[0][0]
__________________________________________________________________________________________________
block3b_se_reduce (Conv2D) (None, 1, 1, 10) 2410 block3b_se_reshape[0][0]
__________________________________________________________________________________________________
block3b_se_expand (Conv2D) (None, 1, 1, 240) 2640 block3b_se_reduce[0][0]
__________________________________________________________________________________________________
block3b_se_excite (Multiply) (None, None, None, 2 0 block3b_activation[0][0]
block3b_se_expand[0][0]
__________________________________________________________________________________________________
block3b_project_conv (Conv2D) (None, None, None, 4 9600 block3b_se_excite[0][0]
__________________________________________________________________________________________________
block3b_project_bn (BatchNormal (None, None, None, 4 160 block3b_project_conv[0][0]
__________________________________________________________________________________________________
block3b_drop (Dropout) (None, None, None, 4 0 block3b_project_bn[0][0]
__________________________________________________________________________________________________
block3b_add (Add) (None, None, None, 4 0 block3b_drop[0][0]
block3a_project_bn[0][0]
__________________________________________________________________________________________________
block4a_expand_conv (Conv2D) (None, None, None, 2 9600 block3b_add[0][0]
__________________________________________________________________________________________________
block4a_expand_bn (BatchNormali (None, None, None, 2 960 block4a_expand_conv[0][0]
__________________________________________________________________________________________________
block4a_expand_activation (Acti (None, None, None, 2 0 block4a_expand_bn[0][0]
__________________________________________________________________________________________________
block4a_dwconv_pad (ZeroPadding (None, None, None, 2 0 block4a_expand_activation[0][0]
__________________________________________________________________________________________________
block4a_dwconv (DepthwiseConv2D (None, None, None, 2 2160 block4a_dwconv_pad[0][0]
__________________________________________________________________________________________________
block4a_bn (BatchNormalization) (None, None, None, 2 960 block4a_dwconv[0][0]
__________________________________________________________________________________________________
block4a_activation (Activation) (None, None, None, 2 0 block4a_bn[0][0]
__________________________________________________________________________________________________
block4a_se_squeeze (GlobalAvera (None, 240) 0 block4a_activation[0][0]
__________________________________________________________________________________________________
block4a_se_reshape (Reshape) (None, 1, 1, 240) 0 block4a_se_squeeze[0][0]
__________________________________________________________________________________________________
block4a_se_reduce (Conv2D) (None, 1, 1, 10) 2410 block4a_se_reshape[0][0]
__________________________________________________________________________________________________
block4a_se_expand (Conv2D) (None, 1, 1, 240) 2640 block4a_se_reduce[0][0]
__________________________________________________________________________________________________
block4a_se_excite (Multiply) (None, None, None, 2 0 block4a_activation[0][0]
block4a_se_expand[0][0]
__________________________________________________________________________________________________
block4a_project_conv (Conv2D) (None, None, None, 8 19200 block4a_se_excite[0][0]
__________________________________________________________________________________________________
block4a_project_bn (BatchNormal (None, None, None, 8 320 block4a_project_conv[0][0]
__________________________________________________________________________________________________
block4b_expand_conv (Conv2D) (None, None, None, 4 38400 block4a_project_bn[0][0]
__________________________________________________________________________________________________
block4b_expand_bn (BatchNormali (None, None, None, 4 1920 block4b_expand_conv[0][0]
__________________________________________________________________________________________________
block4b_expand_activation (Acti (None, None, None, 4 0 block4b_expand_bn[0][0]
__________________________________________________________________________________________________
block4b_dwconv (DepthwiseConv2D (None, None, None, 4 4320 block4b_expand_activation[0][0]
__________________________________________________________________________________________________
block4b_bn (BatchNormalization) (None, None, None, 4 1920 block4b_dwconv[0][0]
__________________________________________________________________________________________________
block4b_activation (Activation) (None, None, None, 4 0 block4b_bn[0][0]
__________________________________________________________________________________________________
block4b_se_squeeze (GlobalAvera (None, 480) 0 block4b_activation[0][0]
__________________________________________________________________________________________________
block4b_se_reshape (Reshape) (None, 1, 1, 480) 0 block4b_se_squeeze[0][0]
__________________________________________________________________________________________________
block4b_se_reduce (Conv2D) (None, 1, 1, 20) 9620 block4b_se_reshape[0][0]
__________________________________________________________________________________________________
block4b_se_expand (Conv2D) (None, 1, 1, 480) 10080 block4b_se_reduce[0][0]
__________________________________________________________________________________________________
block4b_se_excite (Multiply) (None, None, None, 4 0 block4b_activation[0][0]
block4b_se_expand[0][0]
__________________________________________________________________________________________________
block4b_project_conv (Conv2D) (None, None, None, 8 38400 block4b_se_excite[0][0]
__________________________________________________________________________________________________
block4b_project_bn (BatchNormal (None, None, None, 8 320 block4b_project_conv[0][0]
__________________________________________________________________________________________________
block4b_drop (Dropout) (None, None, None, 8 0 block4b_project_bn[0][0]
__________________________________________________________________________________________________
block4b_add (Add) (None, None, None, 8 0 block4b_drop[0][0]
block4a_project_bn[0][0]
__________________________________________________________________________________________________
block4c_expand_conv (Conv2D) (None, None, None, 4 38400 block4b_add[0][0]
__________________________________________________________________________________________________
block4c_expand_bn (BatchNormali (None, None, None, 4 1920 block4c_expand_conv[0][0]
__________________________________________________________________________________________________
block4c_expand_activation (Acti (None, None, None, 4 0 block4c_expand_bn[0][0]
__________________________________________________________________________________________________
block4c_dwconv (DepthwiseConv2D (None, None, None, 4 4320 block4c_expand_activation[0][0]
__________________________________________________________________________________________________
block4c_bn (BatchNormalization) (None, None, None, 4 1920 block4c_dwconv[0][0]
__________________________________________________________________________________________________
block4c_activation (Activation) (None, None, None, 4 0 block4c_bn[0][0]
__________________________________________________________________________________________________
block4c_se_squeeze (GlobalAvera (None, 480) 0 block4c_activation[0][0]
__________________________________________________________________________________________________
block4c_se_reshape (Reshape) (None, 1, 1, 480) 0 block4c_se_squeeze[0][0]
__________________________________________________________________________________________________
block4c_se_reduce (Conv2D) (None, 1, 1, 20) 9620 block4c_se_reshape[0][0]
__________________________________________________________________________________________________
block4c_se_expand (Conv2D) (None, 1, 1, 480) 10080 block4c_se_reduce[0][0]
__________________________________________________________________________________________________
block4c_se_excite (Multiply) (None, None, None, 4 0 block4c_activation[0][0]
block4c_se_expand[0][0]
__________________________________________________________________________________________________
block4c_project_conv (Conv2D) (None, None, None, 8 38400 block4c_se_excite[0][0]
__________________________________________________________________________________________________
block4c_project_bn (BatchNormal (None, None, None, 8 320 block4c_project_conv[0][0]
__________________________________________________________________________________________________
block4c_drop (Dropout) (None, None, None, 8 0 block4c_project_bn[0][0]
__________________________________________________________________________________________________
block4c_add (Add) (None, None, None, 8 0 block4c_drop[0][0]
block4b_add[0][0]
__________________________________________________________________________________________________
block5a_expand_conv (Conv2D) (None, None, None, 4 38400 block4c_add[0][0]
__________________________________________________________________________________________________
block5a_expand_bn (BatchNormali (None, None, None, 4 1920 block5a_expand_conv[0][0]
__________________________________________________________________________________________________
block5a_expand_activation (Acti (None, None, None, 4 0 block5a_expand_bn[0][0]
__________________________________________________________________________________________________
block5a_dwconv (DepthwiseConv2D (None, None, None, 4 12000 block5a_expand_activation[0][0]
__________________________________________________________________________________________________
block5a_bn (BatchNormalization) (None, None, None, 4 1920 block5a_dwconv[0][0]
__________________________________________________________________________________________________
block5a_activation (Activation) (None, None, None, 4 0 block5a_bn[0][0]
__________________________________________________________________________________________________
block5a_se_squeeze (GlobalAvera (None, 480) 0 block5a_activation[0][0]
__________________________________________________________________________________________________
block5a_se_reshape (Reshape) (None, 1, 1, 480) 0 block5a_se_squeeze[0][0]
__________________________________________________________________________________________________
block5a_se_reduce (Conv2D) (None, 1, 1, 20) 9620 block5a_se_reshape[0][0]
__________________________________________________________________________________________________
block5a_se_expand (Conv2D) (None, 1, 1, 480) 10080 block5a_se_reduce[0][0]
__________________________________________________________________________________________________
block5a_se_excite (Multiply) (None, None, None, 4 0 block5a_activation[0][0]
block5a_se_expand[0][0]
__________________________________________________________________________________________________
block5a_project_conv (Conv2D) (None, None, None, 1 53760 block5a_se_excite[0][0]
__________________________________________________________________________________________________
block5a_project_bn (BatchNormal (None, None, None, 1 448 block5a_project_conv[0][0]
__________________________________________________________________________________________________
block5b_expand_conv (Conv2D) (None, None, None, 6 75264 block5a_project_bn[0][0]
__________________________________________________________________________________________________
block5b_expand_bn (BatchNormali (None, None, None, 6 2688 block5b_expand_conv[0][0]
__________________________________________________________________________________________________
block5b_expand_activation (Acti (None, None, None, 6 0 block5b_expand_bn[0][0]
__________________________________________________________________________________________________
block5b_dwconv (DepthwiseConv2D (None, None, None, 6 16800 block5b_expand_activation[0][0]
__________________________________________________________________________________________________
block5b_bn (BatchNormalization) (None, None, None, 6 2688 block5b_dwconv[0][0]
__________________________________________________________________________________________________
block5b_activation (Activation) (None, None, None, 6 0 block5b_bn[0][0]
__________________________________________________________________________________________________
block5b_se_squeeze (GlobalAvera (None, 672) 0 block5b_activation[0][0]
__________________________________________________________________________________________________
block5b_se_reshape (Reshape) (None, 1, 1, 672) 0 block5b_se_squeeze[0][0]
__________________________________________________________________________________________________
block5b_se_reduce (Conv2D) (None, 1, 1, 28) 18844 block5b_se_reshape[0][0]
__________________________________________________________________________________________________
block5b_se_expand (Conv2D) (None, 1, 1, 672) 19488 block5b_se_reduce[0][0]
__________________________________________________________________________________________________
block5b_se_excite (Multiply) (None, None, None, 6 0 block5b_activation[0][0]
block5b_se_expand[0][0]
__________________________________________________________________________________________________
block5b_project_conv (Conv2D) (None, None, None, 1 75264 block5b_se_excite[0][0]
__________________________________________________________________________________________________
block5b_project_bn (BatchNormal (None, None, None, 1 448 block5b_project_conv[0][0]
__________________________________________________________________________________________________
block5b_drop (Dropout) (None, None, None, 1 0 block5b_project_bn[0][0]
__________________________________________________________________________________________________
block5b_add (Add) (None, None, None, 1 0 block5b_drop[0][0]
block5a_project_bn[0][0]
__________________________________________________________________________________________________
block5c_expand_conv (Conv2D) (None, None, None, 6 75264 block5b_add[0][0]
__________________________________________________________________________________________________
block5c_expand_bn (BatchNormali (None, None, None, 6 2688 block5c_expand_conv[0][0]
__________________________________________________________________________________________________
block5c_expand_activation (Acti (None, None, None, 6 0 block5c_expand_bn[0][0]
__________________________________________________________________________________________________
block5c_dwconv (DepthwiseConv2D (None, None, None, 6 16800 block5c_expand_activation[0][0]
__________________________________________________________________________________________________
block5c_bn (BatchNormalization) (None, None, None, 6 2688 block5c_dwconv[0][0]
__________________________________________________________________________________________________
block5c_activation (Activation) (None, None, None, 6 0 block5c_bn[0][0]
__________________________________________________________________________________________________
block5c_se_squeeze (GlobalAvera (None, 672) 0 block5c_activation[0][0]
__________________________________________________________________________________________________
block5c_se_reshape (Reshape) (None, 1, 1, 672) 0 block5c_se_squeeze[0][0]
__________________________________________________________________________________________________
block5c_se_reduce (Conv2D) (None, 1, 1, 28) 18844 block5c_se_reshape[0][0]
__________________________________________________________________________________________________
block5c_se_expand (Conv2D) (None, 1, 1, 672) 19488 block5c_se_reduce[0][0]
__________________________________________________________________________________________________
block5c_se_excite (Multiply) (None, None, None, 6 0 block5c_activation[0][0]
block5c_se_expand[0][0]
__________________________________________________________________________________________________
block5c_project_conv (Conv2D) (None, None, None, 1 75264 block5c_se_excite[0][0]
__________________________________________________________________________________________________
block5c_project_bn (BatchNormal (None, None, None, 1 448 block5c_project_conv[0][0]
__________________________________________________________________________________________________
block5c_drop (Dropout) (None, None, None, 1 0 block5c_project_bn[0][0]
__________________________________________________________________________________________________
block5c_add (Add) (None, None, None, 1 0 block5c_drop[0][0]
block5b_add[0][0]
__________________________________________________________________________________________________
block6a_expand_conv (Conv2D) (None, None, None, 6 75264 block5c_add[0][0]
__________________________________________________________________________________________________
block6a_expand_bn (BatchNormali (None, None, None, 6 2688 block6a_expand_conv[0][0]
__________________________________________________________________________________________________
block6a_expand_activation (Acti (None, None, None, 6 0 block6a_expand_bn[0][0]
__________________________________________________________________________________________________
block6a_dwconv_pad (ZeroPadding (None, None, None, 6 0 block6a_expand_activation[0][0]
__________________________________________________________________________________________________
block6a_dwconv (DepthwiseConv2D (None, None, None, 6 16800 block6a_dwconv_pad[0][0]
__________________________________________________________________________________________________
block6a_bn (BatchNormalization) (None, None, None, 6 2688 block6a_dwconv[0][0]
__________________________________________________________________________________________________
block6a_activation (Activation) (None, None, None, 6 0 block6a_bn[0][0]
__________________________________________________________________________________________________
block6a_se_squeeze (GlobalAvera (None, 672) 0 block6a_activation[0][0]
__________________________________________________________________________________________________
block6a_se_reshape (Reshape) (None, 1, 1, 672) 0 block6a_se_squeeze[0][0]
__________________________________________________________________________________________________
block6a_se_reduce (Conv2D) (None, 1, 1, 28) 18844 block6a_se_reshape[0][0]
__________________________________________________________________________________________________
block6a_se_expand (Conv2D) (None, 1, 1, 672) 19488 block6a_se_reduce[0][0]
__________________________________________________________________________________________________
block6a_se_excite (Multiply) (None, None, None, 6 0 block6a_activation[0][0]
block6a_se_expand[0][0]
__________________________________________________________________________________________________
block6a_project_conv (Conv2D) (None, None, None, 1 129024 block6a_se_excite[0][0]
__________________________________________________________________________________________________
block6a_project_bn (BatchNormal (None, None, None, 1 768 block6a_project_conv[0][0]
__________________________________________________________________________________________________
block6b_expand_conv (Conv2D) (None, None, None, 1 221184 block6a_project_bn[0][0]
__________________________________________________________________________________________________
block6b_expand_bn (BatchNormali (None, None, None, 1 4608 block6b_expand_conv[0][0]
__________________________________________________________________________________________________
block6b_expand_activation (Acti (None, None, None, 1 0 block6b_expand_bn[0][0]
__________________________________________________________________________________________________
block6b_dwconv (DepthwiseConv2D (None, None, None, 1 28800 block6b_expand_activation[0][0]
__________________________________________________________________________________________________
block6b_bn (BatchNormalization) (None, None, None, 1 4608 block6b_dwconv[0][0]
__________________________________________________________________________________________________
block6b_activation (Activation) (None, None, None, 1 0 block6b_bn[0][0]
__________________________________________________________________________________________________
block6b_se_squeeze (GlobalAvera (None, 1152) 0 block6b_activation[0][0]
__________________________________________________________________________________________________
block6b_se_reshape (Reshape) (None, 1, 1, 1152) 0 block6b_se_squeeze[0][0]
__________________________________________________________________________________________________
block6b_se_reduce (Conv2D) (None, 1, 1, 48) 55344 block6b_se_reshape[0][0]
__________________________________________________________________________________________________
block6b_se_expand (Conv2D) (None, 1, 1, 1152) 56448 block6b_se_reduce[0][0]
__________________________________________________________________________________________________
block6b_se_excite (Multiply) (None, None, None, 1 0 block6b_activation[0][0]
block6b_se_expand[0][0]
__________________________________________________________________________________________________
block6b_project_conv (Conv2D) (None, None, None, 1 221184 block6b_se_excite[0][0]
__________________________________________________________________________________________________
block6b_project_bn (BatchNormal (None, None, None, 1 768 block6b_project_conv[0][0]
__________________________________________________________________________________________________
block6b_drop (Dropout) (None, None, None, 1 0 block6b_project_bn[0][0]
__________________________________________________________________________________________________
block6b_add (Add) (None, None, None, 1 0 block6b_drop[0][0]
block6a_project_bn[0][0]
__________________________________________________________________________________________________
block6c_expand_conv (Conv2D) (None, None, None, 1 221184 block6b_add[0][0]
__________________________________________________________________________________________________
block6c_expand_bn (BatchNormali (None, None, None, 1 4608 block6c_expand_conv[0][0]
__________________________________________________________________________________________________
block6c_expand_activation (Acti (None, None, None, 1 0 block6c_expand_bn[0][0]
__________________________________________________________________________________________________
block6c_dwconv (DepthwiseConv2D (None, None, None, 1 28800 block6c_expand_activation[0][0]
__________________________________________________________________________________________________
block6c_bn (BatchNormalization) (None, None, None, 1 4608 block6c_dwconv[0][0]
__________________________________________________________________________________________________
block6c_activation (Activation) (None, None, None, 1 0 block6c_bn[0][0]
__________________________________________________________________________________________________
block6c_se_squeeze (GlobalAvera (None, 1152) 0 block6c_activation[0][0]
__________________________________________________________________________________________________
block6c_se_reshape (Reshape) (None, 1, 1, 1152) 0 block6c_se_squeeze[0][0]
__________________________________________________________________________________________________
block6c_se_reduce (Conv2D) (None, 1, 1, 48) 55344 block6c_se_reshape[0][0]
__________________________________________________________________________________________________
block6c_se_expand (Conv2D) (None, 1, 1, 1152) 56448 block6c_se_reduce[0][0]
__________________________________________________________________________________________________
block6c_se_excite (Multiply) (None, None, None, 1 0 block6c_activation[0][0]
block6c_se_expand[0][0]
__________________________________________________________________________________________________
block6c_project_conv (Conv2D) (None, None, None, 1 221184 block6c_se_excite[0][0]
__________________________________________________________________________________________________
block6c_project_bn (BatchNormal (None, None, None, 1 768 block6c_project_conv[0][0]
__________________________________________________________________________________________________
block6c_drop (Dropout) (None, None, None, 1 0 block6c_project_bn[0][0]
__________________________________________________________________________________________________
block6c_add (Add) (None, None, None, 1 0 block6c_drop[0][0]
block6b_add[0][0]
__________________________________________________________________________________________________
block6d_expand_conv (Conv2D) (None, None, None, 1 221184 block6c_add[0][0]
__________________________________________________________________________________________________
block6d_expand_bn (BatchNormali (None, None, None, 1 4608 block6d_expand_conv[0][0]
__________________________________________________________________________________________________
block6d_expand_activation (Acti (None, None, None, 1 0 block6d_expand_bn[0][0]
__________________________________________________________________________________________________
block6d_dwconv (DepthwiseConv2D (None, None, None, 1 28800 block6d_expand_activation[0][0]
__________________________________________________________________________________________________
block6d_bn (BatchNormalization) (None, None, None, 1 4608 block6d_dwconv[0][0]
__________________________________________________________________________________________________
block6d_activation (Activation) (None, None, None, 1 0 block6d_bn[0][0]
__________________________________________________________________________________________________
block6d_se_squeeze (GlobalAvera (None, 1152) 0 block6d_activation[0][0]
__________________________________________________________________________________________________
block6d_se_reshape (Reshape) (None, 1, 1, 1152) 0 block6d_se_squeeze[0][0]
__________________________________________________________________________________________________
block6d_se_reduce (Conv2D) (None, 1, 1, 48) 55344 block6d_se_reshape[0][0]
__________________________________________________________________________________________________
block6d_se_expand (Conv2D) (None, 1, 1, 1152) 56448 block6d_se_reduce[0][0]
__________________________________________________________________________________________________
block6d_se_excite (Multiply) (None, None, None, 1 0 block6d_activation[0][0]
block6d_se_expand[0][0]
__________________________________________________________________________________________________
block6d_project_conv (Conv2D) (None, None, None, 1 221184 block6d_se_excite[0][0]
__________________________________________________________________________________________________
block6d_project_bn (BatchNormal (None, None, None, 1 768 block6d_project_conv[0][0]
__________________________________________________________________________________________________
block6d_drop (Dropout) (None, None, None, 1 0 block6d_project_bn[0][0]
__________________________________________________________________________________________________
block6d_add (Add) (None, None, None, 1 0 block6d_drop[0][0]
block6c_add[0][0]
__________________________________________________________________________________________________
block7a_expand_conv (Conv2D) (None, None, None, 1 221184 block6d_add[0][0]
__________________________________________________________________________________________________
block7a_expand_bn (BatchNormali (None, None, None, 1 4608 block7a_expand_conv[0][0]
__________________________________________________________________________________________________
block7a_expand_activation (Acti (None, None, None, 1 0 block7a_expand_bn[0][0]
__________________________________________________________________________________________________
block7a_dwconv (DepthwiseConv2D (None, None, None, 1 10368 block7a_expand_activation[0][0]
__________________________________________________________________________________________________
block7a_bn (BatchNormalization) (None, None, None, 1 4608 block7a_dwconv[0][0]
__________________________________________________________________________________________________
block7a_activation (Activation) (None, None, None, 1 0 block7a_bn[0][0]
__________________________________________________________________________________________________
block7a_se_squeeze (GlobalAvera (None, 1152) 0 block7a_activation[0][0]
__________________________________________________________________________________________________
block7a_se_reshape (Reshape) (None, 1, 1, 1152) 0 block7a_se_squeeze[0][0]
__________________________________________________________________________________________________
block7a_se_reduce (Conv2D) (None, 1, 1, 48) 55344 block7a_se_reshape[0][0]
__________________________________________________________________________________________________
block7a_se_expand (Conv2D) (None, 1, 1, 1152) 56448 block7a_se_reduce[0][0]
__________________________________________________________________________________________________
block7a_se_excite (Multiply) (None, None, None, 1 0 block7a_activation[0][0]
block7a_se_expand[0][0]
__________________________________________________________________________________________________
block7a_project_conv (Conv2D) (None, None, None, 3 368640 block7a_se_excite[0][0]
__________________________________________________________________________________________________
block7a_project_bn (BatchNormal (None, None, None, 3 1280 block7a_project_conv[0][0]
__________________________________________________________________________________________________
top_conv (Conv2D) (None, None, None, 1 409600 block7a_project_bn[0][0]
__________________________________________________________________________________________________
top_bn (BatchNormalization) (None, None, None, 1 5120 top_conv[0][0]
__________________________________________________________________________________________________
top_activation (Activation) (None, None, None, 1 0 top_bn[0][0]
==================================================================================================
Total params: 4,049,571
Trainable params: 0
Non-trainable params: 4,049,571
__________________________________________________________________________________________________
Explore the whole Model¶
In [ ]:
Copied!
model.summary()
model.summary()
Model: "model_1" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= input_layer (InputLayer) [(None, 224, 224, 3)] 0 _________________________________________________________________ efficientnetb0 (Functional) (None, None, None, 1280) 4049571 _________________________________________________________________ global_avg_pooling_layer (Gl (None, 1280) 0 _________________________________________________________________ output_layer (Dense) (None, 10) 12810 ================================================================= Total params: 4,062,381 Trainable params: 12,810 Non-trainable params: 4,049,571 _________________________________________________________________
In [ ]:
Copied!
for layer in model.layers:
print(layer.name, layer.trainable)
for layer in model.layers:
print(layer.name, layer.trainable)
input_layer True efficientnetb0 False global_avg_pooling_layer True output_layer True
Exploring GlobalAveragePool2D¶
- Averaging accross inner dimensions
- Denoising the input
- Reducing the dimensionality
- Selecting the most important features (for each depth channel)
In [ ]:
Copied!
# Create a random tensor
input_shape = (1, 4, 4, 3)
input_tensor = tf.random.normal(input_shape)
print(f'Shape of the input tensor {input_tensor.shape}')
# Apply GlobalAveragePool2D
global_avg_pooled_tensor = layers.GlobalAveragePooling2D()(input_tensor)
print(f'Shape after GlobalAveragePooling2D: {global_avg_pooled_tensor.shape}')
# Create a random tensor
input_shape = (1, 4, 4, 3)
input_tensor = tf.random.normal(input_shape)
print(f'Shape of the input tensor {input_tensor.shape}')
# Apply GlobalAveragePool2D
global_avg_pooled_tensor = layers.GlobalAveragePooling2D()(input_tensor)
print(f'Shape after GlobalAveragePooling2D: {global_avg_pooled_tensor.shape}')
Shape of the input tensor (1, 4, 4, 3) Shape after GlobalAveragePooling2D: (1, 3)
In [ ]:
Copied!
# Same as GlobalAveragePooling2D
tf.reduce_mean(input_tensor, axis=[1, 2]).shape
# Same as GlobalAveragePooling2D
tf.reduce_mean(input_tensor, axis=[1, 2]).shape
Out[ ]:
TensorShape([1, 3])
Data Augmentation Layer¶
In [ ]:
Copied!
from tensorflow.keras.layers.experimental import preprocessing as KerasPreprocessing
from tensorflow.keras.layers.experimental import preprocessing as KerasPreprocessing
In [ ]:
Copied!
data_augmentation = tf.keras.models.Sequential([
KerasPreprocessing.RandomFlip('horizontal'),
KerasPreprocessing.RandomRotation(0.2),
KerasPreprocessing.RandomZoom(0.2),
KerasPreprocessing.RandomHeight(0.2),
KerasPreprocessing.RandomWidth(0.2),
KerasPreprocessing.Resizing(height=IMAGE_DIM[0], width=IMAGE_DIM[1]),
KerasPreprocessing.Rescaling(1/255.)
], name='data_augmentation')
data_augmentation = tf.keras.models.Sequential([
KerasPreprocessing.RandomFlip('horizontal'),
KerasPreprocessing.RandomRotation(0.2),
KerasPreprocessing.RandomZoom(0.2),
KerasPreprocessing.RandomHeight(0.2),
KerasPreprocessing.RandomWidth(0.2),
KerasPreprocessing.Resizing(height=IMAGE_DIM[0], width=IMAGE_DIM[1]),
KerasPreprocessing.Rescaling(1/255.)
], name='data_augmentation')
Compare normal vs augmentation image¶
In [ ]:
Copied!
batch = imgdir.load(1)
imgen = next(batch)
rand_image, label = np.squeeze(imgen.train_images), np.squeeze(imgen.train_labels)
aug_image = data_augmentation(tf.expand_dims(rand_image, axis=0))
fig, axn = plt.subplots(1, 2, figsize=(12, 4))
for ax, img, img_type in zip(axn, [rand_image, aug_image], ['Normal', 'Augmented']):
ax.imshow(np.squeeze(img))
ax.set_axis_off()
ax.set_title(img_type, fontdict=dict(size=20, weight='bold'))
batch = imgdir.load(1)
imgen = next(batch)
rand_image, label = np.squeeze(imgen.train_images), np.squeeze(imgen.train_labels)
aug_image = data_augmentation(tf.expand_dims(rand_image, axis=0))
fig, axn = plt.subplots(1, 2, figsize=(12, 4))
for ax, img, img_type in zip(axn, [rand_image, aug_image], ['Normal', 'Augmented']):
ax.imshow(np.squeeze(img))
ax.set_axis_off()
ax.set_title(img_type, fontdict=dict(size=20, weight='bold'))
Running a series of transfer learning experiments¶
Experiment with:
- Three different proportions of data:
1_percent,10_percent,100_percent - Two kinds of transfer learning:
Feature Extraction,Fine Tuning - Two models:
EfficientNetB0,ResNet50V2
In [ ]:
Copied!
from src.utils import create_tensorboard_callback
import os
from src.utils import create_tensorboard_callback
import os
In [ ]:
Copied!
class FoodClassificationExperiment:
DATA_DIR = {'1_percent': '../data/10_food_classes_1_percent',
'10_percent': '../data/10_food_classes_10_percent',
'100_percent': '../data/10_food_classes_all_data'}
IMAGE_DIM = (224, 224)
NUM_CHANNELS = 3
INPUT_SHAPE = (*IMAGE_DIM, NUM_CHANNELS)
BATCH_SIZE = 32
VALIDATION_SPLIT = 0.2
CLASSIFICATION_TYPE = 'categorical'
SCALING_FACTOR = 1/255.
SEED = 42
IMAGE_DATA_DIRECTORY = {data_pct: ClassicImageDataDirectory(data_dir, IMAGE_DIM, np.uint8) for data_pct, data_dir in DATA_DIR.items()}
BASE_EXPERIMENT = 'transfer_learning_experiments'
TASK = '10_food_multiclass_classification'
TFHUB_LOG_DIR = os.path.abspath('../tensorboard_logs/')
EXPERIMENT_COUNTER = 1
def __init__(self, data_pct, pretrained_model, data_augment=True, scale=True, name=None):
self.data_pct = data_pct
self.pretrained_model = tf.keras.models.clone_model(pretrained_model)
self.pretrained_model.set_weights(pretrained_model.get_weights())
self.data_augment = data_augment
self.scale = scale
self.data_dir = self.DATA_DIR[data_pct]
self.imgdir = self.IMAGE_DATA_DIRECTORY[data_pct]
self.mode = 'feature_extraction'
self.training_history = []
self._assign_name(name)
def _assign_name(self, name):
exp_id = self.EXPERIMENT_COUNTER
if name is None:
name = f'experiment_{exp_id}'
self.name = name
self.__class__.EXPERIMENT_COUNTER += 1
def preprocess_data(self):
self.train_datagen = ImageDataGenerator(validation_split=self.VALIDATION_SPLIT)
self.test_datagen = ImageDataGenerator()
def setup_directories(self):
self.train_dir = self.imgdir.train['dir']
self.test_dir = self.imgdir.test['dir']
def import_data_from_directories(self):
self.train_data = self.train_datagen.flow_from_directory(directory=self.train_dir, target_size=self.IMAGE_DIM,
batch_size=self.BATCH_SIZE, class_mode=self.CLASSIFICATION_TYPE,
subset='training', seed=self.SEED)
self.validation_data = self.train_datagen.flow_from_directory(directory=self.train_dir, target_size=self.IMAGE_DIM,
batch_size=self.BATCH_SIZE, class_mode=self.CLASSIFICATION_TYPE,
subset='validation', seed=self.SEED)
self.test_data = self.test_datagen.flow_from_directory(directory=self.test_dir, target_size=self.IMAGE_DIM,
batch_size=self.BATCH_SIZE, class_mode=self.CLASSIFICATION_TYPE,
shuffle=False)
def create_model(self):
inputs = layers.Input(shape=self.INPUT_SHAPE, name='input_layer')
if self.data_augment:
x = self._get_data_augment_layer()(inputs)
else:
x = inputs
if self.scale:
x = KerasPreprocessing.Rescaling(self.SCALING_FACTOR)(x)
self.pretrained_model.trainable = False
features = self.pretrained_model(x)
pooled = layers.GlobalAvgPool2D(name='global_avg_pooling_layer')(features)
outputs = layers.Dense(self.imgdir.n_classes, activation='softmax', name='output_layer')(pooled)
self.model = tf.keras.models.Model(inputs, outputs)
def compile_model(self):
self.model.compile(loss=losses.categorical_crossentropy, optimizer=optimizers.Adam(),
metrics=[KerasMetrics.f1, 'accuracy'])
def _get_data_augment_layer(self):
data_augment_layer = tf.keras.models.Sequential([
KerasPreprocessing.RandomFlip('horizontal'),
KerasPreprocessing.RandomRotation(0.2),
KerasPreprocessing.RandomZoom(0.2),
KerasPreprocessing.RandomHeight(0.2),
KerasPreprocessing.Resizing(height=self.IMAGE_DIM[0], width=self.IMAGE_DIM[1])
])
return data_augment_layer
def set_training_mode(self, mode, unfreeze_last_n=None):
if mode == 'feature_extraction':
assert unfreeze_last_n is None, 'This argument is only valid for mode == "fine_tuning"'
self.pretrained_model.trainable = False
elif mode == 'fine_tuning':
self.pretrained_model.trainable = True
for layer in self.pretrained_model.layers[:-unfreeze_last_n]:
layer.trainable = False
self.mode = mode
def run(self, epochs=10, tfhub_log=False):
if tfhub_log:
exp_name = os.path.join(self.BASE_EXPERIMENT, self.name)
model_callbacks = [create_tensorboard_callback(exp_name, self.TASK, self.TFHUB_LOG_DIR)]
else:
model_callbacks = []
try:
initial_epoch = self.model.history.epoch[-1]
except (IndexError, AttributeError):
initial_epoch = 0
self.model.fit(self.train_data, steps_per_epoch=len(self.train_data),
validation_data=self.validation_data, validation_steps=len(self.validation_data),
epochs=epochs+initial_epoch, initial_epoch=initial_epoch, callbacks=model_callbacks)
history_dict = self.model.history.history.copy()
history_dict['epoch'] = self.model.history.epoch
self.training_history.append((self.mode, history_dict))
def plot_learning_curve(self, metric='loss'):
fig, ax = plt.subplots()
max_epoch = self.training_history[-1][1]['epoch'][-1]
for i, (mode, history_dict) in enumerate(exp.training_history):
ax.plot(history_dict['epoch'], history_dict[metric], color='blue', label=metric)
ax.plot(history_dict['epoch'], history_dict[f'val_{metric}'], color='orange', label=f'val_{metric}')
epochs = history_dict['epoch']
mid_epoch = (epochs[0] + epochs[-1])/2 - 0.5
ax.text(mid_epoch/(max_epoch+1), 0.5, mode, transform=ax.transAxes)
if i == 0:
ax.set_xlim(0, max_epoch)
else:
ax.axvline(x=history_dict['epoch'][0], color='green')
ax.set_xlabel('epoch')
ax.set_ylabel(metric)
return ax
class FoodClassificationExperiment:
DATA_DIR = {'1_percent': '../data/10_food_classes_1_percent',
'10_percent': '../data/10_food_classes_10_percent',
'100_percent': '../data/10_food_classes_all_data'}
IMAGE_DIM = (224, 224)
NUM_CHANNELS = 3
INPUT_SHAPE = (*IMAGE_DIM, NUM_CHANNELS)
BATCH_SIZE = 32
VALIDATION_SPLIT = 0.2
CLASSIFICATION_TYPE = 'categorical'
SCALING_FACTOR = 1/255.
SEED = 42
IMAGE_DATA_DIRECTORY = {data_pct: ClassicImageDataDirectory(data_dir, IMAGE_DIM, np.uint8) for data_pct, data_dir in DATA_DIR.items()}
BASE_EXPERIMENT = 'transfer_learning_experiments'
TASK = '10_food_multiclass_classification'
TFHUB_LOG_DIR = os.path.abspath('../tensorboard_logs/')
EXPERIMENT_COUNTER = 1
def __init__(self, data_pct, pretrained_model, data_augment=True, scale=True, name=None):
self.data_pct = data_pct
self.pretrained_model = tf.keras.models.clone_model(pretrained_model)
self.pretrained_model.set_weights(pretrained_model.get_weights())
self.data_augment = data_augment
self.scale = scale
self.data_dir = self.DATA_DIR[data_pct]
self.imgdir = self.IMAGE_DATA_DIRECTORY[data_pct]
self.mode = 'feature_extraction'
self.training_history = []
self._assign_name(name)
def _assign_name(self, name):
exp_id = self.EXPERIMENT_COUNTER
if name is None:
name = f'experiment_{exp_id}'
self.name = name
self.__class__.EXPERIMENT_COUNTER += 1
def preprocess_data(self):
self.train_datagen = ImageDataGenerator(validation_split=self.VALIDATION_SPLIT)
self.test_datagen = ImageDataGenerator()
def setup_directories(self):
self.train_dir = self.imgdir.train['dir']
self.test_dir = self.imgdir.test['dir']
def import_data_from_directories(self):
self.train_data = self.train_datagen.flow_from_directory(directory=self.train_dir, target_size=self.IMAGE_DIM,
batch_size=self.BATCH_SIZE, class_mode=self.CLASSIFICATION_TYPE,
subset='training', seed=self.SEED)
self.validation_data = self.train_datagen.flow_from_directory(directory=self.train_dir, target_size=self.IMAGE_DIM,
batch_size=self.BATCH_SIZE, class_mode=self.CLASSIFICATION_TYPE,
subset='validation', seed=self.SEED)
self.test_data = self.test_datagen.flow_from_directory(directory=self.test_dir, target_size=self.IMAGE_DIM,
batch_size=self.BATCH_SIZE, class_mode=self.CLASSIFICATION_TYPE,
shuffle=False)
def create_model(self):
inputs = layers.Input(shape=self.INPUT_SHAPE, name='input_layer')
if self.data_augment:
x = self._get_data_augment_layer()(inputs)
else:
x = inputs
if self.scale:
x = KerasPreprocessing.Rescaling(self.SCALING_FACTOR)(x)
self.pretrained_model.trainable = False
features = self.pretrained_model(x)
pooled = layers.GlobalAvgPool2D(name='global_avg_pooling_layer')(features)
outputs = layers.Dense(self.imgdir.n_classes, activation='softmax', name='output_layer')(pooled)
self.model = tf.keras.models.Model(inputs, outputs)
def compile_model(self):
self.model.compile(loss=losses.categorical_crossentropy, optimizer=optimizers.Adam(),
metrics=[KerasMetrics.f1, 'accuracy'])
def _get_data_augment_layer(self):
data_augment_layer = tf.keras.models.Sequential([
KerasPreprocessing.RandomFlip('horizontal'),
KerasPreprocessing.RandomRotation(0.2),
KerasPreprocessing.RandomZoom(0.2),
KerasPreprocessing.RandomHeight(0.2),
KerasPreprocessing.Resizing(height=self.IMAGE_DIM[0], width=self.IMAGE_DIM[1])
])
return data_augment_layer
def set_training_mode(self, mode, unfreeze_last_n=None):
if mode == 'feature_extraction':
assert unfreeze_last_n is None, 'This argument is only valid for mode == "fine_tuning"'
self.pretrained_model.trainable = False
elif mode == 'fine_tuning':
self.pretrained_model.trainable = True
for layer in self.pretrained_model.layers[:-unfreeze_last_n]:
layer.trainable = False
self.mode = mode
def run(self, epochs=10, tfhub_log=False):
if tfhub_log:
exp_name = os.path.join(self.BASE_EXPERIMENT, self.name)
model_callbacks = [create_tensorboard_callback(exp_name, self.TASK, self.TFHUB_LOG_DIR)]
else:
model_callbacks = []
try:
initial_epoch = self.model.history.epoch[-1]
except (IndexError, AttributeError):
initial_epoch = 0
self.model.fit(self.train_data, steps_per_epoch=len(self.train_data),
validation_data=self.validation_data, validation_steps=len(self.validation_data),
epochs=epochs+initial_epoch, initial_epoch=initial_epoch, callbacks=model_callbacks)
history_dict = self.model.history.history.copy()
history_dict['epoch'] = self.model.history.epoch
self.training_history.append((self.mode, history_dict))
def plot_learning_curve(self, metric='loss'):
fig, ax = plt.subplots()
max_epoch = self.training_history[-1][1]['epoch'][-1]
for i, (mode, history_dict) in enumerate(exp.training_history):
ax.plot(history_dict['epoch'], history_dict[metric], color='blue', label=metric)
ax.plot(history_dict['epoch'], history_dict[f'val_{metric}'], color='orange', label=f'val_{metric}')
epochs = history_dict['epoch']
mid_epoch = (epochs[0] + epochs[-1])/2 - 0.5
ax.text(mid_epoch/(max_epoch+1), 0.5, mode, transform=ax.transAxes)
if i == 0:
ax.set_xlim(0, max_epoch)
else:
ax.axvline(x=history_dict['epoch'][0], color='green')
ax.set_xlabel('epoch')
ax.set_ylabel(metric)
return ax
Hold the experiments¶
In [ ]:
Copied!
experiments = {}
experiments = {}
Experiment 1: EfficientNetB0 feature_extraction with 1_percent data + data augmentation¶
In [ ]:
Copied!
exp = FoodClassificationExperiment(data_pct='1_percent', pretrained_model=tf.keras.applications.EfficientNetB0(include_top=False),
data_augment=True, scale=False, name='experiment_1')
exp = FoodClassificationExperiment(data_pct='1_percent', pretrained_model=tf.keras.applications.EfficientNetB0(include_top=False),
data_augment=True, scale=False, name='experiment_1')
In [ ]:
Copied!
exp.preprocess_data()
exp.setup_directories()
exp.import_data_from_directories()
exp.create_model()
exp.compile_model()
exp.run(10, tfhub_log=True)
exp.preprocess_data()
exp.setup_directories()
exp.import_data_from_directories()
exp.create_model()
exp.compile_model()
exp.run(10, tfhub_log=True)
Found 60 images belonging to 10 classes. Found 10 images belonging to 10 classes. Found 2500 images belonging to 10 classes. Saving TensorBoard log files to " C:\Users\bhati\Google Drive\projects\Tensorflow-tutorial-Daniel-Bourke\tensorboard_logs\10_food_multiclass_classification\transfer_learning_experiments\experiment_1\20210531-215453" Epoch 1/10 2/2 [==============================] - 8s 2s/step - loss: 2.3442 - f1: 0.0000e+00 - accuracy: 0.2069 - val_loss: 2.2104 - val_f1: 0.0000e+00 - val_accuracy: 0.2000 Epoch 2/10 2/2 [==============================] - 1s 466ms/step - loss: 2.2137 - f1: 0.0000e+00 - accuracy: 0.1587 - val_loss: 2.1108 - val_f1: 0.0000e+00 - val_accuracy: 0.3000 Epoch 3/10 2/2 [==============================] - 1s 414ms/step - loss: 2.0551 - f1: 0.0000e+00 - accuracy: 0.2722 - val_loss: 2.0228 - val_f1: 0.0000e+00 - val_accuracy: 0.4000 Epoch 4/10 2/2 [==============================] - 1s 480ms/step - loss: 1.9150 - f1: 0.0000e+00 - accuracy: 0.3857 - val_loss: 1.9450 - val_f1: 0.0000e+00 - val_accuracy: 0.6000 Epoch 5/10 2/2 [==============================] - 1s 452ms/step - loss: 1.7686 - f1: 0.0000e+00 - accuracy: 0.5326 - val_loss: 1.8684 - val_f1: 0.0000e+00 - val_accuracy: 0.6000 Epoch 6/10 2/2 [==============================] - 1s 706ms/step - loss: 1.6451 - f1: 0.0000e+00 - accuracy: 0.6476 - val_loss: 1.7902 - val_f1: 0.0000e+00 - val_accuracy: 0.6000 Epoch 7/10 2/2 [==============================] - 1s 643ms/step - loss: 1.5067 - f1: 0.0000e+00 - accuracy: 0.7841 - val_loss: 1.7111 - val_f1: 0.0000e+00 - val_accuracy: 0.6000 Epoch 8/10 2/2 [==============================] - 1s 600ms/step - loss: 1.3895 - f1: 0.1091 - accuracy: 0.7492 - val_loss: 1.6340 - val_f1: 0.0000e+00 - val_accuracy: 0.6000 Epoch 9/10 2/2 [==============================] - 1s 534ms/step - loss: 1.3290 - f1: 0.0784 - accuracy: 0.8910 - val_loss: 1.5600 - val_f1: 0.0000e+00 - val_accuracy: 0.6000 Epoch 10/10 2/2 [==============================] - 1s 552ms/step - loss: 1.2008 - f1: 0.1790 - accuracy: 0.9206 - val_loss: 1.4910 - val_f1: 0.0000e+00 - val_accuracy: 0.6000
In [ ]:
Copied!
exp.plot_learning_curve()
exp.plot_learning_curve()
Out[ ]:
<AxesSubplot:xlabel='epoch', ylabel='loss'>
In [ ]:
Copied!
experiments[exp.name] = exp
experiments[exp.name] = exp
Experiment 2: EfficientNetB0 feature_extraction with 10_percent data + data augmentation¶
In [ ]:
Copied!
exp = FoodClassificationExperiment(data_pct='10_percent', pretrained_model=tf.keras.applications.EfficientNetB0(include_top=False),
data_augment=True, scale=False, name='experiment_2')
exp = FoodClassificationExperiment(data_pct='10_percent', pretrained_model=tf.keras.applications.EfficientNetB0(include_top=False),
data_augment=True, scale=False, name='experiment_2')
In [ ]:
Copied!
exp.preprocess_data()
exp.setup_directories()
exp.import_data_from_directories()
exp.create_model()
exp.compile_model()
exp.run(10, tfhub_log=True)
exp.preprocess_data()
exp.setup_directories()
exp.import_data_from_directories()
exp.create_model()
exp.compile_model()
exp.run(10, tfhub_log=True)
Found 600 images belonging to 10 classes. Found 150 images belonging to 10 classes. Found 2500 images belonging to 10 classes. Saving TensorBoard log files to " C:\Users\bhati\Google Drive\projects\Tensorflow-tutorial-Daniel-Bourke\tensorboard_logs\10_food_multiclass_classification\transfer_learning_experiments\experiment_2\20210531-215538" Epoch 1/10 19/19 [==============================] - 16s 575ms/step - loss: 2.2551 - f1: 0.0023 - accuracy: 0.1934 - val_loss: 1.7777 - val_f1: 0.0000e+00 - val_accuracy: 0.4933 Epoch 2/10 19/19 [==============================] - 9s 493ms/step - loss: 1.6195 - f1: 0.0286 - accuracy: 0.5473 - val_loss: 1.4013 - val_f1: 0.1932 - val_accuracy: 0.6400 Epoch 3/10 19/19 [==============================] - 8s 428ms/step - loss: 1.2548 - f1: 0.3256 - accuracy: 0.6761 - val_loss: 1.2038 - val_f1: 0.4538 - val_accuracy: 0.6600 Epoch 4/10 19/19 [==============================] - 10s 519ms/step - loss: 1.0732 - f1: 0.4765 - accuracy: 0.7444 - val_loss: 1.0909 - val_f1: 0.5652 - val_accuracy: 0.6733 Epoch 5/10 19/19 [==============================] - 9s 452ms/step - loss: 0.9269 - f1: 0.6447 - accuracy: 0.7510 - val_loss: 1.0148 - val_f1: 0.6332 - val_accuracy: 0.7000 Epoch 6/10 19/19 [==============================] - 9s 458ms/step - loss: 0.8278 - f1: 0.6832 - accuracy: 0.7659 - val_loss: 0.9686 - val_f1: 0.6536 - val_accuracy: 0.6800 Epoch 7/10 19/19 [==============================] - 9s 449ms/step - loss: 0.8072 - f1: 0.7057 - accuracy: 0.7839 - val_loss: 0.9338 - val_f1: 0.6565 - val_accuracy: 0.7067 Epoch 8/10 19/19 [==============================] - 10s 524ms/step - loss: 0.6548 - f1: 0.7584 - accuracy: 0.8402 - val_loss: 0.9066 - val_f1: 0.6766 - val_accuracy: 0.6933 Epoch 9/10 19/19 [==============================] - 7s 368ms/step - loss: 0.6256 - f1: 0.7719 - accuracy: 0.8464 - val_loss: 0.8906 - val_f1: 0.6872 - val_accuracy: 0.7067 Epoch 10/10 19/19 [==============================] - 7s 360ms/step - loss: 0.5610 - f1: 0.8203 - accuracy: 0.8720 - val_loss: 0.8636 - val_f1: 0.7125 - val_accuracy: 0.7067
In [ ]:
Copied!
experiments[exp.name] = exp
experiments[exp.name] = exp
Experiment 3: EfficientNetB0 fine_tuning with 10_percent data¶
In [ ]:
Copied!
exp = FoodClassificationExperiment(data_pct='10_percent', pretrained_model=tf.keras.applications.EfficientNetB0(include_top=False),
data_augment=True, scale=False, name='experiment_3')
exp = FoodClassificationExperiment(data_pct='10_percent', pretrained_model=tf.keras.applications.EfficientNetB0(include_top=False),
data_augment=True, scale=False, name='experiment_3')
In [ ]:
Copied!
exp.preprocess_data()
exp.setup_directories()
exp.import_data_from_directories()
exp.create_model()
exp.compile_model()
exp.run(10, tfhub_log=True)
exp.preprocess_data()
exp.setup_directories()
exp.import_data_from_directories()
exp.create_model()
exp.compile_model()
exp.run(10, tfhub_log=True)
Found 600 images belonging to 10 classes. Found 150 images belonging to 10 classes. Found 2500 images belonging to 10 classes. Saving TensorBoard log files to " C:\Users\bhati\Google Drive\projects\Tensorflow-tutorial-Daniel-Bourke\tensorboard_logs\10_food_multiclass_classification\transfer_learning_experiments\experiment_3\20210531-220126" Epoch 1/10 19/19 [==============================] - 14s 469ms/step - loss: 2.2169 - f1: 0.0000e+00 - accuracy: 0.2080 - val_loss: 1.7888 - val_f1: 0.0000e+00 - val_accuracy: 0.4667 Epoch 2/10 19/19 [==============================] - 7s 340ms/step - loss: 1.5668 - f1: 0.0320 - accuracy: 0.6085 - val_loss: 1.4146 - val_f1: 0.2309 - val_accuracy: 0.5867 Epoch 3/10 19/19 [==============================] - 6s 336ms/step - loss: 1.2338 - f1: 0.3146 - accuracy: 0.7225 - val_loss: 1.2164 - val_f1: 0.4433 - val_accuracy: 0.6733 Epoch 4/10 19/19 [==============================] - 7s 341ms/step - loss: 1.0010 - f1: 0.5577 - accuracy: 0.7746 - val_loss: 1.1004 - val_f1: 0.5648 - val_accuracy: 0.6933 Epoch 5/10 19/19 [==============================] - 7s 339ms/step - loss: 0.8962 - f1: 0.6440 - accuracy: 0.7848 - val_loss: 1.0225 - val_f1: 0.6207 - val_accuracy: 0.6933 Epoch 6/10 19/19 [==============================] - 7s 352ms/step - loss: 0.8064 - f1: 0.7054 - accuracy: 0.8069 - val_loss: 0.9713 - val_f1: 0.6309 - val_accuracy: 0.7133 Epoch 7/10 19/19 [==============================] - 8s 416ms/step - loss: 0.7201 - f1: 0.7522 - accuracy: 0.8013 - val_loss: 0.9285 - val_f1: 0.6530 - val_accuracy: 0.7067 Epoch 8/10 19/19 [==============================] - 9s 440ms/step - loss: 0.6193 - f1: 0.7980 - accuracy: 0.8422 - val_loss: 0.8998 - val_f1: 0.6499 - val_accuracy: 0.7467 Epoch 9/10 19/19 [==============================] - 8s 405ms/step - loss: 0.6166 - f1: 0.7904 - accuracy: 0.8646 - val_loss: 0.8838 - val_f1: 0.6517 - val_accuracy: 0.7467 Epoch 10/10 19/19 [==============================] - 7s 366ms/step - loss: 0.6810 - f1: 0.7664 - accuracy: 0.8030 - val_loss: 0.8621 - val_f1: 0.6713 - val_accuracy: 0.7267
Now change training mode¶
In [ ]:
Copied!
exp.set_training_mode('fine_tuning', unfreeze_last_n=10)
exp.set_training_mode('fine_tuning', unfreeze_last_n=10)
In [ ]:
Copied!
exp.run(5, tfhub_log=True)
exp.run(5, tfhub_log=True)
Saving TensorBoard log files to " C:\Users\bhati\Google Drive\projects\Tensorflow-tutorial-Daniel-Bourke\tensorboard_logs\10_food_multiclass_classification\transfer_learning_experiments\experiment_3\20210531-220356" Epoch 10/14 19/19 [==============================] - 7s 382ms/step - loss: 0.5710 - f1: 0.8201 - accuracy: 0.8417 - val_loss: 0.8458 - val_f1: 0.6822 - val_accuracy: 0.7467 Epoch 11/14 19/19 [==============================] - 8s 409ms/step - loss: 0.5596 - f1: 0.8215 - accuracy: 0.8517 - val_loss: 0.8417 - val_f1: 0.7059 - val_accuracy: 0.7400 Epoch 12/14 19/19 [==============================] - 9s 463ms/step - loss: 0.5547 - f1: 0.8142 - accuracy: 0.8567 - val_loss: 0.8341 - val_f1: 0.6968 - val_accuracy: 0.7600 Epoch 13/14 19/19 [==============================] - 8s 421ms/step - loss: 0.5057 - f1: 0.8320 - accuracy: 0.8733 - val_loss: 0.8218 - val_f1: 0.7145 - val_accuracy: 0.7600 Epoch 14/14 19/19 [==============================] - 9s 457ms/step - loss: 0.4708 - f1: 0.8545 - accuracy: 0.8883 - val_loss: 0.8118 - val_f1: 0.7195 - val_accuracy: 0.7733
In [ ]:
Copied!
exp.plot_learning_curve();
exp.plot_learning_curve();
In [ ]:
Copied!
experiments[exp.name] = exp
experiments[exp.name] = exp
Experiment 4: EfficientNetB0 fine_tuning with 100_percent data¶
In [ ]:
Copied!
exp = FoodClassificationExperiment(data_pct='100_percent', pretrained_model=tf.keras.applications.EfficientNetB0(include_top=False),
data_augment=True, scale=False, name='experiment_4')
exp = FoodClassificationExperiment(data_pct='100_percent', pretrained_model=tf.keras.applications.EfficientNetB0(include_top=False),
data_augment=True, scale=False, name='experiment_4')
In [ ]:
Copied!
exp.preprocess_data()
exp.setup_directories()
exp.import_data_from_directories()
exp.create_model()
exp.compile_model()
exp.run(10, tfhub_log=True)
exp.preprocess_data()
exp.setup_directories()
exp.import_data_from_directories()
exp.create_model()
exp.compile_model()
exp.run(10, tfhub_log=True)
Found 6000 images belonging to 10 classes. Found 1500 images belonging to 10 classes. Found 2500 images belonging to 10 classes. Saving TensorBoard log files to " C:\Users\bhati\Google Drive\projects\Tensorflow-tutorial-Daniel-Bourke\tensorboard_logs\10_food_multiclass_classification\transfer_learning_experiments\experiment_4\20210531-220615" Epoch 1/10 188/188 [==============================] - 87s 436ms/step - loss: 1.5972 - f1: 0.2263 - accuracy: 0.4928 - val_loss: 0.6914 - val_f1: 0.7670 - val_accuracy: 0.7907 Epoch 2/10 188/188 [==============================] - 77s 410ms/step - loss: 0.8005 - f1: 0.7165 - accuracy: 0.7610 - val_loss: 0.5780 - val_f1: 0.8037 - val_accuracy: 0.8153 Epoch 3/10 188/188 [==============================] - 66s 349ms/step - loss: 0.6893 - f1: 0.7698 - accuracy: 0.7835 - val_loss: 0.5549 - val_f1: 0.8174 - val_accuracy: 0.8273 Epoch 4/10 188/188 [==============================] - 76s 403ms/step - loss: 0.6296 - f1: 0.7945 - accuracy: 0.7968 - val_loss: 0.5281 - val_f1: 0.8280 - val_accuracy: 0.8300 Epoch 5/10 188/188 [==============================] - 76s 403ms/step - loss: 0.6103 - f1: 0.7964 - accuracy: 0.8101 - val_loss: 0.5016 - val_f1: 0.8418 - val_accuracy: 0.8433 Epoch 6/10 188/188 [==============================] - 76s 402ms/step - loss: 0.5755 - f1: 0.8073 - accuracy: 0.8156 - val_loss: 0.4848 - val_f1: 0.8505 - val_accuracy: 0.8453 Epoch 7/10 188/188 [==============================] - 70s 372ms/step - loss: 0.5419 - f1: 0.8209 - accuracy: 0.8367 - val_loss: 0.4848 - val_f1: 0.8514 - val_accuracy: 0.8467 Epoch 8/10 188/188 [==============================] - 79s 419ms/step - loss: 0.5496 - f1: 0.8188 - accuracy: 0.8227 - val_loss: 0.4788 - val_f1: 0.8493 - val_accuracy: 0.8500 Epoch 9/10 188/188 [==============================] - 81s 432ms/step - loss: 0.4939 - f1: 0.8406 - accuracy: 0.8453 - val_loss: 0.4869 - val_f1: 0.8443 - val_accuracy: 0.8440 Epoch 10/10 188/188 [==============================] - 82s 438ms/step - loss: 0.4918 - f1: 0.8317 - accuracy: 0.8383 - val_loss: 0.4736 - val_f1: 0.8585 - val_accuracy: 0.8513
Now change training mode¶
In [ ]:
Copied!
exp.set_training_mode('fine_tuning', unfreeze_last_n=10)
exp.set_training_mode('fine_tuning', unfreeze_last_n=10)
In [ ]:
Copied!
exp.run(5, tfhub_log=True)
exp.run(5, tfhub_log=True)
Saving TensorBoard log files to " C:\Users\bhati\Google Drive\projects\Tensorflow-tutorial-Daniel-Bourke\tensorboard_logs\10_food_multiclass_classification\transfer_learning_experiments\experiment_4\20210531-222113" Epoch 10/14 188/188 [==============================] - 77s 411ms/step - loss: 0.4873 - f1: 0.8399 - accuracy: 0.8460 - val_loss: 0.4706 - val_f1: 0.8548 - val_accuracy: 0.8473 Epoch 11/14 188/188 [==============================] - 74s 392ms/step - loss: 0.4652 - f1: 0.8487 - accuracy: 0.8520 - val_loss: 0.4665 - val_f1: 0.8530 - val_accuracy: 0.8487 Epoch 12/14 188/188 [==============================] - 70s 374ms/step - loss: 0.4757 - f1: 0.8495 - accuracy: 0.8547 - val_loss: 0.4667 - val_f1: 0.8590 - val_accuracy: 0.8533 Epoch 13/14 188/188 [==============================] - 79s 420ms/step - loss: 0.4517 - f1: 0.8533 - accuracy: 0.8550 - val_loss: 0.4653 - val_f1: 0.8589 - val_accuracy: 0.8513 Epoch 14/14 188/188 [==============================] - 84s 447ms/step - loss: 0.4548 - f1: 0.8511 - accuracy: 0.8537 - val_loss: 0.4704 - val_f1: 0.8542 - val_accuracy: 0.8493
In [ ]:
Copied!
exp.plot_learning_curve()
exp.plot_learning_curve()
Out[ ]:
<AxesSubplot:xlabel='epoch', ylabel='loss'>
In [ ]:
Copied!
experiments[exp.name] = exp
experiments[exp.name] = exp
Save the models¶
In [ ]:
Copied!
exp.pretraiened_model
exp.pretraiened_model
In [ ]:
Copied!
experiment_model_names = {exp_name: '_'.join([exp.pretrained_model.name, exp.mode, exp.data_pct]) for exp_name, exp in experiments.items()}
experiment_model_names = {exp_name: '_'.join([exp.pretrained_model.name, exp.mode, exp.data_pct]) for exp_name, exp in experiments.items()}
In [ ]:
Copied!
for exp_name in experiments:
exp = experiments[exp_name]
model_name = experiment_model_names[exp_name]
model = exp.model
tf.keras.models.save_model(model, os.path.abspath(f'../models/10_food_multiclass_classification/{model_name}'))
for exp_name in experiments:
exp = experiments[exp_name]
model_name = experiment_model_names[exp_name]
model = exp.model
tf.keras.models.save_model(model, os.path.abspath(f'../models/10_food_multiclass_classification/{model_name}'))
INFO:tensorflow:Assets written to: C:\Users\bhati\Google Drive\projects\Tensorflow-tutorial-Daniel-Bourke\models\10_food_multiclass_classification\efficientnetb0_feature_extraction_1_percent\assets INFO:tensorflow:Assets written to: C:\Users\bhati\Google Drive\projects\Tensorflow-tutorial-Daniel-Bourke\models\10_food_multiclass_classification\efficientnetb0_feature_extraction_10_percent\assets INFO:tensorflow:Assets written to: C:\Users\bhati\Google Drive\projects\Tensorflow-tutorial-Daniel-Bourke\models\10_food_multiclass_classification\efficientnetb0_fine_tuning_10_percent\assets INFO:tensorflow:Assets written to: C:\Users\bhati\Google Drive\projects\Tensorflow-tutorial-Daniel-Bourke\models\10_food_multiclass_classification\efficientnetb0_fine_tuning_100_percent\assets
Compare Models¶
In [ ]:
Copied!
experiments
experiments
Out[ ]:
{'experiment_1': <__main__.FoodClassificationExperiment at 0x154ffd3cc08>,
'experiment_2': <__main__.FoodClassificationExperiment at 0x15533332c08>,
'experiment_3': <__main__.FoodClassificationExperiment at 0x1550ae6a7c8>,
'experiment_4': <__main__.FoodClassificationExperiment at 0x15503968dc8>}
In [ ]:
Copied!
compdf = []
for exp_name in experiments:
exp = experiments[exp_name]
model = exp.model
model_name = experiment_model_names[exp_name]
y_pred_prob_test = model.predict(test_data)
y_pred_test = y_pred_prob_test.argmax(axis=1)
crdf = pd.DataFrame(skmetrics.classification_report(test_data.labels, y_pred_test, target_names=test_data.class_indices, output_dict=True))
crdf['model'] = model_name
compdf.append(crdf)
compdf = pd.concat(compdf)
compdf.index.name = 'metric'
compdf.reset_index(inplace=True)
compdf = []
for exp_name in experiments:
exp = experiments[exp_name]
model = exp.model
model_name = experiment_model_names[exp_name]
y_pred_prob_test = model.predict(test_data)
y_pred_test = y_pred_prob_test.argmax(axis=1)
crdf = pd.DataFrame(skmetrics.classification_report(test_data.labels, y_pred_test, target_names=test_data.class_indices, output_dict=True))
crdf['model'] = model_name
compdf.append(crdf)
compdf = pd.concat(compdf)
compdf.index.name = 'metric'
compdf.reset_index(inplace=True)
In [ ]:
Copied!
compdf_small = compdf.loc[compdf['metric'] != 'support', ['metric', 'weighted avg', 'model']].rename(columns={'weighted avg': 'value'})
acc_df = compdf.loc[compdf['metric'] != 'support', ['model', 'accuracy']].drop_duplicates().melt(id_vars='model', var_name='metric')
compdf_small = compdf.loc[compdf['metric'] != 'support', ['metric', 'weighted avg', 'model']].rename(columns={'weighted avg': 'value'})
acc_df = compdf.loc[compdf['metric'] != 'support', ['model', 'accuracy']].drop_duplicates().melt(id_vars='model', var_name='metric')
In [ ]:
Copied!
compdf_small = pd.concat([compdf_small, acc_df]).sort_values('model')
compdf_small = pd.concat([compdf_small, acc_df]).sort_values('model')
In [ ]:
Copied!
import seaborn as sns
import seaborn as sns
In [ ]:
Copied!
plt.figure(figsize=(8, 4))
sns.barplot(x='metric', y='value', hue='model', data=compdf_small)
plt.legend(bbox_to_anchor=[1.01, 0.6])
plt.figure(figsize=(8, 4))
sns.barplot(x='metric', y='value', hue='model', data=compdf_small)
plt.legend(bbox_to_anchor=[1.01, 0.6])
Out[ ]:
<matplotlib.legend.Legend at 0x15594bb04c8>
