4.5 - Transfer learning#
!wget --no-cache -O init.py -q https://raw.githubusercontent.com/rramosp/2020.deeplearning/master/init.py
from init import init; init(force_download=False)
replicating local resources
import tensorflow as tf
from time import time
import pandas as pd
import matplotlib.pyplot as plt
import numpy as np
from local.lib import mlutils
import pickle, bz2
%matplotlib inline
%load_ext tensorboard
Observe AlexNet filters for the first two convolutional layers#
Pretrained Network: https://www.cs.toronto.edu/~guerzhoy/tf_alexnet/
Paper: ImageNet Classification with Deep Convolutional Neural Networks
Source images: Large Scale Visual Recognition Challenge 2012 ImageNet
Local file with weights from first two layers only
!ls -lash local/data/alexnet_conv1_conv2.pklz
1.4M -rw-r--r-- 1 root root 1.4M Sep 3 22:51 local/data/alexnet_conv1_conv2.pklz
alex_c12 = pickle.load(bz2.BZ2File("local/data/alexnet_conv1_conv2.pklz", "r"), encoding="bytes")
for k,(w,b) in alex_c12.items():
print(k, b.shape, w.shape)
b'conv2' (256,) (5, 5, 48, 256)
b'conv1' (96,) (11, 11, 3, 96)
we use for the moment only the first layer
alex_w1 = alex_c12[b"conv1"][0]
print(alex_w1.shape)
mlutils.display_imgs(alex_w1)
(11, 11, 3, 96)
We will use them as starting point for training our network#
!wget -nc https://s3.amazonaws.com/rlx/mini_cifar.h5
--2025-09-03 22:51:35-- https://s3.amazonaws.com/rlx/mini_cifar.h5
Resolving s3.amazonaws.com (s3.amazonaws.com)... 52.217.117.16, 3.5.12.130, 3.5.8.74, ...
Connecting to s3.amazonaws.com (s3.amazonaws.com)|52.217.117.16|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 14803609 (14M) [binary/octet-stream]
Saving to: ‘mini_cifar.h5’
mini_cifar.h5 100%[===================>] 14.12M 18.9MB/s in 0.7s
2025-09-03 22:51:36 (18.9 MB/s) - ‘mini_cifar.h5’ saved [14803609/14803609]
import h5py
with h5py.File('mini_cifar.h5','r') as h5f:
x_cifar = h5f["x"][:]
y_cifar = h5f["y"][:]
mlutils.show_labeled_image_mosaic(x_cifar, y_cifar)
from sklearn.model_selection import train_test_split
x_train, x_test, y_train, y_test = train_test_split(x_cifar, y_cifar, test_size=.25)
print(x_train.shape, y_train.shape, x_test.shape, y_test.shape)
print("\ndistribution of train classes")
print(pd.Series(y_train).value_counts())
print("\ndistribution of test classes")
print(pd.Series(y_test).value_counts())
(2253, 32, 32, 3) (2253,) (751, 32, 32, 3) (751,)
distribution of train classes
2 773
0 745
1 735
Name: count, dtype: int64
distribution of test classes
0 260
2 252
1 239
Name: count, dtype: int64
def get_model(num_classes, img_size=32, compile=True):
print("using",num_classes,"classes")
inputs = tf.keras.Input(shape=(img_size,img_size,3), name="input_1")
layers = tf.keras.layers.Conv2D(96,(11,11), activation="relu")(inputs)
layers = tf.keras.layers.MaxPool2D((2,2))(layers)
layers = tf.keras.layers.Conv2D(60,(11,11), activation="relu")(layers)
layers = tf.keras.layers.Flatten()(layers)
layers = tf.keras.layers.Dropout(0.4)(layers)
layers = tf.keras.layers.Dense(16, activation=tf.nn.relu)(layers)
layers = tf.keras.layers.Dropout(0.4)(layers)
predictions = tf.keras.layers.Dense(num_classes, activation=tf.nn.softmax, name="output_1")(layers)
model = tf.keras.Model(inputs = inputs, outputs=predictions)
if compile:
model.compile(optimizer='adam',
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
return model
def train(model, batch_size, epochs, model_name=""):
tensorboard = tf.keras.callbacks.TensorBoard(log_dir="logs/"+model_name+"_"+"{}".format(time()))
model.fit(x_train, y_train, epochs=epochs, callbacks=[tensorboard],
batch_size=batch_size,
validation_data=(x_test, y_test))
metrics = model.evaluate(x_test, y_test)
return {k:v for k,v in zip (model.metrics_names, metrics)}
num_classes = len(np.unique(y_cifar))
model = get_model(num_classes)
model.summary()
using 3 classes
Model: "functional_1"
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓ ┃ Layer (type) ┃ Output Shape ┃ Param # ┃ ┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩ │ input_1 (InputLayer) │ (None, 32, 32, 3) │ 0 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ conv2d_2 (Conv2D) │ (None, 22, 22, 96) │ 34,944 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ max_pooling2d_1 (MaxPooling2D) │ (None, 11, 11, 96) │ 0 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ conv2d_3 (Conv2D) │ (None, 1, 1, 60) │ 697,020 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ flatten_1 (Flatten) │ (None, 60) │ 0 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ dropout_2 (Dropout) │ (None, 60) │ 0 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ dense_1 (Dense) │ (None, 16) │ 976 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ dropout_3 (Dropout) │ (None, 16) │ 0 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ output_1 (Dense) │ (None, 3) │ 51 │ └─────────────────────────────────┴────────────────────────┴───────────────┘
Total params: 732,991 (2.80 MB)
Trainable params: 732,991 (2.80 MB)
Non-trainable params: 0 (0.00 B)
Train from scratch#
in some runs might take a while to start reducing loss
train(model, batch_size=16, epochs=15, model_name="alexnet_fintuned")
Epoch 1/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 13s 66ms/step - accuracy: 0.3740 - loss: 1.1126 - val_accuracy: 0.5193 - val_loss: 0.9987
Epoch 2/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 9s 63ms/step - accuracy: 0.4912 - loss: 1.0197 - val_accuracy: 0.5300 - val_loss: 0.9636
Epoch 3/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 8s 55ms/step - accuracy: 0.5245 - loss: 0.9775 - val_accuracy: 0.5606 - val_loss: 0.9048
Epoch 4/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 10s 68ms/step - accuracy: 0.5524 - loss: 0.9097 - val_accuracy: 0.5979 - val_loss: 0.8650
Epoch 5/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 10s 64ms/step - accuracy: 0.5700 - loss: 0.9040 - val_accuracy: 0.5992 - val_loss: 0.8447
Epoch 6/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 11s 67ms/step - accuracy: 0.6119 - loss: 0.8549 - val_accuracy: 0.6232 - val_loss: 0.8197
Epoch 7/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 10s 69ms/step - accuracy: 0.5971 - loss: 0.8328 - val_accuracy: 0.6099 - val_loss: 0.8470
Epoch 8/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 9s 59ms/step - accuracy: 0.6235 - loss: 0.8068 - val_accuracy: 0.6431 - val_loss: 0.8111
Epoch 9/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 10s 71ms/step - accuracy: 0.6290 - loss: 0.7944 - val_accuracy: 0.6192 - val_loss: 0.8518
Epoch 10/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 10s 67ms/step - accuracy: 0.6314 - loss: 0.7915 - val_accuracy: 0.6218 - val_loss: 0.8088
Epoch 11/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 8s 58ms/step - accuracy: 0.6417 - loss: 0.7811 - val_accuracy: 0.5979 - val_loss: 0.8322
Epoch 12/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 11s 61ms/step - accuracy: 0.6726 - loss: 0.7235 - val_accuracy: 0.6218 - val_loss: 0.8021
Epoch 13/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 11s 67ms/step - accuracy: 0.6703 - loss: 0.7187 - val_accuracy: 0.6152 - val_loss: 0.8122
Epoch 14/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 10s 67ms/step - accuracy: 0.6770 - loss: 0.7149 - val_accuracy: 0.6471 - val_loss: 0.8125
Epoch 15/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 9s 67ms/step - accuracy: 0.6825 - loss: 0.6860 - val_accuracy: 0.6232 - val_loss: 0.8849
24/24 ━━━━━━━━━━━━━━━━━━━━ 1s 25ms/step - accuracy: 0.6121 - loss: 0.9606
{'loss': 0.8849406838417053, 'compile_metrics': 0.6231691241264343}
test_preds = model.predict(x_test).argmax(axis=1)
mlutils.plot_confusion_matrix(y_test, test_preds, classes=np.r_[0,1,2], normalize=True)
24/24 ━━━━━━━━━━━━━━━━━━━━ 1s 26ms/step
Normalized confusion matrix
[[0.55769231 0.2 0.24230769]
[0.0292887 0.92050209 0.05020921]
[0.26190476 0.32936508 0.40873016]]
<Axes: title={'center': 'Normalized confusion matrix'}, xlabel='Predicted label', ylabel='True label'>
Transfer learning!!!#
Keep first level filters to compare later
model = get_model(num_classes)
w = model.get_weights()
for i in w:
print(i.shape)
using 3 classes
(11, 11, 3, 96)
(96,)
(11, 11, 96, 60)
(60,)
(60, 16)
(16,)
(16, 3)
(3,)
w[0] = alex_w1
model.set_weights(w)
mlutils.display_imgs(model.get_weights()[0])
train(model, batch_size=16, epochs=15, model_name="alexnet_fintuned")
Epoch 1/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 11s 62ms/step - accuracy: 0.4418 - loss: 1.1226 - val_accuracy: 0.6085 - val_loss: 0.9157
Epoch 2/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 10s 59ms/step - accuracy: 0.5626 - loss: 0.9337 - val_accuracy: 0.6378 - val_loss: 0.7739
Epoch 3/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 11s 65ms/step - accuracy: 0.6051 - loss: 0.8411 - val_accuracy: 0.6405 - val_loss: 0.7577
Epoch 4/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 11s 71ms/step - accuracy: 0.6368 - loss: 0.8208 - val_accuracy: 0.6858 - val_loss: 0.7430
Epoch 5/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 9s 66ms/step - accuracy: 0.6822 - loss: 0.7065 - val_accuracy: 0.7044 - val_loss: 0.7052
Epoch 6/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 8s 56ms/step - accuracy: 0.7372 - loss: 0.6665 - val_accuracy: 0.7230 - val_loss: 0.6823
Epoch 7/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 12s 67ms/step - accuracy: 0.7453 - loss: 0.6464 - val_accuracy: 0.7230 - val_loss: 0.6871
Epoch 8/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 10s 70ms/step - accuracy: 0.7457 - loss: 0.6077 - val_accuracy: 0.7257 - val_loss: 0.6831
Epoch 9/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 10s 70ms/step - accuracy: 0.7700 - loss: 0.5462 - val_accuracy: 0.7284 - val_loss: 0.6776
Epoch 10/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 8s 55ms/step - accuracy: 0.7877 - loss: 0.5261 - val_accuracy: 0.7656 - val_loss: 0.6554
Epoch 11/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 11s 59ms/step - accuracy: 0.7920 - loss: 0.4987 - val_accuracy: 0.7510 - val_loss: 0.6628
Epoch 12/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 10s 70ms/step - accuracy: 0.8247 - loss: 0.4640 - val_accuracy: 0.7150 - val_loss: 0.7567
Epoch 13/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 9s 66ms/step - accuracy: 0.8227 - loss: 0.4630 - val_accuracy: 0.7443 - val_loss: 0.6733
Epoch 14/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 10s 65ms/step - accuracy: 0.8522 - loss: 0.4137 - val_accuracy: 0.7510 - val_loss: 0.6674
Epoch 15/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 9s 57ms/step - accuracy: 0.8489 - loss: 0.3744 - val_accuracy: 0.7297 - val_loss: 0.7401
24/24 ━━━━━━━━━━━━━━━━━━━━ 1s 41ms/step - accuracy: 0.7210 - loss: 0.7579
{'loss': 0.7401493787765503, 'compile_metrics': 0.7296937704086304}
test_preds = model.predict(x_test).argmax(axis=1)
mlutils.plot_confusion_matrix(y_test, test_preds, classes=np.r_[0,1,2], normalize=True)
24/24 ━━━━━━━━━━━━━━━━━━━━ 1s 26ms/step
Normalized confusion matrix
[[0.63076923 0.06153846 0.30769231]
[0.06694561 0.74476987 0.18828452]
[0.09920635 0.08333333 0.81746032]]
<Axes: title={'center': 'Normalized confusion matrix'}, xlabel='Predicted label', ylabel='True label'>
first level filters changed but not much
w0 = model.get_weights()[0]
np.mean(np.abs(w0-alex_w1))
np.float32(0.02194051)
mlutils.display_imgs(w[0])
model.save("alexnet_finetuned_minicifar.keras")
observe that you can also save separately the architecture
with open('alexnet_finetuned_minicifar.architecture.json', 'w') as f:
f.write(model.to_json())
!ls -lash alexnet_finetuned_minicifar*
8.0K -rw-r--r-- 1 root root 7.2K Sep 3 22:59 alexnet_finetuned_minicifar.architecture.json
8.5M -rw-r--r-- 1 root root 8.5M Sep 3 22:59 alexnet_finetuned_minicifar.keras
!head alexnet_finetuned_minicifar.architecture.json
{"module": "keras.src.models.functional", "class_name": "Functional", "config": {"name": "functional_2", "trainable": true, "layers": [{"module": "keras.layers", "class_name": "InputLayer", "config": {"batch_shape": [null, 32, 32, 3], "dtype": "float32", "sparse": false, "ragged": false, "name": "input_1"}, "registered_name": null, "name": "input_1", "inbound_nodes": []}, {"module": "keras.layers", "class_name": "Conv2D", "config": {"name": "conv2d_4", "trainable": true, "dtype": {"module": "keras", "class_name": "DTypePolicy", "config": {"name": "float32"}, "registered_name": null}, "filters": 96, "kernel_size": [11, 11], "strides": [1, 1], "padding": "valid", "data_format": "channels_last", "dilation_rate": [1, 1], "groups": 1, "activation": "relu", "use_bias": true, "kernel_initializer": {"module": "keras.initializers", "class_name": "GlorotUniform", "config": {"seed": null}, "registered_name": null}, "bias_initializer": {"module": "keras.initializers", "class_name": "Zeros", "config": {}, "registered_name": null}, "kernel_regularizer": null, "bias_regularizer": null, "activity_regularizer": null, "kernel_constraint": null, "bias_constraint": null}, "registered_name": null, "build_config": {"input_shape": [null, 32, 32, 3]}, "name": "conv2d_4", "inbound_nodes": [{"args": [{"class_name": "__keras_tensor__", "config": {"shape": [null, 32, 32, 3], "dtype": "float32", "keras_history": ["input_1", 0, 0]}}], "kwargs": {}}]}, {"module": "keras.layers", "class_name": "MaxPooling2D", "config": {"name": "max_pooling2d_2", "trainable": true, "dtype": {"module": "keras", "class_name": "DTypePolicy", "config": {"name": "float32"}, "registered_name": null}, "pool_size": [2, 2], "padding": "valid", "strides": [2, 2], "data_format": "channels_last"}, "registered_name": null, "name": "max_pooling2d_2", "inbound_nodes": [{"args": [{"class_name": "__keras_tensor__", "config": {"shape": [null, 22, 22, 96], "dtype": "float32", "keras_history": ["conv2d_4", 0, 0]}}], "kwargs": {}}]}, {"module": "keras.layers", "class_name": "Conv2D", "config": {"name": "conv2d_5", "trainable": true, "dtype": {"module": "keras", "class_name": "DTypePolicy", "config": {"name": "float32"}, "registered_name": null}, "filters": 60, "kernel_size": [11, 11], "strides": [1, 1], "padding": "valid", "data_format": "channels_last", "dilation_rate": [1, 1], "groups": 1, "activation": "relu", "use_bias": true, "kernel_initializer": {"module": "keras.initializers", "class_name": "GlorotUniform", "config": {"seed": null}, "registered_name": null}, "bias_initializer": {"module": "keras.initializers", "class_name": "Zeros", "config": {}, "registered_name": null}, "kernel_regularizer": null, "bias_regularizer": null, "activity_regularizer": null, "kernel_constraint": null, "bias_constraint": null}, "registered_name": null, "build_config": {"input_shape": [null, 11, 11, 96]}, "name": "conv2d_5", "inbound_nodes": [{"args": [{"class_name": "__keras_tensor__", "config": {"shape": [null, 11, 11, 96], "dtype": "float32", "keras_history": ["max_pooling2d_2", 0, 0]}}], "kwargs": {}}]}, {"module": "keras.layers", "class_name": "Flatten", "config": {"name": "flatten_2", "trainable": true, "dtype": {"module": "keras", "class_name": "DTypePolicy", "config": {"name": "float32"}, "registered_name": null}, "data_format": "channels_last"}, "registered_name": null, "build_config": {"input_shape": [null, 1, 1, 60]}, "name": "flatten_2", "inbound_nodes": [{"args": [{"class_name": "__keras_tensor__", "config": {"shape": [null, 1, 1, 60], "dtype": "float32", "keras_history": ["conv2d_5", 0, 0]}}], "kwargs": {}}]}, {"module": "keras.layers", "class_name": "Dropout", "config": {"name": "dropout_4", "trainable": true, "dtype": {"module": "keras", "class_name": "DTypePolicy", "config": {"name": "float32"}, "registered_name": null}, "rate": 0.4, "seed": null, "noise_shape": null}, "registered_name": null, "name": "dropout_4", "inbound_nodes": [{"args": [{"class_name": "__keras_tensor__", "config": {"shape": [null, 60], "dtype": "float32", "keras_history": ["flatten_2", 0, 0]}}], "kwargs": {"training": false}}]}, {"module": "keras.layers", "class_name": "Dense", "config": {"name": "dense_2", "trainable": true, "dtype": {"module": "keras", "class_name": "DTypePolicy", "config": {"name": "float32"}, "registered_name": null}, "units": 16, "activation": "relu", "use_bias": true, "kernel_initializer": {"module": "keras.initializers", "class_name": "GlorotUniform", "config": {"seed": null}, "registered_name": null}, "bias_initializer": {"module": "keras.initializers", "class_name": "Zeros", "config": {}, "registered_name": null}, "kernel_regularizer": null, "bias_regularizer": null, "kernel_constraint": null, "bias_constraint": null}, "registered_name": null, "build_config": {"input_shape": [null, 60]}, "name": "dense_2", "inbound_nodes": [{"args": [{"class_name": "__keras_tensor__", "config": {"shape": [null, 60], "dtype": "float32", "keras_history": ["dropout_4", 0, 0]}}], "kwargs": {}}]}, {"module": "keras.layers", "class_name": "Dropout", "config": {"name": "dropout_5", "trainable": true, "dtype": {"module": "keras", "class_name": "DTypePolicy", "config": {"name": "float32"}, "registered_name": null}, "rate": 0.4, "seed": null, "noise_shape": null}, "registered_name": null, "name": "dropout_5", "inbound_nodes": [{"args": [{"class_name": "__keras_tensor__", "config": {"shape": [null, 16], "dtype": "float32", "keras_history": ["dense_2", 0, 0]}}], "kwargs": {"training": false}}]}, {"module": "keras.layers", "class_name": "Dense", "config": {"name": "output_1", "trainable": true, "dtype": {"module": "keras", "class_name": "DTypePolicy", "config": {"name": "float32"}, "registered_name": null}, "units": 3, "activation": {"module": "builtins", "class_name": "function", "config": "softmax_v2", "registered_name": "function"}, "use_bias": true, "kernel_initializer": {"module": "keras.initializers", "class_name": "GlorotUniform", "config": {"seed": null}, "registered_name": null}, "bias_initializer": {"module": "keras.initializers", "class_name": "Zeros", "config": {}, "registered_name": null}, "kernel_regularizer": null, "bias_regularizer": null, "kernel_constraint": null, "bias_constraint": null}, "registered_name": null, "build_config": {"input_shape": [null, 16]}, "name": "output_1", "inbound_nodes": [{"args": [{"class_name": "__keras_tensor__", "config": {"shape": [null, 16], "dtype": "float32", "keras_history": ["dropout_5", 0, 0]}}], "kwargs": {}}]}], "input_layers": [["input_1", 0, 0]], "output_layers": [["output_1", 0, 0]]}, "registered_name": "Functional", "build_config": {"input_shape": null}, "compile_config": {"optimizer": {"module": "keras.optimizers", "class_name": "Adam", "config": {"name": "adam", "learning_rate": 0.0010000000474974513, "weight_decay": null, "clipnorm": null, "global_clipnorm": null, "clipvalue": null, "use_ema": false, "ema_momentum": 0.99, "ema_overwrite_frequency": null, "loss_scale_factor": null, "gradient_accumulation_steps": null, "beta_1": 0.9, "beta_2": 0.999, "epsilon": 1e-07, "amsgrad": false}, "registered_name": null}, "loss": "sparse_categorical_crossentropy", "loss_weights": null, "metrics": ["accuracy"], "weighted_metrics": null, "run_eagerly": false, "steps_per_execution": 1, "jit_compile": false}}
Could also freeze weights to avoid being changed by training#
def get_modelT(num_classes, img_size=32, compile=True):
print("using",num_classes,"classes")
inputs = tf.keras.Input(shape=(img_size,img_size,3), name="input_1")
layers = tf.keras.layers.Conv2D(96,(11,11), activation="relu", trainable=False)(inputs)
layers = tf.keras.layers.MaxPool2D((2,2))(layers)
layers = tf.keras.layers.Conv2D(60,(11,11), activation="relu")(layers)
layers = tf.keras.layers.Flatten()(layers)
layers = tf.keras.layers.Dropout(0.4)(layers)
layers = tf.keras.layers.Dense(16, activation=tf.nn.relu)(layers)
layers = tf.keras.layers.Dropout(0.4)(layers)
predictions = tf.keras.layers.Dense(num_classes, activation=tf.nn.softmax, name="output_1")(layers)
model = tf.keras.Model(inputs = inputs, outputs=predictions)
if compile:
model.compile(optimizer='adam',
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
return model
model = get_modelT(num_classes)
w = model.get_weights()
w[0] = alex_w1
model.set_weights(w)
using 3 classes
train(model, batch_size=16, epochs=15, model_name="alexnet_fintuned")
Epoch 1/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 6s 33ms/step - accuracy: 0.4030 - loss: 1.2848 - val_accuracy: 0.5686 - val_loss: 0.8850
Epoch 2/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 5s 34ms/step - accuracy: 0.5342 - loss: 0.9216 - val_accuracy: 0.6192 - val_loss: 0.8577
Epoch 3/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 4s 27ms/step - accuracy: 0.5781 - loss: 0.9034 - val_accuracy: 0.7017 - val_loss: 0.7563
Epoch 4/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 4s 31ms/step - accuracy: 0.6045 - loss: 0.8554 - val_accuracy: 0.7084 - val_loss: 0.7891
Epoch 5/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 5s 35ms/step - accuracy: 0.6349 - loss: 0.8040 - val_accuracy: 0.7443 - val_loss: 0.6905
Epoch 6/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 5s 32ms/step - accuracy: 0.6771 - loss: 0.7610 - val_accuracy: 0.7563 - val_loss: 0.6842
Epoch 7/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 6s 35ms/step - accuracy: 0.6954 - loss: 0.7413 - val_accuracy: 0.7523 - val_loss: 0.6848
Epoch 8/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 4s 27ms/step - accuracy: 0.6853 - loss: 0.7366 - val_accuracy: 0.7310 - val_loss: 0.6791
Epoch 9/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 4s 31ms/step - accuracy: 0.6737 - loss: 0.7366 - val_accuracy: 0.7443 - val_loss: 0.6842
Epoch 10/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 6s 39ms/step - accuracy: 0.7206 - loss: 0.6676 - val_accuracy: 0.7803 - val_loss: 0.6347
Epoch 11/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 4s 31ms/step - accuracy: 0.7241 - loss: 0.6732 - val_accuracy: 0.7830 - val_loss: 0.6340
Epoch 12/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 4s 27ms/step - accuracy: 0.7170 - loss: 0.6976 - val_accuracy: 0.7563 - val_loss: 0.6471
Epoch 13/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 6s 39ms/step - accuracy: 0.6831 - loss: 0.7246 - val_accuracy: 0.7417 - val_loss: 0.7406
Epoch 14/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 4s 31ms/step - accuracy: 0.7192 - loss: 0.6706 - val_accuracy: 0.7563 - val_loss: 0.6628
Epoch 15/15
141/141 ━━━━━━━━━━━━━━━━━━━━ 5s 32ms/step - accuracy: 0.7179 - loss: 0.6655 - val_accuracy: 0.7843 - val_loss: 0.6524
24/24 ━━━━━━━━━━━━━━━━━━━━ 1s 28ms/step - accuracy: 0.7709 - loss: 0.7269
{'loss': 0.6523975729942322, 'compile_metrics': 0.7842876315116882}
test_preds = model.predict(x_test).argmax(axis=1)
mlutils.plot_confusion_matrix(y_test, test_preds, classes=np.r_[0,1,2], normalize=True)
24/24 ━━━━━━━━━━━━━━━━━━━━ 1s 27ms/step
Normalized confusion matrix
[[0.77307692 0.04230769 0.18461538]
[0.10041841 0.76987448 0.12970711]
[0.12698413 0.06349206 0.80952381]]
<Axes: title={'center': 'Normalized confusion matrix'}, xlabel='Predicted label', ylabel='True label'>
check filters did not change
w0 = model.get_weights()[0]
np.mean(np.abs(w0-alex_w1))
np.float32(0.0)