diff --git a/GTSRB/03-Tracking-and-visualizing.ipynb b/GTSRB/03-Tracking-and-visualizing.ipynb
index 2aa572ff0c0cae4f81d24bffb8f899e5300d1536..c211ca36f6ccbe6843c878fd7b4572f537dc8bd8 100644
--- a/GTSRB/03-Tracking-and-visualizing.ipynb
+++ b/GTSRB/03-Tracking-and-visualizing.ipynb
@@ -165,10 +165,8 @@
]
},
{
- "cell_type": "code",
- "execution_count": null,
+ "cell_type": "raw",
"metadata": {},
- "outputs": [],
"source": [
"%%bash\n",
"# To clean old logs and saved model, run this cell\n",
diff --git a/GTSRB/README.ipynb b/GTSRB/README.ipynb
index b5fba4ba692e6222ee88b61e49deafddd7dc7093..163fe70440ce929dd5fa2fbcb7f9e3ae99e20f70 100644
--- a/GTSRB/README.ipynb
+++ b/GTSRB/README.ipynb
@@ -18,24 +18,6 @@
"Traffic sign classification with **CNN**, using Tensorflow and **Keras** \n",
"\n",
"\n",
- "Prerequisite\n",
- "------------\n",
- "\n",
- "Environment, with the following packages :\n",
- " - Python 3.6\n",
- " - numpy\n",
- " - Tensorflow 2.0\n",
- " - scikit-image\n",
- " - scikit-learn\n",
- " - Matplotlib\n",
- " - seaborn\n",
- " \n",
- "You can create it from the `environment.yml` file :\n",
- "```\n",
- "# conda env create -f environment.yml\n",
- "```\n",
- "To manage conda environment see [there](https://docs.conda.io/projects/conda/en/latest/user-guide/tasks/manage-environments.html#) \n",
- "\n",
"About the dataset\n",
"-----------------\n",
"\n",
diff --git a/README.md b/README.md
index 146861fcd8998f71db32de2aece738cb039453e3..eb7f172abaae82952ec4f9532a2544bbb2038dda 100644
--- a/README.md
+++ b/README.md
@@ -1,16 +1,30 @@
-German Traffic Sign Recognition Benchmark (GTSRB)
-=================================================
+FIDLE - Formation Introduction au Deep Learning
+===============================================
---
-Introduction au Deep Learning (IDLE) - S. Arias, E. Maldonado, JL. Parouty - CNRS/SARI/DEVLOG - 2020
+S. Arias, E. Maldonado, JL. Parouty - CNRS/SARI/DEVLOG - 2020
## 1/ Environment
-To install your conda environment :
+To run this examples, you need an environment with the following packages :
+ - Python 3.6
+ - numpy
+ - Tensorflow 2.0
+ - scikit-image
+ - scikit-learn
+ - Matplotlib
+ - seaborn
+ - pyplot
+
+You can install such a predefined environment :
```
conda env create -f environment.yml
```
+To manage conda environment see [there](https://docs.conda.io/projects/conda/en/latest/user-guide/tasks/manage-environments.html#)
+
+
+
## 4/ Misc
To update an existing environment :
```
diff --git a/VAE/01-VAE with MNIST.ipynb b/VAE/01-VAE with MNIST.ipynb
index 4a2a0dba8571dc9c3f28e029ef06f5bdad5b1e63..5c0c2b9dc4d1a5701459445e2b8ed00134d1c9f5 100644
--- a/VAE/01-VAE with MNIST.ipynb
+++ b/VAE/01-VAE with MNIST.ipynb
@@ -4,8 +4,8 @@
"cell_type": "markdown",
"metadata": {},
"source": [
- "Text Embedding - IMDB dataset\n",
- "=============================\n",
+ "Variational AutoEncoder\n",
+ "=======================\n",
"---\n",
"Formation Introduction au Deep Learning (FIDLE) - S. Arias, E. Maldonado, JL. Parouty - CNRS/SARI/DEVLOG - 2020 \n",
"\n",
@@ -22,31 +22,18 @@
},
{
"cell_type": "code",
- "execution_count": 1,
+ "execution_count": null,
"metadata": {},
- "outputs": [
- {
- "name": "stdout",
- "output_type": "stream",
- "text": [
- "IDLE 2020 - Practical Work Module\n",
- " Version : 0.2.4\n",
- " Run time : Sunday 2 February 2020, 19:30:36\n",
- " Matplotlib style : ../fidle/talk.mplstyle\n",
- " TensorFlow version : 2.0.0\n",
- " Keras version : 2.2.4-tf\n"
- ]
- }
- ],
+ "outputs": [],
"source": [
"import numpy as np\n",
"\n",
"import tensorflow as tf\n",
"import tensorflow.keras as keras\n",
- "import tensorflow.keras.datasets.imdb as imdb\n",
+ "import tensorflow.keras.datasets.mnist as mnist\n",
"\n",
- "import models.VAE\n",
- "from models.VAE import VariationalAutoencoder\n",
+ "import modules.vae\n",
+ "# from modules.vae import VariationalAutoencoder\n",
"\n",
"import matplotlib.pyplot as plt\n",
"import matplotlib\n",
@@ -56,23 +43,57 @@
"\n",
"from importlib import reload\n",
"\n",
- "\n",
- "\n",
"sys.path.append('..')\n",
"import fidle.pwk as ooo\n",
"\n",
+ "reload(ooo)\n",
"ooo.init()"
]
},
+ {
+ "cell_type": "markdown",
+ "metadata": {},
+ "source": [
+ "## Step 2 - Get data"
+ ]
+ },
{
"cell_type": "code",
- "execution_count": 20,
+ "execution_count": null,
"metadata": {},
"outputs": [],
"source": [
- "reload(models.VAE)\n",
+ "(x_train, y_train), (x_test, y_test) = mnist.load_data()\n",
+ "\n",
+ "x_train = x_train.astype('float32') / 255.\n",
+ "x_train = np.expand_dims(x_train, axis=3)\n",
+ "x_test = x_test.astype('float32') / 255.\n",
+ "x_test = np.expand_dims(x_test, axis=3)\n",
+ "print(x_train.shape)\n",
+ "print(x_test.shape)"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {},
+ "source": [
+ "## Step 3 - Get VAE model"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "reload(modules.vae)\n",
+ "reload(modules.callbacks)\n",
+ "\n",
+ "tag = '000'\n",
+ "\n",
"input_shape = (28,28,1)\n",
"z_dim = 2\n",
+ "verbose = 0\n",
"\n",
"encoder= [ {'type':'Conv2D', 'filters':32, 'kernel_size':(3,3), 'strides':1, 'padding':'same', 'activation':'relu'},\n",
" {'type':'Conv2D', 'filters':64, 'kernel_size':(3,3), 'strides':2, 'padding':'same', 'activation':'relu'},\n",
@@ -86,112 +107,77 @@
" {'type':'Conv2DT', 'filters':1, 'kernel_size':(3,3), 'strides':1, 'padding':'same', 'activation':'sigmoid'}\n",
" ]\n",
"\n",
- "vae = models.VAE.VariationalAutoencoder(input_shape, encoder, decoder, z_dim)\n"
+ "vae = modules.vae.VariationalAutoencoder(input_shape = input_shape, \n",
+ " encoder_layers = encoder, \n",
+ " decoder_layers = decoder,\n",
+ " z_dim = z_dim, \n",
+ " verbose = verbose,\n",
+ " run_tag = tag)"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {},
+ "source": [
+ "## Step 4 - Compile it"
]
},
{
"cell_type": "code",
- "execution_count": 21,
+ "execution_count": null,
"metadata": {},
- "outputs": [
- {
- "name": "stdout",
- "output_type": "stream",
- "text": [
- "Model: \"model_29\"\n",
- "__________________________________________________________________________________________________\n",
- "Layer (type) Output Shape Param # Connected to \n",
- "==================================================================================================\n",
- "encoder_input (InputLayer) [(None, 28, 28, 1)] 0 \n",
- "__________________________________________________________________________________________________\n",
- "Layer_1 (Conv2D) (None, 28, 28, 32) 320 encoder_input[0][0] \n",
- "__________________________________________________________________________________________________\n",
- "Layer_2 (Conv2D) (None, 14, 14, 64) 18496 Layer_1[0][0] \n",
- "__________________________________________________________________________________________________\n",
- "Layer_3 (Conv2D) (None, 7, 7, 64) 36928 Layer_2[0][0] \n",
- "__________________________________________________________________________________________________\n",
- "Layer_4 (Conv2D) (None, 7, 7, 64) 36928 Layer_3[0][0] \n",
- "__________________________________________________________________________________________________\n",
- "flatten_7 (Flatten) (None, 3136) 0 Layer_4[0][0] \n",
- "__________________________________________________________________________________________________\n",
- "mu (Dense) (None, 2) 6274 flatten_7[0][0] \n",
- "__________________________________________________________________________________________________\n",
- "log_var (Dense) (None, 2) 6274 flatten_7[0][0] \n",
- "__________________________________________________________________________________________________\n",
- "encoder_output (Lambda) (None, 2) 0 mu[0][0] \n",
- " log_var[0][0] \n",
- "==================================================================================================\n",
- "Total params: 105,220\n",
- "Trainable params: 105,220\n",
- "Non-trainable params: 0\n",
- "__________________________________________________________________________________________________\n"
- ]
- }
- ],
+ "outputs": [],
"source": [
- "vae.encoder.summary()"
+ "learning_rate = 0.0005\n",
+ "r_loss_factor = 1000\n",
+ "\n",
+ "vae.compile(learning_rate, r_loss_factor)"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {},
+ "source": [
+ "## Step 5 - Train"
]
},
{
"cell_type": "code",
- "execution_count": 22,
+ "execution_count": null,
"metadata": {},
- "outputs": [
- {
- "name": "stdout",
- "output_type": "stream",
- "text": [
- "Model: \"model_30\"\n",
- "_________________________________________________________________\n",
- "Layer (type) Output Shape Param # \n",
- "=================================================================\n",
- "decoder_input (InputLayer) [(None, 2)] 0 \n",
- "_________________________________________________________________\n",
- "dense_7 (Dense) (None, 3136) 9408 \n",
- "_________________________________________________________________\n",
- "reshape_7 (Reshape) (None, 7, 7, 64) 0 \n",
- "_________________________________________________________________\n",
- "Layer_1 (Conv2DTranspose) (None, 7, 7, 64) 36928 \n",
- "_________________________________________________________________\n",
- "Layer_2 (Conv2DTranspose) (None, 14, 14, 64) 36928 \n",
- "_________________________________________________________________\n",
- "Layer_3 (Conv2DTranspose) (None, 28, 28, 32) 18464 \n",
- "_________________________________________________________________\n",
- "Layer_4 (Conv2DTranspose) (None, 28, 28, 1) 289 \n",
- "=================================================================\n",
- "Total params: 102,017\n",
- "Trainable params: 102,017\n",
- "Non-trainable params: 0\n",
- "_________________________________________________________________\n"
- ]
- }
- ],
+ "outputs": [],
"source": [
- "vae.decoder.summary()"
+ "batch_size = 100\n",
+ "epochs = 200\n",
+ "batch_periodicity = 1000\n",
+ "initial_epoch = 0\n",
+ "dataset_size = 0.1"
]
},
{
"cell_type": "code",
- "execution_count": 12,
+ "execution_count": null,
"metadata": {},
- "outputs": [
- {
- "data": {
- "text/plain": [
- "{'b': 2}"
- ]
- },
- "execution_count": 12,
- "metadata": {},
- "output_type": "execute_result"
- }
- ],
+ "outputs": [],
"source": [
- "a={'a':1, 'b':2}\n",
- "del a['a']\n",
- "a"
+ "vae.train(x_train,\n",
+ " x_test,\n",
+ " batch_size = batch_size, \n",
+ " epochs = epochs,\n",
+ " batch_periodicity = batch_periodicity,\n",
+ " initial_epoch = initial_epoch,\n",
+ " dataset_size = dataset_size,\n",
+ " lr_decay = 1\n",
+ " )"
]
},
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "metadata": {},
+ "outputs": [],
+ "source": []
+ },
{
"cell_type": "code",
"execution_count": null,
diff --git a/VAE/modules/callbacks.py b/VAE/modules/callbacks.py
new file mode 100644
index 0000000000000000000000000000000000000000..a7bbe1a178d4f98616a9213dc3794207e813fad3
--- /dev/null
+++ b/VAE/modules/callbacks.py
@@ -0,0 +1,32 @@
+from tensorflow.keras.callbacks import Callback
+import numpy as np
+import matplotlib.pyplot as plt
+import os
+
+class ImagesCallback(Callback):
+
+ def __init__(self, initial_epoch=0, batch_periodicity=1000, vae=None):
+ self.epoch = initial_epoch
+ self.batch_periodicity = batch_periodicity
+ self.vae = vae
+ self.images_dir = vae.run_directory+'/images'
+
+ def on_train_batch_end(self, batch, logs={}):
+
+ if batch % self.batch_periodicity == 0:
+ # ---- Get a random latent point
+ z_new = np.random.normal(size = (1,self.vae.z_dim))
+ # ---- Predict an image
+ image = self.vae.decoder.predict(np.array(z_new))[0]
+ # ---- Squeeze it if monochrome : (lx,ly,1) -> (lx,ly)
+ image = image.squeeze()
+ # ---- Save it
+ filename=f'{self.images_dir}/img_{self.epoch:05d}_{batch:06d}.jpg'
+ if len(image.shape) == 2:
+ plt.imsave(filename, image, cmap='gray_r')
+ else:
+ plt.imsave(filename, image)
+
+ def on_epoch_begin(self, epoch, logs={}):
+ self.epoch += 1
+
diff --git a/VAE/models/VAE.py b/VAE/modules/vae.py
similarity index 51%
rename from VAE/models/VAE.py
rename to VAE/modules/vae.py
index 380c3469b0978078132315eff05e6e3a52276b0a..8f09c47cc91cd039774f9222bc5c2bec67b87446 100644
--- a/VAE/models/VAE.py
+++ b/VAE/modules/vae.py
@@ -3,24 +3,36 @@ import numpy as np
import tensorflow as tf
import tensorflow.keras as keras
from tensorflow.keras import backend as K
-from tensorflow.keras.layers import Input, Conv2D, Flatten, Dense, Conv2DTranspose, Reshape, Lambda, Activation, BatchNormalization, LeakyReLU, Dropout
+from tensorflow.keras.layers import Input, Conv2D, Flatten, Dense, Conv2DTranspose, Reshape, Lambda
+from tensorflow.keras.layers import Activation, BatchNormalization, LeakyReLU, Dropout
from tensorflow.keras.models import Model
+from tensorflow.keras.callbacks import ModelCheckpoint
+from tensorflow.keras.optimizers import Adam
+from tensorflow.keras.utils import plot_model
import tensorflow.keras.datasets.imdb as imdb
+import modules.callbacks
+import os
class VariationalAutoencoder():
- def __init__(self, input_shape, encoder_layers, decoder_layers, z_dim):
+ def __init__(self, input_shape=None, encoder_layers=None, decoder_layers=None, z_dim=None, run_tag='default', verbose=0):
self.name = 'Variational AutoEncoder'
self.input_shape = input_shape
self.encoder_layers = encoder_layers
self.decoder_layers = decoder_layers
self.z_dim = z_dim
+ self.verbose = verbose
+ self.run_directory = f'./run/{run_tag}'
+
+ # ---- Create run directories
+ for d in ('','/models','/figs','/logs','/images'):
+ os.makedirs(self.run_directory+d, mode=0o750, exist_ok=True)
# ==== Encoder ================================================================
@@ -44,7 +56,7 @@ class VariationalAutoencoder():
shape_before_flattening = K.int_shape(x)[1:]
x = Flatten()(x)
- # ---- mu / log_var
+ # ---- mu <-> log_var
self.mu = Dense(self.z_dim, name='mu')(x)
self.log_var = Dense(self.z_dim, name='log_var')(x)
@@ -92,6 +104,19 @@ class VariationalAutoencoder():
self.model = Model(model_input, model_output)
+ # ==== Verbosity ==============================================================
+
+ print('Model initialized.')
+ print('Outputs will be in : ',self.run_directory)
+
+ if verbose>0 :
+ print('\n','-'*10,'Encoder','-'*50,'\n')
+ self.encoder.summary()
+ print('\n','-'*10,'Encoder','-'*50,'\n')
+ self.decoder.summary()
+ self.plot_model()
+
+
def compile(self, learning_rate, r_loss_factor):
self.learning_rate = learning_rate
@@ -111,4 +136,50 @@ class VariationalAutoencoder():
return r_loss + kl_loss
optimizer = Adam(lr=learning_rate)
- self.model.compile(optimizer=optimizer, loss = vae_loss, metrics = [vae_r_loss, vae_kl_loss], experimental_run_tf_function=False)
+ self.model.compile(optimizer=optimizer,
+ loss = vae_loss,
+ metrics = [vae_r_loss, vae_kl_loss],
+ experimental_run_tf_function=False)
+
+
+ def train(self,
+ x_train,x_test,
+ batch_size=32, epochs=200,
+ batch_periodicity=100,
+ initial_epoch=0,
+ dataset_size=1,
+ lr_decay=1):
+
+ # ---- Dataset size
+ n_train = int(x_train.shape[0] * dataset_size)
+ n_test = int(x_test.shape[0] * dataset_size)
+
+ # ---- Callbacks
+ images_callback = modules.callbacks.ImagesCallback(initial_epoch, batch_periodicity, self)
+
+# lr_sched = step_decay_schedule(initial_lr=self.learning_rate, decay_factor=lr_decay, step_size=1)
+
+ filename1 = self.run_directory+"/models/model-{epoch:03d}-{loss:.2f}.h5"
+ batch_per_epoch = int(len(x_train)/batch_size)
+ checkpoint1 = ModelCheckpoint(filename1, save_freq=batch_per_epoch*5,verbose=0)
+
+ filename2 = self.run_directory+"/models/best_model.h5"
+ checkpoint2 = ModelCheckpoint(filename2, save_best_only=True, mode='min',monitor='val_loss',verbose=0)
+
+ callbacks_list = [checkpoint1, checkpoint2, images_callback]
+
+ self.model.fit(x_train[:n_train], x_train[:n_train],
+ batch_size = batch_size,
+ shuffle = True,
+ epochs = epochs,
+ initial_epoch = initial_epoch,
+ callbacks = callbacks_list,
+ validation_data = (x_test[:n_test], x_test[:n_test])
+ )
+
+
+ def plot_model(self):
+ d=self.run_directory+'/figs'
+ plot_model(self.model, to_file=f'{d}/model.png', show_shapes = True, show_layer_names = True, expand_nested=True)
+ plot_model(self.encoder, to_file=f'{d}/encoder.png', show_shapes = True, show_layer_names = True)
+ plot_model(self.decoder, to_file=f'{d}/decoder.png', show_shapes = True, show_layer_names = True)
diff --git a/fidle/pwk.py b/fidle/pwk.py
index fd2939d33d14ecf0fbb1fa3a73854388be62b46f..13cc6dfdca25dc3d8b89e43a65859f23f1289468 100644
--- a/fidle/pwk.py
+++ b/fidle/pwk.py
@@ -31,7 +31,7 @@ import seaborn as sn
from IPython.display import display, Markdown
-VERSION='0.2.4'
+VERSION='0.2.5'
# -------------------------------------------------------------
@@ -351,4 +351,7 @@ def plot_donut(values, labels, colors=["lightsteelblue","coral"], figsize=(6,6),
# Equal aspect ratio ensures that pie is drawn as a circle
plt.axis('equal')
plt.tight_layout()
- plt.show()
\ No newline at end of file
+ plt.show()
+
+def display_md(md_text):
+ display(Markdown(md_text))
\ No newline at end of file