├── README.md
└── model_deXpression.py


/README.md:
--------------------------------------------------------------------------------
1 | # DeXpression
2 | TFLearn Implementation of [DeXpression architecture](https://arxiv.org/abs/1509.05371). 
3 | 
4 | Batch normalization is used instead of LRN. 
5 | 
6 | Gives a precision of 99.3 percent, recall of 99.2 percent and f1-score of 99.2 percent on CKPlus Dataset for human emotion recognition from frontal facial images.
7 | 
8 | Please refer to this GitHub [Issue #1](https://github.com/arundasan91/DeXpression/issues/1) for info regarding the dataset creation.
9 | 


--------------------------------------------------------------------------------
/model_deXpression.py:
--------------------------------------------------------------------------------
 1 | import numpy as np
 2 | import tflearn
 3 | import tflearn.activations as activations
 4 | # Data loading and preprocessing
 5 | from tflearn.activations import relu
 6 | from tflearn.data_utils import shuffle, to_categorical
 7 | from tflearn.layers.conv import avg_pool_2d, conv_2d, max_pool_2d
 8 | from tflearn.layers.core import dropout, flatten, fully_connected, input_data
 9 | from tflearn.layers.merge_ops import merge
10 | from tflearn.layers.normalization import local_response_normalization
11 | from tflearn.layers.normalization import batch_normalization
12 | 
13 | # Give a run ID here. Change it to flags (arguments) in version 2.
14 | ID = '4_1'
15 | RUNID = 'DeXpression_run_' + ID
16 | 
17 | # Give a dropout if required (change to True and define the dropout percentage).
18 | dropout = False
19 | dropout_keep_prob=0.5
20 | 
21 | # Load data
22 | X = np.load('CKP_X.npy')
23 | Y = np.load('CKP_Y.npy')
24 | 
25 | # Define number of output classes.
26 | num_classes = 7
27 | 
28 | # Define padding scheme.
29 | padding = 'VALID'
30 | 
31 | # Model Architecture
32 | network = input_data(shape=[None, 224, 224, 1])
33 | conv_1 = relu(conv_2d(network, 64, 7, strides=2, bias=True, padding=padding, activation=None, name='Conv2d_1'))
34 | maxpool_1 = batch_normalization(max_pool_2d(conv_1, 3, strides=2, padding=padding, name='MaxPool_1'))
35 | #LRN_1 = local_response_normalization(maxpool_1, name='LRN_1')
36 | # FeatEX-1
37 | conv_2a = relu(conv_2d(maxpool_1, 96, 1, strides=1, padding=padding, name='Conv_2a_FX1'))
38 | maxpool_2a = max_pool_2d(maxpool_1, 3, strides=1, padding=padding, name='MaxPool_2a_FX1')
39 | conv_2b = relu(conv_2d(conv_2a, 208, 3, strides=1, padding=padding, name='Conv_2b_FX1'))
40 | conv_2c = relu(conv_2d(maxpool_2a, 64, 1, strides=1, padding=padding, name='Conv_2c_FX1'))
41 | FX1_out = merge([conv_2b, conv_2c], mode='concat', axis=3, name='FX1_out')
42 | # FeatEX-2
43 | conv_3a = relu(conv_2d(FX1_out, 96, 1, strides=1, padding=padding, name='Conv_3a_FX2'))
44 | maxpool_3a = max_pool_2d(FX1_out, 3, strides=1, padding=padding, name='MaxPool_3a_FX2')
45 | conv_3b = relu(conv_2d(conv_3a, 208, 3, strides=1, padding=padding, name='Conv_3b_FX2'))
46 | conv_3c = relu(conv_2d(maxpool_3a, 64, 1, strides=1, padding=padding, name='Conv_3c_FX2'))
47 | FX2_out = merge([conv_3b, conv_3c], mode='concat', axis=3, name='FX2_out')
48 | net = flatten(FX2_out)
49 | if dropout:
50 |     net = dropout(net, dropout_keep_prob)
51 | loss = fully_connected(net, num_classes,activation='softmax')
52 | 
53 | # Compile the model and define the hyperparameters
54 | network = tflearn.regression(loss, optimizer='Adam',
55 |                      loss='categorical_crossentropy',
56 |                      learning_rate=0.0001)
57 | 
58 | # Final definition of model checkpoints and other configurations
59 | model = tflearn.DNN(network, checkpoint_path='/home/cc/DeXpression/DeXpression_checkpoints',
60 |                     max_checkpoints=1, tensorboard_verbose=2, tensorboard_dir="./tflearn_logs/")
61 | 
62 | # Fit the model, train for 20 epochs. (Change all parameters to flags (arguments) on version 2.)
63 | model.fit(X, Y, n_epoch=20, validation_set=0.1, shuffle=True, show_metric=True, batch_size=350, snapshot_step=2000,snapshot_epoch=True, run_id=RUNID)
64 | 
65 | # Save the model
66 | model.save('./DeXpression_checkpoints/' + RUNID + '.model')
67 | 
68 | # Load the model if required, later.
69 | #model.load('./DeXpression_checkpoints/' + RUNID + '.model')
70 | 


--------------------------------------------------------------------------------