Tutorial: Implementation of Siamese Network on Caffe, Torch, Tensorflow

  1 ---
  2 title: Siamese Network Tutorial
  3 description: Train and test a siamese network on MNIST data.
  4 category: example
  5 include_in_docs: true
  6 layout: default
  7 priority: 100
  8 ---
  9 
 10 # Siamese Network Training with Caffe
 11 This example shows how you can use weight sharing and a contrastive loss
 12 function to learn a model using a siamese network in Caffe.
 13 
 14 We will assume that you have caffe successfully compiled. If not, please refer
 15 to the [Installation page](../../installation.html). This example builds on the
 16 [MNIST tutorial](mnist.html) so it would be a good idea to read that before
 17 continuing.
 18 
 19 *The guide specifies all paths and assumes all commands are executed from the
 20 root caffe directory*
 21 
 22 ## Prepare Datasets
 23 
 24 You will first need to download and convert the data from the MNIST
 25 website. To do this, simply run the following commands:
 26 
 27     ./data/mnist/get_mnist.sh
 28     ./examples/siamese/create_mnist_siamese.sh
 29 
 30 After running the script there should be two datasets,
 31 `./examples/siamese/mnist_siamese_train_leveldb`, and
 32 `./examples/siamese/mnist_siamese_test_leveldb`.
 33 
 34 ## The Model
 35 First, we will define the model that we want to train using the siamese network.
 36 We will use the convolutional net defined in
 37 `./examples/siamese/mnist_siamese.prototxt`. This model is almost
 38 exactly the same as the [LeNet model](mnist.html), the only difference is that
 39 we have replaced the top layers that produced probabilities over the 10 digit
 40 classes with a linear "feature" layer that produces a 2 dimensional vector.
 41 
 42     layer {
 43       name: "feat"
 44       type: "InnerProduct"
 45       bottom: "ip2"
 46       top: "feat"
 47       param {
 48         name: "feat_w"
 49         lr_mult: 1
 50       }
 51       param {
 52         name: "feat_b"
 53         lr_mult: 2
 54       }
 55       inner_product_param {
 56         num_output: 2
 57       }
 58     }
 59 
 60 ## Define the Siamese Network
 61 
 62 In this section we will define the siamese network used for training. The
 63 resulting network is defined in
 64 `./examples/siamese/mnist_siamese_train_test.prototxt`.
 65 
 66 ### Reading in the Pair Data
 67 
 68 We start with a data layer that reads from the LevelDB database we created
 69 earlier. Each entry in this database contains the image data for a pair of
 70 images (`pair_data`) and a binary label saying if they belong to the same class
 71 or different classes (`sim`).
 72 
 73     layer {
 74       name: "pair_data"
 75       type: "Data"
 76       top: "pair_data"
 77       top: "sim"
 78       include { phase: TRAIN }
 79       transform_param {
 80         scale: 0.00390625
 81       }
 82       data_param {
 83         source: "examples/siamese/mnist_siamese_train_leveldb"
 84         batch_size: 64
 85       }
 86     }
 87 
 88 In order to pack a pair of images into the same blob in the database we pack one
 89 image per channel. We want to be able to work with these two images separately,
 90 so we add a slice layer after the data layer. This takes the `pair_data` and
 91 slices it along the channel dimension so that we have a single image in `data`
 92 and its paired image in `data_p.`
 93 
 94     layer {
 95       name: "slice_pair"
 96       type: "Slice"
 97       bottom: "pair_data"
 98       top: "data"
 99       top: "data_p"
100       slice_param {
101         slice_dim: 1
102         slice_point: 1
103       }
104     }
105 
106 ### Building the First Side of the Siamese Net
107 
108 Now we can specify the first side of the siamese net. This side operates on
109 `data` and produces `feat`. Starting from the net in
110 `./examples/siamese/mnist_siamese.prototxt` we add default weight fillers. Then
111 we name the parameters of the convolutional and inner product layers. Naming the
112 parameters allows Caffe to share the parameters between layers on both sides of
113 the siamese net. In the definition this looks like:
114 
115     ...
116     param { name: "conv1_w" ...  }
117     param { name: "conv1_b" ...  }
118     ...
119     param { name: "conv2_w" ...  }
120     param { name: "conv2_b" ...  }
121     ...
122     param { name: "ip1_w" ...  }
123     param { name: "ip1_b" ...  }
124     ...
125     param { name: "ip2_w" ...  }
126     param { name: "ip2_b" ...  }
127     ...
128 
129 ### Building the Second Side of the Siamese Net
130 
131 Now we need to create the second path that operates on `data_p` and produces
132 `feat_p`. This path is exactly the same as the first. So we can just copy and
133 paste it. Then we change the name of each layer, input, and output by appending
134 `_p` to differentiate the "paired" layers from the originals.
135 
136 ### Adding the Contrastive Loss Function
137 
138 To train the network we will optimize a contrastive loss function proposed in:
139 Raia Hadsell, Sumit Chopra, and Yann LeCun "Dimensionality Reduction by Learning
140 an Invariant Mapping". This loss function encourages matching pairs to be close
141 together in feature space while pushing non-matching pairs apart. This cost
142 function is implemented with the `CONTRASTIVE_LOSS` layer:
143 
144     layer {
145         name: "loss"
146         type: "ContrastiveLoss"
147         contrastive_loss_param {
148             margin: 1.0
149         }
150         bottom: "feat"
151         bottom: "feat_p"
152         bottom: "sim"
153         top: "loss"
154     }
155 
156 ## Define the Solver
157 
158 Nothing special needs to be done to the solver besides pointing it at the
159 correct model file. The solver is defined in
160 `./examples/siamese/mnist_siamese_solver.prototxt`.
161 
162 ## Training and Testing the Model
163 
164 Training the model is simple after you have written the network definition
165 protobuf and solver protobuf files. Simply run
166 `./examples/siamese/train_mnist_siamese.sh`:
167 
168     ./examples/siamese/train_mnist_siamese.sh
169 
170 # Plotting the results
171 
172 First, we can draw the model and siamese networks by running the following
173 commands that draw the DAGs defined in the .prototxt files:
174 
175     ./python/draw_net.py \
176         ./examples/siamese/mnist_siamese.prototxt \
177         ./examples/siamese/mnist_siamese.png
178 
179     ./python/draw_net.py \
180         ./examples/siamese/mnist_siamese_train_test.prototxt \
181         ./examples/siamese/mnist_siamese_train_test.png
182 
183 Second, we can load the learned model and plot the features using the iPython
184 notebook:
185 
186     ipython notebook ./examples/siamese/mnist_siamese.ipynb

  1 # Run on GPU: THEANO_FLAGS=mode=FAST_RUN,device=gpu,floatX=float32 python mnist_siamese_graph.py
  2 from __future__ import print_function
  3  
  4 import sys
  5 import os
  6 import time
  7 import numpy as np
  8 import theano
  9 import theano.tensor as T
 10 import lasagne
 11 import utils 
 12 from progressbar import AnimatedMarker, Bar, BouncingBar, Counter, ETA, \
 13     FileTransferSpeed, FormatLabel, Percentage, \
 14     ProgressBar, ReverseBar, RotatingMarker, \
 15     SimpleProgress, Timer
 16 import matplotlib.pyplot as plt
 17 from matplotlib import gridspec
 18 import cPickle as pickle
 19 import time
 20 from sklearn import metrics
 21 from scipy import interpolate
 22 from lasagne.regularization import regularize_layer_params_weighted, l2, l1
 23 from lasagne.regularization import regularize_layer_params
 24   
 25 NUM_EPOCHS = 40
 26 BATCH_SIZE = 100
 27 LEARNING_RATE = 0.001
 28 MOMENTUM = 0.9
 29  
 30 # def build_cnn(input_var=None):
 31 #     net = lasagne.layers.InputLayer(shape=(None, 1, 64, 64),
 32 #                                         input_var=input_var)
 33 #     cnn1 = lasagne.layers.Conv2DLayer(
 34 #             net, num_filters=96, filter_size=(7, 7),
 35 #             nonlinearity=lasagne.nonlinearities.rectify,
 36 #             W=lasagne.init.GlorotNormal())
 37 #     pool1 = lasagne.layers.MaxPool2DLayer(cnn1, pool_size=(2, 2))
 38 #     cnn2 = lasagne.layers.Conv2DLayer(
 39 #             pool1, num_filters=64, filter_size=(6, 6),
 40 #             nonlinearity=lasagne.nonlinearities.rectify,
 41 #             W=lasagne.init.GlorotNormal())
 42 #     fc1 = lasagne.layers.DenseLayer(cnn2, num_units=128)
 43 #     # network = lasagne.layers.FlattenLayer(fc1)
 44 #     return fc1
 45  
 46 def build_cnn(input_var=None):
 47     net = lasagne.layers.InputLayer(shape=(None, 1, 64, 64),
 48                                         input_var=input_var)
 49     cnn1 = lasagne.layers.Conv2DLayer(
 50             net, num_filters=96, filter_size=(7, 7),
 51             nonlinearity=lasagne.nonlinearities.rectify,
 52             stride = (3,3),
 53             W=lasagne.init.GlorotNormal())
 54     pool1 = lasagne.layers.MaxPool2DLayer(cnn1, pool_size=(2, 2))
 55     cnn2 = lasagne.layers.Conv2DLayer(
 56             pool1, num_filters=192, filter_size=(5, 5),
 57             nonlinearity=lasagne.nonlinearities.rectify,
 58             W=lasagne.init.GlorotNormal())
 59     pool2 = lasagne.layers.MaxPool2DLayer(cnn2, pool_size=(2, 2))
 60     cnn3 = lasagne.layers.Conv2DLayer(
 61             pool2, num_filters=256, filter_size=(3, 3),
 62             nonlinearity=lasagne.nonlinearities.rectify,
 63             W=lasagne.init.GlorotNormal())
 64     # fc1 = lasagne.layers.DenseLayer(cnn2, num_units=128)
 65     network = lasagne.layers.FlattenLayer(cnn3)
 66     return network
 67  
 68 def init_data(train,test):
 69     dtrain = utils.load_brown_dataset("/home/vassilis/Datasets/"+train+"/")
 70     dtest = utils.load_brown_dataset("/home/vassilis/Datasets/"+test+"/")
 71  
 72     dtrain['patches'] = dtrain['patches'].astype('float32')
 73     dtest['patches'] = dtest['patches'].astype('float32')
 74  
 75     dtrain['patches'] /= 255
 76     dtest['patches'] /= 255
 77  
 78     mu = dtrain['patches'].mean()
 79     dtrain['patches'] = dtrain['patches'] - mu
 80     dtest['patches'] = dtest['patches'] - mu
 81     return dtrain,dtest
 82  
 83 def eval_test(net,d):
 84     bs = 100
 85     pb = np.array_split(d['patches'],bs)
 86     descrs = []
 87     for i,minib in enumerate(pb):
 88         dd = lasagne.layers.get_output(net,minib).eval()
 89         descrs.append(dd)
 90  
 91     descrs = np.vstack(descrs)
 92     dists = np.zeros(100000,)
 93     lbls = np.zeros(100000,)
 94      
 95     for i in range(100000):
 96         idx1 = d['testgt'][i][0]
 97         idx2 = d['testgt'][i][1]
 98         lbl = d['testgt'][i][2]
 99         dists[i] = np.linalg.norm(descrs[idx1]-descrs[idx2])
100         lbls[i] = lbl
101         #print(dists[i],lbls[i])
102     fpr, tpr, thresholds = metrics.roc_curve(lbls, -dists, pos_label=1)
103     f = interpolate.interp1d(tpr, fpr)
104     fpr95 = f(0.95)
105     print('fpr95-> '+str(fpr95))
106  
107 def main(num_epochs=NUM_EPOCHS):
108     widgets = ['Mini-batch training: ', Percentage(), ' ', Bar(),
109              ' ', ETA(), ' ']
110     print("> Loading data...")
111     dtrain,dtest = init_data('liberty','notredame')
112     net = build_cnn()
113  
114     dtr = utils.gen_pairs(dtrain,1200000)
115     ntr = dtr.shape[0]
116  
117     X = T.tensor4()
118     y = T.ivector()
119     a = lasagne.layers.get_output(net,X)
120  
121     fx1 = a[1::2, :]
122     fx2 = a[::2, :]
123     d = T.sum(( fx1- fx2)**2, -1)
124  
125     l2_penalty = regularize_layer_params(net, l2) * 1e-3
126  
127     loss = T.mean(y * d +
128                   (1 - y) * T.maximum(0, 1 - d))+l2_penalty
129  
130     all_params = lasagne.layers.get_all_params(net)
131     updates = lasagne.updates.nesterov_momentum(
132         loss, all_params, LEARNING_RATE, MOMENTUM)
133  
134     trainf = theano.function([X, y], loss,updates=updates)    
135  
136     num_batches = ntr // BATCH_SIZE
137     print(num_batches)
138     print("> Done loading data...")
139     print("> Started learning with "+str(num_batches)+" batches")
140  
141     shuf = np.random.permutation(ntr)
142  
143     X_tr = np.zeros((BATCH_SIZE*2,1,64,64)).astype('float32')
144     y_tr = np.zeros(BATCH_SIZE).astype('int32')
145  
146     for epoch in range(NUM_EPOCHS):
147         batch_train_losses = []
148         pbar = ProgressBar(widgets=widgets, maxval=num_batches).start()
149         for k in range(num_batches):
150             sh = shuf[k*BATCH_SIZE:k*BATCH_SIZE+BATCH_SIZE]
151             pbar.update(k)
152             # fill batch here
153             for s in range(0,BATCH_SIZE*2,2):
154                 # idx1 = dtrain['traingt'][sh[s/2],0]
155                 # idx2 = dtrain['traingt'][sh[s/2],1]
156                 # lbl = dtrain['traingt'][sh[s/2],2]
157  
158                 idx1 = dtr[sh[s/2]][0]
159                 idx2 = dtr[sh[s/2]][1]
160                 lbl = dtr[sh[s/2]][2]
161                  
162                 X_tr[s] = dtrain['patches'][idx1]
163                 X_tr[s+1] = dtrain['patches'][idx2]
164                 y_tr[s/2] = lbl
165  
166             batch_train_loss = trainf(X_tr,y_tr)
167             batch_train_losses.append(batch_train_loss)
168         avg_train_loss = np.mean(batch_train_losses)
169         pbar.finish()
170         print("> Epoch " + str(epoch) + ", loss: "+str(avg_train_loss))
171  
172         eval_test(net,dtest)
173  
174         with open('net.pickle', 'wb') as f:
175             pickle.dump(net, f, -1)
176      
177         # netlayers = lasagne.layers.get_all_layers(net)
178         # print(netlayers)
179         # layer = netlayers[1]
180         # print(layer)
181         # print(layer.num_filters)
182         # W = layer.W.get_value()
183         # b = layer.b.get_value()
184         # f = [w + bb for w, bb in zip(W, b)]
185         # gs = gridspec.GridSpec(8, 12)
186         # for i in range(layer.num_filters):
187         #     g = gs[i]
188         #     ax = plt.subplot(g)
189         #     ax.grid()
190         #     ax.set_xticks([])
191         #     ax.set_yticks([])
192         #     ax.imshow(f[i][0])
193         # plt.show()
194          
195  
196 if __name__ == '__main__':
197    main(sys.argv[1])