CNNCreator_CifarClassifierNetwork.py 18.3 KB
Newer Older
1
from caffe2.python import workspace, core, model_helper, brew, optimizer
2 3 4
from caffe2.python.predictor import mobile_exporter
from caffe2.proto import caffe2_pb2
import numpy as np
5
import math
6
import logging
7
import os
8
import sys
9
import lmdb
10 11 12
class CNNCreator_CifarClassifierNetwork:

    module = None
13 14 15
    _current_dir_ = os.path.join('./')
    _data_dir_    = os.path.join(_current_dir_, 'data', 'CifarClassifierNetwork')
    _model_dir_   = os.path.join(_current_dir_, 'model', 'CifarClassifierNetwork')
16

17 18
    INIT_NET    = os.path.join(_model_dir_, 'init_net.pb')
    PREDICT_NET = os.path.join(_model_dir_, 'predict_net.pb')
19

20 21 22 23 24 25 26 27 28 29 30
    def get_total_num_iter(self, num_epoch, batch_size, dataset_size):
        #Force floating point calculation
        batch_size_float = float(batch_size)
        dataset_size_float = float(dataset_size)

        iterations_float = math.ceil(num_epoch*(dataset_size_float/batch_size_float))
        iterations_int = int(iterations_float)

        return iterations_int


31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48
    def add_input(self, model, batch_size, db, db_type, device_opts):
        with core.DeviceScope(device_opts):
            # load the data
            data_uint8, label = brew.db_input(
                model,
                blobs_out=["data_uint8", "label"],
                batch_size=batch_size,
                db=db,
                db_type=db_type,
            )
            # cast the data to float
            data = model.Cast(data_uint8, "data", to=core.DataType.FLOAT)

            # scale data from [0,255] down to [0,1]
            data = model.Scale(data, data, scale=float(1./256))

            # don't need the gradient for the backward pass
            data = model.StopGradient(data, data)
49 50 51 52

            dataset_size = int (lmdb.open(db).stat()['entries'])

            return data, label, dataset_size
53 54 55 56 57 58

    def create_model(self, model, data, device_opts):
    	with core.DeviceScope(device_opts):

    		data = data
    		# data, output shape: {[3,32,32]}
59
      		conv2_1_ = brew.conv(model, data, 'conv2_1_', dim_in=3, dim_out=8, kernel=3, stride=1)
60 61 62 63 64 65 66 67 68 69
    		# conv2_1_, output shape: {[8,32,32]}
    		batchnorm2_1_ = mx.symbol.BatchNorm(data=conv2_1_,
    		    fix_gamma=True,
    		    name="batchnorm2_1_")
    		relu2_1_ = brew.relu(model, batchnorm2_1_, batchnorm2_1_)
      		conv3_1_ = brew.conv(model, relu2_1_, 'conv3_1_', dim_in=8, dim_out=8, kernel=3, stride=1)
    		# conv3_1_, output shape: {[8,32,32]}
    		batchnorm3_1_ = mx.symbol.BatchNorm(data=conv3_1_,
    		    fix_gamma=True,
    		    name="batchnorm3_1_")
70
    		conv2_2_ = brew.conv(model, data, 'conv2_2_', dim_in=3, dim_out=8, kernel=1, stride=1)
71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276
    		# conv2_2_, output shape: {[8,32,32]}
    		batchnorm2_2_ = mx.symbol.BatchNorm(data=conv2_2_,
    		    fix_gamma=True,
    		    name="batchnorm2_2_")
    		add4_ = batchnorm3_1_ + batchnorm2_2_
    		# add4_, output shape: {[8,32,32]}
    		relu4_ = brew.relu(model, add4_, add4_)
      		conv5_1_ = brew.conv(model, relu4_, 'conv5_1_', dim_in=8, dim_out=16, kernel=3, stride=2)
    		# conv5_1_, output shape: {[16,16,16]}
    		batchnorm5_1_ = mx.symbol.BatchNorm(data=conv5_1_,
    		    fix_gamma=True,
    		    name="batchnorm5_1_")
    		relu5_1_ = brew.relu(model, batchnorm5_1_, batchnorm5_1_)
      		conv6_1_ = brew.conv(model, relu5_1_, 'conv6_1_', dim_in=16, dim_out=16, kernel=3, stride=1)
    		# conv6_1_, output shape: {[16,16,16]}
    		batchnorm6_1_ = mx.symbol.BatchNorm(data=conv6_1_,
    		    fix_gamma=True,
    		    name="batchnorm6_1_")
    		conv5_2_ = brew.conv(model, relu4_, 'conv5_2_', dim_in=8, dim_out=16, kernel=1, stride=2)
    		# conv5_2_, output shape: {[16,16,16]}
    		batchnorm5_2_ = mx.symbol.BatchNorm(data=conv5_2_,
    		    fix_gamma=True,
    		    name="batchnorm5_2_")
    		add7_ = batchnorm6_1_ + batchnorm5_2_
    		# add7_, output shape: {[16,16,16]}
    		relu7_ = brew.relu(model, add7_, add7_)
      		conv8_1_ = brew.conv(model, relu7_, 'conv8_1_', dim_in=16, dim_out=16, kernel=3, stride=1)
    		# conv8_1_, output shape: {[16,16,16]}
    		batchnorm8_1_ = mx.symbol.BatchNorm(data=conv8_1_,
    		    fix_gamma=True,
    		    name="batchnorm8_1_")
    		relu8_1_ = brew.relu(model, batchnorm8_1_, batchnorm8_1_)
      		conv9_1_ = brew.conv(model, relu8_1_, 'conv9_1_', dim_in=16, dim_out=16, kernel=3, stride=1)
    		# conv9_1_, output shape: {[16,16,16]}
    		batchnorm9_1_ = mx.symbol.BatchNorm(data=conv9_1_,
    		    fix_gamma=True,
    		    name="batchnorm9_1_")
    		add10_ = batchnorm9_1_ + relu7_
    		# add10_, output shape: {[16,16,16]}
    		relu10_ = brew.relu(model, add10_, add10_)
      		conv11_1_ = brew.conv(model, relu10_, 'conv11_1_', dim_in=16, dim_out=16, kernel=3, stride=1)
    		# conv11_1_, output shape: {[16,16,16]}
    		batchnorm11_1_ = mx.symbol.BatchNorm(data=conv11_1_,
    		    fix_gamma=True,
    		    name="batchnorm11_1_")
    		relu11_1_ = brew.relu(model, batchnorm11_1_, batchnorm11_1_)
      		conv12_1_ = brew.conv(model, relu11_1_, 'conv12_1_', dim_in=16, dim_out=16, kernel=3, stride=1)
    		# conv12_1_, output shape: {[16,16,16]}
    		batchnorm12_1_ = mx.symbol.BatchNorm(data=conv12_1_,
    		    fix_gamma=True,
    		    name="batchnorm12_1_")
    		add13_ = batchnorm12_1_ + relu10_
    		# add13_, output shape: {[16,16,16]}
    		relu13_ = brew.relu(model, add13_, add13_)
      		conv14_1_ = brew.conv(model, relu13_, 'conv14_1_', dim_in=16, dim_out=32, kernel=3, stride=2)
    		# conv14_1_, output shape: {[32,8,8]}
    		batchnorm14_1_ = mx.symbol.BatchNorm(data=conv14_1_,
    		    fix_gamma=True,
    		    name="batchnorm14_1_")
    		relu14_1_ = brew.relu(model, batchnorm14_1_, batchnorm14_1_)
      		conv15_1_ = brew.conv(model, relu14_1_, 'conv15_1_', dim_in=32, dim_out=32, kernel=3, stride=1)
    		# conv15_1_, output shape: {[32,8,8]}
    		batchnorm15_1_ = mx.symbol.BatchNorm(data=conv15_1_,
    		    fix_gamma=True,
    		    name="batchnorm15_1_")
    		conv14_2_ = brew.conv(model, relu13_, 'conv14_2_', dim_in=16, dim_out=32, kernel=1, stride=2)
    		# conv14_2_, output shape: {[32,8,8]}
    		batchnorm14_2_ = mx.symbol.BatchNorm(data=conv14_2_,
    		    fix_gamma=True,
    		    name="batchnorm14_2_")
    		add16_ = batchnorm15_1_ + batchnorm14_2_
    		# add16_, output shape: {[32,8,8]}
    		relu16_ = brew.relu(model, add16_, add16_)
      		conv17_1_ = brew.conv(model, relu16_, 'conv17_1_', dim_in=32, dim_out=32, kernel=3, stride=1)
    		# conv17_1_, output shape: {[32,8,8]}
    		batchnorm17_1_ = mx.symbol.BatchNorm(data=conv17_1_,
    		    fix_gamma=True,
    		    name="batchnorm17_1_")
    		relu17_1_ = brew.relu(model, batchnorm17_1_, batchnorm17_1_)
      		conv18_1_ = brew.conv(model, relu17_1_, 'conv18_1_', dim_in=32, dim_out=32, kernel=3, stride=1)
    		# conv18_1_, output shape: {[32,8,8]}
    		batchnorm18_1_ = mx.symbol.BatchNorm(data=conv18_1_,
    		    fix_gamma=True,
    		    name="batchnorm18_1_")
    		add19_ = batchnorm18_1_ + relu16_
    		# add19_, output shape: {[32,8,8]}
    		relu19_ = brew.relu(model, add19_, add19_)
      		conv20_1_ = brew.conv(model, relu19_, 'conv20_1_', dim_in=32, dim_out=32, kernel=3, stride=1)
    		# conv20_1_, output shape: {[32,8,8]}
    		batchnorm20_1_ = mx.symbol.BatchNorm(data=conv20_1_,
    		    fix_gamma=True,
    		    name="batchnorm20_1_")
    		relu20_1_ = brew.relu(model, batchnorm20_1_, batchnorm20_1_)
      		conv21_1_ = brew.conv(model, relu20_1_, 'conv21_1_', dim_in=32, dim_out=32, kernel=3, stride=1)
    		# conv21_1_, output shape: {[32,8,8]}
    		batchnorm21_1_ = mx.symbol.BatchNorm(data=conv21_1_,
    		    fix_gamma=True,
    		    name="batchnorm21_1_")
    		add22_ = batchnorm21_1_ + relu19_
    		# add22_, output shape: {[32,8,8]}
    		relu22_ = brew.relu(model, add22_, add22_)
      		conv23_1_ = brew.conv(model, relu22_, 'conv23_1_', dim_in=32, dim_out=64, kernel=3, stride=2)
    		# conv23_1_, output shape: {[64,4,4]}
    		batchnorm23_1_ = mx.symbol.BatchNorm(data=conv23_1_,
    		    fix_gamma=True,
    		    name="batchnorm23_1_")
    		relu23_1_ = brew.relu(model, batchnorm23_1_, batchnorm23_1_)
      		conv24_1_ = brew.conv(model, relu23_1_, 'conv24_1_', dim_in=64, dim_out=64, kernel=3, stride=1)
    		# conv24_1_, output shape: {[64,4,4]}
    		batchnorm24_1_ = mx.symbol.BatchNorm(data=conv24_1_,
    		    fix_gamma=True,
    		    name="batchnorm24_1_")
    		conv23_2_ = brew.conv(model, relu22_, 'conv23_2_', dim_in=32, dim_out=64, kernel=1, stride=2)
    		# conv23_2_, output shape: {[64,4,4]}
    		batchnorm23_2_ = mx.symbol.BatchNorm(data=conv23_2_,
    		    fix_gamma=True,
    		    name="batchnorm23_2_")
    		add25_ = batchnorm24_1_ + batchnorm23_2_
    		# add25_, output shape: {[64,4,4]}
    		relu25_ = brew.relu(model, add25_, add25_)
      		conv26_1_ = brew.conv(model, relu25_, 'conv26_1_', dim_in=64, dim_out=64, kernel=3, stride=1)
    		# conv26_1_, output shape: {[64,4,4]}
    		batchnorm26_1_ = mx.symbol.BatchNorm(data=conv26_1_,
    		    fix_gamma=True,
    		    name="batchnorm26_1_")
    		relu26_1_ = brew.relu(model, batchnorm26_1_, batchnorm26_1_)
      		conv27_1_ = brew.conv(model, relu26_1_, 'conv27_1_', dim_in=64, dim_out=64, kernel=3, stride=1)
    		# conv27_1_, output shape: {[64,4,4]}
    		batchnorm27_1_ = mx.symbol.BatchNorm(data=conv27_1_,
    		    fix_gamma=True,
    		    name="batchnorm27_1_")
    		add28_ = batchnorm27_1_ + relu25_
    		# add28_, output shape: {[64,4,4]}
    		relu28_ = brew.relu(model, add28_, add28_)
      		conv29_1_ = brew.conv(model, relu28_, 'conv29_1_', dim_in=64, dim_out=64, kernel=3, stride=1)
    		# conv29_1_, output shape: {[64,4,4]}
    		batchnorm29_1_ = mx.symbol.BatchNorm(data=conv29_1_,
    		    fix_gamma=True,
    		    name="batchnorm29_1_")
    		relu29_1_ = brew.relu(model, batchnorm29_1_, batchnorm29_1_)
      		conv30_1_ = brew.conv(model, relu29_1_, 'conv30_1_', dim_in=64, dim_out=64, kernel=3, stride=1)
    		# conv30_1_, output shape: {[64,4,4]}
    		batchnorm30_1_ = mx.symbol.BatchNorm(data=conv30_1_,
    		    fix_gamma=True,
    		    name="batchnorm30_1_")
    		add31_ = batchnorm30_1_ + relu28_
    		# add31_, output shape: {[64,4,4]}
    		relu31_ = brew.relu(model, add31_, add31_)
    		globalpooling31_ = mx.symbol.Pooling(data=relu31_,
    		    global_pool=True,
    		    kernel=(1,1),
    		    pool_type="avg",
    		    name="globalpooling31_")
    		# globalpooling31_, output shape: {[64,1,1]}
    		fc31_ = brew.fc(model, globalpooling31_, 'fc31_', dim_in=64, dim_out=128)
    		# fc31_, output shape: {[128,1,1]}
    		dropout31_ = mx.symbol.Dropout(data=fc31_,
    		    p=0.5,
    		    name="dropout31_")
    		fc32_ = brew.fc(model, dropout31_, 'fc32_', dim_in=128, dim_out=10)
    		# fc32_, output shape: {[10,1,1]}
    		softmax = brew.softmax(model, fc32_, 'softmax')

    		return softmax

    # this adds the loss and optimizer
    def add_training_operators(self, model, output, label, device_opts, opt_type, base_learning_rate, policy, stepsize, epsilon, beta1, beta2, gamma, momentum) :
    	with core.DeviceScope(device_opts):
    		xent = model.LabelCrossEntropy([output, label], 'xent')
    		loss = model.AveragedLoss(xent, "loss")

    		model.AddGradientOperators([loss])

    		if opt_type == 'adam':
    		    if policy == 'step':
    		        opt = optimizer.build_adam(model, base_learning_rate=base_learning_rate, policy=policy, stepsize=stepsize, beta1=beta1, beta2=beta2, epsilon=epsilon)
    		    elif policy == 'fixed' or policy == 'inv':
    		        opt = optimizer.build_adam(model, base_learning_rate=base_learning_rate, policy=policy, beta1=beta1, beta2=beta2, epsilon=epsilon)
    		    print("adam optimizer selected")
    		elif opt_type == 'sgd':
    		    if policy == 'step':
    		        opt = optimizer.build_sgd(model, base_learning_rate=base_learning_rate, policy=policy, stepsize=stepsize, gamma=gamma, momentum=momentum)
    		    elif policy == 'fixed' or policy == 'inv':
    		        opt = optimizer.build_sgd(model, base_learning_rate=base_learning_rate, policy=policy, gamma=gamma, momentum=momentum)
    		    print("sgd optimizer selected")
    		elif opt_type == 'rmsprop':
    		    if policy == 'step':
    		        opt = optimizer.build_rms_prop(model, base_learning_rate=base_learning_rate, policy=policy, stepsize=stepsize, decay=gamma, momentum=momentum, epsilon=epsilon)
    		    elif policy == 'fixed' or policy == 'inv':
    		        opt = optimizer.build_rms_prop(model, base_learning_rate=base_learning_rate, policy=policy, decay=gamma, momentum=momentum, epsilon=epsilon)
    		    print("rmsprop optimizer selected")
    		elif opt_type == 'adagrad':
    		    if policy == 'step':
    		        opt = optimizer.build_adagrad(model, base_learning_rate=base_learning_rate, policy=policy, stepsize=stepsize, decay=gamma, epsilon=epsilon)
    		    elif policy == 'fixed' or policy == 'inv':
    		        opt = optimizer.build_adagrad(model, base_learning_rate=base_learning_rate, policy=policy, decay=gamma, epsilon=epsilon)
    		    print("adagrad optimizer selected")

    def add_accuracy(self, model, output, label, device_opts, eval_metric):
        with core.DeviceScope(device_opts):
            if eval_metric == 'accuracy':
                accuracy = brew.accuracy(model, [output, label], "accuracy")
            elif eval_metric == 'top_k_accuracy':
                accuracy = brew.accuracy(model, [output, label], "accuracy", top_k=3)
            return accuracy

277 278
    def train(self, num_epoch=1000, batch_size=64, context='gpu', eval_metric='accuracy', opt_type='adam', base_learning_rate=0.001, weight_decay=0.001, policy='fixed', stepsize=1, epsilon=1E-8, beta1=0.9, beta2=0.999, gamma=0.999, momentum=0.9) :
        if context == 'cpu':
279 280
            device_opts = core.DeviceOption(caffe2_pb2.CPU, 0)
            print("CPU mode selected")
281
        elif context == 'gpu':
282 283 284
            device_opts = core.DeviceOption(caffe2_pb2.CUDA, 0)
            print("GPU mode selected")

285
    	workspace.ResetWorkspace(self._model_dir_)
286 287 288 289

    	arg_scope = {"order": "NCHW"}
    	# == Training model ==
    	train_model= model_helper.ModelHelper(name="train_net", arg_scope=arg_scope)
290
    	data, label, train_dataset_size = self.add_input(train_model, batch_size=batch_size, db=os.path.join(self._data_dir_, 'mnist-train-nchw-lmdb'), db_type='lmdb', device_opts=device_opts)
291 292 293 294 295 296 297 298 299 300 301
    	softmax = self.create_model(train_model, data, device_opts=device_opts)
    	self.add_training_operators(train_model, softmax, label, device_opts, opt_type, base_learning_rate, policy, stepsize, epsilon, beta1, beta2, gamma, momentum)
    	self.add_accuracy(train_model, softmax, label, device_opts, eval_metric)
    	with core.DeviceScope(device_opts):
    		brew.add_weight_decay(train_model, weight_decay)

    	# Initialize and create the training network
    	workspace.RunNetOnce(train_model.param_init_net)
    	workspace.CreateNet(train_model.net, overwrite=True)

    	# Main Training Loop
302 303 304
    	iterations = self.get_total_num_iter(num_epoch, batch_size, train_dataset_size)
        print("** Starting Training for " + str(num_epoch) + " epochs = " + str(iterations) + " iterations **")
    	for i in range(iterations):
305
    		workspace.RunNet(train_model.net)
306 307
    		if i % 50 == 0:
    			print 'Iter ' + str(i) + ': ' + 'Loss ' + str(workspace.FetchBlob("loss")) + ' - ' + 'Accuracy ' + str(workspace.FetchBlob('accuracy'))
308 309 310 311 312
    	print("Training done")

    	print("== Running Test model ==")
    	# == Testing model. ==
    	test_model= model_helper.ModelHelper(name="test_net", arg_scope=arg_scope, init_params=False)
313
    	data, label, test_dataset_size = self.add_input(test_model, batch_size=batch_size, db=os.path.join(self._data_dir_, 'mnist-test-nchw-lmdb'), db_type='lmdb', device_opts=device_opts)
314 315 316 317 318 319
    	softmax = self.create_model(test_model, data, device_opts=device_opts)
    	self.add_accuracy(test_model, predictions, label, device_opts, eval_metric)
    	workspace.RunNetOnce(test_model.param_init_net)
    	workspace.CreateNet(test_model.net, overwrite=True)

    	# Main Testing Loop
320 321
    	test_accuracy = np.zeros(test_dataset_size/batch_size)
    	for i in range(test_dataset_size/batch_size):
322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344
    		# Run a forward pass of the net on the current batch
    		workspace.RunNet(test_model.net)
    		# Collect the batch accuracy from the workspace
    		test_accuracy[i] = workspace.FetchBlob('accuracy')

    	print('Test_accuracy: {:.4f}'.format(test_accuracy.mean()))

    	# == Deployment model. ==
    	# We simply need the main AddModel part.
    	deploy_model = model_helper.ModelHelper(name="deploy_net", arg_scope=arg_scope, init_params=False)
    	self.create_model(deploy_model, "data", device_opts)

    	print("Saving deploy model")
    	self.save_net(self.INIT_NET, self.PREDICT_NET, deploy_model)

    def save_net(self, init_net_path, predict_net_path, model):

    	init_net, predict_net = mobile_exporter.Export(
    		workspace,
    		model.net,
    		model.params
    	)

345
        try:
346
            os.makedirs(self._model_dir_)
347
        except OSError:
348
            if not os.path.isdir(self._model_dir_):
349 350
                raise

351
    	print("Save the model to init_net.pb and predict_net.pb")
352
    	with open(predict_net_path, 'wb') as f:
353
    		f.write(model.net._net.SerializeToString())
354
    	with open(init_net_path, 'wb') as f:
355 356 357
    		f.write(init_net.SerializeToString())

    	print("Save the model to init_net.pbtxt and predict_net.pbtxt")
358 359

    	with open(init_net_path.replace('.pb','.pbtxt'), 'w') as f:
360
    		f.write(str(init_net))
361
    	with open(predict_net_path.replace('.pb','.pbtxt'), 'w') as f:
362 363 364 365
    		f.write(str(predict_net))
    	print("== Saved init_net and predict_net ==")

    def load_net(self, init_net_path, predict_net_path, device_opts):
366 367 368 369 370 371 372 373 374
        if not os.path.isfile(init_net_path):
            logging.error("Network loading failure. File '" + os.path.abspath(init_net_path) + "' does not exist.")
            sys.exit(1)
        elif not os.path.isfile(predict_net_path):
            logging.error("Network loading failure. File '" + os.path.abspath(predict_net_path) + "' does not exist.")
            sys.exit(1)

        init_def = caffe2_pb2.NetDef()
    	with open(init_net_path, 'rb') as f:
375 376 377 378 379
    		init_def.ParseFromString(f.read())
    		init_def.device_option.CopyFrom(device_opts)
    		workspace.RunNetOnce(init_def.SerializeToString())

    	net_def = caffe2_pb2.NetDef()
380
    	with open(predict_net_path, 'rb') as f:
381 382 383
    		net_def.ParseFromString(f.read())
    		net_def.device_option.CopyFrom(device_opts)
    		workspace.CreateNet(net_def.SerializeToString(), overwrite=True)
384
    	print("== Loaded init_net and predict_net ==")