CNNSupervisedTrainer.ftl 7.7 KB
Newer Older
Nicola Gatto's avatar
Nicola Gatto committed
1
2
3
4
5
6
7
8
import mxnet as mx
import logging
import numpy as np
import time
import os
import shutil
from mxnet import gluon, autograd, nd

Eyüp Harputlu's avatar
Eyüp Harputlu committed
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
class CrossEntropyLoss(gluon.loss.Loss):
    def __init__(self, axis=-1, sparse_label=True, weight=None, batch_axis=0, **kwargs):
        super(CrossEntropyLoss, self).__init__(weight, batch_axis, **kwargs)
        self._axis = axis
        self._sparse_label = sparse_label

    def hybrid_forward(self, F, pred, label, sample_weight=None):
        pred = F.log(pred)
        if self._sparse_label:
            loss = -F.pick(pred, label, axis=self._axis, keepdims=True)
        else:
            label = gluon.loss._reshape_like(F, label, pred)
            loss = -F.sum(pred * label, axis=self._axis, keepdims=True)
        loss = gluon.loss._apply_weighting(F, loss, self._weight, sample_weight)
        return F.mean(loss, axis=self._batch_axis, exclude=True)

Nicola Gatto's avatar
Nicola Gatto committed
25
26
27
28
29
30
31
32
33
class CNNSupervisedTrainer(object):
    def __init__(self, data_loader, net_constructor, net=None):
        self._data_loader = data_loader
        self._net_creator = net_constructor
        self._net = net

    def train(self, batch_size=64,
              num_epoch=10,
              eval_metric='acc',
Eyüp Harputlu's avatar
Eyüp Harputlu committed
34
35
              loss ='softmax_cross_entropy',
              loss_params={},
Nicola Gatto's avatar
Nicola Gatto committed
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
              optimizer='adam',
              optimizer_params=(('learning_rate', 0.001),),
              load_checkpoint=True,
              context='gpu',
              checkpoint_period=5,
              normalize=True):
        if context == 'gpu':
            mx_context = mx.gpu()
        elif context == 'cpu':
            mx_context = mx.cpu()
        else:
            logging.error("Context argument is '" + context + "'. Only 'cpu' and 'gpu are valid arguments'.")

        if 'weight_decay' in optimizer_params:
            optimizer_params['wd'] = optimizer_params['weight_decay']
            del optimizer_params['weight_decay']
        if 'learning_rate_decay' in optimizer_params:
            min_learning_rate = 1e-08
            if 'learning_rate_minimum' in optimizer_params:
                min_learning_rate = optimizer_params['learning_rate_minimum']
                del optimizer_params['learning_rate_minimum']
            optimizer_params['lr_scheduler'] = mx.lr_scheduler.FactorScheduler(
                                                   optimizer_params['step_size'],
                                                   factor=optimizer_params['learning_rate_decay'],
                                                   stop_factor_lr=min_learning_rate)
            del optimizer_params['step_size']
            del optimizer_params['learning_rate_decay']


        train_iter, test_iter, data_mean, data_std = self._data_loader.load_data(batch_size)
        if self._net is None:
            if normalize:
                self._net_creator.construct(
                    context=mx_context, data_mean=nd.array(data_mean), data_std=nd.array(data_std))
            else:
                self._net_creator.construct(context=mx_context)

        begin_epoch = 0
        if load_checkpoint:
            begin_epoch = self._net_creator.load(mx_context)
        else:
            if os.path.isdir(self._net_creator._model_dir_):
                shutil.rmtree(self._net_creator._model_dir_)

        self._net = self._net_creator.net

        try:
            os.makedirs(self._net_creator._model_dir_)
        except OSError:
            if not os.path.isdir(self._net_creator._model_dir_):
                raise

        trainer = mx.gluon.Trainer(self._net.collect_params(), optimizer, optimizer_params)

Eyüp Harputlu's avatar
Eyüp Harputlu committed
90
91
92
93
94
95
        margin = loss_params['margin'] if 'margin' in loss_params else 1.0
        sparseLabel = loss_params['sparse_label'] if 'sparse_label' in loss_params else True
        if loss == 'softmax_cross_entropy':
            fromLogits = loss_params['from_logits'] if 'from_logits' in loss_params else False
            loss_function = mx.gluon.loss.SoftmaxCrossEntropyLoss(from_logits=fromLogits, sparse_label=sparseLabel)
        elif loss == 'sigmoid_binary_cross_entropy':
Nicola Gatto's avatar
Nicola Gatto committed
96
            loss_function = mx.gluon.loss.SigmoidBinaryCrossEntropyLoss()
Eyüp Harputlu's avatar
Eyüp Harputlu committed
97
98
99
        elif loss == 'cross_entropy':
            loss_function = CrossEntropyLoss(sparse_label=sparseLabel)
        elif loss == 'l2':
Nicola Gatto's avatar
Nicola Gatto committed
100
            loss_function = mx.gluon.loss.L2Loss()
Eyüp Harputlu's avatar
Eyüp Harputlu committed
101
        elif loss == 'l1':
Nicola Gatto's avatar
Nicola Gatto committed
102
            loss_function = mx.gluon.loss.L2Loss()
Eyüp Harputlu's avatar
Eyüp Harputlu committed
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
        elif loss == 'huber':
            rho = loss_params['rho'] if 'rho' in loss_params else 1
            loss_function = mx.gluon.loss.HuberLoss(rho=rho)
        elif loss == 'hinge':
            loss_function = mx.gluon.loss.HingeLoss(margin=margin)
        elif loss == 'squared_hinge':
            loss_function = mx.gluon.loss.SquaredHingeLoss(margin=margin)
        elif loss == 'logistic':
            labelFormat = loss_params['label_format'] if 'label_format' in loss_params else 'signed'
            loss_function = mx.gluon.loss.LogisticLoss(label_format=labelFormat)
        elif loss == 'kullback_leibler':
            fromLogits = loss_params['from_logits'] if 'from_logits' in loss_params else True
            loss_function = mx.gluon.loss.KLDivLoss(from_logits=fromLogits)
        else:
            logging.error("Invalid loss parameter.")
Nicola Gatto's avatar
Nicola Gatto committed
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

        speed_period = 50
        tic = None

        for epoch in range(begin_epoch, begin_epoch + num_epoch):
            train_iter.reset()
            for batch_i, batch in enumerate(train_iter):
                data = batch.data[0].as_in_context(mx_context)
                label = batch.label[0].as_in_context(mx_context)
                with autograd.record():
                    output = self._net(data)
                    loss = loss_function(output, label)

                loss.backward()
                trainer.step(batch_size)

                if tic is None:
                    tic = time.time()
                else:
                    if batch_i % speed_period == 0:
                        try:
                            speed = speed_period * batch_size / (time.time() - tic)
                        except ZeroDivisionError:
                            speed = float("inf")

                        logging.info("Epoch[%d] Batch[%d] Speed: %.2f samples/sec" % (epoch, batch_i, speed))

                        tic = time.time()

            tic = None

            train_iter.reset()
            metric = mx.metric.create(eval_metric)
            for batch_i, batch in enumerate(train_iter):
                data = batch.data[0].as_in_context(mx_context)
                label = batch.label[0].as_in_context(mx_context)
                output = self._net(data)
                predictions = mx.nd.argmax(output, axis=1)
                metric.update(preds=predictions, labels=label)
            train_metric_score = metric.get()[1]

            test_iter.reset()
            metric = mx.metric.create(eval_metric)
            for batch_i, batch in enumerate(test_iter):
                data = batch.data[0].as_in_context(mx_context)
                label = batch.label[0].as_in_context(mx_context)
                output = self._net(data)
                predictions = mx.nd.argmax(output, axis=1)
                metric.update(preds=predictions, labels=label)
            test_metric_score = metric.get()[1]

            logging.info("Epoch[%d] Train: %f, Test: %f" % (epoch, train_metric_score, test_metric_score))

            if (epoch - begin_epoch) % checkpoint_period == 0:
                self._net.save_parameters(self.parameter_path() + '-' + str(epoch).zfill(4) + '.params')

        self._net.save_parameters(self.parameter_path() + '-' + str(num_epoch + begin_epoch).zfill(4) + '.params')
        self._net.export(self.parameter_path() + '_newest', epoch=0)

    def parameter_path(self):
        return self._net_creator._model_dir_ + self._net_creator._model_prefix_