remain mnist demo

v1_api_demo/mnist/.gitignore

+data/raw_data
+data/*.list
+mnist_vgg_model
+plot.png
+train.log
+*pyc
+.ipynb_checkpoints
+params.pkl
+params.tar
+params.tar.gz

v1_api_demo/mnist/api_train.py

+"""
+A very basic example for how to use current Raw SWIG API to train mnist network.
+
+Current implementation uses Raw SWIG, which means the API call is directly \
+passed to C++ side of Paddle.
+
+The user api could be simpler and carefully designed.
+"""
+import random
+
+import numpy as np
+import paddle.v2 as paddle_v2
+import py_paddle.swig_paddle as api
+from paddle.trainer_config_helpers import *
+from py_paddle import DataProviderConverter
+
+from mnist_util import read_from_mnist
+
+
+def init_parameter(network):
+    assert isinstance(network, api.GradientMachine)
+    for each_param in network.getParameters():
+        assert isinstance(each_param, api.Parameter)
+        array_size = len(each_param)
+        array = np.random.uniform(-1.0, 1.0, array_size).astype('float32')
+        each_param.getBuf(api.PARAMETER_VALUE).copyFromNumpyArray(array)
+
+
+def generator_to_batch(generator, batch_size):
+    ret_val = list()
+    for each_item in generator:
+        ret_val.append(each_item)
+        if len(ret_val) == batch_size:
+            yield ret_val
+            ret_val = list()
+    if len(ret_val) != 0:
+        yield ret_val
+
+
+class BatchPool(object):
+    def __init__(self, generator, batch_size):
+        self.data = list(generator)
+        self.batch_size = batch_size
+
+    def __call__(self):
+        random.shuffle(self.data)
+        for offset in xrange(0, len(self.data), self.batch_size):
+            limit = min(offset + self.batch_size, len(self.data))
+            yield self.data[offset:limit]
+
+
+def input_order_converter(generator):
+    for each_item in generator:
+        yield each_item['pixel'], each_item['label']
+
+
+def main():
+    api.initPaddle("-use_gpu=false", "-trainer_count=4")  # use 4 cpu cores
+
+    optimizer = paddle_v2.optimizer.Adam(
+        learning_rate=1e-4,
+        batch_size=1000,
+        model_average=ModelAverage(average_window=0.5),
+        regularization=L2Regularization(rate=0.5))
+
+    # Create Local Updater. Local means not run in cluster.
+    # For a cluster training, here we can change to createRemoteUpdater
+    # in future.
+    updater = optimizer.create_local_updater()
+    assert isinstance(updater, api.ParameterUpdater)
+
+    # define network
+    images = paddle_v2.layer.data(
+        name='pixel', type=paddle_v2.data_type.dense_vector(784))
+    label = paddle_v2.layer.data(
+        name='label', type=paddle_v2.data_type.integer_value(10))
+    hidden1 = paddle_v2.layer.fc(input=images, size=200)
+    hidden2 = paddle_v2.layer.fc(input=hidden1, size=200)
+    inference = paddle_v2.layer.fc(input=hidden2,
+                                   size=10,
+                                   act=paddle_v2.activation.Softmax())
+    cost = paddle_v2.layer.classification_cost(input=inference, label=label)
+
+    # Create Simple Gradient Machine.
+    model_config = paddle_v2.layer.parse_network(cost)
+    m = api.GradientMachine.createFromConfigProto(model_config,
+                                                  api.CREATE_MODE_NORMAL,
+                                                  optimizer.enable_types())
+
+    # This type check is not useful. Only enable type hint in IDE.
+    # Such as PyCharm
+    assert isinstance(m, api.GradientMachine)
+
+    # Initialize Parameter by numpy.
+    init_parameter(network=m)
+
+    # Initialize ParameterUpdater.
+    updater.init(m)
+
+    # DataProvider Converter is a utility convert Python Object to Paddle C++
+    # Input. The input format is as same as Paddle's DataProvider.
+    converter = DataProviderConverter(input_types=[images.type, label.type])
+
+    train_file = './data/raw_data/train'
+    test_file = './data/raw_data/t10k'
+
+    # start gradient machine.
+    # the gradient machine must be started before invoke forward/backward.
+    # not just for training, but also for inference.
+    m.start()
+
+    # evaluator can print error rate, etc. It is a C++ class.
+    batch_evaluator = m.makeEvaluator()
+    test_evaluator = m.makeEvaluator()
+
+    # Get Train Data.
+    # TrainData will stored in a data pool. Currently implementation is not care
+    # about memory, speed. Just a very naive implementation.
+    train_data_generator = input_order_converter(read_from_mnist(train_file))
+    train_data = BatchPool(train_data_generator, 512)
+
+    # outArgs is Neural Network forward result. Here is not useful, just passed
+    # to gradient_machine.forward
+    outArgs = api.Arguments.createArguments(0)
+
+    for pass_id in xrange(2):  # we train 2 passes.
+        updater.startPass()
+
+        for batch_id, data_batch in enumerate(train_data()):
+            # data_batch is input images.
+            # here, for online learning, we could get data_batch from network.
+
+            # Start update one batch.
+            pass_type = updater.startBatch(len(data_batch))
+
+            # Start BatchEvaluator.
+            # batch_evaluator can be used between start/finish.
+            batch_evaluator.start()
+
+            # forwardBackward is a shortcut for forward and backward.
+            # It is sometimes faster than invoke forward/backward separately,
+            # because in GradientMachine, it may be async.
+            m.forwardBackward(converter(data_batch), outArgs, pass_type)
+
+            for each_param in m.getParameters():
+                updater.update(each_param)
+
+            # Get cost. We use numpy to calculate total cost for this batch.
+            cost_vec = outArgs.getSlotValue(0)
+            cost_vec = cost_vec.copyToNumpyMat()
+            cost = cost_vec.sum() / len(data_batch)
+
+            # Make evaluator works.
+            m.eval(batch_evaluator)
+
+            # Print logs.
+            print 'Pass id', pass_id, 'Batch id', batch_id, 'with cost=', \
+                cost, batch_evaluator
+
+            batch_evaluator.finish()
+            # Finish batch.
+            #  * will clear gradient.
+            #  * ensure all values should be updated.
+            updater.finishBatch(cost)
+
+        # testing stage. use test data set to test current network.
+        updater.apply()
+        test_evaluator.start()
+        test_data_generator = input_order_converter(read_from_mnist(test_file))
+        for data_batch in generator_to_batch(test_data_generator, 512):
+            # in testing stage, only forward is needed.
+            m.forward(converter(data_batch), outArgs, api.PASS_TEST)
+            m.eval(test_evaluator)
+
+        # print error rate for test data set
+        print 'Pass', pass_id, ' test evaluator: ', test_evaluator
+        test_evaluator.finish()
+        updater.restore()
+
+        updater.catchUpWith()
+        params = m.getParameters()
+        for each_param in params:
+            assert isinstance(each_param, api.Parameter)
+            value = each_param.getBuf(api.PARAMETER_VALUE)
+            value = value.copyToNumpyArray()
+
+            # Here, we could save parameter to every where you want
+            print each_param.getName(), value
+
+        updater.finishPass()
+
+    m.finish()
+
+
+if __name__ == '__main__':
+    main()

v1_api_demo/mnist/data/generate_list.py

+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+o = open("./" + "train.list", "w")
+o.write("./data/raw_data/train" + "\n")
+o.close()
+
+o = open("./" + "test.list", "w")
+o.write("./data/raw_data/t10k" + "\n")
+o.close()

v1_api_demo/mnist/data/get_mnist_data.sh

+#!/usr/bin/env sh
+# This scripts downloads the mnist data and unzips it.
+set -e
+DIR="$( cd "$(dirname "$0")" ; pwd -P )"
+rm -rf "$DIR/raw_data"
+mkdir "$DIR/raw_data"
+cd "$DIR/raw_data"
+
+echo "Downloading..."
+
+for fname in train-images-idx3-ubyte train-labels-idx1-ubyte t10k-images-idx3-ubyte t10k-labels-idx1-ubyte
+do
+    if [ ! -e $fname ]; then
+        wget --no-check-certificate http://yann.lecun.com/exdb/mnist/${fname}.gz
+        gunzip ${fname}.gz
+    fi
+done
+
+cd $DIR
+rm -f *.list
+python generate_list.py

v1_api_demo/mnist/light_mnist.py

+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from paddle.trainer_config_helpers import *
+
+is_predict = get_config_arg("is_predict", bool, False)
+
+####################Data Configuration ##################
+
+if not is_predict:
+    data_dir = './data/'
+    define_py_data_sources2(
+        train_list=data_dir + 'train.list',
+        test_list=data_dir + 'test.list',
+        module='mnist_provider',
+        obj='process')
+
+######################Algorithm Configuration #############
+settings(batch_size=50, learning_rate=0.001, learning_method=AdamOptimizer())
+
+#######################Network Configuration #############
+
+data_size = 1 * 28 * 28
+label_size = 10
+img = data_layer(name='pixel', size=data_size)
+
+
+# light cnn
+# A shallower cnn model: [CNN, BN, ReLU, Max-Pooling] x4 + FC x1
+# Easier to train for mnist dataset and quite efficient
+# Final performance is close to deeper ones on tasks such as digital and character classification 
+def light_cnn(input_image, num_channels, num_classes):
+    def __light__(ipt,
+                  num_filter=128,
+                  times=1,
+                  conv_filter_size=3,
+                  dropouts=0,
+                  num_channels_=None):
+        return img_conv_group(
+            input=ipt,
+            num_channels=num_channels_,
+            pool_size=2,
+            pool_stride=2,
+            conv_padding=0,
+            conv_num_filter=[num_filter] * times,
+            conv_filter_size=conv_filter_size,
+            conv_act=ReluActivation(),
+            conv_with_batchnorm=True,
+            conv_batchnorm_drop_rate=dropouts,
+            pool_type=MaxPooling())
+
+    tmp = __light__(input_image, num_filter=128, num_channels_=num_channels)
+    tmp = __light__(tmp, num_filter=128)
+    tmp = __light__(tmp, num_filter=128)
+    tmp = __light__(tmp, num_filter=128, conv_filter_size=1)
+
+    tmp = fc_layer(input=tmp, size=num_classes, act=SoftmaxActivation())
+    return tmp
+
+
+predict = light_cnn(input_image=img, num_channels=1, num_classes=label_size)
+
+if not is_predict:
+    lbl = data_layer(name="label", size=label_size)
+    inputs(img, lbl)
+    outputs(classification_cost(input=predict, label=lbl))
+else:
+    outputs(predict)

v1_api_demo/mnist/mnist_provider.py

+from paddle.trainer.PyDataProvider2 import *
+from mnist_util import read_from_mnist
+
+
+# Define a py data provider
+@provider(
+    input_types={'pixel': dense_vector(28 * 28),
+                 'label': integer_value(10)},
+    cache=CacheType.CACHE_PASS_IN_MEM)
+def process(settings, filename):  # settings is not used currently.
+    for each in read_from_mnist(filename):
+        yield each

v1_api_demo/mnist/mnist_util.py

+import numpy
+
+__all__ = ['read_from_mnist']
+
+
+def read_from_mnist(filename):
+    imgf = filename + "-images-idx3-ubyte"
+    labelf = filename + "-labels-idx1-ubyte"
+    f = open(imgf, "rb")
+    l = open(labelf, "rb")
+
+    f.read(16)
+    l.read(8)
+
+    # Define number of samples for train/test
+    if "train" in filename:
+        n = 60000
+    else:
+        n = 10000
+
+    images = numpy.fromfile(
+        f, 'ubyte', count=n * 28 * 28).reshape((n, 28 * 28)).astype('float32')
+    images = images / 255.0 * 2.0 - 1.0
+    labels = numpy.fromfile(l, 'ubyte', count=n).astype("int")
+
+    for i in xrange(n):
+        yield {"pixel": images[i, :], 'label': labels[i]}
+
+    f.close()
+    l.close()

v1_api_demo/mnist/train.sh

+#!/bin/bash
+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+set -e
+config=vgg_16_mnist.py
+output=./mnist_vgg_model
+log=train.log
+
+paddle train \
+--config=$config \
+--dot_period=10 \
+--log_period=100 \
+--test_all_data_in_one_period=1 \
+--use_gpu=0 \
+--trainer_count=1 \
+--num_passes=100 \
+--save_dir=$output \
+2>&1 | tee $log
+paddle usage -l $log -e $? -n "mnist_train" >/dev/null 2>&1
+
+python -m paddle.utils.plotcurve -i $log > plot.png

v1_api_demo/mnist/vgg_16_mnist.py

+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from paddle.trainer_config_helpers import *
+
+is_predict = get_config_arg("is_predict", bool, False)
+
+####################Data Configuration ##################
+
+if not is_predict:
+    data_dir = './data/'
+    define_py_data_sources2(
+        train_list=data_dir + 'train.list',
+        test_list=data_dir + 'test.list',
+        module='mnist_provider',
+        obj='process')
+
+######################Algorithm Configuration #############
+settings(
+    batch_size=128,
+    learning_rate=0.1 / 128.0,
+    learning_method=MomentumOptimizer(0.9),
+    regularization=L2Regularization(0.0005 * 128))
+
+#######################Network Configuration #############
+
+data_size = 1 * 28 * 28
+label_size = 10
+img = data_layer(name='pixel', size=data_size)
+
+# small_vgg is predined in trainer_config_helpers.network
+predict = small_vgg(input_image=img, num_channels=1, num_classes=label_size)
+
+if not is_predict:
+    lbl = data_layer(name="label", size=label_size)
+    inputs(img, lbl)
+    outputs(classification_cost(input=predict, label=lbl))
+else:
+    outputs(predict)