示例#1
0
文件: trainer.py 项目: wqren/striate
    def train(self):
        MiniBatchTrainer.train(self)

        for i, cate in enumerate(self.num_caterange_list):
            self.set_category_range(cate)
            self.curr_batch = self.curr_epoch = 0
            self.num_batch = self.num_batch_list[i]

            model = self.checkpoint_dumper.get_checkpoint()
            layers = model['model_state']['layers']

            fc = layers[-2]
            fc['weight'] = None
            fc['bias'] = None
            fc['weightIncr'] = None
            fc['biasIncr'] = None
            #for l in layers:
            #  if l['type'] == 'fc':
            #    l['weight'] = None
            #    l['bias'] = None
            #    l['weightIncr'] = None
            #    l['biasIncr'] = None

            #fc = layers[-2]
            fc['outputSize'] = cate

            self.learning_rate = self.learning_rate_list[i]
            self.net = FastNet(self.learning_rate,
                               self.image_shape,
                               init_model=model)

            self.net.clear_weight_incr()
            MiniBatchTrainer.train(self)
示例#2
0
  def __init__(self, test_id, data_dir, provider, checkpoint_dir, train_range, test_range, test_freq,
      save_freq, batch_size, num_epoch, image_size, image_color, learning_rate,  params):

    self.origin_test_range = test_range
    if len(test_range) != 1:
      test_range = [test_range[0]]
    AutoStopTrainer.__init__(self, test_id, data_dir, provider, checkpoint_dir, train_range, test_range, test_freq,
        save_freq, batch_size, num_epoch, image_size, image_color, learning_rate, False)

    self.conv_params = []
    self.fc_params = []
    self.softmax_param = None

    self.params = params

    conv = True
    for ld in self.params:
      if ld['type'] in ['conv', 'rnorm', 'pool', 'neuron'] and conv:
        self.conv_params.append(ld)
      elif ld['type'] == 'fc' or (not conv and ld['type'] == 'neuron'):
        self.fc_params.append(ld)
        conv = False
      else:
        self.softmax_param = ld

    self.conv_stack = FastNet.split_conv_to_stack(self.conv_params)
    self.fc_stack = FastNet.split_fc_to_stack(self.fc_params)

    pprint.pprint(self.conv_stack)
    pprint.pprint(self.fc_stack)

    self.fakefc_param = self.fc_stack[-1][0]
示例#3
0
文件: trainer.py 项目: wqren/striate
    def train(self):
        MiniBatchTrainer.train(self)

        for i, group in enumerate(self.num_group_list):
            self.set_num_group(group)
            self.curr_batch = self.curr_epoch = 0
            self.num_batch = self.num_batch_list[i]

            model = self.checkpoint_dumper.get_checkpoint()
            layers = model['model_state']['layers']

            fc = layers[-2]
            fc['outputSize'] = group
            fc['weight'] = None
            fc['bias'] = None
            fc['weightIncr'] = None
            fc['biasIncr'] = None

            self.learning_rate = self.learning_rate_list[i]
            self.net = FastNet(self.learning_rate,
                               self.image_shape,
                               init_model=model)

            self.net.clear_weight_incr()
            MiniBatchTrainer.train(self)
示例#4
0
class ImageNetCatewisedTrainer(MiniBatchTrainer):
  def __init__(self, test_id, data_dir, data_provider, checkpoint_dir, train_range, test_range,
      test_freq, save_freq, batch_size, num_minibatch, image_size, image_color, learning_rate,
      init_model, num_caterange_list, adjust_freq = 100, factor = 1.0):
    # no meaning
    assert len(num_caterange_list) == len(num_minibatch) and num_caterange_list[-1] == 1000

    self.init_output = num_caterange_list[0]
    self.range_list = num_caterange_list[1:]
    self.train_minibatch_list  = num_minibatch[1:]

    fc = init_model[-2]
    fc['outputSize'] = self.init_output

    self.learning_rate = learning_rate[0]
    self.learning_rate_list = learning_rate[1:]

    MiniBatchTrainer.__init__(self, test_id, data_dir, data_provider, checkpoint_dir, train_range,
        test_range, test_freq, save_freq, batch_size, num_minibatch[0], image_size, image_color,
        self.learning_rate,  init_model = init_model)

  def init_data_provider(self):
    ''' we begin with 100 categories'''
    self.set_category_range(self.init_output)

  def set_category_range(self, r):
    dp = DataProvider.get_by_name(self.data_provider)
    self.train_dp = dp(self.data_dir, self.train_range, category_range = range(r))
    self.test_dp = dp(self.data_dir, self.test_range, category_range = range(r))


  def train(self):
    MiniBatchTrainer.train(self)

    for i, cate in enumerate(self.range_list):
      self.set_category_range(cate)
      self.num_batch = self.curr_epoch = self.curr_batch = 0
      self.curr_minibatch = 0
      self.num_minibatch = self.train_minibatch_list[i]

      model = load(self.checkpoint_file)
      layers = model['model_state']['layers']

      for l in layers:
        if l['type'] == 'fc':
          l['weight'] = None
          l['bias'] = None
          l['weightIncr'] = None
          l['biasIncr'] = None

      fc = layers[-2]
      fc['outputSize'] = cate

      self.learning_rate = self.learning_rate_list[i]
      self.net = FastNet(self.learning_rate, self.image_shape, self.n_out, init_model = model)

      self.net.clear_weight_incr()
      MiniBatchTrainer.train(self)
示例#5
0
文件: trainer.py 项目: wqren/striate
class ImageNetCatewisedTrainer(MiniBatchTrainer):
  def _finish_init(self):
    assert len(self.num_caterange_list) == len(self.num_batch) and self.num_caterange_list[-1] == 1000
    self.num_batch_list  = self.num_batch[1:]
    self.num_batch = self.num_batch[0]

    init_output = self.num_caterange_list[0]
    self.num_caterange_list = self.num_caterange_list[1:]

    fc = self.init_model[-2]
    fc['outputSize'] = init_output

    self.learning_rate_list = self.learning_rate[1:]
    self.learning_rate = self.learning_rate[0]

    self.set_category_range(init_output)
    self.net = FastNet(self.learning_rate, self.image_shape, init_model = self.init_model)
    MiniBatchTrainer._finish_init(self)


  def set_category_range(self, r):
    dp = DataProvider.get_by_name(self.data_provider)
    self.train_dp = dp(self.data_dir, self.train_range, category_range = range(r))
    self.test_dp = dp(self.data_dir, self.test_range, category_range = range(r))


  def train(self):
    MiniBatchTrainer.train(self)

    for i, cate in enumerate(self.num_caterange_list):
      self.set_category_range(cate)
      self.curr_batch = self.curr_epoch = 0
      self.num_batch = self.num_batch_list[i]

      model = self.checkpoint_dumper.get_checkpoint()
      layers = model['model_state']['layers']

      fc = layers[-2]
      fc['weight'] = None
      fc['bias'] = None
      fc['weightIncr'] = None
      fc['biasIncr'] = None
      #for l in layers:
      #  if l['type'] == 'fc':
      #    l['weight'] = None
      #    l['bias'] = None
      #    l['weightIncr'] = None
      #    l['biasIncr'] = None

      #fc = layers[-2]
      fc['outputSize'] = cate

      self.learning_rate = self.learning_rate_list[i]
      self.net = FastNet(self.learning_rate, self.image_shape, init_model = model)

      self.net.clear_weight_incr()
      MiniBatchTrainer.train(self)
示例#6
0
class ImageNetCateGroupTrainer(MiniBatchTrainer):
  def __init__(self, test_id, data_dir, data_provider, checkpoint_dir, train_range, test_range,
      test_freq, save_freq, batch_size, num_minibatch, image_size, image_color, learning_rate,
      num_group_list, init_model, adjust_freq = 100, factor = 1.0):

    self.train_minibatch_list = num_minibatch[1:]
    self.num_group_list = num_group_list[1:]
    self.learning_rate_list = learning_rate[1:]

    layers = init_model
    fc = layers[-2]
    fc['outputSize'] = num_group_list[0]

    MiniBatchTrainer.__init__(self, test_id, data_dir, data_provider, checkpoint_dir, train_range, test_range,
        test_freq, save_freq, batch_size, num_minibatch[0], image_size, image_color, learning_rate[0], init_model = init_model)


  def set_num_group(self, n):
    dp = DataProvider.get_by_name(self.data_provider)
    self.train_dp = dp(self.data_dir, self.train_range, n)
    self.test_dp = dp(self.data_dir, self.test_range, n)

  def init_data_provider(self):
    self.set_num_group(self.n_out)

  def train(self):
    MiniBatchTrainer.train(self)

    for i, group in enumerate(self.num_group_list):
      self.set_num_group(group)
      self.num_batch = self.curr_epoch = self.curr_batch = 0
      self.curr_minibatch = 0
      self.num_minibatch = self.train_minibatch_list[i]

      model = load(self.checkpoint_file)
      layers = model['model_state']['layers']

      fc = layers[-2]
      fc['outputSize'] = group
      fc['weight'] = None
      fc['bias'] = None
      fc['weightIncr'] = None
      fc['biasIncr'] = None

      self.learning_rate = self.learning_rate_list[i]
      self.net = FastNet(self.learning_rate, self.image_shape, self.n_out, init_model = model)

      self.net.clear_weight_incr()
      MiniBatchTrainer.train(self)
示例#7
0
文件: trainer.py 项目: wqren/striate
class ImageNetCateGroupTrainer(MiniBatchTrainer):
    def _finish_init(self):
        self.num_batch_list = self.num_batch[1:]
        self.num_batch = self.num_batch[0]
        self.learning_rate_list = self.learning_rate[1:]
        self.learning_rate = self.learning_rate[0]

        layers = self.init_model
        fc = layers[-2]
        fc['outputSize'] = self.num_group_list[0]
        self.num_group_list = self.num_group_list[1:]

        self.set_num_group(fc['outputSize'])
        self.net = FastNet(self.learning_rate,
                           self.image_shape,
                           init_model=self.init_model)
        MiniBatchTrainer._finish_init(self)

    def set_num_group(self, n):
        dp = DataProvider.get_by_name(self.data_provider)
        self.train_dp = dp(self.data_dir, self.train_range, n)
        self.test_dp = dp(self.data_dir, self.test_range, n)

    def train(self):
        MiniBatchTrainer.train(self)

        for i, group in enumerate(self.num_group_list):
            self.set_num_group(group)
            self.curr_batch = self.curr_epoch = 0
            self.num_batch = self.num_batch_list[i]

            model = self.checkpoint_dumper.get_checkpoint()
            layers = model['model_state']['layers']

            fc = layers[-2]
            fc['outputSize'] = group
            fc['weight'] = None
            fc['bias'] = None
            fc['weightIncr'] = None
            fc['biasIncr'] = None

            self.learning_rate = self.learning_rate_list[i]
            self.net = FastNet(self.learning_rate,
                               self.image_shape,
                               init_model=model)

            self.net.clear_weight_incr()
            MiniBatchTrainer.train(self)
示例#8
0
文件: trainer.py 项目: wqren/striate
  def train(self):
    MiniBatchTrainer.train(self)

    for i, cate in enumerate(self.num_caterange_list):
      self.set_category_range(cate)
      self.curr_batch = self.curr_epoch = 0
      self.num_batch = self.num_batch_list[i]

      model = self.checkpoint_dumper.get_checkpoint()
      layers = model['model_state']['layers']

      fc = layers[-2]
      fc['weight'] = None
      fc['bias'] = None
      fc['weightIncr'] = None
      fc['biasIncr'] = None
      #for l in layers:
      #  if l['type'] == 'fc':
      #    l['weight'] = None
      #    l['bias'] = None
      #    l['weightIncr'] = None
      #    l['biasIncr'] = None

      #fc = layers[-2]
      fc['outputSize'] = cate

      self.learning_rate = self.learning_rate_list[i]
      self.net = FastNet(self.learning_rate, self.image_shape, init_model = model)

      self.net.clear_weight_incr()
      MiniBatchTrainer.train(self)
示例#9
0
文件: trainer.py 项目: wqren/striate
    def _finish_init(self):
        self.num_batch_list = self.num_batch[1:]
        self.num_batch = self.num_batch[0]
        self.learning_rate_list = self.learning_rate[1:]
        self.learning_rate = self.learning_rate[0]

        layers = self.init_model
        fc = layers[-2]
        fc['outputSize'] = self.num_group_list[0]
        self.num_group_list = self.num_group_list[1:]

        self.set_num_group(fc['outputSize'])
        self.net = FastNet(self.learning_rate,
                           self.image_shape,
                           init_model=self.init_model)
        MiniBatchTrainer._finish_init(self)
示例#10
0
文件: trainer.py 项目: wqren/striate
class ImageNetCateGroupTrainer(MiniBatchTrainer):
  def _finish_init(self):
    self.num_batch_list = self.num_batch[1:]
    self.num_batch = self.num_batch[0]
    self.learning_rate_list = self.learning_rate[1:]
    self.learning_rate = self.learning_rate[0]

    layers = self.init_model
    fc = layers[-2]
    fc['outputSize'] = self.num_group_list[0]
    self.num_group_list = self.num_group_list[1:]

    self.set_num_group(fc['outputSize'])
    self.net = FastNet(self.learning_rate, self.image_shape, init_model = self.init_model)
    MiniBatchTrainer._finish_init(self)

  def set_num_group(self, n):
    dp = DataProvider.get_by_name(self.data_provider)
    self.train_dp = dp(self.data_dir, self.train_range, n)
    self.test_dp = dp(self.data_dir, self.test_range, n)


  def train(self):
    MiniBatchTrainer.train(self)

    for i, group in enumerate(self.num_group_list):
      self.set_num_group(group)
      self.curr_batch = self.curr_epoch = 0
      self.num_batch = self.num_batch_list[i]

      model = self.checkpoint_dumper.get_checkpoint()
      layers = model['model_state']['layers']

      fc = layers[-2]
      fc['outputSize'] = group
      fc['weight'] = None
      fc['bias'] = None
      fc['weightIncr'] = None
      fc['biasIncr'] = None

      self.learning_rate = self.learning_rate_list[i]
      self.net = FastNet(self.learning_rate, self.image_shape, init_model = model)

      self.net.clear_weight_incr()
      MiniBatchTrainer.train(self)
示例#11
0
  def __init__(self, test_id, data_dir, data_provider, checkpoint_dir, train_range, test_range, test_freq, save_freq, batch_size, num_epoch, image_size,
               image_color, learning_rate, auto_init=False, init_model=None, adjust_freq=1, factor=1.0):
    self.test_id = test_id
    self.data_dir = data_dir
    self.data_provider = data_provider
    self.checkpoint_dir = checkpoint_dir
    self.train_range = train_range
    self.test_range = test_range
    self.test_freq = test_freq
    self.save_freq = save_freq
    self.batch_size = batch_size
    self.num_epoch = num_epoch
    self.image_size = image_size
    self.image_color = image_color
    self.learning_rate = learning_rate
    # doesn't matter anymore
    self.n_out = 10
    self.factor = factor
    self.adjust_freq = adjust_freq
    self.regex = re.compile('^test%d-(\d+)\.(\d+)$' % self.test_id)

    self.init_data_provider()
    self.image_shape = (self.batch_size, self.image_color, self.image_size, self.image_size)

    if init_model is not None and 'model_state' in init_model:
      self.train_outputs = init_model['model_state']['train_outputs']
      self.test_outputs = init_model['model_state']['test_outputs']
    else:
      self.train_outputs = []
      self.test_outputs = []

    self.curr_minibatch = self.num_batch = self.curr_epoch = self.curr_batch = 0
    self.net = FastNet(self.learning_rate, self.image_shape, self.n_out, init_model=init_model)

    self.train_data = None
    self.test_data = None

    self.num_train_minibatch = 0
    self.num_test_minibatch = 0
    self.checkpoint_file = ''
    
    self.train_dumper = None #DataDumper('/scratch1/imagenet-pickle/train-data.pickle')
    self.test_dumper = None #DataDumper('/scratch1/imagenet-pickle/test-data.pickle')
    self.input = None
示例#12
0
  def train(self):
    AutoStopTrainer.train(self)

    if self.layerwised:
      for i in range(len(self.n_filters) - 1):
        next_n_filter = [self.n_filters[i + 1]]
        next_size_filter = [self.size_filters[i + 1]]
        model = load(self.checkpoint_file)
        self.net = FastNet(self.learning_rate, self.image_shape, 0, initModel=model)
        self.net.del_layer()
        self.net.del_layer()
        self.net.disable_bprop()

        self.net.add_parameterized_layers(next_n_filter, next_size_filter, self.fc_nouts)
        self.init_data_provider()
        self.scheduler = Scheduler(self)
        self.test_outputs = []
        self.train_outputs = []
        AutoStopTrainer.train(self)
示例#13
0
文件: trainer.py 项目: wqren/striate
    def _finish_init(self):
        assert len(self.num_caterange_list) == len(
            self.num_batch) and self.num_caterange_list[-1] == 1000
        self.num_batch_list = self.num_batch[1:]
        self.num_batch = self.num_batch[0]

        init_output = self.num_caterange_list[0]
        self.num_caterange_list = self.num_caterange_list[1:]

        fc = self.init_model[-2]
        fc['outputSize'] = init_output

        self.learning_rate_list = self.learning_rate[1:]
        self.learning_rate = self.learning_rate[0]

        self.set_category_range(init_output)
        self.net = FastNet(self.learning_rate,
                           self.image_shape,
                           init_model=self.init_model)
        MiniBatchTrainer._finish_init(self)
示例#14
0
文件: trainer.py 项目: wqren/striate
  def _finish_init(self):
    self.num_batch_list = self.num_batch[1:]
    self.num_batch = self.num_batch[0]
    self.learning_rate_list = self.learning_rate[1:]
    self.learning_rate = self.learning_rate[0]

    layers = self.init_model
    fc = layers[-2]
    fc['outputSize'] = self.num_group_list[0]
    self.num_group_list = self.num_group_list[1:]

    self.set_num_group(fc['outputSize'])
    self.net = FastNet(self.learning_rate, self.image_shape, init_model = self.init_model)
    MiniBatchTrainer._finish_init(self)
示例#15
0
文件: trainer.py 项目: wqren/striate
    def _finish_init(self):
        self.curr_model = []
        self.complete_model = self.init_model
        self.fc_params = []
        self.conv_params = []
        self.final_num_epoch = self.num_epoch

        conv = True
        for ld in self.init_model:
            if ld['type'] in ['conv', 'rnorm', 'pool', 'neuron'] and conv:
                #self.conv_params.append(ld)
                self.curr_model.append(ld)
            elif ld['type'] == 'fc' or (not conv and ld['type'] == 'neuron'):
                self.fc_params.append(ld)
                conv = False
            else:
                self.softmax_param = ld

        #self.conv_stack = FastNet.split_conv_to_stack(self.conv_params)
        #for i in range(3):
        #  self.curr_model.extend(self.conv_stack[i])

        self.fc_stack = FastNet.split_fc_to_stack(self.fc_params)
        #tmp = self.conv_stack[3:]
        #tmp.extend(self.fc_stack)
        #self.stack = tmp
        self.stack = self.fc_stack

        self.curr_model.append(self.stack[-1][0])
        self.curr_model.append(self.softmax_param)
        del self.stack[-1]
        pprint.pprint(self.stack)

        self.layerwised = True
        self.num_epoch = 1
        self.net = FastNet(self.learning_rate, self.image_shape,
                           self.curr_model)
示例#16
0
文件: trainer.py 项目: wqren/striate
  def _finish_init(self):
    self.curr_model = []
    self.complete_model = self.init_model
    self.fc_params = []
    self.conv_params = []
    self.final_num_epoch = self.num_epoch

    conv = True
    for ld in self.init_model:
      if ld['type'] in ['conv', 'rnorm', 'pool', 'neuron'] and conv:
        #self.conv_params.append(ld)
        self.curr_model.append(ld)
      elif ld['type'] == 'fc' or (not conv and ld['type'] == 'neuron'):
        self.fc_params.append(ld)
        conv = False
      else:
        self.softmax_param = ld

    #self.conv_stack = FastNet.split_conv_to_stack(self.conv_params)
    #for i in range(3):
    #  self.curr_model.extend(self.conv_stack[i])

    self.fc_stack = FastNet.split_fc_to_stack(self.fc_params)
    #tmp = self.conv_stack[3:]
    #tmp.extend(self.fc_stack)
    #self.stack = tmp
    self.stack = self.fc_stack

    self.curr_model.append(self.stack[-1][0])
    self.curr_model.append(self.softmax_param)
    del self.stack[-1]
    pprint.pprint(self.stack)

    self.layerwised = True
    self.num_epoch = 1
    self.net = FastNet(self.learning_rate, self.image_shape, self.curr_model)
示例#17
0
文件: trainer.py 项目: wqren/striate
    def get_trainer_by_name(name, param_dict):
        net = FastNet(param_dict['learning_rate'],
                      param_dict['image_shape'],
                      init_model=None)
        param_dict['net'] = net
        if name == 'layerwise':
            return ImageNetLayerwisedTrainer(**param_dict)

        if name == 'catewise':
            return ImageNetCatewisedTrainer(**param_dict)

        if name == 'categroup':
            return ImageNetCateGroupTrainer(**param_dict)

        net = FastNet(param_dict['learning_rate'], param_dict['image_shape'],
                      param_dict['init_model'])
        param_dict['net'] = net
        if name == 'normal':
            return Trainer(**param_dict)

        if name == 'minibatch':
            return MiniBatchTrainer(**param_dict)

        raise Exception, 'No trainer found for name: %s' % name
示例#18
0
  def train(self):
    # train conv stack layer by layer
    for i, stack in enumerate(self.conv_stack):
      if self.checkpoint_file != '':
        model = load(self.checkpoint_file)
        self.net = FastNet(self.learning_rate, self.image_shape, self.n_out, initModel=model)
        # delete softmax layer
        self.net.del_layer()
        self.net.del_layer()

        # for i in range(len(self.fc_params)):
        #  self.net.del_layer()

        self.net.disable_bprop()

      layerParam = stack + [self.fakefc_param, self.softmax_param]
      self.net.append_layers_from_dict(layerParam)

      self.init_data_provider()
      self.scheduler.reset()
      self.scheduler.set_level(i)
      self.test_outputs = []
      self.train_output = []
      AutoStopTrainer.train(self)

    # train fc layer
    for i, stack in enumerate(self.fc_stack):
      model = load(self.checkpoint_file)
      self.net = FastNet(self.learning_rate, self.image_shape, self.n_out, initModel=model)
      self.net.del_layer()
      self.net.del_layer()

      self.net.disable_bprop()

      if i == len(self.fc_stack) - 1:
        layerParam = stack + [self.softmax_param]
      else:
        layerParam = stack + [self.fakefc_param, self.softmax_param]
      self.net.append_layers_from_dict(layerParam)

      self.init_data_provider()
      self.scheduler.reset()
      self.scheduler.set_level(i)
      self.test_outputs = []
      self.train_output = []
      AutoStopTrainer.train(self)

    model = load(self.checkpoint_file)
    self.test_id += 1
    self.net = FastNet(self.learning_rate, self.image_shape, self.n_out, initModel=model)
    self.test_range = self.origin_test_range
    self.init_data_provider()
    self.scheduler = Scheduler(self)
    self.num_epoch /= 2
    AutoStopTrainer.train(self)
示例#19
0
文件: trainer.py 项目: wqren/striate
  def _finish_init(self):
    assert len(self.num_caterange_list) == len(self.num_batch) and self.num_caterange_list[-1] == 1000
    self.num_batch_list  = self.num_batch[1:]
    self.num_batch = self.num_batch[0]

    init_output = self.num_caterange_list[0]
    self.num_caterange_list = self.num_caterange_list[1:]

    fc = self.init_model[-2]
    fc['outputSize'] = init_output

    self.learning_rate_list = self.learning_rate[1:]
    self.learning_rate = self.learning_rate[0]

    self.set_category_range(init_output)
    self.net = FastNet(self.learning_rate, self.image_shape, init_model = self.init_model)
    MiniBatchTrainer._finish_init(self)
示例#20
0
文件: trainer.py 项目: wqren/striate
  def train(self):
    MiniBatchTrainer.train(self)

    for i, group in enumerate(self.num_group_list):
      self.set_num_group(group)
      self.curr_batch = self.curr_epoch = 0
      self.num_batch = self.num_batch_list[i]

      model = self.checkpoint_dumper.get_checkpoint()
      layers = model['model_state']['layers']

      fc = layers[-2]
      fc['outputSize'] = group
      fc['weight'] = None
      fc['bias'] = None
      fc['weightIncr'] = None
      fc['biasIncr'] = None

      self.learning_rate = self.learning_rate_list[i]
      self.net = FastNet(self.learning_rate, self.image_shape, init_model = model)

      self.net.clear_weight_incr()
      MiniBatchTrainer.train(self)
示例#21
0
class LayerwisedTrainer(AutoStopTrainer):
  def __init__(self, test_id, data_dir, provider, checkpoint_dir, train_range, test_range, test_freq,
      save_freq, batch_size, num_epoch, image_size, image_color, learning_rate, n_filters,
      size_filters, fc_nouts):
    AutoStopTrainer.__init__(self, test_id, data_dir,provider,  checkpoint_dir, train_range, test_range, test_freq,
        save_freq, batch_size, num_epoch, image_size, image_color, learning_rate, 0, False)
    if len(n_filters) == 1:
      self.layerwised = False
    else:
      self.layerwised = True


    self.n_filters = n_filters
    self.size_filters = size_filters
    self.fc_nouts = fc_nouts

    init_n_filter = [self.n_filters[0]]
    init_size_filter = [self.size_filters[0]]

    self.net.add_parameterized_layers(init_n_filter, init_size_filter, self.fc_nouts)

  def train(self):
    AutoStopTrainer.train(self)

    if self.layerwised:
      for i in range(len(self.n_filters) - 1):
        next_n_filter = [self.n_filters[i + 1]]
        next_size_filter = [self.size_filters[i + 1]]
        model = load(self.checkpoint_file)
        self.net = FastNet(self.learning_rate, self.image_shape, 0, initModel=model)
        self.net.del_layer()
        self.net.del_layer()
        self.net.disable_bprop()

        self.net.add_parameterized_layers(next_n_filter, next_size_filter, self.fc_nouts)
        self.init_data_provider()
        self.scheduler = Scheduler(self)
        self.test_outputs = []
        self.train_outputs = []
        AutoStopTrainer.train(self)
        
  def add_parameterized_layers(self, net, n_filters=None, size_filters=None, fc_nout=[10]):
    for i in range(len(n_filters)):
      prev = n_filters[i - 1] if i > 0 else net.imgShapes[-1][1]
      filter_shape = (n_filters[i], prev, size_filters[i], size_filters[i])
      conv = layer.ConvLayer('conv' + str(net.numConv), filter_shape, net.imgShapes[-1])
      net.append_layer(conv)

      neuron = layer.NeuronLayer('neuron' + str(net.numConv), net.imgShapes[-1], type='tanh')
      net.append_layer(neuron)

      pool = layer.MaxPoolLayer('pool' + str(net.numConv), net.imgShapes[-1])
      net.append_layer(pool)

      rnorm = layer.ResponseNormLayer('rnorm' + str(net.numConv), net.imgShapes[-1])
      net.append_layer(rnorm)

    for i in range(len(fc_nout)):
      fc = layer.FCLayer('fc' + str(i + 1), net.inputShapes[-1], fc_nout[-1])
      net.append_layer(fc)

    net.append_layer(layer.SoftmaxLayer('softmax', net.inputShapes[-1]))
示例#22
0
文件: trainer.py 项目: wqren/striate
  def train(self):
    Trainer.train(self)
    for i, stack in enumerate(self.stack):
      pprint.pprint(stack)
      self.curr_model = self.checkpoint_dumper.get_checkpoint()
      self.curr_batch = self.curr_epoch =  0

      l = self.curr_model['model_state']['layers'][-2]
      assert l['type'] == 'fc'

      l['weight'] = None
      l['bias'] = None
      l['weightIncr'] = None
      l['biasIncr'] = None

      if i == len(self.stack) - 1:
        self.num_epoch = self.final_num_epoch

      layers = self.curr_model['model_state']['layers']
      stack[0]['epsW'] *= self.learning_rate
      stack[0]['epsB'] *= self.learning_rate
      model = [stack[0], stack[1], layers[-2], layers[-1]]

      train_dp_old = self.train_dp
      test_dp_old = self.test_dp
      self.init_subnet_data_provider()

      self.train_dumper = None
      self.test_dumper = None

      image_shape_old = self.image_shape
      shape = self.curr_model['model_state']['layers'][-3]['outputShape']
      size= shape[0] * shape[1] * shape[2]
      self.image_shape = (size, 1, 1, self.batch_size)
      self.net = FastNet(1.0, self.image_shape, init_model = model)

      old_num_epoch = self.num_epoch
      self.num_epoch = 1
      Trainer.train(self)

      self.curr_batch = self.curr_epoch = 0

      self.num_epoch = old_num_epoch

      self.image_shape = image_shape_old
      del layers[-1], layers[-1]
      layers.extend(self.net.get_dumped_layers())

      self.train_dp = train_dp_old
      self.test_dp = test_dp_old

      #for layer in self.curr_model['model_state']['layers'][:-2]:
      #  layer['disableBprop'] = True

      #stack[0]['epsW'] *= self.learning_rate
      #stack[0]['epsB'] *= self.learning_rate
      #self.curr_model['model_state']['layers'].insert(-2, stack[0])
      #self.curr_model['model_state']['layers'].insert(-2, stack[1])


      self.init_output_dumper()
      self.init_data_provider()
      self.net = FastNet(self.learning_rate, self.image_shape,  init_model = self.curr_model)
      Trainer.train(self)
示例#23
0
class ImageNetLayerwisedTrainer(AutoStopTrainer):
  def __init__(self, test_id, data_dir, provider, checkpoint_dir, train_range, test_range, test_freq,
      save_freq, batch_size, num_epoch, image_size, image_color, learning_rate,  params):

    self.origin_test_range = test_range
    if len(test_range) != 1:
      test_range = [test_range[0]]
    AutoStopTrainer.__init__(self, test_id, data_dir, provider, checkpoint_dir, train_range, test_range, test_freq,
        save_freq, batch_size, num_epoch, image_size, image_color, learning_rate, False)

    self.conv_params = []
    self.fc_params = []
    self.softmax_param = None

    self.params = params

    conv = True
    for ld in self.params:
      if ld['type'] in ['conv', 'rnorm', 'pool', 'neuron'] and conv:
        self.conv_params.append(ld)
      elif ld['type'] == 'fc' or (not conv and ld['type'] == 'neuron'):
        self.fc_params.append(ld)
        conv = False
      else:
        self.softmax_param = ld

    self.conv_stack = FastNet.split_conv_to_stack(self.conv_params)
    self.fc_stack = FastNet.split_fc_to_stack(self.fc_params)

    pprint.pprint(self.conv_stack)
    pprint.pprint(self.fc_stack)

    self.fakefc_param = self.fc_stack[-1][0]

  def report(self):
    pass

  def init_data_provider(self):
    self.train_dp = ImageNetDataProvider(self.data_dir, self.train_range)
    self.test_dp = ImageNetDataProvider(self.data_dir, self.test_range)

  def train(self):
    # train conv stack layer by layer
    for i, stack in enumerate(self.conv_stack):
      if self.checkpoint_file != '':
        model = load(self.checkpoint_file)
        self.net = FastNet(self.learning_rate, self.image_shape, self.n_out, initModel=model)
        # delete softmax layer
        self.net.del_layer()
        self.net.del_layer()

        # for i in range(len(self.fc_params)):
        #  self.net.del_layer()

        self.net.disable_bprop()

      layerParam = stack + [self.fakefc_param, self.softmax_param]
      self.net.append_layers_from_dict(layerParam)

      self.init_data_provider()
      self.scheduler.reset()
      self.scheduler.set_level(i)
      self.test_outputs = []
      self.train_output = []
      AutoStopTrainer.train(self)

    # train fc layer
    for i, stack in enumerate(self.fc_stack):
      model = load(self.checkpoint_file)
      self.net = FastNet(self.learning_rate, self.image_shape, self.n_out, initModel=model)
      self.net.del_layer()
      self.net.del_layer()

      self.net.disable_bprop()

      if i == len(self.fc_stack) - 1:
        layerParam = stack + [self.softmax_param]
      else:
        layerParam = stack + [self.fakefc_param, self.softmax_param]
      self.net.append_layers_from_dict(layerParam)

      self.init_data_provider()
      self.scheduler.reset()
      self.scheduler.set_level(i)
      self.test_outputs = []
      self.train_output = []
      AutoStopTrainer.train(self)

    model = load(self.checkpoint_file)
    self.test_id += 1
    self.net = FastNet(self.learning_rate, self.image_shape, self.n_out, initModel=model)
    self.test_range = self.origin_test_range
    self.init_data_provider()
    self.scheduler = Scheduler(self)
    self.num_epoch /= 2
    AutoStopTrainer.train(self)
示例#24
0
class Trainer:
  CHECKPOINT_REGEX = None
  def __init__(self, test_id, data_dir, data_provider, checkpoint_dir, train_range, test_range, test_freq, save_freq, batch_size, num_epoch, image_size,
               image_color, learning_rate, auto_init=False, init_model=None, adjust_freq=1, factor=1.0):
    self.test_id = test_id
    self.data_dir = data_dir
    self.data_provider = data_provider
    self.checkpoint_dir = checkpoint_dir
    self.train_range = train_range
    self.test_range = test_range
    self.test_freq = test_freq
    self.save_freq = save_freq
    self.batch_size = batch_size
    self.num_epoch = num_epoch
    self.image_size = image_size
    self.image_color = image_color
    self.learning_rate = learning_rate
    # doesn't matter anymore
    self.n_out = 10
    self.factor = factor
    self.adjust_freq = adjust_freq
    self.regex = re.compile('^test%d-(\d+)\.(\d+)$' % self.test_id)

    self.init_data_provider()
    self.image_shape = (self.batch_size, self.image_color, self.image_size, self.image_size)

    if init_model is not None and 'model_state' in init_model:
      self.train_outputs = init_model['model_state']['train_outputs']
      self.test_outputs = init_model['model_state']['test_outputs']
    else:
      self.train_outputs = []
      self.test_outputs = []

    self.curr_minibatch = self.num_batch = self.curr_epoch = self.curr_batch = 0
    self.net = FastNet(self.learning_rate, self.image_shape, self.n_out, init_model=init_model)

    self.train_data = None
    self.test_data = None

    self.num_train_minibatch = 0
    self.num_test_minibatch = 0
    self.checkpoint_file = ''
    
    self.train_dumper = None #DataDumper('/scratch1/imagenet-pickle/train-data.pickle')
    self.test_dumper = None #DataDumper('/scratch1/imagenet-pickle/test-data.pickle')
    self.input = None


  def init_data_provider(self):
    dp = DataProvider.get_by_name(self.data_provider)
    self.train_dp = dp(self.data_dir, self.train_range)
    self.test_dp = dp(self.data_dir, self.test_range)


  def get_next_minibatch(self, i, train=TRAIN):
    if train == TRAIN:
      data = self.train_data
    else:
      data = self.test_data

    batch_data = data.data
    batch_label = data.labels
    batch_size = self.batch_size

    mini_data = batch_data[:, i * batch_size: (i + 1) * batch_size]
    locked_data = driver.pagelocked_empty(mini_data.shape, mini_data.dtype, order='C',
                                          mem_flags=driver.host_alloc_flags.PORTABLE)
    locked_data[:] = mini_data

    if self.input is not None and locked_data.shape == self.input.shape:
      self.input.set(locked_data)
    else:
      self.input = gpuarray.to_gpu(locked_data)
    
    label = batch_label[i * batch_size : (i + 1) * batch_size]
    #label = gpuarray.to_gpu(label)

    #label = gpuarray.to_gpu(np.require(batch_label[i * batch_size : (i + 1) * batch_size],  dtype =
    #  np.float, requirements = 'C'))

    return self.input, label


  def save_checkpoint(self):
    model = {}
    model['batchnum'] = self.train_dp.get_batch_num()
    model['epoch'] = self.num_epoch + 1
    model['layers'] = self.net.get_dumped_layers()

    model['train_outputs'] = self.train_outputs
    model['test_outputs'] = self.test_outputs

    dic = {'model_state': model, 'op':None}
    self.print_net_summary()
    
    if not os.path.exists(self.checkpoint_dir):
      os.system('mkdir -p \'%s\'' % self.checkpoint_dir)
    
    saved_filename = [f for f in os.listdir(self.checkpoint_dir) if self.regex.match(f)]
    for f in saved_filename:
      os.remove(os.path.join(self.checkpoint_dir, f))
    checkpoint_filename = "test%d-%d.%d" % (self.test_id, self.curr_epoch, self.curr_batch)
    checkpoint_file_path = os.path.join(self.checkpoint_dir, checkpoint_filename)
    self.checkpoint_file = checkpoint_file_path
    print >> sys.stderr,  checkpoint_file_path
    with open(checkpoint_file_path, 'w') as f:
      cPickle.dump(dic, f, protocol=-1)
    util.log('save file finished')

  def get_test_error(self):
    start = time.time()
    self.test_data = self.test_dp.get_next_batch()

    self.num_test_minibatch = divup(self.test_data.data.shape[1], self.batch_size)
    for i in range(self.num_test_minibatch):
      input, label = self.get_next_minibatch(i, TEST)
      self.net.train_batch(input, label, TEST)
      self._capture_test_data()
    
    cost , correct, numCase, = self.net.get_batch_information()
    self.test_outputs += [({'logprob': [cost, 1 - correct]}, numCase, time.time() - start)]
    print >> sys.stderr,  '[%d] error: %f logreg: %f time: %f' % (self.test_data.batchnum, 1 - correct, cost, time.time() - start)

  def print_net_summary(self):
    print >> sys.stderr,  '--------------------------------------------------------------'
    for s in self.net.get_summary():
      name = s[0]
      values = s[1]
      print >> sys.stderr,  "Layer '%s' weight: %e [%e]" % (name, values[0], values[1])
      print >> sys.stderr,  "Layer '%s' bias: %e [%e]" % (name, values[2], values[3])


  def should_continue_training(self):
    return self.curr_epoch <= self.num_epoch

  def check_test_data(self):
    return self.num_batch % self.test_freq == 0

  def check_save_checkpoint(self):
    return self.num_batch % self.save_freq == 0

  def check_adjust_lr(self):
    return self.num_batch % self.adjust_freq == 0
  
  def _finished_training(self):
    if self.train_dumper is not None:
      self.train_dumper.flush()
    
    if self.test_dumper is not None:
      self.test_dumper.flush()
      
  def _capture_training_data(self):
    if not self.train_dumper:
      return

    self.train_dumper.add({'labels' : self.net.label.get(),
                           'fc' : self.net.outputs[-3].get().transpose() })
    
  def _capture_test_data(self):
    if not self.test_dumper:
      return
    self.test_dumper.add({'labels' : self.net.label.get(),
                           'fc' : self.net.outputs[-3].get().transpose() })

  def train(self):
    self.print_net_summary()
    util.log('Starting training...')
    while self.should_continue_training():
      self.train_data = self.train_dp.get_next_batch()  # self.train_dp.wait()
      self.curr_epoch = self.train_data.epoch
      self.curr_batch = self.train_data.batchnum

      start = time.time()
      self.num_train_minibatch = divup(self.train_data.data.shape[1], self.batch_size)
      t = 0
      
      for i in range(self.num_train_minibatch):
        input, label = self.get_next_minibatch(i)
        stime = time.time()
        self.net.train_batch(input, label)
        self._capture_training_data()
        t += time.time() - stime
        self.curr_minibatch += 1

      cost , correct, numCase = self.net.get_batch_information()
      self.train_outputs += [({'logprob': [cost, 1 - correct]}, numCase, time.time() - start)]
      print >> sys.stderr,  '%d.%d: error: %f logreg: %f time: %f' % (self.curr_epoch, self.curr_batch, 1 - correct, cost, time.time() - start)

      self.num_batch += 1
      if self.check_test_data():
        print >> sys.stderr,  '---- test ----'
        self.get_test_error()
        print >> sys.stderr,  '------------'

      if self.factor != 1.0 and self.check_adjust_lr():
        print >> sys.stderr,  '---- adjust learning rate ----'
        self.net.adjust_learning_rate(self.factor)
        print >> sys.stderr,  '--------'

      if self.check_save_checkpoint():
        print >> sys.stderr,  '---- save checkpoint ----'
        self.save_checkpoint()
        print >> sys.stderr,  '------------'

      wait_time = time.time()

      #print 'waitting', time.time() - wait_time, 'secs to load'
      #print 'time to train a batch file is', time.time() - start)


    self.get_test_error()
    self.save_checkpoint()
    self.report()
    self._finished_training()

  def predict(self, save_layers = None, filename = None):
    self.net.save_layerouput(save_layers)
    self.print_net_summary()
    util.log('Starting predict...')
    save_output = []
    while self.curr_epoch < 2:
      start = time.time()
      self.test_data = self.test_dp.get_next_batch()
      self.curr_epoch = self.test_data.epoch
      self.curr_batch = self.test_data.batchnum

      self.num_test_minibatch = divup(self.test_data.data.shape[1], self.batch_size)
      for i in range(self.num_test_minibatch):
        input, label = self.get_next_minibatch(i, TEST)
        self.net.train_batch(input, label, TEST)
      cost , correct, numCase = self.net.get_batch_information()
      print >> sys.stderr,  '%d.%d: error: %f logreg: %f time: %f' % (self.curr_epoch, self.curr_batch, 1 - correct, cost, time.time() - start)
      if save_layers is not None:
        save_output.extend(self.net.get_save_output())

    if save_layers is not None:
      if filename is not None:
        with open(filename, 'w') as f:
          cPickle.dump(save_output, f, protocol = -1)
        util.log('save layer output finished')


  def report(self):
    rep = self.net.get_report()
    if rep is not None:
      print rep
示例#25
0
文件: trainer.py 项目: wqren/striate
class ImageNetCatewisedTrainer(MiniBatchTrainer):
    def _finish_init(self):
        assert len(self.num_caterange_list) == len(
            self.num_batch) and self.num_caterange_list[-1] == 1000
        self.num_batch_list = self.num_batch[1:]
        self.num_batch = self.num_batch[0]

        init_output = self.num_caterange_list[0]
        self.num_caterange_list = self.num_caterange_list[1:]

        fc = self.init_model[-2]
        fc['outputSize'] = init_output

        self.learning_rate_list = self.learning_rate[1:]
        self.learning_rate = self.learning_rate[0]

        self.set_category_range(init_output)
        self.net = FastNet(self.learning_rate,
                           self.image_shape,
                           init_model=self.init_model)
        MiniBatchTrainer._finish_init(self)

    def set_category_range(self, r):
        dp = DataProvider.get_by_name(self.data_provider)
        self.train_dp = dp(self.data_dir,
                           self.train_range,
                           category_range=range(r))
        self.test_dp = dp(self.data_dir,
                          self.test_range,
                          category_range=range(r))

    def train(self):
        MiniBatchTrainer.train(self)

        for i, cate in enumerate(self.num_caterange_list):
            self.set_category_range(cate)
            self.curr_batch = self.curr_epoch = 0
            self.num_batch = self.num_batch_list[i]

            model = self.checkpoint_dumper.get_checkpoint()
            layers = model['model_state']['layers']

            fc = layers[-2]
            fc['weight'] = None
            fc['bias'] = None
            fc['weightIncr'] = None
            fc['biasIncr'] = None
            #for l in layers:
            #  if l['type'] == 'fc':
            #    l['weight'] = None
            #    l['bias'] = None
            #    l['weightIncr'] = None
            #    l['biasIncr'] = None

            #fc = layers[-2]
            fc['outputSize'] = cate

            self.learning_rate = self.learning_rate_list[i]
            self.net = FastNet(self.learning_rate,
                               self.image_shape,
                               init_model=model)

            self.net.clear_weight_incr()
            MiniBatchTrainer.train(self)
示例#26
0
文件: trainer.py 项目: wqren/striate
    def train(self):
        Trainer.train(self)
        for i, stack in enumerate(self.stack):
            pprint.pprint(stack)
            self.curr_model = self.checkpoint_dumper.get_checkpoint()
            self.curr_batch = self.curr_epoch = 0

            l = self.curr_model['model_state']['layers'][-2]
            assert l['type'] == 'fc'

            l['weight'] = None
            l['bias'] = None
            l['weightIncr'] = None
            l['biasIncr'] = None

            if i == len(self.stack) - 1:
                self.num_epoch = self.final_num_epoch

            layers = self.curr_model['model_state']['layers']
            stack[0]['epsW'] *= self.learning_rate
            stack[0]['epsB'] *= self.learning_rate
            model = [stack[0], stack[1], layers[-2], layers[-1]]

            train_dp_old = self.train_dp
            test_dp_old = self.test_dp
            self.init_subnet_data_provider()

            self.train_dumper = None
            self.test_dumper = None

            image_shape_old = self.image_shape
            shape = self.curr_model['model_state']['layers'][-3]['outputShape']
            size = shape[0] * shape[1] * shape[2]
            self.image_shape = (size, 1, 1, self.batch_size)
            self.net = FastNet(1.0, self.image_shape, init_model=model)

            old_num_epoch = self.num_epoch
            self.num_epoch = 1
            Trainer.train(self)

            self.curr_batch = self.curr_epoch = 0

            self.num_epoch = old_num_epoch

            self.image_shape = image_shape_old
            del layers[-1], layers[-1]
            layers.extend(self.net.get_dumped_layers())

            self.train_dp = train_dp_old
            self.test_dp = test_dp_old

            #for layer in self.curr_model['model_state']['layers'][:-2]:
            #  layer['disableBprop'] = True

            #stack[0]['epsW'] *= self.learning_rate
            #stack[0]['epsB'] *= self.learning_rate
            #self.curr_model['model_state']['layers'].insert(-2, stack[0])
            #self.curr_model['model_state']['layers'].insert(-2, stack[1])

            self.init_output_dumper()
            self.init_data_provider()
            self.net = FastNet(self.learning_rate,
                               self.image_shape,
                               init_model=self.curr_model)
            Trainer.train(self)
示例#27
0
文件: trainer.py 项目: wqren/striate
class ImageNetLayerwisedTrainer(Trainer):
    def _finish_init(self):
        self.curr_model = []
        self.complete_model = self.init_model
        self.fc_params = []
        self.conv_params = []
        self.final_num_epoch = self.num_epoch

        conv = True
        for ld in self.init_model:
            if ld['type'] in ['conv', 'rnorm', 'pool', 'neuron'] and conv:
                #self.conv_params.append(ld)
                self.curr_model.append(ld)
            elif ld['type'] == 'fc' or (not conv and ld['type'] == 'neuron'):
                self.fc_params.append(ld)
                conv = False
            else:
                self.softmax_param = ld

        #self.conv_stack = FastNet.split_conv_to_stack(self.conv_params)
        #for i in range(3):
        #  self.curr_model.extend(self.conv_stack[i])

        self.fc_stack = FastNet.split_fc_to_stack(self.fc_params)
        #tmp = self.conv_stack[3:]
        #tmp.extend(self.fc_stack)
        #self.stack = tmp
        self.stack = self.fc_stack

        self.curr_model.append(self.stack[-1][0])
        self.curr_model.append(self.softmax_param)
        del self.stack[-1]
        pprint.pprint(self.stack)

        self.layerwised = True
        self.num_epoch = 1
        self.net = FastNet(self.learning_rate, self.image_shape,
                           self.curr_model)

    def report(self):
        pass

    def should_continue_training(self):
        #if self.layerwised and self.curr_epoch == 2:
        #  self.net.enable_bprop()
        return self.curr_epoch <= self.num_epoch

    def init_subnet_data_provider(self):
        dp = DataProvider.get_by_name('intermediate')
        count = self.train_dumper.get_count()
        self.train_dp = dp(self.train_output_filename, range(0, count), 'fc')
        count = self.test_dumper.get_count()
        self.test_dp = dp(self.test_output_filename, range(0, count), 'fc')
        #dp = DataProvider.get_by_name('memory')
        #self.train_dp = dp(self.train_dumper)
        #self.test_dp = dp(self.test_dumper)

    def train(self):
        Trainer.train(self)
        for i, stack in enumerate(self.stack):
            pprint.pprint(stack)
            self.curr_model = self.checkpoint_dumper.get_checkpoint()
            self.curr_batch = self.curr_epoch = 0

            l = self.curr_model['model_state']['layers'][-2]
            assert l['type'] == 'fc'

            l['weight'] = None
            l['bias'] = None
            l['weightIncr'] = None
            l['biasIncr'] = None

            if i == len(self.stack) - 1:
                self.num_epoch = self.final_num_epoch

            layers = self.curr_model['model_state']['layers']
            stack[0]['epsW'] *= self.learning_rate
            stack[0]['epsB'] *= self.learning_rate
            model = [stack[0], stack[1], layers[-2], layers[-1]]

            train_dp_old = self.train_dp
            test_dp_old = self.test_dp
            self.init_subnet_data_provider()

            self.train_dumper = None
            self.test_dumper = None

            image_shape_old = self.image_shape
            shape = self.curr_model['model_state']['layers'][-3]['outputShape']
            size = shape[0] * shape[1] * shape[2]
            self.image_shape = (size, 1, 1, self.batch_size)
            self.net = FastNet(1.0, self.image_shape, init_model=model)

            old_num_epoch = self.num_epoch
            self.num_epoch = 1
            Trainer.train(self)

            self.curr_batch = self.curr_epoch = 0

            self.num_epoch = old_num_epoch

            self.image_shape = image_shape_old
            del layers[-1], layers[-1]
            layers.extend(self.net.get_dumped_layers())

            self.train_dp = train_dp_old
            self.test_dp = test_dp_old

            #for layer in self.curr_model['model_state']['layers'][:-2]:
            #  layer['disableBprop'] = True

            #stack[0]['epsW'] *= self.learning_rate
            #stack[0]['epsB'] *= self.learning_rate
            #self.curr_model['model_state']['layers'].insert(-2, stack[0])
            #self.curr_model['model_state']['layers'].insert(-2, stack[1])

            self.init_output_dumper()
            self.init_data_provider()
            self.net = FastNet(self.learning_rate,
                               self.image_shape,
                               init_model=self.curr_model)
            Trainer.train(self)
示例#28
0
文件: trainer.py 项目: wqren/striate
class ImageNetLayerwisedTrainer(Trainer):
  def _finish_init(self):
    self.curr_model = []
    self.complete_model = self.init_model
    self.fc_params = []
    self.conv_params = []
    self.final_num_epoch = self.num_epoch

    conv = True
    for ld in self.init_model:
      if ld['type'] in ['conv', 'rnorm', 'pool', 'neuron'] and conv:
        #self.conv_params.append(ld)
        self.curr_model.append(ld)
      elif ld['type'] == 'fc' or (not conv and ld['type'] == 'neuron'):
        self.fc_params.append(ld)
        conv = False
      else:
        self.softmax_param = ld

    #self.conv_stack = FastNet.split_conv_to_stack(self.conv_params)
    #for i in range(3):
    #  self.curr_model.extend(self.conv_stack[i])

    self.fc_stack = FastNet.split_fc_to_stack(self.fc_params)
    #tmp = self.conv_stack[3:]
    #tmp.extend(self.fc_stack)
    #self.stack = tmp
    self.stack = self.fc_stack

    self.curr_model.append(self.stack[-1][0])
    self.curr_model.append(self.softmax_param)
    del self.stack[-1]
    pprint.pprint(self.stack)

    self.layerwised = True
    self.num_epoch = 1
    self.net = FastNet(self.learning_rate, self.image_shape, self.curr_model)

  def report(self):
    pass

  def should_continue_training(self):
    #if self.layerwised and self.curr_epoch == 2:
    #  self.net.enable_bprop()
    return self.curr_epoch <= self.num_epoch

  def init_subnet_data_provider(self):
    dp = DataProvider.get_by_name('intermediate')
    count = self.train_dumper.get_count()
    self.train_dp = dp(self.train_output_filename,  range(0, count), 'fc')
    count = self.test_dumper.get_count()
    self.test_dp = dp(self.test_output_filename, range(0, count), 'fc')
    #dp = DataProvider.get_by_name('memory')
    #self.train_dp = dp(self.train_dumper)
    #self.test_dp = dp(self.test_dumper)

  def train(self):
    Trainer.train(self)
    for i, stack in enumerate(self.stack):
      pprint.pprint(stack)
      self.curr_model = self.checkpoint_dumper.get_checkpoint()
      self.curr_batch = self.curr_epoch =  0

      l = self.curr_model['model_state']['layers'][-2]
      assert l['type'] == 'fc'

      l['weight'] = None
      l['bias'] = None
      l['weightIncr'] = None
      l['biasIncr'] = None

      if i == len(self.stack) - 1:
        self.num_epoch = self.final_num_epoch

      layers = self.curr_model['model_state']['layers']
      stack[0]['epsW'] *= self.learning_rate
      stack[0]['epsB'] *= self.learning_rate
      model = [stack[0], stack[1], layers[-2], layers[-1]]

      train_dp_old = self.train_dp
      test_dp_old = self.test_dp
      self.init_subnet_data_provider()

      self.train_dumper = None
      self.test_dumper = None

      image_shape_old = self.image_shape
      shape = self.curr_model['model_state']['layers'][-3]['outputShape']
      size= shape[0] * shape[1] * shape[2]
      self.image_shape = (size, 1, 1, self.batch_size)
      self.net = FastNet(1.0, self.image_shape, init_model = model)

      old_num_epoch = self.num_epoch
      self.num_epoch = 1
      Trainer.train(self)

      self.curr_batch = self.curr_epoch = 0

      self.num_epoch = old_num_epoch

      self.image_shape = image_shape_old
      del layers[-1], layers[-1]
      layers.extend(self.net.get_dumped_layers())

      self.train_dp = train_dp_old
      self.test_dp = test_dp_old

      #for layer in self.curr_model['model_state']['layers'][:-2]:
      #  layer['disableBprop'] = True

      #stack[0]['epsW'] *= self.learning_rate
      #stack[0]['epsB'] *= self.learning_rate
      #self.curr_model['model_state']['layers'].insert(-2, stack[0])
      #self.curr_model['model_state']['layers'].insert(-2, stack[1])


      self.init_output_dumper()
      self.init_data_provider()
      self.net = FastNet(self.learning_rate, self.image_shape,  init_model = self.curr_model)
      Trainer.train(self)