#
test_iterator = None
#############################
rng = sparnn.utils.quick_npy_rng()
theano_rng = sparnn.utils.quick_theano_rng(rng)
############################# interface layer
param = {
"id": "ucf101-flow-vgg16-fc7",
"use_mask": True,
"input_ndim": 3,
"output_ndim": 2,
"output_data_type": "int64"
}
interface_layer = InterfaceLayer(param)
x = interface_layer.input
mask = interface_layer.mask
y = interface_layer.output
timesteps = x.shape[0]
minibatch_size = x.shape[1]
feature_dim = 4096
hidden_dim = 512
out_dim = 1024
actions = 101
data_dim = (feature_dim, )
logger.info("Data Dim:" + str(data_dim))
def __init__(self, param, config):
self.param = param
self.config = config
# 2x2 小碎片是2*2的小图片（叫做patch，或者掩模）
feature_num = config['patch_size'] * config['patch_size']
row_num = int(config['size'][0] / config['patch_size'])
col_num = int(config['size'][1] / config['patch_size'])
data_dim = (feature_num, row_num, col_num)
kernel_size = config['kernel_size']
kernel_num = config['kernel_num']
logger.debug('build model:' + str(feature_num) + ' ' + str(data_dim) +
' ' + str(kernel_size) + ' ' + str(kernel_num))
param = {
"id": "DeepWeather-interface",
"use_input_mask": False,
"input_ndim": 5,
"output_ndim": 5
}
self.interface_layer = InterfaceLayer(param)
# interface_layer.input: (1, 100, 100), reshape_input: (16, 25, 25), data_dim: (16, 25, 25)
self.reshape_input = quick_reshape_patch(self.interface_layer.input,
config['patch_size'])
self.reshape_output = quick_reshape_patch(self.interface_layer.output,
config['patch_size'])
minibatch_size = self.interface_layer.input.shape[1]
self.middle_layers = []
param = {
"id": 0,
"rng": rng,
"theano_rng": theano_rng,
"dim_in": data_dim,
"dim_out": (kernel_num[0], row_num, col_num),
"input_receptive_field": (kernel_size, kernel_size),
"transition_receptive_field": (kernel_size, kernel_size),
"minibatch_size": minibatch_size,
"input": self.reshape_input,
"n_steps": config['input_seq_length']
}
self.middle_layers.append(ConvLSTMLayer(param))
param = {
"id": 1,
"rng": rng,
"theano_rng": theano_rng,
"dim_in": (kernel_num[0], row_num, col_num),
"dim_out": (kernel_num[1], row_num, col_num),
"input_receptive_field": (kernel_size, kernel_size),
"transition_receptive_field": (kernel_size, kernel_size),
"minibatch_size": minibatch_size,
"input": self.middle_layers[0].output,
"n_steps": config['input_seq_length']
}
self.middle_layers.append(ConvLSTMLayer(param))
param = {
"id": 2,
"rng": rng,
"theano_rng": theano_rng,
"dim_in": (kernel_num[1], row_num, col_num),
"dim_out": (kernel_num[2], row_num, col_num),
"input_receptive_field": (kernel_size, kernel_size),
"transition_receptive_field": (kernel_size, kernel_size),
"minibatch_size": minibatch_size,
"input": self.middle_layers[1].output,
"n_steps": config['input_seq_length']
}
self.middle_layers.append(ConvLSTMLayer(param))
param = {
"id": 3,
"rng": rng,
"theano_rng": theano_rng,
"dim_in": data_dim,
"dim_out": (kernel_num[0], row_num, col_num),
"input_receptive_field": (kernel_size, kernel_size),
"transition_receptive_field": (kernel_size, kernel_size),
"init_hidden_state": self.middle_layers[0].output[-1],
"init_cell_state": self.middle_layers[0].cell_output[-1],
"minibatch_size": minibatch_size,
"input": None,
"n_steps": config['output_seq_length'] - 1
}
self.middle_layers.append(ConvLSTMLayer(param))
param = {
"id": 4,
"rng": rng,
"theano_rng": theano_rng,
"dim_in": (kernel_num[0], row_num, col_num),
"dim_out": (kernel_num[1], row_num, col_num),
"input_receptive_field": (kernel_size, kernel_size),
"transition_receptive_field": (kernel_size, kernel_size),
"init_hidden_state": self.middle_layers[1].output[-1],
"init_cell_state": self.middle_layers[1].cell_output[-1],
"minibatch_size": minibatch_size,
"input": self.middle_layers[3].output,
"n_steps": config['output_seq_length'] - 1
}
self.middle_layers.append(ConvLSTMLayer(param))
param = {
"id": 5,
"rng": rng,
"theano_rng": theano_rng,
"dim_in": (kernel_num[1], row_num, col_num),
"dim_out": (kernel_num[2], row_num, col_num),
"input_receptive_field": (kernel_size, kernel_size),
"transition_receptive_field": (kernel_size, kernel_size),
"init_hidden_state": self.middle_layers[2].output[-1],
"init_cell_state": self.middle_layers[2].cell_output[-1],
"minibatch_size": minibatch_size,
"input": self.middle_layers[4].output,
"n_steps": config['output_seq_length'] - 1
}
self.middle_layers.append(ConvLSTMLayer(param))
param = {
"id":
6,
"rng":
rng,
"theano_rng":
theano_rng,
"dim_in":
(kernel_num[0] + kernel_num[1] + kernel_num[2], row_num, col_num),
"dim_out":
data_dim,
"input_receptive_field": (1, 1),
"input_stride": (1, 1),
"activation":
"sigmoid",
"minibatch_size":
minibatch_size,
"conv_type":
"same",
"input":
TT.concatenate([
TT.concatenate([
self.middle_layers[0].output[-1:],
self.middle_layers[1].output[-1:],
self.middle_layers[2].output[-1:]
],
axis=2),
TT.concatenate([
self.middle_layers[3].output, self.middle_layers[4].output,
self.middle_layers[5].output
],
axis=2)
])
}
self.middle_layers.append(ConvForwardLayer(param))
# dim_in = (16, 25, 25)
param = {
"id": "cost",
"rng": rng,
"theano_rng": theano_rng, # "cost_func": "BinaryCrossEntropy",
"cost_func": config['cost_func'],
"dim_in": data_dim,
"dim_out": (1, 1, 1),
"minibatch_size": minibatch_size,
"input": self.middle_layers[-1].output,
"target": self.reshape_output
}
self.cost_layer = ElementwiseCostLayer(param)
self.outputs = [{
"name": "prediction",
"value": self.middle_layers[-1].output
}]
model_param = {
'interface_layer': self.interface_layer,
'middle_layers': self.middle_layers,
'cost_layer': self.cost_layer,
#'combine_interface_layer': self.combine_interface_layer,
'outputs': self.outputs,
'errors': None,
'name': "DeepWeatherModel-test",
'problem_type': "regression"
}
super(FactorWeatherModel, self).__init__(model_param)