def __init__(self, input_dim, f1, f2, d1, num_classes):
'''
Parameters
----------
input_dim:
size of input based on training data
n1 : int
The number of neurons in the first hidden layer
num_classes : int
The number of classes predicted by the model'''
init_kwargs = {'gain': np.sqrt(2)}
self.conv1 = conv(input_dim,
f1,
5,
5,
weight_initializer=glorot_uniform,
weight_kwargs=init_kwargs)
self.conv2 = conv(f1,
f2,
5,
5,
weight_initializer=glorot_uniform,
weight_kwargs=init_kwargs)
self.dense1 = dense(f2 * 37 * 37,
d1,
weight_initializer=glorot_uniform,
weight_kwargs=init_kwargs)
self.dense2 = dense(d1,
num_classes,
weight_initializer=glorot_uniform,
weight_kwargs=init_kwargs)
def __init__(self):
self.conv1 = conv(1,
5,
2,
2,
stride=2,
padding=0,
weight_initializer=glorot_uniform)
self.conv2 = conv(5,
10,
2,
2,
stride=1,
padding=0,
weight_initializer=glorot_uniform)
self.dense1 = dense(360, 300, weight_initializer=glorot_uniform)
self.dense2 = dense(300, 300, weight_initializer=glorot_uniform)
self.dense3 = dense(300, 5, weight_initializer=glorot_uniform)
self.layers = (self.conv1, self.conv2, self.dense1, self.dense2,
self.dense3)
self.tensors = []
for layer in self.layers:
for parameter in layer.parameters:
self.tensors.append(parameter)
self.weights = [parameter.data for parameter in self.tensors]
def __init__(self):
params = np.load("params.npy")
#this gain is a parameter for the weight initializiation function glorot_uniform
#which you can read more about in the documentation, but it isn't crucial for now
#If you would like to read more about how Xavier Glorot explains the rationalization behind these weight initializations,
#look here for his paper written with Yoshua Bengio. (http://proceedings.mlr.press/v9/glorot10a/glorot10a.pdf)
init_kwargs = {'gain': np.sqrt(2)}
#We will use a dropout probability of 0.5 so that values are randomly set to 0 in our data
self.dropout_prob = 0.5
#initialize your two dense and convolution layers as class attributes using the functions imported from MyNN
#We will use weight_initializer=glorot_uniform for all 4 layers
#You know the input size of your first convolution layer. Try messing around with the output size but make sure that the following
#layers dimmensions line up. For your first convolution layer start with input = 1, output = 20, filter_dims = 5, stride = 5,
#padding = 0
self.dense1 = dense(180,
200,
weight_initializer=glorot_uniform,
weight_kwargs=init_kwargs)
self.dense2 = dense(200,
10,
weight_initializer=glorot_uniform,
weight_kwargs=init_kwargs)
self.conv1 = conv(1,
20, (5, 5),
stride=1,
padding=0,
weight_initializer=glorot_uniform,
weight_kwargs=init_kwargs)
self.conv2 = conv(20,
20, (2, 2),
stride=2,
padding=0,
weight_initializer=glorot_uniform,
weight_kwargs=init_kwargs)
self.dropout = dropout(self.dropout_prob)
self.conv1.weight = Tensor(params[0])
self.conv1.bias = Tensor(params[1])
self.conv2.weight = Tensor(params[2])
self.conv2.bias = Tensor(params[3])
self.dense1.weight = Tensor(params[4])
self.dense1.bias = Tensor(params[5])
self.dense2.weight = Tensor(params[6])
self.dense2.bias = Tensor(params[7])
def __init__(self):
#Check this: the 3s-- kernel size
gain = {'gain': np.sqrt(2)}
self.conv1 = conv(3, 20, (5,5), weight_initializer=glorot_uniform,
weight_kwargs=gain)
self.conv2 = conv(20, 10, (5,5), weight_initializer=glorot_uniform,
weight_kwargs=gain)
#Check the dimensions on this:
self.dense3 = dense(49000 , 232, weight_initializer=glorot_uniform,
weight_kwargs=gain)
def __init__(self):
self.conv1 = conv(1,
10,
5,
padding=0,
weight_initializer=glorot_uniform)
self.conv2 = conv(5,
20,
5,
padding=0,
weight_initializer=glorot_uniform)
self.dense1 = dense(290, 20, weight_initializer=glorot_uniform)
self.dense2 = dense(20, 2, weight_initializer=glorot_uniform)
def __init__(self):
""" Initializes model layers and weights. """
# <COGINST>
init_kwargs = {'gain': np.sqrt(2)}
self.conv1 = conv(200, 250, 2, stride = 1, weight_initializer = glorot_normal, weight_kwargs = init_kwargs)
self.dense1 = dense(250, 250, weight_initializer = glorot_normal, weight_kwargs = init_kwargs)
self.dense2 = dense(250,1, weight_initializer = glorot_normal, weight_kwargs = init_kwargs)
def __init__(self):
self.conv1 = conv(1,
50,
3,
3,
stride=1,
weight_initializer=glorot_uniform)
self.conv2 = conv(50,
20,
3,
3,
stride=1,
weight_initializer=glorot_uniform)
self.dense1 = dense(180, 50, weight_initializer=glorot_uniform)
self.dense2 = dense(50, 29, weight_initializer=glorot_uniform)
pass
def __init__(self):
self.conv1 = conv(1,
50, (5, 5),
stride=5,
weight_initializer=glorot_uniform,
weight_kwargs=gain)
self.conv2 = conv(50,
20, (2, 2),
stride=2,
weight_initializer=glorot_uniform,
weight_kwargs=gain)
self.dense1 = dense(500,
50,
weight_initializer=glorot_uniform,
weight_kwargs=gain)
self.dense2 = dense(50,
29,
weight_initializer=glorot_uniform,
weight_kwargs=gain)
pass
def __init__(self, dim_in=48, num_out=7, load=False):
self.conv1 = conv(1, 32, 2, 2, stride=1, padding=0, weight_initializer=Model.init)
self.conv2 = conv(32, 64, 3, 3, stride=2, padding=1, weight_initializer=Model.init)
self.conv3 = conv(64, 128, 2, 2, stride=2, weight_initializer=Model.init)
self.conv4 = conv(128, 256, 3, 3, stride=3, weight_initializer=Model.init)
self.dense1 = dense(256, 512, weight_initializer=Model.init)
self.dense2 = dense(512, num_out, weight_initializer=Model.init)
if (load):
data = np.load("data/npmodelParam.npz")
self.conv1.weight = data["l1w"]
self.conv1.bias = data["l1b"]
self.conv2.weight = data["l2w"]
self.conv2.bias = data["l2b"]
self.conv3.weight = data["l3w"]
self.conv3.bias = data["l3b"]
self.conv4.weight = data["l4w"]
self.conv4.bias = data["l4b"]
self.dense1.weight = data["l5w"]
self.dense1.bias = data["l5b"]
self.dense2.weight = data["l6w"]
self.dense2.bias = data["l6b"]
def policyForward(self, data):
data = mg.Tensor(data)
conv1 = conv(1,
20,
5,
5,
stride=1,
padding=0,
weight_initializer=glorot_uniform)
for i in range(len(self.weights) - 1):
data = mg.matmul(data, self.weights[i]) # hidden layers
data[data < 0] = 0 # ReLU
return mg.matmul(data, self.weights[-1]) # outNeurons
