def conv_3x1_1x1_max_2x2_no_relu():
"""Baseline model without relu."""
graph_gen = gg.GraphGenerator()
graph_gen.add(gg.make_conv_quant, "conv1", 1, 4, (3, 1, 0))
graph_gen.add(gg.make_pool_max, "max1", 2, 2)
graph_gen.add(gg.make_conv_quant, "conv2", 4, 8, (1, 1, 0))
graph_gen.add(gg.make_pool_ave, "ave1")
return graph_gen.get_model("cnn", (1, 1, 6, 6), (1, 8, 1, 1))
def conv_3x1_1x1_max_2x2_one_channel():
"""Baseline model with only one channel in every layer."""
graph_gen = gg.GraphGenerator()
graph_gen.add(gg.make_conv_quant, "conv1", 1, 1, (3, 1, 0))
graph_gen.add(gg.make_relu, "relu1")
graph_gen.add(gg.make_pool_max, "max1", 2, 2)
graph_gen.add(gg.make_conv_quant, "conv2", 1, 1, (1, 1, 0))
graph_gen.add(gg.make_relu, "relu2")
graph_gen.add(gg.make_pool_ave, "ave1")
return graph_gen.get_model("cnn", (1, 1, 6, 6), (1, 1, 1, 1))
def conv_3x1_1x1_max_2x2():
"""Baseline model. size: 6x6 -> 4x4 -> 2x2"""
graph_gen = gg.GraphGenerator()
graph_gen.add(gg.make_conv_quant, "conv1", 1, 4, (3, 1, 0))
graph_gen.add(gg.make_relu, "relu1")
graph_gen.add(gg.make_pool_max, "max1", 2, 2)
graph_gen.add(gg.make_conv_quant, "conv2", 4, 8, (1, 1, 0))
graph_gen.add(gg.make_relu, "relu2")
graph_gen.add(gg.make_pool_ave, "ave1")
return graph_gen.get_model("cnn", (1, 1, 6, 6), (1, 8, 1, 1))
def conv_3x1_1x1_max_2x2_colored_input():
"""Baseline model with a colored input, i. e. three input channel."""
graph_gen = gg.GraphGenerator()
graph_gen.add(gg.make_conv_quant, "conv1", 3, 4, (3, 1, 0))
graph_gen.add(gg.make_relu, "relu1")
graph_gen.add(gg.make_pool_max, "max1", 2, 2)
graph_gen.add(gg.make_conv_quant, "conv2", 4, 8, (1, 1, 0))
graph_gen.add(gg.make_relu, "relu2")
graph_gen.add(gg.make_pool_ave, "ave1")
return graph_gen.get_model("cnn", (1, 3, 6, 6), (1, 8, 1, 1))
def conv_3x1_1x1_max_2x2_odd_input():
"""Baseline model with an odd input."""
graph_gen = gg.GraphGenerator()
graph_gen.add(gg.make_conv_quant, "conv1", 1, 4, (3, 1, 0))
graph_gen.add(gg.make_relu, "relu1")
graph_gen.add(gg.make_pool_max, "max1", 2, 2)
graph_gen.add(gg.make_conv_quant, "conv2", 4, 8, (1, 1, 0))
graph_gen.add(gg.make_relu, "relu2")
graph_gen.add(gg.make_pool_ave, "ave1")
return graph_gen.get_model("cnn", (1, 1, 7, 7), (1, 8, 1, 1))
def conv_3x1_1x1_max_3x3():
"""size: 14x14 -> 12x12 -> 4x4"""
graph_gen = gg.GraphGenerator()
graph_gen.add(gg.make_conv_quant, "conv1", 1, 4, (3, 1, 0))
graph_gen.add(gg.make_relu, "relu1")
graph_gen.add(gg.make_pool_max, "max1", 3, 3)
graph_gen.add(gg.make_conv_quant, "conv2", 4, 8, (1, 1, 0))
graph_gen.add(gg.make_relu, "relu2")
graph_gen.add(gg.make_pool_ave, "ave1")
return graph_gen.get_model("cnn", (1, 1, 14, 14), (1, 8, 1, 1))
def conv_2x1_1x1_max_3x2():
"""size: 16x16 -> 15x15 -> 7x7"""
graph_gen = gg.GraphGenerator()
graph_gen.add(gg.make_conv_quant, "conv1", 1, 4, (2, 1, 0))
graph_gen.add(gg.make_relu, "relu1")
graph_gen.add(gg.make_pool_max, "max1", 3, 2)
graph_gen.add(gg.make_conv_quant, "conv2", 4, 8, (1, 1, 0))
graph_gen.add(gg.make_relu, "relu2")
graph_gen.add(gg.make_pool_ave, "ave1")
return graph_gen.get_model("cnn", (1, 1, 16, 16), (1, 8, 1, 1))
def conv_3x2_1x1_max_2x1_padding():
"""size: 15x15 -> 17x17 -> 8x8 -> 7x7"""
graph_gen = gg.GraphGenerator()
graph_gen.add(gg.make_conv_quant, "conv1", 1, 4, (3, 2, 1))
graph_gen.add(gg.make_relu, "relu1")
graph_gen.add(gg.make_pool_max, "max1", 2, 1)
graph_gen.add(gg.make_conv_quant, "conv2", 4, 8, (1, 1, 0))
graph_gen.add(gg.make_relu, "relu2")
graph_gen.add(gg.make_pool_ave, "ave1")
return graph_gen.get_model("cnn", (1, 1, 15, 15), (1, 8, 1, 1))
def conv_3x1_1x1_max_2x1():
"""size: 12x12 -> 10x10 -> 9x9"""
graph_gen = gg.GraphGenerator()
graph_gen.add(gg.make_conv_quant, "conv1", 1, 4, (3, 1, 0))
graph_gen.add(gg.make_relu, "relu1")
graph_gen.add(gg.make_pool_max, "max1", 2, 1)
graph_gen.add(gg.make_conv_quant, "conv2", 4, 8, (1, 1, 0))
graph_gen.add(gg.make_relu, "relu2")
graph_gen.add(gg.make_pool_ave, "ave1")
return graph_gen.get_model("cnn", (1, 1, 12, 12), (1, 8, 1, 1))
def conv_3x1_1x1_max_2x2_odd_channel():
"""Baseline model with an odd number of channel. The channel depth is
specified on purpose. There was a bug with channel depth = 2^x+1."""
graph_gen = gg.GraphGenerator()
graph_gen.add(gg.make_conv_quant, "conv1", 1, 5, (3, 1, 0))
graph_gen.add(gg.make_relu, "relu1")
graph_gen.add(gg.make_pool_max, "max1", 2, 2)
graph_gen.add(gg.make_conv_quant, "conv2", 5, 9, (1, 1, 0))
graph_gen.add(gg.make_relu, "relu2")
graph_gen.add(gg.make_pool_ave, "ave1")
return graph_gen.get_model("cnn", (1, 1, 6, 6), (1, 9, 1, 1))
def conv_3x3_2x2_1x1():
"""size: 12x12 -> 4x4 -> 2x2"""
graph_gen = gg.GraphGenerator()
graph_gen.add(gg.make_conv_quant, "conv1", 1, 4, (3, 3, 0))
graph_gen.add(gg.make_relu, "relu1")
graph_gen.add(gg.make_conv_quant, "conv2", 4, 6, (2, 2, 0))
graph_gen.add(gg.make_relu, "relu2")
graph_gen.add(gg.make_conv_quant, "conv3", 6, 8, (1, 1, 0))
graph_gen.add(gg.make_relu, "relu3")
graph_gen.add(gg.make_pool_ave, "ave1")
return graph_gen.get_model("cnn", (1, 1, 8, 8), (1, 8, 1, 1))
def conv_2x_3x1_1x1_max_2x2_padding():
"""size: 8x8 -> 8x8 -> 4x4 -> 4x4 -> 2x2"""
graph_gen = gg.GraphGenerator()
graph_gen.add(gg.make_conv_quant, "conv1", 1, 8, (3, 1, 1))
graph_gen.add(gg.make_relu, "relu1")
graph_gen.add(gg.make_pool_max, "max1", 2, 2)
graph_gen.add(gg.make_conv_quant, "conv2", 8, 16, (3, 1, 1))
graph_gen.add(gg.make_relu, "relu2")
graph_gen.add(gg.make_pool_max, "max2", 2, 2)
graph_gen.add(gg.make_conv_quant, "conv3", 16, 32, (1, 1, 0))
graph_gen.add(gg.make_relu, "relu3")
graph_gen.add(gg.make_conv_quant, "conv4", 32, 8, (1, 1, 0))
graph_gen.add(gg.make_relu, "relu4")
graph_gen.add(gg.make_pool_ave, "ave1")
return graph_gen.get_model("cnn", (1, 1, 14, 14), (1, 8, 1, 1))
def conv_4x3x1_1x1():
"""size: 10x10 -> 8x8 -> 6x6 -> 4x4 -> 2x2"""
graph_gen = gg.GraphGenerator()
graph_gen.add(gg.make_conv_quant, "conv1", 1, 8, (3, 1, 0))
graph_gen.add(gg.make_relu, "relu1")
graph_gen.add(gg.make_conv_quant, "conv2", 8, 10, (3, 1, 0))
graph_gen.add(gg.make_relu, "relu2")
graph_gen.add(gg.make_conv_quant, "conv3", 10, 12, (3, 1, 0))
graph_gen.add(gg.make_relu, "relu3")
graph_gen.add(gg.make_conv_quant, "conv4", 12, 14, (3, 1, 0))
graph_gen.add(gg.make_relu, "relu4")
graph_gen.add(gg.make_conv_quant, "conv5", 14, 16, (1, 1, 0))
graph_gen.add(gg.make_relu, "relu5")
graph_gen.add(gg.make_pool_ave, "ave1")
return graph_gen.get_model("cnn", (1, 1, 10, 10), (1, 16, 1, 1))
def conv_2x_3x1_1x1_max_2x2_mt():
"""Model of my master thesis, for comparison.
size: 48x24 -> 24x12 -> 12x6"""
graph_gen = gg.GraphGenerator()
graph_gen.add(gg.make_conv_quant, "conv1", 1, 16, (3, 1, 1))
graph_gen.add(gg.make_relu, "relu1")
graph_gen.add(gg.make_pool_max, "max1", 2, 2)
graph_gen.add(gg.make_conv_quant, "conv2", 16, 32, (3, 1, 1))
graph_gen.add(gg.make_relu, "relu2")
graph_gen.add(gg.make_pool_max, "max2", 2, 2)
graph_gen.add(gg.make_conv_quant, "conv3", 32, 64, (1, 1, 0))
graph_gen.add(gg.make_relu, "relu3")
graph_gen.add(gg.make_conv_quant, "conv4", 64, 2, (1, 1, 0))
graph_gen.add(gg.make_relu, "relu4")
graph_gen.add(gg.make_pool_ave, "ave1")
return graph_gen.get_model("cnn", (1, 1, 48, 24), (1, 2, 1, 1))
