add files

flops.py

@@ -0,0 +1,32 @@
+import torch
+# from ptflops import get_model_complexity_info
+# from model import NetworkCIFAR as Network
+from thop import profile
+from model_search import Network as Network
+from model import NetworkCIFAR as Network_CIFAR
+import genotypes
+
+# torch.cuda.set_device(0)
+init_channels = 36
+CIFAR_CLASSES = 10
+layers = 20
+auxiliary = True
+input = torch.randn(1, 3, 32, 32)
+# input = input.cuda()
+genotype = eval("genotypes.%s" % 'SWD_NAS')
+
+# genotype = eval("Genotype(normal=[('sep_conv_3x3', 0), ('sep_conv_3x3', 1), ('sep_conv_3x3', 0), ('sep_conv_3x3', 1), ('sep_conv_3x3', 0), ('skip_connect', 0), ('skip_connect', 0), ('dil_conv_5x5', 2)], normal_concat=range(2, 6), reduce=[('max_pool_3x3', 0), ('max_pool_3x3', 1), ('skip_connect', 2), ('max_pool_3x3', 0), ('max_pool_3x3', 0), ('skip_connect', 2), ('skip_connect', 2), ('avg_pool_3x3', 1)], reduce_concat=range(2, 6))")
+
+criterion = torch.nn.CrossEntropyLoss()
+criterion = criterion.cuda()
+
+with torch.cuda.device(0):
+    model = Network_CIFAR(init_channels, CIFAR_CLASSES, layers, auxiliary, genotype)
+    # net = Network(init_channels, CIFAR_CLASSES, layers, criterion)
+    # model = model.cuda()
+    model.drop_path_prob = 0.2
+    flops, params = profile(model, inputs=(input,))
+    # macs, params = get_model_complexity_info(net, (3, 32, 32), as_strings=True,
+    #                                        print_per_layer_stat=True, verbose=True)
+print(f"FLOPs: {flops / 1e6} M FLOPs")
+print(f"Number of Parameters: {params}")

genotypes.py

@@ -0,0 +1,95 @@
+from collections import namedtuple
+
+Genotype = namedtuple('Genotype', 'normal normal_concat reduce reduce_concat')
+Normal_Genotype = namedtuple('Normal_Genotype', 'normal normal_concat')
+Reduce_Genotype = namedtuple('Reduce_Genotype', 'reduce reduce_concat')
+
+PARAMS = {'conv_3x1_1x3':864, 'conv_7x1_1x7':2016, 'sep_conv_7x7': 1464, 'conv 3x3':1296, 'sep_conv_5x5': 888, 'sep_conv_3x3':504, 'dil_conv_5x5': 444, 'conv 1x1':144, 'dil_conv_3x3':252, 'skip_connect':0, 'none':0, 'max_pool_3x3':0, 'avg_pool_3x3':0}
+
+PRIMITIVES = [
+    'sep_conv_3x3',
+    'sep_conv_5x5',
+    'dil_conv_3x3',
+    'dil_conv_5x5',
+    'skip_connect',
+    'avg_pool_3x3',
+    'max_pool_3x3',
+    'none'
+]
+
+NASNet = Genotype(
+  normal = [
+    ('sep_conv_5x5', 1),
+    ('sep_conv_3x3', 0),
+    ('sep_conv_5x5', 0),
+    ('sep_conv_3x3', 0),
+    ('avg_pool_3x3', 1),
+    ('skip_connect', 0),
+    ('avg_pool_3x3', 0),
+    ('avg_pool_3x3', 0),
+    ('sep_conv_3x3', 1),
+    ('skip_connect', 1),
+  ],
+  normal_concat = [2, 3, 4, 5, 6],
+  reduce = [
+    ('sep_conv_5x5', 1),
+    ('sep_conv_7x7', 0),
+    ('max_pool_3x3', 1),
+    ('sep_conv_7x7', 0),
+    ('avg_pool_3x3', 1),
+    ('sep_conv_5x5', 0),
+    ('skip_connect', 3),
+    ('avg_pool_3x3', 2),
+    ('sep_conv_3x3', 2),
+    ('max_pool_3x3', 1),
+  ],
+  reduce_concat = [4, 5, 6],
+)
+
+AmoebaNet = Genotype(
+  normal = [
+    ('avg_pool_3x3', 0),
+    ('max_pool_3x3', 1),
+    ('sep_conv_3x3', 0),
+    ('sep_conv_5x5', 2),
+    ('sep_conv_3x3', 0),
+    ('avg_pool_3x3', 3),
+    ('sep_conv_3x3', 1),
+    ('skip_connect', 1),
+    ('skip_connect', 0),
+    ('avg_pool_3x3', 1),
+    ],
+  normal_concat = [4, 5, 6],
+  reduce = [
+    ('avg_pool_3x3', 0),
+    ('sep_conv_3x3', 1),
+    ('max_pool_3x3', 0),
+    ('sep_conv_7x7', 2),
+    ('sep_conv_7x7', 0),
+    ('avg_pool_3x3', 1),
+    ('max_pool_3x3', 0),
+    ('max_pool_3x3', 1),
+    ('conv_7x1_1x7', 0),
+    ('sep_conv_3x3', 5),
+  ],
+  reduce_concat = [3, 4, 6]
+)
+
+DARTS_V1 = Genotype(normal=[('sep_conv_3x3', 1), ('sep_conv_3x3', 0), ('skip_connect', 0), ('sep_conv_3x3', 1), ('skip_connect', 0), ('sep_conv_3x3', 1), ('sep_conv_3x3', 0), ('skip_connect', 2)], normal_concat=[2, 3, 4, 5], reduce=[('max_pool_3x3', 0), ('max_pool_3x3', 1), ('skip_connect', 2), ('max_pool_3x3', 0), ('max_pool_3x3', 0), ('skip_connect', 2), ('skip_connect', 2), ('avg_pool_3x3', 0)], reduce_concat=[2, 3, 4, 5])
+DARTS_V2 = Genotype(normal=[('sep_conv_3x3', 0), ('sep_conv_3x3', 1), ('sep_conv_3x3', 0), ('sep_conv_3x3', 1), ('sep_conv_3x3', 1), ('skip_connect', 0), ('skip_connect', 0), ('dil_conv_3x3', 2)], normal_concat=[2, 3, 4, 5], reduce=[('max_pool_3x3', 0), ('max_pool_3x3', 1), ('skip_connect', 2), ('max_pool_3x3', 1), ('max_pool_3x3', 0), ('skip_connect', 2), ('skip_connect', 2), ('max_pool_3x3', 1)], reduce_concat=[2, 3, 4, 5])
+
+DARTS = DARTS_V1
+
+Cell_0 = Normal_Genotype(normal=[('sep_conv_3x3', 0), ('dil_conv_3x3', 1), ('dil_conv_5x5', 0), ('skip_connect', 1), ('sep_conv_3x3', 0), ('sep_conv_5x5', 3), ('dil_conv_5x5', 0), ('sep_conv_3x3', 3)], normal_concat=range(2, 6))
+Cell_1 = Reduce_Genotype(reduce=[('max_pool_3x3', 0), ('sep_conv_5x5', 1), ('skip_connect', 0), ('sep_conv_5x5', 1), ('max_pool_3x3', 0), ('skip_connect', 1), ('sep_conv_5x5', 0), ('sep_conv_3x3', 1)], reduce_concat=range(2, 6))
+Cell_2 = Normal_Genotype(normal=[('skip_connect', 0), ('sep_conv_3x3', 1), ('dil_conv_5x5', 0), ('sep_conv_3x3', 1), ('dil_conv_3x3', 0), ('sep_conv_5x5', 1), ('sep_conv_3x3', 1), ('dil_conv_5x5', 4)], normal_concat=range(2, 6))
+Cell_3 = Reduce_Genotype(reduce=[('skip_connect', 1), ('skip_connect', 0), ('dil_conv_3x3', 1), ('dil_conv_3x3', 0), ('avg_pool_3x3', 0), ('max_pool_3x3', 1), ('avg_pool_3x3', 0), ('max_pool_3x3', 1)], reduce_concat=range(2, 6))
+Cell_4 = Normal_Genotype(normal=[('skip_connect', 0), ('dil_conv_5x5', 1), ('sep_conv_3x3', 0), ('sep_conv_3x3', 1), ('sep_conv_5x5', 0), ('max_pool_3x3', 2), ('sep_conv_5x5', 0), ('sep_conv_3x3', 1)], normal_concat=range(2, 6))
+
+Cell_0 = Normal_Genotype(normal=[('skip_connect', 0), ('sep_conv_5x5', 1), ('max_pool_3x3', 0), ('dil_conv_3x3', 1), ('sep_conv_3x3', 0), ('dil_conv_5x5', 1), ('sep_conv_5x5', 0), ('avg_pool_3x3', 1)], normal_concat=range(2, 6))
+Cell_1 = Reduce_Genotype(reduce=[('dil_conv_3x3', 0), ('dil_conv_3x3', 1), ('max_pool_3x3', 0), ('sep_conv_5x5', 1), ('dil_conv_3x3', 1), ('sep_conv_5x5', 3), ('sep_conv_5x5', 0), ('skip_connect', 1)], reduce_concat=range(2, 6))
+Cell_2 = Normal_Genotype(normal=[('skip_connect', 0), ('sep_conv_3x3', 1), ('avg_pool_3x3', 0), ('sep_conv_3x3', 2), ('sep_conv_5x5', 0), ('sep_conv_3x3', 1), ('avg_pool_3x3', 0), ('skip_connect', 1)], normal_concat=range(2, 6))
+Cell_3 = Reduce_Genotype(reduce=[('avg_pool_3x3', 0), ('sep_conv_3x3', 1), ('skip_connect', 0), ('avg_pool_3x3', 1), ('dil_conv_3x3', 0), ('dil_conv_3x3', 1), ('sep_conv_3x3', 0), ('dil_conv_5x5', 1)], reduce_concat=range(2, 6))
+Cell_4 = Normal_Genotype(normal=[('sep_conv_3x3', 0), ('sep_conv_5x5', 1), ('skip_connect', 0), ('skip_connect', 1), ('sep_conv_5x5', 0), ('dil_conv_5x5', 1), ('dil_conv_5x5', 0), ('max_pool_3x3', 1)], normal_concat=range(2, 6))
+
+SWD_NAS = [Cell_0, Cell_1, Cell_2, Cell_3, Cell_4]

inference_time.py

@@ -0,0 +1,27 @@
+import torch
+from model_search import Network as Network
+
+iteration = 50
+model = Network(16, 10, 5, None)
+# model = model.cuda()
+
+input = torch.randn(1, 3, 32, 32)
+# input = input.cuda()
+
+starter, ender = torch.cuda.Event(enable_timing=True), torch.cuda.Event(enable_timing=True)
+for _ in range(50):
+    _ = model(input)
+
+times = torch.zeros(iteration)
+with torch.no_grad():
+    for iter in range(iteration):
+        starter.record()
+        _ = model(input)
+        ender.record()
+        # Waits for everything to finish running
+        torch.cuda.synchronize()
+        times[iter] = starter.elapsed_time(ender)
+
+mean_time = times.mean().item()
+print("Inference time: {:.6f}, FPS: {} ".format(mean_time, 1000 / mean_time))
+