LaNAS/Distributed_LaNAS/clientX/nasnet_set.py [64:310]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - def gen_code_from_list(sample, node_num=6): node = [[-1 for col in range(4)] for row in range(node_num)] for i in range(node_num): for j in range(4): if j <= 1: node[i][j] = sample[i * 2 + j] else: node[i][j] = sample[i * 2 + j + (node_num - 1) * 2] return node operations = ['sep_conv_3x3', 'max_pool_3x3', 'skip_connect', 'sep_conv_5x5'] def gen_all_nets(max_node=Max_node): block = [] all_connections = [] for i in range(max_node): connections = [] for j in range(i+2): connections.append(j) all_connections.append(connections) b = [[] for i in range(max_node)] for i in range(max_node): for left_type in range(len(operations)): for right_type in range(len(operations)): for left_input in all_connections[i]: for right_input in all_connections[i]: if [right_type, left_type, left_input, right_input] not in b[i]: b[i].append([left_type, right_type, left_input, right_input]) sample_nums = 1 for i in range(max_node): sample_nums *= len(b[i]) print(sample_nums) for o in b[0]: for t in b[1]: for th in b[2]: for f in b[3]: if edit_distance([o, t, th, f], root) == 6: block.append([o, t, th, f]) print(len(block)) return block def edit_distance(net_1, net_2): distance = 0 for i in range(len(net_1)): if [net_1[i][2], net_1[i][3]] == [net_2[i][2], net_2[i][3]]: if net_1[i][0] != net_2[i][0]: distance += 1 if net_1[i][1] != net_2[i][1]: distance += 1 elif [net_1[i][2], net_1[i][3]] == [net_2[i][3], net_2[i][2]]: if net_1[i][0] != net_2[i][1]: distance += 1 if net_1[i][1] != net_2[i][0]: distance += 1 elif [net_1[i][0], net_1[i][2]] == [net_2[i][0], net_2[i][2]]: if net_1[i][1] != net_2[i][1]: distance += 1 if net_1[i][3] != net_2[i][3]: distance += 1 elif [net_1[i][0], net_1[i][2]] == [net_2[i][1], net_2[i][3]]: if net_1[i][1] != net_2[i][0]: distance += 1 if net_1[i][3] != net_2[i][2]: distance += 1 elif [net_1[i][1], net_1[i][3]] == [net_2[i][1], net_2[i][3]]: if net_1[i][0] != net_2[i][0]: distance += 1 if net_1[i][2] != net_2[i][2]: distance += 1 elif [net_1[i][1], net_1[i][3]] == [net_2[i][0], net_2[i][2]]: if net_1[i][0] != net_2[i][1]: distance += 1 if net_1[i][2] != net_2[i][3]: distance += 1 else: if net_1[i][0] in net_2[i][:2]: if net_1[i][0] == net_2[i][0]: if net_1[i][2] != net_2[i][2]: distance += 1 if net_1[i][1] != net_2[i][1]: distance += 1 if net_1[i][3] != net_2[i][3]: distance += 1 else: if net_1[i][2] != net_2[i][3]: distance += 1 if net_1[i][1] != net_2[i][0]: distance += 1 if net_1[i][3] != net_2[i][2]: distance += 1 elif net_1[i][1] in net_2[i][:2]: if net_1[i][1] == net_2[i][1]: if net_1[i][2] != net_2[i][2]: distance += 1 if net_1[i][0] != net_2[i][0]: distance += 1 if net_1[i][3] != net_2[i][3]: distance += 1 else: if net_1[i][2] != net_2[i][3]: distance += 1 if net_1[i][0] != net_2[i][1]: distance += 1 if net_1[i][3] != net_2[i][2]: distance += 1 else: distance += 2 if net_1[i][2] in net_2[i][-2:]: if net_1[i][2] == net_2[i][2]: if net_1[i][3] != net_2[i][3]: distance += 1 elif net_1[i][2] == net_2[i][3]: if net_1[i][3] != net_2[i][2]: distance += 1 elif net_1[i][3] in net_2[i][-2:]: if net_1[i][3] == net_2[i][3]: if net_1[i][2] != net_2[i][3]: distance += 1 elif net_1[i][3] == net_2[i][2]: if net_1[i][2] != net_2[i][3]: distance += 1 else: distance += 2 return distance def gen_code(max_node=Max_node): normal_node = [[-1 for col in range(4)] for row in range(max_node)] reduction_node = [[-1 for col in range(4)] for row in range(max_node)] for i in range(max_node): for j in range(4): if j <= 1: normal_node[i][j] = random.randint(0, len(operations) - 1) reduction_node[i][j] = random.randint(0, len(operations) - 1) else: normal_node[i][j] = random.randint(0, i+1) reduction_node[i][j] = random.randint(0, i+1) concat_code = [normal_node, reduction_node] return concat_code def get_node_depth(cur_node, node_code): ## get the depth for current node assert cur_node[-1] != -1 assert cur_node[-2] != -1 depth_left = 1 depth_right = 1 if cur_node[-1] == 0 or cur_node[-1] == 1: if cur_node[-2] == 0 or cur_node[-2] == 1: return 1 if not (cur_node[-2] == 0 or cur_node[-2] == 1): depth_left += get_node_depth(cur_node=node_code[cur_node[-2] - 2], node_code=node_code) if not (cur_node[-1] == 0 or cur_node[-1] == 1): depth_right += get_node_depth(cur_node=node_code[cur_node[-1] - 2], node_code=node_code) return max(depth_left, depth_right) def split_and_output(net_code, splitted_num=500, to_file=False): sample = int(len(net_code) / splitted_num) gpu_nodes = [[] for i in range(splitted_num)] for i in range(splitted_num): gpu_nodes[i] = net_code[i * sample:(i + 1) * sample] remainer = net_code[splitted_num * sample:] for i in range(len(remainer)): gpu_nodes[i].append(remainer[i]) for i in range(splitted_num): path = './gpu_files/' + 'gpu' + str(i) if not os.path.exists(path): os.makedirs(path) if to_file: for i in range(splitted_num): outfile = './gpu_files/' + 'gpu' + str(i) + '/splitted_lstm_dataset' with open(outfile, 'w') as json_data: json.dump(gpu_nodes[i], json_data) def load_from_file(id=0): outfile = 'gpu' + str(id) + '.txt' net_list = [] with open(outfile, 'w') as f: while True: line = f.readline().strip('\n') if line is None: break net_list.append([json.loads(line), json.loads(line)]) return net_list def translator(code, max_node=6): # input: code type # output: geno type n = 0 normal = [] normal_concat = [] reduce_concat = [] for i in range(max_node+2): normal_concat.append(i) reduce_concat.append(i) reduce = [] for cell in range(len(code)): if cell == 0: # for normal cell for block in range(len(code[cell])): normal.append((operations[code[cell][block][0]], code[cell][block][2])) normal.append((operations[code[cell][block][1]], code[cell][block][3])) if code[cell][block][2] in normal_concat: normal_concat.remove(code[cell][block][2]) if code[cell][block][3] in normal_concat: normal_concat.remove(code[cell][block][3]) else: # for reduction cell for block in range(len(code[cell])): reduce.append((operations[code[cell][block][0]], code[cell][block][2])) reduce.append((operations[code[cell][block][1]], code[cell][block][3])) if code[cell][block][2] in reduce_concat: reduce_concat.remove(code[cell][block][2]) if code[cell][block][3] in reduce_concat: reduce_concat.remove(code[cell][block][3]) if 0 in reduce_concat: reduce_concat.remove(0) if 1 in reduce_concat: reduce_concat.remove(1) gen_type = Genotype(normal=normal, normal_concat=normal_concat, reduce=reduce, reduce_concat=reduce_concat) # print(normal, normal_concat, reduce, reduce_concat) return gen_type - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - LaNAS/LaNet/CIFAR10/nasnet_set.py [20:271]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - def gen_code_from_list(sample, node_num=6): node = [[-1 for col in range(4)] for row in range(node_num)] for i in range(node_num): for j in range(4): if j <= 1: node[i][j] = sample[i * 2 + j] else: node[i][j] = sample[i * 2 + j + (node_num - 1) * 2] return node operations = ['sep_conv_3x3', 'max_pool_3x3', 'skip_connect', 'sep_conv_5x5'] def gen_all_nets(max_node=Max_node): block = [] all_connections = [] for i in range(max_node): connections = [] for j in range(i+2): connections.append(j) all_connections.append(connections) b = [[] for i in range(max_node)] for i in range(max_node): for left_type in range(len(operations)): for right_type in range(len(operations)): for left_input in all_connections[i]: for right_input in all_connections[i]: if [right_type, left_type, left_input, right_input] not in b[i]: b[i].append([left_type, right_type, left_input, right_input]) sample_nums = 1 for i in range(max_node): sample_nums *= len(b[i]) print(sample_nums) for o in b[0]: for t in b[1]: for th in b[2]: for f in b[3]: if edit_distance([o, t, th, f], root) == 6: block.append([o, t, th, f]) # for i in block: # print(i) print(len(block)) return block def edit_distance(net_1, net_2): distance = 0 for i in range(len(net_1)): if [net_1[i][2], net_1[i][3]] == [net_2[i][2], net_2[i][3]]: if net_1[i][0] != net_2[i][0]: distance += 1 if net_1[i][1] != net_2[i][1]: distance += 1 elif [net_1[i][2], net_1[i][3]] == [net_2[i][3], net_2[i][2]]: if net_1[i][0] != net_2[i][1]: distance += 1 if net_1[i][1] != net_2[i][0]: distance += 1 elif [net_1[i][0], net_1[i][2]] == [net_2[i][0], net_2[i][2]]: if net_1[i][1] != net_2[i][1]: distance += 1 if net_1[i][3] != net_2[i][3]: distance += 1 elif [net_1[i][0], net_1[i][2]] == [net_2[i][1], net_2[i][3]]: if net_1[i][1] != net_2[i][0]: distance += 1 if net_1[i][3] != net_2[i][2]: distance += 1 elif [net_1[i][1], net_1[i][3]] == [net_2[i][1], net_2[i][3]]: if net_1[i][0] != net_2[i][0]: distance += 1 if net_1[i][2] != net_2[i][2]: distance += 1 elif [net_1[i][1], net_1[i][3]] == [net_2[i][0], net_2[i][2]]: if net_1[i][0] != net_2[i][1]: distance += 1 if net_1[i][2] != net_2[i][3]: distance += 1 else: if net_1[i][0] in net_2[i][:2]: if net_1[i][0] == net_2[i][0]: if net_1[i][2] != net_2[i][2]: distance += 1 if net_1[i][1] != net_2[i][1]: distance += 1 if net_1[i][3] != net_2[i][3]: distance += 1 else: if net_1[i][2] != net_2[i][3]: distance += 1 if net_1[i][1] != net_2[i][0]: distance += 1 if net_1[i][3] != net_2[i][2]: distance += 1 elif net_1[i][1] in net_2[i][:2]: if net_1[i][1] == net_2[i][1]: if net_1[i][2] != net_2[i][2]: distance += 1 if net_1[i][0] != net_2[i][0]: distance += 1 if net_1[i][3] != net_2[i][3]: distance += 1 else: if net_1[i][2] != net_2[i][3]: distance += 1 if net_1[i][0] != net_2[i][1]: distance += 1 if net_1[i][3] != net_2[i][2]: distance += 1 else: distance += 2 if net_1[i][2] in net_2[i][-2:]: if net_1[i][2] == net_2[i][2]: if net_1[i][3] != net_2[i][3]: distance += 1 elif net_1[i][2] == net_2[i][3]: if net_1[i][3] != net_2[i][2]: distance += 1 elif net_1[i][3] in net_2[i][-2:]: if net_1[i][3] == net_2[i][3]: if net_1[i][2] != net_2[i][3]: distance += 1 elif net_1[i][3] == net_2[i][2]: if net_1[i][2] != net_2[i][3]: distance += 1 else: distance += 2 return distance def gen_code(max_node=Max_node): normal_node = [[-1 for col in range(4)] for row in range(max_node)] reduction_node = [[-1 for col in range(4)] for row in range(max_node)] for i in range(max_node): for j in range(4): if j <= 1: normal_node[i][j] = random.randint(0, len(operations) - 1) reduction_node[i][j] = random.randint(0, len(operations) - 1) else: normal_node[i][j] = random.randint(0, i+1) reduction_node[i][j] = random.randint(0, i+1) # print(get_node_depth(normal_node[i], normal_node)) concat_code = [normal_node, reduction_node] return concat_code def get_node_depth(cur_node, node_code): ## get the depth for current node assert cur_node[-1] != -1 assert cur_node[-2] != -1 depth_left = 1 depth_right = 1 if cur_node[-1] == 0 or cur_node[-1] == 1: if cur_node[-2] == 0 or cur_node[-2] == 1: return 1 if not (cur_node[-2] == 0 or cur_node[-2] == 1): depth_left += get_node_depth(cur_node=node_code[cur_node[-2] - 2], node_code=node_code) if not (cur_node[-1] == 0 or cur_node[-1] == 1): depth_right += get_node_depth(cur_node=node_code[cur_node[-1] - 2], node_code=node_code) return max(depth_left, depth_right) def split_and_output(net_code, splitted_num=500, to_file=False): sample = int(len(net_code) / splitted_num) gpu_nodes = [[] for i in range(splitted_num)] for i in range(splitted_num): gpu_nodes[i] = net_code[i * sample:(i + 1) * sample] remainer = net_code[splitted_num * sample:] for i in range(len(remainer)): gpu_nodes[i].append(remainer[i]) for i in range(splitted_num): path = './gpu_files/' + 'gpu' + str(i) if not os.path.exists(path): os.makedirs(path) if to_file: for i in range(splitted_num): outfile = './gpu_files/' + 'gpu' + str(i) + '/splitted_lstm_dataset' with open(outfile, 'w') as json_data: json.dump(gpu_nodes[i], json_data) def load_from_file(id=0): outfile = 'gpu' + str(id) + '.txt' net_list = [] with open(outfile, 'w') as f: while True: line = f.readline().strip('\n') if line is None: break net_list.append([json.loads(line), json.loads(line)]) return net_list def translator(code, max_node=6): # input: code type # output: geno type n = 0 normal = [] normal_concat = [] reduce_concat = [] for i in range(max_node+2): normal_concat.append(i) reduce_concat.append(i) reduce = [] for cell in range(len(code)): if cell == 0: # for normal cell for block in range(len(code[cell])): normal.append((operations[code[cell][block][0]], code[cell][block][2])) normal.append((operations[code[cell][block][1]], code[cell][block][3])) if code[cell][block][2] in normal_concat: normal_concat.remove(code[cell][block][2]) if code[cell][block][3] in normal_concat: normal_concat.remove(code[cell][block][3]) else: # for reduction cell for block in range(len(code[cell])): reduce.append((operations[code[cell][block][0]], code[cell][block][2])) reduce.append((operations[code[cell][block][1]], code[cell][block][3])) if code[cell][block][2] in reduce_concat: reduce_concat.remove(code[cell][block][2]) if code[cell][block][3] in reduce_concat: reduce_concat.remove(code[cell][block][3]) if 0 in reduce_concat: reduce_concat.remove(0) if 1 in reduce_concat: reduce_concat.remove(1) gen_type = Genotype(normal=normal, normal_concat=normal_concat, reduce=reduce, reduce_concat=reduce_concat) # print(normal, normal_concat, reduce, reduce_concat) return gen_type - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -