def __init__()

in src/models.py [0:0]

• 53 lines of code
• 23 McCabe index (conditional complexity)

    def __init__(self, D, W, skip, n_in, n_out, config, net_idx):
        super(BaseNet, self).__init__()
        self.net_idx = net_idx
        if "auto" in skip:

            skip_layer = 7

            if len(skip) > 4:
                skip_layer = int(skip[4:])

            if config.posEnc and config.posEnc[net_idx] and "RayMarch" in config.inFeatures[net_idx]:
                if config.posEnc[net_idx] == "nerf":
                    freq = config.posEncArgs[net_idx].split("-")
                    posInputs = (int(freq[0]))*6 + 3
                    dirInputs = (int(freq[1]))*6 + 2
                    # NOTE: this assumes 8 layers I guess
                    skip = f"0::{posInputs}-{D*skip_layer//8}:{posInputs}:"
            # did not override so set to nothing
            if "auto" in skip:
                print("Warning auto skip setup not detectable, using no skip connections")
                skip = ""
        self.name = f"relu{self.net_idx}({W}x{D}{skip.replace(':','.') if skip else ''})"
        self.D = D
        self.W = W
        self.inputLocations = {0:(0,n_in)}
        if skip:
            self.inputLocations = dict()
            decode_skips = [p for p in skip.split('-')]
            for s in decode_skips:
                match = re.search('^([0-9]+)(:?)([0-9]*)(:?)([0-9]*)$', s)
                if not match:
                    raise Exception("could not decode skip info")
                loc = match.group(1)
                has_first = match.group(2)
                start_feat = match.group(3)
                has_inbetween = match.group(4)
                end_feat = match.group(5)
                if has_first == '' and has_inbetween == '':
                    #all
                    self.inputLocations[int(loc)] = (0,n_in)
                elif has_first == ':' and has_inbetween == '':
                    single = int(start_feat+end_feat)
                    self.inputLocations[int(loc)] = (single, single+1)
                else:
                    istart = int(start_feat) if start_feat != '' else 0
                    iend =   int(end_feat) if end_feat != '' else n_in
                    self.inputLocations[int(loc)] = (istart, iend)

            if 0 not in self.inputLocations:
                self.inputLocations[0] = (0,n_in)
        self.n_in = n_in
        self.n_out = n_out
        layers = [nn.Linear(self.inputLocations[0][1]-self.inputLocations[0][0], self.W)]
        for i in range(1, self.D):
            layers.append(nn.Linear(self.inputLocations[i][1]-self.inputLocations[i][0] + self.W if i in self.inputLocations else self.W, self.W if i != self.D - 1 else self.n_out))
        # layers.append(nn.Linear(self.W, self.n_out))
        self.layers = nn.ModuleList(layers)

        self.activation = F.relu

        for i, l in enumerate(self.layers):
            nn.init.kaiming_normal_(l.weight)

        self.init_weights()