015b435553d1591378233dfe8792027d910d56eb,batchflow/models/eager_torch/resnet.py,ResBlock,__init__,#ResBlock#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#,69

Before Change

        num_convs = sum([letter in CONV_LETTERS for letter in layout])

        vars = {key: get_num_channels(inputs) for key in ["S", "same"]}
        filters = [filters] * num_convs if isinstance(filters, (int, str)) else filters
        filters = [eval(str(item), {}, vars) for item in filters]

        kernel_size = [kernel_size] * num_convs if isinstance(kernel_size, int) else kernel_size
        strides = [strides] * num_convs if isinstance(strides, int) else strides
        strides_d = list(strides)
        groups = [groups] * num_convs
        side_branch_stride = np.prod(strides)
        side_branch_stride_d = int(side_branch_stride)

        if downsample:
            downsample = 2 if downsample is True else downsample
            strides_d[0] *= downsample
            side_branch_stride_d *= downsample
        if bottleneck:
            bottleneck = 4 if bottleneck is True else bottleneck
            layout = "cna" + layout + "cna"
            kernel_size = [1] + kernel_size + [1]
            strides = [1] + strides + [1]
            strides_d = [1] + strides_d + [1]
            groups = [1] + groups + [1]
            filters = [filters[0] // bottleneck] + filters + [filters[0]]
        if se:
            layout += "S*"
        layout = "B" + layout + op

        layer_params = [{"strides": strides_d, "side_branch/strides": side_branch_stride_d}] + [{}]*(n_reps-1)
        self.block = ConvBlock(*layer_params, inputs=inputs, layout=layout, filters=filters,
                               kernel_size=kernel_size, strides=strides, groups=groups,
                               side_branch={"layout": "c", "filters": filters[-1], "strides": side_branch_stride},
                               **kwargs)

After Change

        num_convs = sum([letter in CONV_LETTERS for letter in layout])

        filters = [filters] * num_convs if isinstance(filters, (int, str)) else filters
        filters = [safe_eval(str(item), get_num_channels(inputs)) for item in filters]

        kernel_size = [kernel_size] * num_convs if isinstance(kernel_size, int) else kernel_size
        strides = [strides] * num_convs if isinstance(strides, int) else strides
        strides_d = list(strides)
        groups = [groups] * num_convs
        side_branch_stride = np.prod(strides)
        side_branch_stride_d = int(side_branch_stride)

        if downsample:
            downsample = 2 if downsample is True else downsample
            strides_d[0] *= downsample
            side_branch_stride_d *= downsample
        if bottleneck:
            bottleneck = 4 if bottleneck is True else bottleneck
            layout = "cna" + layout + "cna"
            kernel_size = [1] + kernel_size + [1]
            strides = [1] + strides + [1]
            strides_d = [1] + strides_d + [1]
            groups = [1] + groups + [1]
            filters = [filters[0] // bottleneck] + filters + [filters[0]]
        if se:
            layout += "S*"
        layout = "B" + layout + op

        layer_params = [{"strides": strides_d, "side_branch/strides": side_branch_stride_d}] + [{}]*(n_reps-1)
        self.block = ConvBlock(*layer_params, inputs=inputs, layout=layout, filters=filters,
                               kernel_size=kernel_size, strides=strides, groups=groups,
                               side_branch={"layout": "c", "filters": filters[-1], "strides": side_branch_stride},
                               **kwargs)

Instances