def parse_deconv(json_param):
params = {}
params["name"] = json_param["name"]
params["top"] = json_param["top"][0]
params["bot"] = json_param["bottom"][0]
params["type"] = "deconv"
json_conv_param = json_param["convolution_param"]
params["additive_conv"] = False
if "pad" in json_conv_param:
params["pad"] = json_conv_param["pad"]
else:
params["pad"] = [0, 0, 0]
params["stride"] = json_conv_param["stride"]
params["ofm"] = json_conv_param["num_output"]
if "group" in json_conv_param:
params["group"] = json_conv_param["group"]
assert (params["group"] == 1 or params["group"] == params["ofm"])
params["json_kernel_size"] = json_conv_param["kernel_size"]
params["kernel"] = "{}_kernel".format(params["name"])
params["bias"] = "{}_bias".format(params["name"])
params["bias_dim"] = [params["ofm"]]
params["bias_size"] = round_to_simd(params["bias_dim"][0])
params["scale_size"] = round_to_simd(params["ofm"])
params["scale"] = "{}_scale".format(params["name"])
return params
def h5ker_to_znnphiker(ifm, ofm, kz, kx, ky):
total_ifms = round_to_simd(kdim[0])
total_ofms = round_to_simd(kdim[1])
offset = ofm/S
offset *= total_ifms/S
offset += ifm/S
offset *= kdim[2]
offset += kz
offset *= kdim[3]
offset += kx
offset *= kdim[4]
offset += ky
offset *= S
offset += ifm % S
offset *= S
offset += ofm % S
return offset
def set_deconv_dim(params, bot_tensor):
params["bn"] = bot_tensor.dim[0]
params["ifm"] = bot_tensor.dim[1]
params["id"] = bot_tensor.dim[2]
params["ihw"] = bot_tensor.dim[3]
params["kernel_dim"] = [params["ifm"], params["ofm"]]
params["kernel_dim"] += params["json_kernel_size"]
params["kernel_size"] = params["kernel_dim"][2] * params["kernel_dim"][3] * params["kernel_dim"][4]
params["kernel_size"] *= round_to_simd(params["kernel_dim"][0])
params["kernel_size"] *= round_to_simd(params["kernel_dim"][1])
top_dim = [-1, -1, -1, -1, -1]
top_dim[0] = bot_tensor.dim[0]
top_dim[1] = params["ofm"]
for i in [2, 3, 4]:
top_dim[i] = params["stride"][i - 2] * (bot_tensor.dim[i] - 1)
top_dim[i] += params["kernel_dim"][i] - 2*params["pad"][i - 2]
params["top_dim"] = top_dim
params["bot_dim"] = bot_tensor.dim
def parse_conv(json_param):
params = {}
params["name"] = json_param["name"]
params["top"] = json_param["top"][0]
params["bot"] = json_param["bottom"][0]
params["type"] = "conv"
json_conv_param = json_param["convolution_param"]
if "pad" in json_conv_param:
params["pad"] = json_conv_param["pad"]
else:
params["pad"] = [0, 0, 0]
params["out_pad"] = [0, 0, 0]
params["stride"] = json_conv_param["stride"]
params["ofm"] = json_conv_param["num_output"]
params["json_kernel_size"] = json_conv_param["kernel_size"]
params["kernel"] = "{}_kernel".format(params["name"])
params["bias"] = "{}_bias".format(params["name"])
params["bias_dim"] = [params["ofm"]]
params["bias_size"] = round_to_simd(params["bias_dim"][0])
params["activation"] = None
params["additive_conv"] = False
params["scale"] = "{}_scale".format(params["name"])
params["scale_size"] = round_to_simd(params["ofm"])
params["scale_data"] = None
params["cores"] = 2
params["ht"] = 2
return params