Load encodec model checkpoint.
(checkpoint, outfile, use_f16)
| 111 | |
| 112 | |
| 113 | def parse_codec_model_weights(checkpoint, outfile, use_f16): |
| 114 | """Load encodec model checkpoint.""" |
| 115 | keys = [k for k in checkpoint.keys() if "codec_model" in k] |
| 116 | |
| 117 | for name in keys: |
| 118 | if "weight_g" in name: |
| 119 | # the tensor has already been parsed with the corresponding "weight_v" |
| 120 | # tensor to form the final weights tensor of the convolution, therefore |
| 121 | # we skip it |
| 122 | continue |
| 123 | |
| 124 | if "inited" in name or "cluster_size" in name or "embed_avg" in name: |
| 125 | # "inited", "cluster_size" and "embed_avg" tensors in quantizer are not used |
| 126 | # for the forward pass |
| 127 | continue |
| 128 | |
| 129 | # Remove prefix from the variable name and the dot |
| 130 | clean_name = name.replace("codec_model.", "") |
| 131 | |
| 132 | var_data = checkpoint[name] |
| 133 | |
| 134 | if not "weight_v" in name: |
| 135 | # if conv kernel, do not squeeze because 3d tensor |
| 136 | var_data = var_data.numpy().squeeze() |
| 137 | else: |
| 138 | # weight_v has its corresponding magnitude tensor to rescale the weights |
| 139 | # of the convolutional layers. We parse both kinds of weights jointly to |
| 140 | # build the final weight tensor of the convolution. |
| 141 | base_name = name.split(".")[:-1] |
| 142 | weight_g_name = ".".join(base_name + ["weight_g"]) |
| 143 | var_data_g = checkpoint[weight_g_name] |
| 144 | |
| 145 | final_var_data = torch._weight_norm(var_data, var_data_g, dim=0) |
| 146 | var_data = final_var_data.numpy() |
| 147 | |
| 148 | name = ".".join(base_name + ["weight"]) |
| 149 | clean_name = name.replace("codec_model.", "") |
| 150 | |
| 151 | if "encoder" in clean_name or "decoder" in clean_name: |
| 152 | if clean_name in DECODER_CONV_TRANSPOSE_LAYERS: |
| 153 | pattern = r"decoder.layers.(\d+).conv\.(bias|weight)$" |
| 154 | replacement = r"decoder.model.\1.convtr.convtr.\2" |
| 155 | clean_name = re.sub(pattern, replacement, clean_name) |
| 156 | elif "conv" in clean_name: |
| 157 | pattern = r"(encoder|decoder).layers.(\d+)(.*?).conv\.(bias|weight)$" |
| 158 | replacement = r"\1.model.\2\3.conv.conv.\4" |
| 159 | clean_name = re.sub(pattern, replacement, clean_name) |
| 160 | elif "lstm" in clean_name: |
| 161 | clean_name = clean_name.replace("layers", "model") |
| 162 | elif "quantizer" in clean_name: |
| 163 | pattern = r"quantizer.layers.(\d+)\.codebook\.(.+)$" |
| 164 | replacement = r"quantizer.vq.layers.\1._codebook.\2" |
| 165 | clean_name = re.sub(pattern, replacement, clean_name) |
| 166 | else: |
| 167 | raise Exception(f"Unrecognized variable name: {clean_name}") |
| 168 | |
| 169 | print(f"Processing variable: {name} with shape: {var_data.shape}") |
| 170 |
no test coverage detected