加载预训练模型后,报RuntimeError: Shape not matching
Created by: gezimonkey
paddlepaddle版本:1.7 使用baidu_cn1.2k_model_fluid作为预训练模型 数据为16k的wav mean_std.npz和vocab.txt是为自己生成的
参数: ----------- Configuration Arguments ----------- augment_conf_path: conf/augmentation.config batch_size: 16 dev_manifest: data/swd/16k/manifest.dev init_from_pretrained_model: models/baidu_cn1.2k_model_fluid/ is_local: True learning_rate: 0.0005 max_duration: 15.0 mean_std_path: data/swd/16k/mean_std.npz min_duration: 0.9 num_conv_layers: 2 num_epoch: 50 num_iter_print: 100 num_rnn_layers: 3 num_samples: 51000 output_model_dir: ./checkpoints/swd rnn_layer_size: 2048 save_epoch: 1 share_rnn_weights: False shuffle_method: batch_shuffle_clipped specgram_type: linear test_off: False train_manifest: data/swd/16k/manifest.train use_gpu: True use_gru: True use_sortagrad: True vocab_path: data/swd/16k/vocab.txt
最后报: RuntimeError: Shape not matching: the Program requires a parameter with a shape of ((4096L, 1133L)), while the loaded parameter (namely [ layer_5_fc_weight ]) has a shape of ((4096, 5304)).