【keras模型查看】模型结构、模型参数、每层输入/输出

查看keras模型结构

from tensorflow import keras
    
# 查看模型层及参数
deepxi.model.summary()

# 将模型结构保存为图片
model_img_name = args.ver + '-' + args.network_type + '.png'
keras.utils.plot_model(deepxi.model, model_img_name, show_shapes=True)

示例：

查看keras模型参数

可参考：【模型参数】tensorflow1.x (slim) 和tensorflow2.x (keras) 的查看模型参数方式：https://blog.csdn.net/u010637291/article/details/108143002

# 给定keras模型，如deepxi.model， deepxi模型可参考：  https://github.com/anicolson/DeepXi

# 查看模型可训练参数
for v in deepxi.model.trainable_variables:
	print(str(v.name) + ', ' + str(v.shape)) 	# 变量名+变量shape
	print(str(v.value))							# 变量值

# 查看所有参数：
model_variables = deepxi.model.variables
for v in model_variables:
	print(str(v.name) + ', ' + str(v.shape))
	

示例：在watch窗口查看deepxi.model.trainable_variables，可查看到共有65个变量，每一个变量的名字、shape和值均可查看：

查看keras模型每层输入/输出

查看模型每层输出：

def print_layer_output(deepxi):
    from tensorflow.keras import backend as K

    inp = deepxi.model.input  # input
    outputs = [layer.output for layer in deepxi.model.layers]  # all layer outputs
    functors = [K.function([inp], [out]) for out in outputs]  # evaluation functions

    # Testing with a wav file
    test_x, test_x_len, _, test_x_base_names = Batch('../deepxi_dataset/deepxi_test_set/test_noisy_speech_1')
    
    print("Processing observations...")
    inp_batch, supplementary_batch, n_frames = deepxi.observation_batch(test_x, test_x_len)

  	# 模型每一层输出
    layer_outs = [func([inp_batch, 1.]) for func in functors]

	# Writing to a file
 	output = open('layer_output.txt', 'w+')
    for i in range(len(layer_outs)):  #
        for j in range(len(layer_outs[i][0][0])):
            for k in range(len(layer_outs[i][0][0][j])):
                # print(layer_outs[i][0][0][j][k], end='\t')
                output.write(str(layer_outs[i][0][0][j][k]) + ' ')
            # print()
            output.write('\n')
        # print()
        # print()
        output.write('\n')
        output.write('\n')
    print('done')
    
    return layer_outs

示例：layer_outs：可查看共有47层，每层的参数值、最大最小值及dtype等均可查看

在此，完成的操作有：
1）针对某一个wav文件，模型的每层输出；
2）针对一个dataset，模型的每层输出。

def print_layer_output_4wav(deepxi, wavfile = '../deepxi_dataset/deepxi_test_set/test_noisy_speech_1/6930-81414-0003_SIGNAL021_0dB.wav'):
    '''
    Generating outputs of all layers
    @param deepxi: our pretrained deepxi
    @param wavfile: a certain wav file with its dir and name.
    @return:
    '''
    from tensorflow.keras import backend as K
    from deepxi.utils import read_wav
    import numpy as np

    input = deepxi.model.input  # input
    outputs = [layer.output for layer in deepxi.model.layers]  # all layer outputs
    functors = [K.function([input], [output]) for output in outputs]  # evaluation functions

    # Reading the wav file
    (wav, _fs) = read_wav(wavfile)

    # Observing
    inp, _supplementary = deepxi.inp_tgt.observation(wav)

    # Batching
    n_frames = deepxi.inp_tgt.n_frames(len(wav))
    inp_batch = np.zeros([1, n_frames, deepxi.inp_tgt.n_feat], np.float32)
    inp_batch[0, :n_frames, :] = inp

    # outputs of all layers
    layer_outs = [func([inp_batch, 1.]) for func in functors]

    return layer_outs

def generate_train_dataset(deepxi):
    from deepxi import utils
    import math
    from deepxi.args import get_args

    args = get_args()

    if args.set_path != "set":
        args.data_path = args.data_path + '/' + args.set_path.rsplit('/', 1)[-1]  # data path.
    train_s_path = args.set_path + '/train_clean_speech'  # path to the clean speech training set.
    train_d_path = args.set_path + '/train_noise'  # path to the noise training set.
    train_s_list = utils.batch_list(train_s_path, 'clean_speech', args.data_path)
    train_d_list = utils.batch_list(train_d_path, 'noise', args.data_path)

    deepxi.train_s_list = train_s_list
    deepxi.train_d_list = train_d_list

    deepxi.mbatch_size = args.mbatch_size
    deepxi.n_examples = len(train_s_list)
    deepxi.n_iter = math.ceil(deepxi.n_examples / deepxi.mbatch_size)

    dataset = deepxi.dataset(n_epochs=200)
    return dataset


def print_layer_output_4dataset(deepxi, dataset):
    '''

    @param deepxi:
    @param dataset:
    @return:
    '''
    from tensorflow.keras import backend as K

    input = deepxi.model.input  # input
    outputs = [layer.output for layer in deepxi.model.layers]  # all layer outputs
    functors = [K.function([input], [output]) for output in outputs]  # evaluation functions

    layer_outs_batch = []
    for (inp_batch, _tgt_batch, _seq_mask_batch) in dataset.as_numpy_iterator():
        layer_outs = [func([inp_batch, 1.]) for func in functors]
        layer_outs_batch.append(layer_outs)

    return layer_outs_batch


if __name__ == '__main__':
    # pretrained model
    from quantization_test.create_load_test_model import create_model, load_variables_for_model
    deepxi = create_model()
    pretrained_deepxi = load_variables_for_model(deepxi)

    # summary model
    summary_save_model(pretrained_deepxi)

    # print variables
    print_model_variables(pretrained_deepxi)

    # print outputs of all layers, according to a wav file
    layer_outputs = print_layer_output_4wav(pretrained_deepxi)

    # print outputs of all layers, according to a dataset
    dataset = generate_train_dataset(pretrained_deepxi)
    layer_outputs_batch = print_layer_output_4dataset(pretrained_deepxi, dataset)

    print('done')

同理，可查看每层的输入：

inp = deepxi.model.input  # input
inputs = [layer.input for layer in deepxi.model.layers if (isinstance(layer, keras.layers.ReLU) or isinstance(layer, keras.layers.Activation))]  # activation layer inputs
outputs = [layer.output for layer in deepxi.model.layers if (isinstance(layer, keras.layers.ReLU) or isinstance(layer, keras.layers.Activation))]
functors_inp = [K.function([inp], [input]) for input in inputs]
functors_outp = [K.function([inp], [output]) for output in outputs]

# Testing
test_x, test_x_len, _, test_x_base_names = Batch(testing_path)
print("Processing observations...")
inp_batch, supplementary_batch, n_frames = deepxi.observation_batch(test_x, test_x_len)

layer_ins = [func(inp_batch) for func in functors_inp]