tensorflow 相关问题

我在Google Colab上遇到了这个问题，TF版本：2.16.1，keras版本：3.3.3 这是错误： ValueError Traceback（最近一次调用最后一次）我在Google Colab上遇到了问题，TF版本：2.16.1，keras版本：3.3.3 这是错误： ValueError Traceback (most recent call last) <ipython-input-46-4d65c8f7b80b> in <cell line: 1>() ----> 1 m3.load_weights("./model_v8.weights.h5") 1 frames /usr/local/lib/python3.10/dist-packages/keras/src/utils/traceback_utils.py in error_handler(*args, **kwargs) 120 # To get the full stack trace, call: 121 # `keras.config.disable_traceback_filtering()` --> 122 raise e.with_traceback(filtered_tb) from None 123 finally: 124 del filtered_tb /usr/local/lib/python3.10/dist-packages/keras/src/saving/saving_lib.py in _raise_loading_failure(error_msgs, warn_only) 293 warnings.warn(msg) 294 else: --> 295 raise ValueError(msg) 296 297 ValueError: A total of 1 objects could not be loaded. Example error message for object <Embedding name=embedding_10, built=True>: Layer 'embedding_10' expected 1 variables, but received 0 variables during loading. Expected: ['embeddings'] List of objects that could not be loaded: [<Embedding name=embedding_10, built=True>] 我正在以 Bahdanau 的注意力实施 seq2seq。权重是昨天从同一模型训练的。在加载之前，我用translate()方法做了一些翻译，因为如果我只是实例化模型实例然后加载，似乎无法加载权重。这是我的模型的代码。编码器： class Encoder(Layer): def __init__(self, tokenizer, embedding_size, hidden_units): """ Encoder Block in seq2seq :param tokenizer: tokenizer of the source language :param embedding_size: dimensionality of the embedding layer :param hidden_units: dimensionality of the output """ super().__init__() self.tokenizer = tokenizer self.embedding_size = embedding_size self.hidden_units = hidden_units self.vocab_size = tokenizer.vocabulary_size() self.embedding = Embedding(input_dim=self.vocab_size, output_dim=embedding_size) self.rnn = Bidirectional( merge_mode="sum", layer=LSTM(units=hidden_units, dropout=DROPOUT, return_sequences=True, return_state=True)) def call(self, x, training=True): """ :param x: [batch, time_steps] :param training: is training or not :return: encoder_hidden_state: [batch, hidden_state_dim] state_h: [batch, hidden_state_dim] state_c: [batch, hidden_state_dim] """ mask = tf.where(x != 0, True, False) x = self.embedding(x) x, forward_h, forward_c, backward_h, backward_c = self.rnn(x, mask=mask, training=training) return x, forward_h + backward_h, forward_c + backward_c 巴赫达瑙注意： class BahdanauAttention(Layer): def __init__(self, hidden_units): super().__init__() self.Va = Dense(1) self.Wa = Dense(hidden_units) self.Ua = Dense(hidden_units) self.norm = LayerNormalization() self.tanh = Activation(tf.keras.activations.tanh) self.add = Add() def call(self, context, x): """ Calculate the context vector based on all encoder hidden states and previous decoder state. :param: context: tensor, all encoder hidden states :param: x: tensor, previous state from Decoder :return: context_vector: tensor, the calculated context vector based on the input parameters """ # Expand dims to ensure scores shape = [batch, Ty, Tx] context = tf.expand_dims(context, axis=1) x = tf.expand_dims(x, axis=2) scores = self.Va(self.tanh(self.add([self.Wa(context), self.Ua(x)]))) scores = tf.squeeze(scores, axis=-1) attn_weights = tf.nn.softmax(scores, axis=-1) # NOTE: context shape = [batch, 1, Tx, feature] so that expand # dim of attention weights context_vector = tf.expand_dims(attn_weights, axis=-1) * context context_vector = tf.reduce_sum(context_vector, axis=-2) context_vector = self.norm(context_vector) context_vector = self.add([context_vector, tf.squeeze(x, -2)]) return context_vector 解码器： class Decoder(Layer): def __init__(self, tokenizer, embedding_size, hidden_units): """ Decoder Block in seq2seq :param tokenizer: tokenizer of the source language :param embedding_size: dimensionality of the embedding layer :param hidden_units: dimensionality of the output """ super().__init__() self.tokenizer = tokenizer self.embedding_size = embedding_size self.hidden_units = hidden_units self.vocab = tokenizer.get_vocabulary() self.vocab_size = tokenizer.vocabulary_size() self.embedding = Embedding(input_dim=self.vocab_size, output_dim=embedding_size) self.rnn = LSTM(units=hidden_units, dropout=DROPOUT, return_sequences=True, return_state=True) self.attention = BahdanauAttention(hidden_units) self.dense = Dense(self.vocab_size) def call(self, context, x, encoder_state, training=True, return_state=False): """ :param context: all encoder states :param x: all initial decoder states :param encoder_state: last state from encoder :param training: :param return_state: :return: logits: state_h: hidden state state_c: cell state """ mask = tf.where(x != 0, True, False) x = self.embedding(x) decoder_outputs, state_h, state_c = self.rnn(x, initial_state=encoder_state, mask=mask, training=training) dense_inputs = self.attention(context, decoder_outputs) logits = self.dense(dense_inputs) if return_state: return logits, state_h, state_c else: return logits 型号： class NMT(Model): @classmethod def add_method(cls, fun): setattr(cls, fun.__name__, fun) return fun def __init__(self, input_tokenizer, output_tokenizer, embedding_size, hidden_units): """ Initialize an instance for Neural Machine Translation Task :param input_tokenizer: tokenizer of the input language :param output_tokenizer: tokenizer of the output language :param embedding_size: dimensionality of embedding layer :param hidden_units: dimensionality of the output """ super().__init__() self.input_tokenizer = input_tokenizer self.output_tokenizer = output_tokenizer self.embedding_size = embedding_size self.hidden_units = hidden_units self.encoder = Encoder(input_tokenizer, embedding_size, hidden_units) self.decoder = Decoder(output_tokenizer, embedding_size, hidden_units) def call(self, inputs): encoder_inputs, decoder_inputs = inputs encoder_outputs, state_h, state_c = self.encoder(encoder_inputs) logits = self.decoder(encoder_outputs, decoder_inputs, [state_h, state_c]) return logits @NMT.add_method def translate(self, next_inputs, maxlen=40): """ """ def sampling(logits): probs = tf.nn.softmax(logits) dist = probs.numpy().squeeze() idx = np.random.choice(range(self.decoder.vocab_size), p=dist) return idx translation = [] next_inputs = expand_contractions(next_inputs.lower(), en_contraction_map) next_idx = np.asarray(self.encoder.tokenizer(next_inputs)) while next_idx.ndim != 2: next_idx = tf.expand_dims(next_idx, axis=0) encoder_outputs, state_h, state_c = self.encoder(next_idx, training=False) next_inputs = "[START]" next_idx = np.asarray(word_to_idx[next_inputs]) for i in range(maxlen): while next_idx.ndim != 2: next_idx = tf.expand_dims(next_idx, axis=0) logits, state_h, state_c = self.decoder(encoder_outputs, next_idx, [state_h, state_c], training=False, return_state=True) next_idx = sampling(logits) next_inputs = self.decoder.vocab[next_idx] if next_inputs == "[END]": break elif next_inputs == "[UNK]": continue else: translation.append(next_inputs) return " ".join(translation) 如果您想查看的话，这是笔记本的链接： https://colab.research.google.com/drive/1EKOm7ULFKEusvEFb8thSGOHTfRnRKru7?usp=sharing 我寻找解决方案，但它们通常是关于检查 tf 和 keras 版本。我用的是h5py，它说我的权重文件中的keras ver是3.3.3，所以看起来没有问题。我还尝试使用 get_config() 和 from_config() 方法中的序列化自定义对象来保存整个模型，但出现相同的错误，并出现有关 Decoder 类的 build() 方法的警告，因此我返回到保存重量，然后加载仅此而已。我还注意到，在 TF 的保存和加载模型教程中，重量是直接在示例中加载的。如何实现这一目标？顺便说一句，是否有关于为自定义层实现 build() 的任何指导，因为我的解码器由其他不同的 TF 层组成，所以我不知道如何实现一个。原来Colab默认的tf版本是2.15.1，当我安装tf_text时没有在notebook的头部指定版本时更新到了2.16.1。因此，我将 tf_text 指定为 2.15.1，权重加载完美。

tensorflow tensorflow2.0

回答 1 投票 0

自定义 TF 模型的权重加载问题（已关闭）

tensorflow tensorflow2.0

回答 1 投票 0

用最少层数训练神经网络以获得绝对函数

我正在尝试训练神经网络来学习 y = |x|功能。我们知道，绝对函数有两条不同的线在零点处相互连接。所以我试图有以下序列...

python tensorflow machine-learning keras deep-learning

回答 1 投票 0

自动编码器和极坐标

自动编码器可以学习极坐标的变换吗？如果一组二维数据近似位于一个圆上，则存在一个由角度参数化的低维流形，描述...

python tensorflow autoencoder polar-coordinates