Код: Выделить всё
import tensorflow as tf
from tensorflow.keras import Model
from tensorflow.keras.layers import Dense
from tensorflow.train import Checkpoint, CheckpointManager
from tensorflow.keras.callbacks import Callback
from encoders.lstm_encoders import StackedBidirectionalLSTMEncoder
from decoders.lstm_decoders import StackedLSTMDecoder
from typing import Optional, Any, Tuple
class RetrosynthesisSeq2SeqModel(Model):
def __init__(self, input_vocab_size: int, output_vocab_size: int, encoder_embedding_dim: int,
decoder_embedding_dim: int, units: int, dropout_rate: float = 0.2, *args, **kwargs):
super(RetrosynthesisSeq2SeqModel, self).__init__(*args, **kwargs)
# Save the number of units (neurons)
self.units: int = units
# Encoder layer
self.encoder: StackedBidirectionalLSTMEncoder = StackedBidirectionalLSTMEncoder(
input_vocab_size, encoder_embedding_dim, units, dropout_rate
)
# Decoder layer
self.decoder: StackedLSTMDecoder = StackedLSTMDecoder(
output_vocab_size, decoder_embedding_dim, units, dropout_rate
)
# Save the vocabulary sizes
self.input_vocab_size: int = input_vocab_size
self.output_vocab_size: int = output_vocab_size
# Mapping encoder final states to decoder initial states
self.enc_state_h: Dense = Dense(units, name='enc_state_h')
self.enc_state_c: Dense = Dense(units, name='enc_state_c')
# Store the data processors (to be set externally)
self.encoder_data_processor: Optional[Any] = None
self.decoder_data_processor: Optional[Any] = None
# Save the dropout rate
self.dropout_rate: float = dropout_rate
def build(self, input_shape):
# Define the input shapes for encoder and decoder
encoder_input_shape, decoder_input_shape = input_shape
# Pass a dummy input through encoder and decoder to initialize weights
encoder_dummy = tf.zeros(encoder_input_shape)
decoder_dummy = tf.zeros(decoder_input_shape)
# Forward pass to build the model
self.call((encoder_dummy, decoder_dummy), training=False)
# Mark the model as built
super(RetrosynthesisSeq2SeqModel, self).build(input_shape)
def call(self, inputs: Tuple[tf.Tensor, tf.Tensor], training: Optional[bool] = None) -> tf.Tensor:
"""
Forward pass of the Seq2Seq model.
Args:
inputs (Tuple[tf.Tensor, tf.Tensor]): Tuple containing encoder and decoder inputs.
training (Optional[bool], optional): Training flag. Defaults to None.
Returns:
tf.Tensor: The output predictions from the decoder.
"""
# Extract encoder and decoder inputs
encoder_input, decoder_input = inputs
# Encoder
encoder_output, state_h, state_c = self.encoder.call(encoder_input, training=training)
# Map encoder final states to decoder initial states
decoder_initial_state_h: tf.Tensor = self.enc_state_h(state_h) # (batch_size, units)
decoder_initial_state_c: tf.Tensor = self.enc_state_c(state_c) # (batch_size, units)
decoder_initial_state: Tuple[tf.Tensor, tf.Tensor] = (decoder_initial_state_h, decoder_initial_state_c)
# Prepare decoder inputs as a tuple
decoder_inputs: Tuple[tf.Tensor, Tuple[tf.Tensor, tf.Tensor], tf.Tensor] = (
decoder_input,
decoder_initial_state,
encoder_output
)
# Extract encoder mask
encoder_mask: Optional[tf.Tensor] = self.encoder.compute_mask(encoder_input)
# Decoder
output: tf.Tensor = self.decoder.call(
decoder_inputs,
training=training,
mask=encoder_mask
)
return output
def get_config(self) -> dict:
"""
Returns the configuration of the layer for serialization.
Returns:
dict: A Python dictionary containing the layer's configuration.
"""
config = super(RetrosynthesisSeq2SeqModel, self).get_config()
config.update({
'units': self.units,
'input_vocab_size': self.input_vocab_size,
'output_vocab_size': self.output_vocab_size,
'encoder_embedding_dim': self.encoder.embedding.output_dim,
'decoder_embedding_dim': self.decoder.embedding.output_dim,
'dropout_rate': self.dropout_rate,
'encoder': tf.keras.layers.serialize(self.encoder),
'decoder': tf.keras.layers.serialize(self.decoder),
'enc_state_h': tf.keras.layers.serialize(self.enc_state_h),
'enc_state_c': tf.keras.layers.serialize(self.enc_state_c)
})
return config
@classmethod
def from_config(cls, config: dict) -> 'RetrosynthesisSeq2SeqModel':
"""
Creates a layer from its config.
Args:
config (dict): A Python dictionary containing the layer's configuration.
Returns:
RetrosynthesisSeq2SeqModel: A new instance of RetrosynthesisSeq2SeqModel configured using the provided
config.
"""
# Deserialize layers
config['encoder'] = tf.keras.layers.deserialize(config['encoder'])
config['decoder'] = tf.keras.layers.deserialize(config['decoder'])
config['enc_state_h'] = tf.keras.layers.deserialize(config['enc_state_h'])
config['enc_state_c'] = tf.keras.layers.deserialize(config['enc_state_c'])
return cls(**config)
Код: Выделить всё
Traceback (most recent call last):
File "/home/Computational-Chemistry/aizynthfinder-project/seq2seq-expansion-strategy/src/__main__.py", line 23, in
main()
File "/home/Computational-Chemistry/aizynthfinder-project/seq2seq-expansion-strategy/src/__main__.py", line 19, in main
trainer.run()
File "/home/Computational-Chemistry/aizynthfinder-project/seq2seq-expansion-strategy/src/trainers/trainer.py", line 325, in run
self.save_model()
File "/home/Computational-Chemistry/aizynthfinder-project/seq2seq-expansion-strategy/src/trainers/trainer.py", line 313, in save_model
self.model.save(os.path.join(model_save_path, 'model.keras'))
File "/home/anaconda3/envs/aizynth-env/lib/python3.10/site-packages/keras/src/utils/traceback_utils.py", line 122, in error_handler
raise e.with_traceback(filtered_tb) from None
File "/home/anaconda3/envs/aizynth-env/lib/python3.10/site-packages/h5py/_hl/group.py", line 483, in __setitem__
ds = self.create_dataset(None, data=obj)
File "/home/anaconda3/envs/aizynth-env/lib/python3.10/site-packages/h5py/_hl/group.py", line 183, in create_dataset
dsid = dataset.make_new_dset(group, shape, dtype, data, name, **kwds)
File "/home/anaconda3/envs/aizynth-env/lib/python3.10/site-packages/h5py/_hl/dataset.py", line 86, in make_new_dset
tid = h5t.py_create(dtype, logical=1)
File "h5py/h5t.pyx", line 1663, in h5py.h5t.py_create
File "h5py/h5t.pyx", line 1687, in h5py.h5t.py_create
File "h5py/h5t.pyx", line 1705, in h5py.h5t.py_create
File "h5py/h5t.pyx", line 1459, in h5py.h5t._c_int
TypeError: Unsupported integer size (0)
Process finished with exit code 1
< h2>Отладка и попытки решения
Чтобы попытаться исправить ситуацию, я реализовал методы get_config() и from_config() в своей модели Seq2Seq и во всех пользовательских слоях. . Код модели можно увидеть в комментарии ниже. Я также позаботился о том, чтобы **kwargs был передан в суперкласс каждого подкласса Layer, чтобы Keras полностью регистрировал все подслои и конфигурации.
Для отладки я добавил следующее рекурсивная функция для проверки слоев и подслоев:
Код: Выделить всё
@staticmethod
def inspect_model_layers(model):
def _inspect_layer(layer, indent=0):
indent_str = " " * indent
print(f"{indent_str}Layer: {layer.name}")
config = layer.get_config()
for key, value in config.items():
print(f"{indent_str} - {key}: {value}")
# Recursively inspect sublayers if any
if hasattr(layer, 'layers'): # For layers like Bidirectional, TimeDistributed, etc.
for sublayer in layer.layers:
_inspect_layer(sublayer, indent + 1)
elif hasattr(layer, 'layer'): # For layers like RNN that contain a single layer
_inspect_layer(layer.layer, indent + 1)
for layer in model.layers:
_inspect_layer(layer)
Код: Выделить всё
Layer: stacked_bidirectional_lstm_encoder
- name: stacked_bidirectional_lstm_encoder
- trainable: True
- dtype: {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}
- vocab_size: 40
- encoder_embedding_dim: 128
- units: 256
- dropout_rate: 0.2
- embedding: {'module': 'keras.layers', 'class_name': 'Embedding', 'config': {'name': 'embedding_1', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'input_dim': 40, 'output_dim': 128, 'embeddings_initializer': {'module': 'keras.initializers', 'class_name': 'RandomUniform', 'config': {'minval': -0.05, 'maxval': 0.05, 'seed': None}, 'registered_name': None}, 'embeddings_regularizer': None, 'activity_regularizer': None, 'embeddings_constraint': None, 'mask_zero': True}, 'registered_name': None, 'build_config': {'input_shape': (32, 140)}}
- bidirectional_lstm_1: {'module': 'keras.layers', 'class_name': 'Bidirectional', 'config': {'name': 'bidirectional_lstm_1', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'merge_mode': 'concat', 'layer': {'module': 'keras.layers', 'class_name': 'LSTM', 'config': {'name': 'forward_lstm', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'return_sequences': True, 'return_state': True, 'go_backwards': False, 'stateful': False, 'unroll': False, 'zero_output_for_mask': True, 'units': 256, 'activation': 'tanh', 'recurrent_activation': 'sigmoid', 'use_bias': True, 'kernel_initializer': {'module': 'keras.initializers', 'class_name': 'GlorotUniform', 'config': {'seed': None}, 'registered_name': None}, 'recurrent_initializer': {'module': 'keras.initializers', 'class_name': 'OrthogonalInitializer', 'config': {'gain': 1.0, 'seed': None}, 'registered_name': None}, 'bias_initializer': {'module': 'keras.initializers', 'class_name': 'Zeros', 'config': {}, 'registered_name': None}, 'unit_forget_bias': True, 'kernel_regularizer': None, 'recurrent_regularizer': None, 'bias_regularizer': None, 'activity_regularizer': None, 'kernel_constraint': None, 'recurrent_constraint': None, 'bias_constraint': None, 'dropout': 0.0, 'recurrent_dropout': 0.0, 'seed': None}, 'registered_name': None, 'build_config': {'input_shape': (32, 140, 128)}}, 'backward_layer': {'module': 'keras.layers', 'class_name': 'LSTM', 'config': {'name': 'backward_lstm', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'return_sequences': True, 'return_state': True, 'go_backwards': True, 'stateful': False, 'unroll': False, 'zero_output_for_mask': True, 'units': 256, 'activation': 'tanh', 'recurrent_activation': 'sigmoid', 'use_bias': True, 'kernel_initializer': {'module': 'keras.initializers', 'class_name': 'GlorotUniform', 'config': {'seed': None}, 'registered_name': None}, 'recurrent_initializer': {'module': 'keras.initializers', 'class_name': 'OrthogonalInitializer', 'config': {'gain': 1.0, 'seed': None}, 'registered_name': None}, 'bias_initializer': {'module': 'keras.initializers', 'class_name': 'Zeros', 'config': {}, 'registered_name': None}, 'unit_forget_bias': True, 'kernel_regularizer': None, 'recurrent_regularizer': None, 'bias_regularizer': None, 'activity_regularizer': None, 'kernel_constraint': None, 'recurrent_constraint': None, 'bias_constraint': None, 'dropout': 0.0, 'recurrent_dropout': 0.0, 'seed': None}, 'registered_name': None, 'build_config': {'input_shape': (32, 140, 128)}}}, 'registered_name': None, 'build_config': {'input_shape': (32, 140, 128)}}
- dropout_1: {'module': 'keras.layers', 'class_name': 'Dropout', 'config': {'name': 'encoder_dropout_1', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'rate': 0.2, 'seed': None, 'noise_shape': None}, 'registered_name': None, 'build_config': {'input_shape': (32, 140, 512)}}
- bidirectional_lstm_2: {'module': 'keras.layers', 'class_name': 'Bidirectional', 'config': {'name': 'bidirectional_lstm_2', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'merge_mode': 'concat', 'layer': {'module': 'keras.layers', 'class_name': 'LSTM', 'config': {'name': 'forward_lstm_1', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'return_sequences': True, 'return_state': True, 'go_backwards': False, 'stateful': False, 'unroll': False, 'zero_output_for_mask': True, 'units': 256, 'activation': 'tanh', 'recurrent_activation': 'sigmoid', 'use_bias': True, 'kernel_initializer': {'module': 'keras.initializers', 'class_name': 'GlorotUniform', 'config': {'seed': None}, 'registered_name': None}, 'recurrent_initializer': {'module': 'keras.initializers', 'class_name': 'OrthogonalInitializer', 'config': {'gain': 1.0, 'seed': None}, 'registered_name': None}, 'bias_initializer': {'module': 'keras.initializers', 'class_name': 'Zeros', 'config': {}, 'registered_name': None}, 'unit_forget_bias': True, 'kernel_regularizer': None, 'recurrent_regularizer': None, 'bias_regularizer': None, 'activity_regularizer': None, 'kernel_constraint': None, 'recurrent_constraint': None, 'bias_constraint': None, 'dropout': 0.0, 'recurrent_dropout': 0.0, 'seed': None}, 'registered_name': None, 'build_config': {'input_shape': (32, 140, 512)}}, 'backward_layer': {'module': 'keras.layers', 'class_name': 'LSTM', 'config': {'name': 'backward_lstm_1', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'return_sequences': True, 'return_state': True, 'go_backwards': True, 'stateful': False, 'unroll': False, 'zero_output_for_mask': True, 'units': 256, 'activation': 'tanh', 'recurrent_activation': 'sigmoid', 'use_bias': True, 'kernel_initializer': {'module': 'keras.initializers', 'class_name': 'GlorotUniform', 'config': {'seed': None}, 'registered_name': None}, 'recurrent_initializer': {'module': 'keras.initializers', 'class_name': 'OrthogonalInitializer', 'config': {'gain': 1.0, 'seed': None}, 'registered_name': None}, 'bias_initializer': {'module': 'keras.initializers', 'class_name': 'Zeros', 'config': {}, 'registered_name': None}, 'unit_forget_bias': True, 'kernel_regularizer': None, 'recurrent_regularizer': None, 'bias_regularizer': None, 'activity_regularizer': None, 'kernel_constraint': None, 'recurrent_constraint': None, 'bias_constraint': None, 'dropout': 0.0, 'recurrent_dropout': 0.0, 'seed': None}, 'registered_name': None, 'build_config': {'input_shape': (32, 140, 512)}}}, 'registered_name': None, 'build_config': {'input_shape': (32, 140, 512)}}
- dropout_2: {'module': 'keras.layers', 'class_name': 'Dropout', 'config': {'name': 'encoder_dropout_2', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'rate': 0.2, 'seed': None, 'noise_shape': None}, 'registered_name': None, 'build_config': {'input_shape': (32, 140, 512)}}
Layer: stacked_lstm_decoder
- name: stacked_lstm_decoder
- trainable: True
- dtype: {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}
- vocab_size: 40
- decoder_embedding_dim: 256
- units: 256
- dropout_rate: 0.2
- embedding: {'module': 'keras.layers', 'class_name': 'Embedding', 'config': {'name': 'embedding_3', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'input_dim': 40, 'output_dim': 256, 'embeddings_initializer': {'module': 'keras.initializers', 'class_name': 'RandomUniform', 'config': {'minval': -0.05, 'maxval': 0.05, 'seed': None}, 'registered_name': None}, 'embeddings_regularizer': None, 'activity_regularizer': None, 'embeddings_constraint': None, 'mask_zero': True}, 'registered_name': None, 'build_config': {'input_shape': (32, 139)}}
- lstm_decoder_1: {'module': 'keras.layers', 'class_name': 'LSTM', 'config': {'name': 'lstm_decoder_1', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'return_sequences': True, 'return_state': True, 'go_backwards': False, 'stateful': False, 'unroll': False, 'zero_output_for_mask': False, 'units': 256, 'activation': 'tanh', 'recurrent_activation': 'sigmoid', 'use_bias': True, 'kernel_initializer': {'module': 'keras.initializers', 'class_name': 'GlorotUniform', 'config': {'seed': None}, 'registered_name': None}, 'recurrent_initializer': {'module': 'keras.initializers', 'class_name': 'OrthogonalInitializer', 'config': {'gain': 1.0, 'seed': None}, 'registered_name': None}, 'bias_initializer': {'module': 'keras.initializers', 'class_name': 'Zeros', 'config': {}, 'registered_name': None}, 'unit_forget_bias': True, 'kernel_regularizer': None, 'recurrent_regularizer': None, 'bias_regularizer': None, 'activity_regularizer': None, 'kernel_constraint': None, 'recurrent_constraint': None, 'bias_constraint': None, 'dropout': 0.0, 'recurrent_dropout': 0.0, 'seed': None}, 'registered_name': None, 'build_config': {'shapes_dict': {'sequences_shape': (32, 139, 256), 'initial_state_shape': ((32, 256), (32, 256))}}}
- dropout_1: {'module': 'keras.layers', 'class_name': 'Dropout', 'config': {'name': 'decoder_dropout_1', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'rate': 0.2, 'seed': None, 'noise_shape': None}, 'registered_name': None, 'build_config': {'input_shape': (32, 139, 256)}}
- lstm_decoder_2: {'module': 'keras.layers', 'class_name': 'LSTM', 'config': {'name': 'lstm_decoder_2', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'return_sequences': True, 'return_state': True, 'go_backwards': False, 'stateful': False, 'unroll': False, 'zero_output_for_mask': False, 'units': 256, 'activation': 'tanh', 'recurrent_activation': 'sigmoid', 'use_bias': True, 'kernel_initializer': {'module': 'keras.initializers', 'class_name': 'GlorotUniform', 'config': {'seed': None}, 'registered_name': None}, 'recurrent_initializer': {'module': 'keras.initializers', 'class_name': 'OrthogonalInitializer', 'config': {'gain': 1.0, 'seed': None}, 'registered_name': None}, 'bias_initializer': {'module': 'keras.initializers', 'class_name': 'Zeros', 'config': {}, 'registered_name': None}, 'unit_forget_bias': True, 'kernel_regularizer': None, 'recurrent_regularizer': None, 'bias_regularizer': None, 'activity_regularizer': None, 'kernel_constraint': None, 'recurrent_constraint': None, 'bias_constraint': None, 'dropout': 0.0, 'recurrent_dropout': 0.0, 'seed': None}, 'registered_name': None, 'build_config': {'input_shape': (32, 139, 256)}}
- dropout_2: {'module': 'keras.layers', 'class_name': 'Dropout', 'config': {'name': 'decoder_dropout_2', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'rate': 0.2, 'seed': None, 'noise_shape': None}, 'registered_name': None, 'build_config': {'input_shape': (32, 139, 256)}}
- lstm_decoder_3: {'module': 'keras.layers', 'class_name': 'LSTM', 'config': {'name': 'lstm_decoder_3', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'return_sequences': True, 'return_state': True, 'go_backwards': False, 'stateful': False, 'unroll': False, 'zero_output_for_mask': False, 'units': 256, 'activation': 'tanh', 'recurrent_activation': 'sigmoid', 'use_bias': True, 'kernel_initializer': {'module': 'keras.initializers', 'class_name': 'GlorotUniform', 'config': {'seed': None}, 'registered_name': None}, 'recurrent_initializer': {'module': 'keras.initializers', 'class_name': 'OrthogonalInitializer', 'config': {'gain': 1.0, 'seed': None}, 'registered_name': None}, 'bias_initializer': {'module': 'keras.initializers', 'class_name': 'Zeros', 'config': {}, 'registered_name': None}, 'unit_forget_bias': True, 'kernel_regularizer': None, 'recurrent_regularizer': None, 'bias_regularizer': None, 'activity_regularizer': None, 'kernel_constraint': None, 'recurrent_constraint': None, 'bias_constraint': None, 'dropout': 0.0, 'recurrent_dropout': 0.0, 'seed': None}, 'registered_name': None, 'build_config': {'input_shape': (32, 139, 256)}}
- dropout_3: {'module': 'keras.layers', 'class_name': 'Dropout', 'config': {'name': 'decoder_dropout_3', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'rate': 0.2, 'seed': None, 'noise_shape': None}, 'registered_name': None, 'build_config': {'input_shape': (32, 139, 256)}}
- lstm_decoder_4: {'module': 'keras.layers', 'class_name': 'LSTM', 'config': {'name': 'lstm_decoder_4', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'return_sequences': True, 'return_state': True, 'go_backwards': False, 'stateful': False, 'unroll': False, 'zero_output_for_mask': False, 'units': 256, 'activation': 'tanh', 'recurrent_activation': 'sigmoid', 'use_bias': True, 'kernel_initializer': {'module': 'keras.initializers', 'class_name': 'GlorotUniform', 'config': {'seed': None}, 'registered_name': None}, 'recurrent_initializer': {'module': 'keras.initializers', 'class_name': 'OrthogonalInitializer', 'config': {'gain': 1.0, 'seed': None}, 'registered_name': None}, 'bias_initializer': {'module': 'keras.initializers', 'class_name': 'Zeros', 'config': {}, 'registered_name': None}, 'unit_forget_bias': True, 'kernel_regularizer': None, 'recurrent_regularizer': None, 'bias_regularizer': None, 'activity_regularizer': None, 'kernel_constraint': None, 'recurrent_constraint': None, 'bias_constraint': None, 'dropout': 0.0, 'recurrent_dropout': 0.0, 'seed': None}, 'registered_name': None, 'build_config': {'input_shape': (32, 139, 256)}}
- dropout_4: {'module': 'keras.layers', 'class_name': 'Dropout', 'config': {'name': 'decoder_dropout_4', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'rate': 0.2, 'seed': None, 'noise_shape': None}, 'registered_name': None, 'build_config': {'input_shape': (32, 139, 256)}}
- attention: {'module': 'attention.attention', 'class_name': 'BahdanauAttention', 'config': {'name': 'bahdanau_attention', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'units': 256, 'attention_dense1': {'module': 'keras.layers', 'class_name': 'Dense', 'config': {'name': 'attention_dense1', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'units': 256, 'activation': 'linear', 'use_bias': True, 'kernel_initializer': {'module': 'keras.initializers', 'class_name': 'GlorotUniform', 'config': {'seed': None}, 'registered_name': None}, 'bias_initializer': {'module': 'keras.initializers', 'class_name': 'Zeros', 'config': {}, 'registered_name': None}, 'kernel_regularizer': None, 'bias_regularizer': None, 'kernel_constraint': None, 'bias_constraint': None}, 'registered_name': None, 'build_config': {'input_shape': (32, 1, 140, 512)}}, 'attention_dense2': {'module': 'keras.layers', 'class_name': 'Dense', 'config': {'name': 'attention_dense2', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'units': 256, 'activation': 'linear', 'use_bias': True, 'kernel_initializer': {'module': 'keras.initializers', 'class_name': 'GlorotUniform', 'config': {'seed': None}, 'registered_name': None}, 'bias_initializer': {'module': 'keras.initializers', 'class_name': 'Zeros', 'config': {}, 'registered_name': None}, 'kernel_regularizer': None, 'bias_regularizer': None, 'kernel_constraint': None, 'bias_constraint': None}, 'registered_name': None, 'build_config': {'input_shape': (32, 139, 1, 256)}}, 'attention_v': {'module': 'keras.layers', 'class_name': 'Dense', 'config': {'name': 'attention_v', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'units': 1, 'activation': 'linear', 'use_bias': True, 'kernel_initializer': {'module': 'keras.initializers', 'class_name': 'GlorotUniform', 'config': {'seed': None}, 'registered_name': None}, 'bias_initializer': {'module': 'keras.initializers', 'class_name': 'Zeros', 'config': {}, 'registered_name': None}, 'kernel_regularizer': None, 'bias_regularizer': None, 'kernel_constraint': None, 'bias_constraint': None}, 'registered_name': None, 'build_config': {'input_shape': (32, 139, 140, 256)}}}, 'registered_name': 'BahdanauAttention', 'build_config': {'input_shape': [(32, 140, 512), (32, 139, 256)]}}
- dense: {'module': 'keras.layers', 'class_name': 'Dense', 'config': {'name': 'dense', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'units': 40, 'activation': 'softmax', 'use_bias': True, 'kernel_initializer': {'module': 'keras.initializers', 'class_name': 'GlorotUniform', 'config': {'seed': None}, 'registered_name': None}, 'bias_initializer': {'module': 'keras.initializers', 'class_name': 'Zeros', 'config': {}, 'registered_name': None}, 'kernel_regularizer': None, 'bias_regularizer': None, 'kernel_constraint': None, 'bias_constraint': None}, 'registered_name': None, 'build_config': {'input_shape': (32, 139, 768)}}
Layer: enc_state_h
- name: enc_state_h
- trainable: True
- dtype: {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}
- units: 256
- activation: linear
- use_bias: True
- kernel_initializer: {'module': 'keras.initializers', 'class_name': 'GlorotUniform', 'config': {'seed': None}, 'registered_name': None}
- bias_initializer: {'module': 'keras.initializers', 'class_name': 'Zeros', 'config': {}, 'registered_name': None}
- kernel_regularizer: None
- bias_regularizer: None
- kernel_constraint: None
- bias_constraint: None
Layer: enc_state_c
- name: enc_state_c
- trainable: True
- dtype: {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}
- units: 256
- activation: linear
- use_bias: True
- kernel_initializer: {'module': 'keras.initializers', 'class_name': 'GlorotUniform', 'config': {'seed': None}, 'registered_name': None}
- bias_initializer: {'module': 'keras.initializers', 'class_name': 'Zeros', 'config': {}, 'registered_name': None}
- kernel_regularizer: None
- bias_regularizer: None
- kernel_constraint: None
- bias_constraint: None
Подробнее здесь: https://stackoverflow.com/questions/790 ... m-keras-mo
