i have an model< and difference input shape, if it's univariate so the inputshape is…
- User
i have an model< and difference input shape, if it's univariate so the input_shape is (128,1)< for multivariate it could be (128,4)
code for model
энкодер трансформера с регуляризацией и батч-нормализацией
def gru_branch(inputs, input_shape, units, dropout, l2_reg=0.01): x = layers.LayerNormalization(epsilon=1e-6)(inputs) x = layers.GRU(units, activation='relu', return_sequences=True, kernel_regularizer=regularizers.l2(l2_reg))(x) x = layers.Dropout(dropout)(x) x = layers.Dense(input_shape[0], activation='relu', kernel_regularizer=regularizers.l2(l2_reg))(x) return x + inputs
def transformer_encoder(inputs, head_size, num_heads, ff_dim, dropout=0): # Normalization and Attention x = layers.LayerNormalization(epsilon=1e-6)(inputs) x = layers.MultiHeadAttention( key_dim=head_size, num_heads=num_heads, dropout=dropout )(x, x) x = layers.Dropout(dropout)(x) res = x + inputs
# Feed Forward Part x = layers.LayerNormalization(epsilon=1e-6)(res) x = layers.Conv1D(filters=ff_dim, kernel_size=1, activation="relu")(x) x = layers.Dropout(dropout)(x) x = layers.Conv1D(filters=inputs.shape[-1], kernel_size=1)(x) return x + resdef build_model( input_shape, head_size, num_heads, ff_dim, num_transformer_blocks, mlp_units, gru_units, l2_reg, dropout, mlp_dropout ): inputs = keras.Input(shape=input_shape) x = inputs y = inputs
for _ in range(num_transformer_blocks): x = transformer_encoder(x, head_size, num_heads, ff_dim, dropout) y = gru_branch(y, input_shape, gru_units, dropout, l2_reg) #new x = layers.Concatenate()([x, y]) #new x = layers.GlobalAveragePooling1D(data_format="channels_first")(x) for dim in mlp_units: x = layers.Dense(dim, activation="relu")(x) x = layers.Dropout(mlp_dropout)(x) #univariate outputs = layers.Dense(64)(x) return keras.Model(inputs, outputs)adapt this code for both multi and uni variate inputs
- Assistant
Certainly! The key to adapting the model for both univariate and multivariate inputs lies in handling the difference in input shapes and ensuring the final layer provides an appropriate output shape for both cases. Below, the provided model is adapted to handle both univariate and multivariate inputs based on a parameter
num_features.from tensorflow.keras import layers, regularizers, models import tensorflow as tf def gru_branch(inputs, input_shape, units, dropout, l2_reg=0.01): x = layers.LayerNormalization(epsilon=1e-6)(inputs) x = layers.GRU(units, activation='relu', return_sequences=True, kernel_regularizer=regularizers.l2(l2_reg))(x) x = layers.Dropout(dropout)(x) x = layers.Dense(input_shape[0], activation='relu', kernel_regularizer=regularizers.l2(l2_reg))(x) return x + inputs def transformer_encoder(inputs, head_size, num_heads, ff_dim, dropout=0): # Normalization and Attention x = layers.LayerNormalization(epsilon=1e-6)(inputs) x = layers.MultiHeadAttention( key_dim=head_size, num_heads=num_heads, dropout=dropout )(x, x) x = layers.Dropout(dropout)(x) res = x + inputs # Feed Forward Part x = layers.LayerNormalization(epsilon=1e-6)(res) x = layers.Conv1D(filters=ff_dim, kernel_size=1, activation="relu")(x) x = layers.Dropout(dropout)(x) x = layers.Conv1D(filters=inputs.shape[-1], kernel_size=1)(x) return x + res def build_model( input_shape, head_size, num_heads, ff_dim, num_transformer_blocks, mlp_units, gru_units, l2_reg, dropout, mlp_dropout, num_outputs, # separate univariate/multivariate by different outputs ): inputs = keras.Input(shape=input_shape) x = inputs y = inputs for _ in range(num_transformer_blocks): x = transformer_encoder(x, head_size, num_heads, ff_dim, dropout) y = gru_branch(y, input_shape, gru_units, dropout, l2_reg) #new x = layers.Concatenate()([x, y]) #new x = layers.GlobalAveragePooling1D(data_format="channels_first")(x) for dim in mlp_units: x = layers.Dense(dim, activation="relu")(x) x = layers.Dropout(mlp_dropout)(x) outputs = layers.Dense(num_outputs)(x) # Adjusted to handle different output dimensions return models.Model(inputs, outputs)Example Usage
For Univariate Input:
univariate_shape = (128, 1) num_outputs = 1 # For univariate time series output model = build_model( input_shape=univariate_shape, head_size=256, num_heads=4, ff_dim=4, num_transformer_blocks=4, mlp_units=[128], gru_units=128, l2_reg=0.01, dropout=0.1, mlp_dropout=0.1, num_outputs=num_outputs )For Multivariate Input:
multivariate_shape = (128, 4) num_outputs = 4 # For multivariate time series output model = build_model( input_shape=multivariate_shape, head_size=256, num_heads=4, ff_dim=4, num_transformer_blocks=4, mlp_units=[128], gru_units=128, l2_reg=0.01, dropout=0.1, mlp_dropout=0.1, num_outputs=num_outputs )By adjusting
input_shapeandnum_outputsas needed, this model can cater to both univariate and multivariate time series data. - User
no outputs don't only depend on num_features, output should be as (forecast_length, num_features)
- Assistant
I understand. The output shape needs to be
(forecast_length, num_features)whereforecast_lengthis the length of the prediction horizon, andnum_featurescorresponds to the number of features in the time series (e.g., 1 for univariate, ornfor multivariate).Below is the adapted code that handles both cases as you described:
from tensorflow.keras import layers, regularizers, models import tensorflow as tf def gru_branch(inputs, input_shape, units, dropout, l2_reg=0.01): x = layers.LayerNormalization(epsilon=1e-6)(inputs) x = layers.GRU(units, activation='relu', return_sequences=True, kernel_regularizer=regularizers.l2(l2_reg))(x) x = layers.Dropout(dropout)(x) x = layers.Dense(input_shape[0], activation='relu', kernel_regularizer=regularizers.l2(l2_reg))(x) return x + inputs def transformer_encoder(inputs, head_size, num_heads, ff_dim, dropout=0): # Normalization and Attention x = layers.LayerNormalization(epsilon=1e-6)(inputs) x = layers.MultiHeadAttention( key_dim=head_size, num_heads=num_heads, dropout=dropout )(x, x) x = layers.Dropout(dropout)(x) res = x + inputs # Feed Forward Part x = layers.LayerNormalization(epsilon=1e-6)(res) x = layers.Conv1D(filters=ff_dim, kernel_size=1, activation="relu")(x) x = layers.Dropout(dropout)(x) x = layers.Conv1D(filters=inputs.shape[-1], kernel_size=1)(x) return x + res def build_model( input_shape, head_size, num_heads, ff_dim, num_transformer_blocks, mlp_units, gru_units, l2_reg, dropout, mlp_dropout, forecast_length ): num_features = input_shape[1] inputs = keras.Input(shape=input_shape) x = inputs y = inputs for _ in range(num_transformer_blocks): x = transformer_encoder(x, head_size, num_heads, ff_dim, dropout) y = gru_branch(y, input_shape, gru_units, dropout, l2_reg) #new x = layers.Concatenate()([x, y]) #new x = layers.GlobalAveragePooling1D(data_format="channels_first")(x) for dim in mlp_units: x = layers.Dense(dim, activation="relu")(x) x = layers.Dropout(mlp_dropout)(x) # Outputs has shape (forecast_length, num_features) output_shape = (forecast_length, num_features) x = layers.RepeatVector(forecast_length)(x) outputs = layers.TimeDistributed(layers.Dense(num_features))(x) return models.Model(inputs, outputs) # Example Usage univariate_shape = (128, 1) multivariate_shape = (128, 4) forecast_length = 10 # Length of the forecast horizon # For Univariate Input univariate_model = build_model( input_shape=univariate_shape, head_size=256, num_heads=4, ff_dim=4, num_transformer_blocks=4, mlp_units=[128], gru_units=128, l2_reg=0.01, dropout=0.1, mlp_dropout=0.1, forecast_length=forecast_length ) # For Multivariate Input multivariate_model = build_model( input_shape=multivariate_shape, head_size=256, num_heads=4, ff_dim=4, num_transformer_blocks=4, mlp_units=[128], gru_units=128, l2_reg=0.01, dropout=0.1, mlp_dropout=0.1, forecast_length=forecast_length ) # Print model summaries for validation univariate_model.summary() multivariate_model.summary()This code constructs a model that can handle both univariate and multivariate inputs and produces outputs with the shape
(forecast_length, num_features). The key is theRepeatVectorandTimeDistributedlayers, which ensure the correct forecasting shape and structure for the multi-step predictions. - User
check what's wrong
энкодер трансформера с регуляризацией и батч-нормализацией
def gru_branch(inputs, input_shape, units, dropout, l2_reg=0.01): x = layers.LayerNormalization(epsilon=1e-6)(inputs) x = layers.GRU(units, activation='tanh', return_sequences=True, kernel_regularizer=regularizers.l2(l2_reg))(x) x = layers.Dropout(dropout)(x) x = layers.Dense(input_shape[0], activation='relu', kernel_regularizer=regularizers.l2(l2_reg))(x) return x + inputs
def transformer_encoder(inputs, head_size, num_heads, ff_dim, dropout=0): # Normalization and Attention x = layers.LayerNormalization(epsilon=1e-6)(inputs) x = layers.MultiHeadAttention( key_dim=head_size, num_heads=num_heads, dropout=dropout )(x, x) x = layers.Dropout(dropout)(x) res = x + inputs
# Feed Forward Part x = layers.LayerNormalization(epsilon=1e-6)(res) x = layers.Conv1D(filters=ff_dim, kernel_size=1, activation="relu")(x) x = layers.Dropout(dropout)(x) x = layers.Conv1D(filters=inputs.shape[-1], kernel_size=1)(x) return x + resэнкодер трансформера с регуляризацией и батч-нормализацией
def gru_branch(inputs, input_shape, units, dropout, l2_reg=0.01): x = layers.LayerNormalization(epsilon=1e-6)(inputs) x = layers.GRU(units, activation='tanh', return_sequences=True, kernel_regularizer=regularizers.l2(l2_reg))(x) x = layers.Dropout(dropout)(x) x = layers.Dense(input_shape[0], activation='relu', kernel_regularizer=regularizers.l2(l2_reg))(x) return x + inputs
def transformer_encoder(inputs, head_size, num_heads, ff_dim, dropout=0): # Normalization and Attention x = layers.LayerNormalization(epsilon=1e-6)(inputs) x = layers.MultiHeadAttention( key_dim=head_size, num_heads=num_heads, dropout=dropout )(x, x) x = layers.Dropout(dropout)(x) res = x + inputs
# Feed Forward Part x = layers.LayerNormalization(epsilon=1e-6)(res) x = layers.Conv1D(filters=ff_dim, kernel_size=1, activation="relu")(x) x = layers.Dropout(dropout)(x) x = layers.Conv1D(filters=inputs.shape[-1], kernel_size=1)(x) return x + resclass CombinedLoss(Loss): def init(self, alpha=0.5): super(CombinedLoss, self).init() self.alpha = alpha
def call(self, y_true, y_pred): epsilon = tf.keras.backend.epsilon() # Малое значение для избежания деления на ноль diff = tf.abs(y_true - y_pred) + epsilon # Разница между предсказанным и истинным значением log_diff = -tf.math.log(diff) # Логарифм от разницы, чтобы минимизировать её log_loss = tf.reduce_mean(log_diff) mse_loss = tf.reduce_mean(tf.square(y_true - y_pred)) # MSE loss combined_loss = self.alpha * log_loss + (1 - self.alpha) * mse_loss return combined_lossclass CombinedLoss(Loss): def init(self, alpha=0.5): super(CombinedLoss, self).init() self.alpha = alpha
def call(self, y_true, y_pred): epsilon = tf.keras.backend.epsilon() # Малое значение для избежания деления на ноль diff = tf.abs(y_true - y_pred) + epsilon # Разница между предсказанным и истинным значением log_diff = -tf.math.log(diff) # Логарифм от разницы, чтобы минимизировать её log_loss = tf.reduce_mean(log_diff) mse_loss = tf.reduce_mean(tf.square(y_true - y_pred)) # MSE loss combined_loss = self.alpha * log_loss + (1 - self.alpha) * mse_loss return combined_lossОбучение
if TRAINING: with tf.device('/GPU:0'): # Старт логирования в MLflow mlflow.keras.autolog()
with mlflow.start_run(experiment_id=experiment_id, run_name=run_name): model.fit( train_dataset, validation_data=val_dataset, epochs=num_epochs, batch_size=batch_size, callbacks=[reduce_lr_callback, early_stopping_callback, checkpoint_callback, lr_scheduler] ) #Сохранение модели в MLflow model.save_weights(checkpoint_filepath) mlflow.log_artifact(checkpoint_filepath) #Завершение логирования в Mlflow mlflow.end_run()InvalidArgumentError Traceback (most recent call last) Cell In[21], line 9 6 mlflow.keras.autolog() 8 with mlflow.start_run(experiment_id=experiment_id, run_name=run_name): ----> 9 model.fit( 10 train_dataset, 11 validation_data=val_dataset, 12 epochs=num_epochs, 13 batch_size=batch_size, 14 callbacks=[reduce_lr_callback, early_stopping_callback, checkpoint_callback, lr_scheduler] 15 ) 17 #Сохранение модели в MLflow 18 model.save_weights(checkpoint_filepath)
File ~\miniconda3\envs\pronin\lib\site-packages\mlflow\utils\autologging_utils\safety.py:574, in safe_patch..safe_patch_function(*args, **kwargs) 564 try_log_autologging_event( 565 AutologgingEventLogger.get_logger().log_patch_function_start, 566 session, (...) 570 kwargs, 571 ) 573 if patch_is_class: --> 574 patch_function.call(call_original, *args, **kwargs) 575 else: 576 patch_function(call_original, *args, **kwargs)
File ~\miniconda3\envs\pronin\lib\site-packages\mlflow\utils\autologging_utils\safety.py:165, in PatchFunction.call(cls, original, *args, **kwargs) 163 @classmethod 164 def call(cls, original, *args, **kwargs): --> 165 return cls().call(original, *args, **kwargs)
File ~\miniconda3\envs\pronin\lib\site-packages\mlflow\utils\autologging_utils\safety.py:176, in PatchFunction.call(self, original, *args, **kwargs) 172 self._on_exception(e) 173 finally: 174 # Regardless of what happens during the
_on_exceptioncallback, reraise 175 # the original implementation exception once the callback completes --> 176 raise eFile ~\miniconda3\envs\pronin\lib\site-packages\mlflow\utils\autologging_utils\safety.py:169, in PatchFunction.call(self, original, *args, **kwargs) 167 def call(self, original, *args, **kwargs): 168 try: --> 169 return self._patch_implementation(original, *args, **kwargs) 170 except (Exception, KeyboardInterrupt) as e: 171 try:
File ~\miniconda3\envs\pronin\lib\site-packages\mlflow\utils\autologging_utils\safety.py:227, in with_managed_run..PatchWithManagedRun._patch_implementation(self, original, *args, **kwargs) 224 if not mlflow.active_run(): 225 self.managed_run = create_managed_run() --> 227 result = super()._patch_implementation(original, *args, **kwargs) 229 if self.managed_run: 230 mlflow.end_run(RunStatus.to_string(RunStatus.FINISHED))
File ~\miniconda3\envs\pronin\lib\site-packages\mlflow\tensorflow_init_.py:1333, in autolog..FitPatch._patch_implementation(self, original, inst, *args, **kwargs) 1326 except Exception as e: 1327 _logger.warning( 1328 "Failed to log training dataset information to " 1329 "MLflow Tracking. Reason: %s", 1330 e, 1331 ) -> 1333 history = original(inst, *args, **kwargs) 1335 if log_models: 1336 _log_keras_model(history, args)
File ~\miniconda3\envs\pronin\lib\site-packages\mlflow\utils\autologging_utils\safety.py:557, in safe_patch..safe_patch_function..call_original(og_args, **og_kwargs) 554 original_result = original(_og_args, **_og_kwargs) 555 return original_result --> 557 return call_original_fn_with_event_logging(_original_fn, og_args, og_kwargs)
File ~\miniconda3\envs\pronin\lib\site-packages\mlflow\utils\autologging_utils\safety.py:492, in safe_patch..safe_patch_function..call_original_fn_with_event_logging(original_fn, og_args, og_kwargs) 483 try: 484 try_log_autologging_event( 485 AutologgingEventLogger.get_logger().log_original_function_start, 486 session, (...) 490 og_kwargs, 491 ) --> 492 original_fn_result = original_fn(*og_args, **og_kwargs) 494 try_log_autologging_event( 495 AutologgingEventLogger.get_logger().log_original_function_success, 496 session, (...) 500 og_kwargs, 501 ) 502 return original_fn_result
File ~\miniconda3\envs\pronin\lib\site-packages\mlflow\utils\autologging_utils\safety.py:554, in safe_patch..safe_patch_function..call_original.._original_fn(_og_args, **_og_kwargs) 546 # Show all non-MLflow warnings as normal (i.e. not as event logs) 547 # during original function execution, even if silent mode is enabled 548 # (
silent=True), since these warnings originate from the ML framework 549 # or one of its dependencies and are likely relevant to the caller 550 with set_non_mlflow_warnings_behavior_for_current_thread( 551 disable_warnings=False, 552 reroute_warnings=False, 553 ): --> 554 original_result = original(_og_args, **_og_kwargs) 555 return original_resultFile ~\miniconda3\envs\pronin\lib\site-packages\keras\utils\traceback_utils.py:70, in filter_traceback..error_handler(*args, **kwargs) 67 filtered_tb = _process_traceback_frames(e.traceback) 68 # To get the full stack trace, call: 69 #
tf.debugging.disable_traceback_filtering()---> 70 raise e.with_traceback(filtered_tb) from None 71 finally: 72 del filtered_tbFile ~\miniconda3\envs\pronin\lib\site-packages\tensorflow\python\eager\execute.py:54, in quick_execute(op_name, num_outputs, inputs, attrs, ctx, name) 52 try: 53 ctx.ensure_initialized() ---> 54 tensors = pywrap_tfe.TFE_Py_Execute(ctx._handle, device_name, op_name, 55 inputs, attrs, num_outputs) 56 except core._NotOkStatusException as e: 57 if name is not None:
InvalidArgumentError: Graph execution error:
Detected at node 'CombinedLoss/sub' defined at (most recent call last): File "C:\Users\tecon\miniconda3\envs\pronin\lib\runpy.py", line 196, in _run_module_as_main return _run_code(code, main_globals, None, File "C:\Users\tecon\miniconda3\envs\pronin\lib\runpy.py", line 86, in _run_code exec(code, run_globals) File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\ipykernel_launcher.py", line 18, in app.launch_new_instance() File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\traitlets\config\application.py", line 1075, in launch_instance app.start() File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\ipykernel\kernelapp.py", line 739, in start self.io_loop.start() File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\tornado\platform\asyncio.py", line 205, in start self.asyncio_loop.run_forever() File "C:\Users\tecon\miniconda3\envs\pronin\lib\asyncio\base_events.py", line 600, in run_forever self._run_once() File "C:\Users\tecon\miniconda3\envs\pronin\lib\asyncio\base_events.py", line 1896, in _run_once handle._run() File "C:\Users\tecon\miniconda3\envs\pronin\lib\asyncio\events.py", line 80, in _run self._context.run(self._callback, *self._args) File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\ipykernel\kernelbase.py", line 545, in dispatch_queue await self.process_one() File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\ipykernel\kernelbase.py", line 534, in process_one await dispatch(*args) File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\ipykernel\kernelbase.py", line 437, in dispatch_shell await result File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\ipykernel\ipkernel.py", line 362, in execute_request await super().execute_request(stream, ident, parent) File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\ipykernel\kernelbase.py", line 778, in execute_request reply_content = await reply_content File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\ipykernel\ipkernel.py", line 449, in do_execute res = shell.run_cell( File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\ipykernel\zmqshell.py", line 549, in run_cell return super().run_cell(*args, **kwargs) File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\IPython\core\interactiveshell.py", line 3075, in run_cell result = self._run_cell( File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\IPython\core\interactiveshell.py", line 3130, in _run_cell result = runner(coro) File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\IPython\core\async_helpers.py", line 129, in pseudo_sync_runner coro.send(None) File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\IPython\core\interactiveshell.py", line 3334, in run_cell_async has_raised = await self.run_ast_nodes(code_ast.body, cell_name, File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\IPython\core\interactiveshell.py", line 3517, in run_ast_nodes if await self.run_code(code, result, async=asy): File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\IPython\core\interactiveshell.py", line 3577, in run_code exec(code_obj, self.user_global_ns, self.user_ns) File "C:\Users\tecon\AppData\Local\Temp\ipykernel_112156\338443566.py", line 9, in model.fit( File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\mlflow\utils\autologging_utils\safety.py", line 574, in safe_patch_function patch_function.call(call_original, *args, **kwargs) File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\mlflow\utils\autologging_utils\safety.py", line 165, in call return cls().call(original, *args, **kwargs) File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\mlflow\utils\autologging_utils\safety.py", line 169, in call return self._patch_implementation(original, *args, **kwargs) File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\mlflow\utils\autologging_utils\safety.py", line 227, in _patch_implementation result = super().patch_implementation(original, *args, **kwargs) File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\mlflow\tensorflow_init.py", line 1333, in _patch_implementation history = original(inst, *args, **kwargs) File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\mlflow\utils\autologging_utils\safety.py", line 557, in call_original return call_original_fn_with_event_logging(_original_fn, og_args, og_kwargs) File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\mlflow\utils\autologging_utils\safety.py", line 492, in call_original_fn_with_event_logging original_fn_result = original_fn(og_args, **og_kwargs) File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\mlflow\utils\autologging_utils\safety.py", line 554, in _original_fn original_result = original(_og_args, **_og_kwargs) File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\keras\utils\traceback_utils.py", line 65, in error_handler return fn(*args, **kwargs) File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\keras\engine\training.py", line 1564, in fit tmp_logs = self.train_function(iterator) File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\keras\engine\training.py", line 1160, in train_function return step_function(self, iterator) File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\keras\engine\training.py", line 1146, in step_function outputs = model.distribute_strategy.run(run_step, args=(data,)) File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\keras\engine\training.py", line 1135, in run_step outputs = model.train_step(data) File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\keras\engine\training.py", line 994, in train_step loss = self.compute_loss(x, y, y_pred, sample_weight) File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\keras\engine\training.py", line 1052, in compute_loss return self.compiled_loss( File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\keras\engine\compile_utils.py", line 265, in call loss_value = loss_obj(y_t, y_p, sample_weight=sw) File "C:\Users\tecon\miniconda3\envs\pronin\lib\site-packages\keras\losses.py", line 152, in call losses = call_fn(y_true, y_pred) File "C:\Users\tecon\AppData\Local\Temp\ipykernel_112156\2206673266.py", line 26, in call diff = tf.abs(y_true - y_pred) + epsilon # Разница между предсказанным и истинным значением Node: 'CombinedLoss/sub' required broadcastable shapes [[{{node CombinedLoss/sub}}]] [Op:__inference_train_function_9647]
- Assistant
It looks like the error you're encountering is related to incompatible shapes during the computation of the loss function. Specifically, the line
diff = tf.abs(y_true - y_pred) + epsilonis causing an issue because the shapes ofy_trueandy_predare not broadcastable (i.e., they can't be aligned for an element-wise subtraction).To diagnose and resolve this, you should:
- Ensure that the shapes of
y_trueandy_predmatch before they are processed in the loss function. - Verify the shapes of the model outputs and the true labels used during training.
Here are the changes needed:
-
Ensure that
y_trueandy_predhave the same shape:- Both
y_trueandy_predshould have the shape(batch_size, forecast_length, num_features).
- Both
-
Ensure correct shape in the model’s build function:
- The final output layer should output a tensor of shape
(batch_size, forecast_length, num_features).
- The final output layer should output a tensor of shape
Let's update the model and the loss function accordingly:
Updated CombinedLoss Class
import tensorflow as tf from tensorflow.keras.losses import Loss class CombinedLoss(Loss): def __init__(self, alpha=0.5): super(CombinedLoss, self).__init__() self.alpha = alpha def call(self, y_true, y_pred): epsilon = tf.keras.backend.epsilon() # Small value to avoid dividing by zero diff = tf.abs(tf.cast(y_true, tf.float32) - tf.cast(y_pred, tf.float32)) + epsilon # Cast to float32 log_diff = -tf.math.log(diff) # Logarithm of the difference to minimize it log_loss = tf.reduce_mean(log_diff) mse_loss = tf.reduce_mean(tf.square(y_true - y_pred)) # MSE loss combined_loss = self.alpha * log_loss + (1 - self.alpha) * mse_loss return combined_lossUpdated build_model Function
from tensorflow.keras import layers, regularizers, models import keras def gru_branch(inputs, input_shape, units, dropout, l2_reg=0.01): x = layers.LayerNormalization(epsilon=1e-6)(inputs) x = layers.GRU(units, activation='tanh', return_sequences=True, kernel_regularizer=regularizers.l2(l2_reg))(x) x = layers.Dropout(dropout)(x) x = layers.Dense(input_shape[0], activation='relu', kernel_regularizer=regularizers.l2(l2_reg))(x) return x + inputs def transformer_encoder(inputs, head_size, num_heads, ff_dim, dropout=0): x = layers.LayerNormalization(epsilon=1e-6)(inputs) x = layers.MultiHeadAttention( key_dim=head_size, num_heads=num_heads, dropout=dropout )(x, x) x = layers.Dropout(dropout)(x) res = x + inputs x = layers.LayerNormalization(epsilon=1e-6)(res) x = layers.Conv1D(filters=ff_dim, kernel_size=1, activation="relu")(x) x = layers.Dropout(dropout)(x) x = layers.Conv1D(filters=inputs.shape[-1], kernel_size=1)(x) return x + res def build_model( input_shape, head_size, num_heads, ff_dim, num_transformer_blocks, mlp_units, gru_units, l2_reg, dropout, mlp_dropout, forecast_length ): num_features = input_shape[1] inputs = keras.Input(shape=input_shape) x = inputs y = inputs for _ in range(num_transformer_blocks): x = transformer_encoder(x, head_size, num_heads, ff_dim, dropout) y = gru_branch(y, input_shape, gru_units, dropout, l2_reg) x = layers.Concatenate()([x, y]) x = layers.GlobalAveragePooling1D()(x) # Note: No data_format argument for dim in mlp_units: x = layers.Dense(dim, activation="relu")(x) x = layers.Dropout(mlp_dropout)(x) x = layers.RepeatVector(forecast_length)(x) outputs = layers.TimeDistributed(layers.Dense(num_features))(x) return models.Model(inputs, outputs) # Example Usage univariate_shape = (128, 1) multivariate_shape = (128, 4) forecast_length = 10 # For Univariate Input univariate_model = build_model( input_shape=univariate_shape, head_size=256, num_heads=4, ff_dim=4, num_transformer_blocks=4, mlp_units=[128], gru_units=128, l2_reg=0.01, dropout=0.1, mlp_dropout=0.1, forecast_length=forecast_length ) # For Multivariate Input multivariate_model = build_model( input_shape=multivariate_shape, head_size=256, num_heads=4, ff_dim=4, num_transformer_blocks=4, mlp_units=[128], gru_units=128, l2_reg=0.01, dropout=0.1, mlp_dropout=0.1, forecast_length=forecast_length ) # Print model summaries for validation univariate_model.summary() multivariate_model.summary()Training Loop
Ensure you compile your model with the correct loss function, and double-check your dataset to ensure shapes match:
from keras import optimizers # Use the custom loss loss_function = CombinedLoss() # Compile the model model.compile( optimizer=optimizers.Adam(), loss=loss_function, metrics=['mae'] ) # Training: Ensure train_dataset and val_dataset have shapes (batch_size, 128, num_features) if TRAINING: with tf.device('/GPU:0'): mlflow.keras.autolog() with mlflow.start_run(experiment_id=experiment_id, run_name=run_name): history = model.fit( train_dataset, validation_data=val_dataset, epochs=num_epochs, batch_size=batch_size, callbacks=[reduce_lr_callback, early_stopping_callback, checkpoint_callback, lr_scheduler] ) model.save_weights(checkpoint_filepath) mlflow.log_artifact(checkpoint_filepath) mlflow.end_run()Ensure your datasets (both
train_datasetandval_dataset) yield data in the shape(batch_size, 128, num_features)where128is the sequence length andnum_featuresis either 1 (univariate) or more (multivariate). This way, during loss computation,y_trueandy_predwill have compatible shapes. - Ensure that the shapes of