modified readme

Readme.md

@@ -36,7 +36,8 @@ Run:
 
 `bash configure_compss.sh`
 
-Modify runcompss.sh to your needs
+Modify runcompss.sh to your needs, for example, if it is desired that some epochs are perfomed asynchronously the following parameter should be added to the runcompss call:
+`--num_async_epochs=5`
 
 Run:
 

docker/pyeddl/eddl_compss_distributed.py

@@ -119,16 +119,3 @@ class Eddl_Compss_Distributed:
         final_parameters = net_parametersToNumpy(eddl.get_parameters(model))
 
         return final_parameters
-
-    
-
-    @constraint(computing_units="${OMP_NUM_THREADS}")
-    @task(accumulated_parameters=COMMUTATIVE, parameters_to_aggregate=IN, mult_factor=IN, target_direction=IN)
-    def aggregate_parameters_async(self, accumulated_parameters, parameters_to_aggregate, mult_factor):
-
-        for i in range(0, len(accumulated_parameters)):
-            for j in range(0, len(accumulated_parameters[i])):
-                accumulated_parameters[i][j] = (
-                            (accumulated_parameters[i][j] + parameters_to_aggregate[i][j]) / 2).astype(np.float32)
-
-        return accumulated_parameters
\ No newline at end of file

docker/pyeddl/eddl_compss_distributed_api.py

@@ -74,50 +74,3 @@ def train_batch(model, x_train, y_train, num_workers, num_epochs_for_param_sync,
 
     # Set the parameters of the model to the aggregated parameters
     eddl.set_parameters(model, net_parametersToTensor(final_weights))
-
-
-def fit_async(model, x_train, y_train, num_workers, num_epochs_for_param_sync, max_num_async_epochs,
-              workers_batch_size):
-    global compss_object
-
-    # Define the number of images corresponding to each computing unit
-    num_total_samples = x_train.shape[0]
-    num_images_per_worker = int(num_total_samples / num_workers)
-
-    # Variable where parameters will be aggregated asynchornously
-    accumulated_parameters = net_parametersToNumpy(eddl.get_parameters(model))
-
-    # Define the parameters for each worker
-    workers_parameters = [net_parametersToNumpy(eddl.get_parameters(model)) for i in range(0, num_workers)]
-    x_blocks = [x[0] for x in paired_partition(x_train, y_train)]
-    y_blocks = [x[1] for x in paired_partition(x_train, y_train)]
-
-    # Until the maximum number of asynchrnous epochs is reached
-    for i in range(0, max_num_async_epochs):
-        # Train and aggregate the parameters asynchronously for each distributed computing unit
-        for j in range(0, num_workers):
-            shuffled_x, shuffled_y = block_shuffle_async(
-                x_blocks[j],
-                y_blocks[j],
-                workers_parameters[j])
-            
-            x_blocks[j], y_blocks[j] = [shuffled_x], [shuffled_y]
-
-            workers_parameters[j] = compss_object.train_batch(
-                x_blocks[j],
-                y_blocks[j],
-                workers_parameters[j],
-                num_images_per_worker,
-                num_epochs_for_param_sync,
-                workers_batch_size)
-            
-            workers_parameters[j] = compss_object.aggregate_parameters_async(
-                accumulated_parameters,
-                workers_parameters[j],
-                1 / num_workers)
-
-    # Wait until every computing unit has aggregated its parameters
-    accumulated_parameters = compss_wait_on(accumulated_parameters)
-
-    # Set the model parameters to the aggregated parameters
-    eddl.set_parameters(model, net_parametersToTensor(accumulated_parameters))