Added AdamW optimizer, misc bug fixes

pom.xml

@@ -18,7 +18,7 @@
     <!-- .. SE-Libraries .................................................. -->
 
       <CNNArch.version>0.3.5-SNAPSHOT</CNNArch.version>
-      <CNNTrain.version>0.3.10-SNAPSHOT</CNNTrain.version>
+      <CNNTrain.version>0.3.11-SNAPSHOT</CNNTrain.version>
       <CNNArch2X.version>0.0.6-SNAPSHOT</CNNArch2X.version>
       <embedded-montiarc-math-opt-generator>0.1.6</embedded-montiarc-math-opt-generator>
       <EMADL2PythonWrapper.version>0.0.2-SNAPSHOT</EMADL2PythonWrapper.version>

src/main/java/de/monticore/lang/monticar/cnnarch/gluongenerator/CNNTrain2Gluon.java

@@ -130,12 +130,25 @@ public class CNNTrain2Gluon extends CNNTrainGenerator {
 
         Map<String, String> fileContentMap = new HashMap<>();
 
+        //Context Information and Optimizer for local adaption during prediction for replay memory layer (the second only applicaple for supervised learning)
+        String cnnTrainLAOptimizerTemplateContent = templateConfiguration.processTemplate(ftlContext, "CNNLAOptimizer.ftl");
+        fileContentMap.put("CNNLAOptimizer_" + getInstanceName() + ".h", cnnTrainLAOptimizerTemplateContent);
+
+        //AdamW optimizer if used for training
+        String optimizerName = configuration.getOptimizer().getName();
+        Optional<OptimizerSymbol> criticOptimizer = configuration.getCriticOptimizer();
+        String criticOptimizerName = "";
+        if (criticOptimizer.isPresent()){
+            criticOptimizerName = criticOptimizer.get().getName();
+        }
+        if (optimizerName.equals("adamw") || criticOptimizerName.equals("adamw")){
+            String adamWContent = templateConfiguration.processTemplate(ftlContext, "Optimizer/AdamW.ftl");
+            fileContentMap.put("AdamW.py", adamWContent);
+        }
+
         if (configData.isSupervisedLearning()) {
             String cnnTrainTrainerTemplateContent = templateConfiguration.processTemplate(ftlContext, "CNNTrainer.ftl");
             fileContentMap.put("CNNTrainer_" + getInstanceName() + ".py", cnnTrainTrainerTemplateContent);
-            //Optimizer for local adaption during prediction for replay memory layer
-            String cnnTrainLAOptimizerTemplateContent = templateConfiguration.processTemplate(ftlContext, "CNNLAOptimizer.ftl");
-            fileContentMap.put("CNNLAOptimizer_" + getInstanceName() + ".h", cnnTrainLAOptimizerTemplateContent);
         } else if (configData.isGan()) {
             final String trainerName = "CNNTrainer_" + getInstanceName();
             if (!configuration.getDiscriminatorNetwork().isPresent()) {
@@ -192,7 +205,6 @@ public class CNNTrain2Gluon extends CNNTrainGenerator {
 
             final String ganTrainerContent = templateConfiguration.processTemplate(ftlContext, "gan/Trainer.ftl");
             fileContentMap.put(trainerName + ".py", ganTrainerContent);
-
         } else if (configData.isReinforcementLearning()) {
             final String trainerName = "CNNTrainer_" + getInstanceName();
             final RLAlgorithm rlAlgorithm = configData.getRlAlgorithm();

src/main/resources/templates/gluon/CNNCreator.ftl

@@ -40,7 +40,7 @@ class ${tc.fileNameWithoutEnding}:
 
             if os.path.isdir(self._model_dir_):
                 for file in os.listdir(self._model_dir_):
-                    if ".params" in file and self._model_prefix_ + "_" + str(i) in file:
+                    if ".params" in file and self._model_prefix_ + "_" + str(i) in file and not "loss" in file:
                         epochStr = file.replace(".params","").replace(self._model_prefix_ + "_" + str(i) + "-","")
                         epoch = int(epochStr)
                         if epoch > lastEpoch:
@@ -69,7 +69,7 @@ class ${tc.fileNameWithoutEnding}:
 
                     for file in os.listdir(self._weights_dir_):
 
-                        if ".params" in file and self._model_prefix_ + "_" + str(i) in file:
+                        if ".params" in file and self._model_prefix_ + "_" + str(i) in file and not "loss" in file:
                             epochStr = file.replace(".params","").replace(self._model_prefix_ + "_" + str(i) + "-","")
                             epoch = int(epochStr)
                             if epoch > lastEpoch:
@@ -87,9 +87,9 @@ class ${tc.fileNameWithoutEnding}:
             warnings.simplefilter("ignore")
             self.networks[${networkInstruction?index}].collect_params().initialize(self.weight_initializer, force_reinit=False, ctx=context)
         self.networks[${networkInstruction?index}].hybridize()
-        self.networks[${networkInstruction?index}](<#list tc.getStreamInputDimensions(networkInstruction.body) as dimensions><#if tc.cutDimensions(dimensions)[tc.cutDimensions(dimensions)?size-1] == "1">mx.nd.zeros((${tc.join(tc.cutDimensions(dimensions), ",")},), ctx=context[0])<#else>mx.nd.zeros((1, ${tc.join(tc.cutDimensions(dimensions), ",")},), ctx=context[0])</#if><#sep>, </#list>)
+        self.networks[${networkInstruction?index}](<#list tc.getStreamInputDimensions(networkInstruction.body) as dimensions><#if tc.cutDimensions(dimensions)[tc.cutDimensions(dimensions)?size-1] == "1" && tc.cutDimensions(dimensions)?size != 1>mx.nd.zeros((${tc.join(tc.cutDimensions(dimensions), ",")},), ctx=context[0])<#else>mx.nd.zeros((1, ${tc.join(tc.cutDimensions(dimensions), ",")},), ctx=context[0])</#if><#sep>, </#list>)
 <#if networkInstruction.body.episodicSubNetworks?has_content>
-        self.networks[0].episodicsubnet0_(<#list tc.getStreamInputDimensions(networkInstruction.body) as dimensions><#if tc.cutDimensions(dimensions)[tc.cutDimensions(dimensions)?size-1] == "1">mx.nd.zeros((${tc.join(tc.cutDimensions(dimensions), ",")},), ctx=context[0])<#else>mx.nd.zeros((1, ${tc.join(tc.cutDimensions(dimensions), ",")},), ctx=context[0])</#if><#sep>, </#list>)
+        self.networks[0].episodicsubnet0_(<#list tc.getStreamInputDimensions(networkInstruction.body) as dimensions><#if tc.cutDimensions(dimensions)[tc.cutDimensions(dimensions)?size-1] == "1" && tc.cutDimensions(dimensions)?size != 1>mx.nd.zeros((${tc.join(tc.cutDimensions(dimensions), ",")},), ctx=context[0])<#else>mx.nd.zeros((1, ${tc.join(tc.cutDimensions(dimensions), ",")},), ctx=context[0])</#if><#sep>, </#list>)
 </#if>
 </#list>
 

src/main/resources/templates/gluon/CNNGanTrainer.ftl

@@ -184,16 +184,16 @@ class ${tc.fileNameWithoutEnding}:
             del discriminator_optimizer_params['learning_rate_decay']
 
         if normalize:
-            self._net_creator_dis.construct(mx_context, data_mean=data_mean, data_std=data_std)
+            self._net_creator_dis.construct([mx_context], data_mean=data_mean, data_std=data_std)
         else:
-            self._net_creator_dis.construct(mx_context)
+            self._net_creator_dis.construct([mx_context])
 
-        self._net_creator_gen.construct(mx_context)
+        self._net_creator_gen.construct([mx_context])
 
         if self.use_qnet:
-            self._net_creator_qnet.construct(mx_context)
+            self._net_creator_qnet.construct([mx_context])
             if load_checkpoint:
-                self._net_creator_qnet.load(mx_context)
+                self._net_creator_qnet.load([mx_context])
             else:
                 if os.path.isdir(self._net_creator_qnet._model_dir_):
                     shutil.rmtree(self._net_creator_qnet._model_dir_)
@@ -206,8 +206,8 @@ class ${tc.fileNameWithoutEnding}:
 
         begin_epoch = 0
         if load_checkpoint:
-            begin_epoch = self._net_creator_dis.load(mx_context)
-            self._net_creator_gen.load(mx_context)
+            begin_epoch = self._net_creator_dis.load([mx_context])
+            self._net_creator_gen.load([mx_context])
         else:
             if os.path.isdir(self._net_creator_dis._model_dir_):
                 shutil.rmtree(self._net_creator_dis._model_dir_)
@@ -255,9 +255,9 @@ class ${tc.fileNameWithoutEnding}:
                 gen_input, exp_qnet_output = create_generator_input(batch)
 
                 with autograd.record():
-                    fake_data = gen_net(*gen_input)
+                    fake_data = gen_net(*gen_input)[0][0]
                     fake_data.detach()
-                    discriminated_fake_dis = dis_net(fake_data, *dis_conditional_input)
+                    discriminated_fake_dis = dis_net(fake_data, *dis_conditional_input)[0][0]
                     if self.use_qnet:
                         discriminated_fake_dis, _ = discriminated_fake_dis
 
@@ -265,7 +265,7 @@ class ${tc.fileNameWithoutEnding}:
                     real_labels = mx.nd.ones(discriminated_fake_dis.shape, ctx=mx_context)
 
                     loss_resultF = dis_loss(discriminated_fake_dis, fake_labels)
-                    discriminated_real_dis = dis_net(real_data, *dis_conditional_input)
+                    discriminated_real_dis = dis_net(real_data, *dis_conditional_input)[0][0]
                     if self.use_qnet:
                         discriminated_real_dis, _ = discriminated_real_dis
                     loss_resultR = dis_loss(discriminated_real_dis, real_labels)
@@ -276,8 +276,8 @@ class ${tc.fileNameWithoutEnding}:
 
                 if batch_i % k_value == 0:
                     with autograd.record():
-                        fake_data = gen_net(*gen_input)
-                        discriminated_fake_gen = dis_net(fake_data, *dis_conditional_input)
+                        fake_data = gen_net(*gen_input)[0][0]
+                        discriminated_fake_gen = dis_net(fake_data, *dis_conditional_input)[0][0]
                         if self.use_qnet:
                             discriminated_fake_gen, features = discriminated_fake_gen
                         loss_resultG = dis_loss(discriminated_fake_gen, real_labels)
@@ -285,7 +285,7 @@ class ${tc.fileNameWithoutEnding}:
                             condition = batch.data[traindata_to_index[generator_target_name + "_"]]
                             loss_resultG = loss_resultG + gen_loss_weight * generator_loss_func(fake_data, condition)
                         if self.use_qnet:
-                            qnet_discriminated = [q_net(features)]
+                            qnet_discriminated = [q_net(features)[0][0]]
                             for i, qnet_out in enumerate(qnet_discriminated):
                                 loss_resultG = loss_resultG + qnet_losses[i](qnet_out, exp_qnet_output[i])
                         loss_resultG.backward()

src/main/resources/templates/gluon/CNNModelLoader.ftl

@@ -82,12 +82,12 @@ public:
             loadComponent(network_json_path, network_param_path, network_symbol_list, network_param_map_list);
         }else{
             for(int i=0; i < num_subnets; i++){
-                network_json_path = file_prefix + "_replay_sub_net_" + std::to_string(i) + "-symbol.json";
-                network_param_path = file_prefix + "_replay_sub_net_" + std::to_string(i) + "-0000.params";
+                network_json_path = file_prefix + "_episodic_sub_net_" + std::to_string(i) + "-symbol.json";
+                network_param_path = file_prefix + "_episodic_sub_net_" + std::to_string(i) + "-0000.params";
                 loadComponent(network_json_path, network_param_path, network_symbol_list, network_param_map_list);
                 if(i >= 1){
-                    query_json_path = file_prefix + "_replay_query_net_" + std::to_string(i) + "-symbol.json";
-                    query_param_path = file_prefix + "_replay_query_net_" + std::to_string(i) + "-0000.params";
+                    query_json_path = file_prefix + "_episodic_query_net_" + std::to_string(i) + "-symbol.json";
+                    query_param_path = file_prefix + "_episodic_query_net_" + std::to_string(i) + "-0000.params";
                     loadComponent(query_json_path, query_param_path, query_symbol_list, query_param_map_list);
                     
                     memory_path = file_prefix + "_replay_memory_" + std::to_string(i);

src/main/resources/templates/gluon/CNNNet.ftl

@@ -378,34 +378,44 @@ class EpisodicMemory(EpisodicReplayMemoryInterface):
         #propagate the input as the rest is only used for replay
         return [args, []]
 
-    def store_samples(self, data, y, query_network, store_prob):
+    def store_samples(self, data, y, query_network, store_prob, context):
         num_pus = len(data)
         sub_batch_sizes = [data[i][0][0].shape[0] for i in range(num_pus)]
         num_inputs = len(data[0][0])
-        num_outputs = len(y[0])
+        num_outputs = len(y)
+        mx_context = context[0]
 
         if len(self.key_memory) == 0:
-            self.key_memory = nd.empty(0, ctx=mx.cpu())
-            self.value_memory = nd.empty(0, ctx=mx.cpu())
-            self.label_memory = nd.empty((num_outputs, 0), ctx=mx.cpu())
+            self.key_memory = nd.empty(0, ctx=mx_context)
+            self.value_memory = nd.empty(0, ctx=mx_context)
+            self.label_memory = nd.empty((num_outputs, 0), ctx=mx_context)
 
-        ind = nd.sample_multinomial(store_prob, sub_batch_sizes)
-        if len(sub_batch_sizes) == 1:
-            ind = nd.reshape(ind, (1,-1))
+        ind = [nd.sample_multinomial(store_prob, sub_batch_sizes[i]).as_in_context(mx_context) for i in range(num_pus)]
 
-        if(nd.max(ind)):
+        max_inds = [nd.max(ind[i]) for i in range(num_pus)]
+        if any(max_inds):
             to_store_values = []
             for i in range(num_inputs):
-                tmp_values = nd.contrib.boolean_mask(data[0][0][i].as_in_context(mx.cpu()), ind[0])
-                for j in range(1, num_pus):
-                    tmp_values = nd.concat(tmp_values, nd.contrib.boolean_mask(data[j][0][i].as_in_context(mx.cpu()), ind[j]), dim=0)
+                tmp_values = []
+                for j in range(0, num_pus):
+                    if max_inds[j]:
+                        #tmp = data[j][0][i].as_in_context(mx_context)
+                        if isinstance(tmp_values, list):
+
+                            tmp_values = nd.contrib.boolean_mask(data[j][0][i].as_in_context(mx_context), ind[j])
+                        else:
+                            tmp_values = nd.concat(tmp_values, nd.contrib.boolean_mask(data[j][0][i].as_in_context(mx_context), ind[j]), dim=0)
                 to_store_values.append(tmp_values)
 
-            to_store_labels = nd.empty((num_outputs, len(to_store_values[0])), ctx=mx.cpu())
+            to_store_labels = nd.empty((num_outputs, len(to_store_values[0])), ctx=mx_context)
             for i in range(num_outputs):
-                tmp_labels = nd.contrib.boolean_mask(y[0][i].as_in_context(mx.cpu()), ind[0])
-                for j in range(1, num_pus):
-                    tmp_labels = nd.concat(tmp_labels, nd.contrib.boolean_mask(y[j][i].as_in_context(mx.cpu()), ind[j]), dim=0)
+                tmp_labels = []
+                for j in range(0, num_pus):
+                    if max_inds[j]:
+                        if isinstance(tmp_labels, list):
+                            tmp_labels = nd.contrib.boolean_mask(y[i][j].as_in_context(mx_context), ind[j])
+                        else:
+                            tmp_labels = nd.concat(tmp_labels, nd.contrib.boolean_mask(y[i][j].as_in_context(mx_context), ind[j]), dim=0)
                 to_store_labels[i] = tmp_labels
 
             to_store_keys = query_network(*to_store_values[0:self.query_net_num_inputs])
@@ -441,7 +451,7 @@ class EpisodicMemory(EpisodicReplayMemoryInterface):
 
         return sample_batches
 
-    def get_query_network(self):
+    def get_query_network(self, context):
         lastEpoch = 0
         for file in os.listdir(self.query_net_dir):
             if self.query_net_prefix in file and ".json" in file:
@@ -462,16 +472,16 @@ class EpisodicMemory(EpisodicReplayMemoryInterface):
                 inputNames.append("data" + str(i))
         with warnings.catch_warnings():
             warnings.simplefilter("ignore")
-            net = mx.gluon.nn.SymbolBlock.imports(self.query_net_dir + symbolFile, inputNames, self.query_net_dir + weightFile, ctx=mx.cpu())
+            net = mx.gluon.nn.SymbolBlock.imports(self.query_net_dir + symbolFile, inputNames, self.query_net_dir + weightFile, ctx=context[0])
         net.hybridize()
         return net
     
     def save_memory(self, path):
-        mem_arr = [("keys", self.key_memory)] + [("labels", self.key_memory)] + [("values_"+str(k),v) for (k,v) in enumerate(self.value_memory)]
+        mem_arr = [("keys", self.key_memory)] + [("labels", self.label_memory)] + [("values_"+str(k),v) for (k,v) in enumerate(self.value_memory)]
         mem_dict = {entry[0]:entry[1] for entry in mem_arr}
         nd.save(path, mem_dict)
-    
-    
+
+
 <#list tc.architecture.networkInstructions as networkInstruction>
 #Stream ${networkInstruction?index}
 <#list networkInstruction.body.episodicSubNetworks as elements>
@@ -530,7 +540,7 @@ class Net_${networkInstruction?index}(gluon.HybridBlock):
 <#else>
 ${tc.include(networkInstruction.body, "ARCHITECTURE_DEFINITION")}
 </#if>    
-        pass
+            pass
 
     def hybrid_forward(self, F, ${tc.join(tc.getStreamInputNames(networkInstruction.body, false), ", ")}):
 <#if networkInstruction.body.episodicSubNetworks?has_content>

src/main/resources/templates/gluon/CNNPredictor.ftl

@@ -98,9 +98,10 @@ public:
             }
         }
 </#if>
-    
+
+        NDArray input_temp;
 <#list tc.getStreamInputNames(networkInstruction.body, false) as variable>
-        NDArray input_temp(network_input_shapes[${variable?index}], ctx, false, dtype);
+        input_temp = NDArray(network_input_shapes[${variable?index}], ctx, false, dtype);
         input_temp.SyncCopyFromCPU(in_${variable}.data(), network_input_sizes[${variable?index}]);
         input_temp.CopyTo(&(network_handles[0]->arg_dict()[network_input_keys[${variable?index}]]));
 </#list>
@@ -177,7 +178,8 @@ public:
         }
     
         for(mx_uint i=0; i < query_num_inputs[net_start_ind-1]; i++){
-            prev_output[I].CopyTo(&(query_handle->arg_dict()["data" + std::to_string(i)]));
+            prev_output[i].CopyTo(&(query_handle->arg_dict()["data" + std::to_string(i)]));
+        }
         NDArray::WaitAll();
 
         query_handle->Forward(false);
@@ -417,9 +419,19 @@ ${tc.include(networkInstruction.body, "PREDICTION_PARAMETER")}
 
         replay_memory = model_loader.GetReplayMemory();
         
+        std::vector<mx_uint> label_memory_shape = replay_memory[0]["labels"].GetShape();
+        std::vector<mx_uint> lab_shape;
+        if(label_memory_shape.size() == 2){
+            lab_shape = {1, 1};
+        }else{
+            lab_shape.push_back(1);
+            for(mx_uint i=2; i<label_memory_shape.size(); i++){
+                lab_shape.push_back(label_memory_shape[i]);
+            }
+        }
         std::vector<std::vector<mx_uint>> label_shapes;
         for(mx_uint i=0; i < num_outputs; i++){
-            label_shapes.push_back({1, 1}); //<--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
+            label_shapes.push_back(lab_shape);
         }
     
         Symbol loss = MakeLoss(model_loader.GetLoss());

src/main/resources/templates/gluon/CNNSupervisedTrainer.ftl

@@ -10,6 +10,11 @@ import math
 import sys
 import inspect
 from mxnet import gluon, autograd, nd
+try:
+    import AdamW
+except:
+    pass
+
 
 class CrossEntropyLoss(gluon.loss.Loss):
     def __init__(self, axis=-1, sparse_label=True, weight=None, batch_axis=0, **kwargs):
@@ -56,7 +61,7 @@ class SoftmaxCrossEntropyLossIgnoreIndices(gluon.loss.Loss):
             loss = -(pred * label).sum(axis=self._axis, keepdims=True)
         # ignore some indices for loss, e.g. <pad> tokens in NLP applications
         for i in self._ignore_indices:
-            loss = loss * mx.nd.logical_not(mx.nd.equal(mx.nd.argmax(pred, axis=1), mx.nd.ones_like(mx.nd.argmax(pred, axis=1))*i) * mx.nd.equal(mx.nd.argmax(pred, axis=1), label))
+            loss = F.broadcast_mul(loss, F.logical_not(F.broadcast_equal(F.argmax(pred, axis=1), F.ones_like(F.argmax(pred, axis=1))*i) * F.broadcast_equal(F.argmax(pred, axis=1), label)))
         return loss.mean(axis=self._batch_axis, exclude=True)
 
 class DiceLoss(gluon.loss.Loss):
@@ -293,6 +298,7 @@ class ${tc.fileNameWithoutEnding}:
             mx_context = [mx.cpu()]
         else:
             logging.error("Context argument is '" + context + "'. Only 'cpu' and 'gpu are valid arguments'.")
+        single_pu_batch_size = int(batch_size/num_pus)
 
         if preprocessing:
             preproc_lib = "CNNPreprocessor_${tc.fileNameWithoutEnding?keep_after("CNNSupervisedTrainer_")}_executor"
@@ -337,7 +343,10 @@ class ${tc.fileNameWithoutEnding}:
             if not os.path.isdir(self._net_creator._model_dir_):
                 raise
 
-        trainers = [mx.gluon.Trainer(network.collect_params(), optimizer, optimizer_params) for network in self._networks.values() if len(network.collect_params().values()) != 0]
+        if optimizer == "adamw":
+            trainers = [mx.gluon.Trainer(network.collect_params(), AdamW.AdamW(**optimizer_params)) for network in self._networks.values() if len(network.collect_params().values()) != 0]
+        else:
+            trainers = [mx.gluon.Trainer(network.collect_params(), optimizer, optimizer_params) for network in self._networks.values() if len(network.collect_params().values()) != 0]
 
         margin = loss_params['margin'] if 'margin' in loss_params else 1.0
         sparseLabel = loss_params['sparse_label'] if 'sparse_label' in loss_params else True
@@ -408,7 +417,7 @@ class ${tc.fileNameWithoutEnding}:
         layer = [param for param in inspect.getmembers(sub_net, lambda x: not(inspect.isroutine(x))) if param[0].startswith("memory")][0][1]
         episodic_layers[0].append(layer)
         episodic_store_buffer[${networkInstruction?index}].append([])
-        episodic_query_networks[0].append(episodic_layers[0][-1].get_query_network())
+        episodic_query_networks[0].append(episodic_layers[0][-1].get_query_network(mx_context))
         store_prob[${networkInstruction?index}].append(nd.array([1-episodic_layers[0][-1].store_prob, episodic_layers[0][-1].store_prob], ctx=mx.cpu()))                                                                                  
 </#if>
 </#list>
@@ -440,7 +449,12 @@ class ${tc.fileNameWithoutEnding}:
                 with autograd.record():
 <#include "pythonExecuteTrain.ftl">
 
-                for loss in losses:
+                    for i in range(num_pus):
+                        losses = [0]*num_pus
+                        for element in lossList[i]:
+                            losses[i] = losses[i] + element
+
+                for loss in losses: 
                     loss.backward()
                     loss_total += loss.sum().asscalar()
                     global_loss_train += loss.sum().asscalar()
@@ -462,7 +476,7 @@ class ${tc.fileNameWithoutEnding}:
                 #storing samples for episodic replay
                 for net_i in range(len(self._networks)):
                     for layer_i, layer in enumerate(episodic_layers[net_i]):
-                        layer.store_samples(episodic_store_buffer[net_i][layer_i], labels, episodic_query_networks[net_i][layer_i], store_prob[net_i][layer_i])
+                        layer.store_samples(episodic_store_buffer[net_i][layer_i], labels, episodic_query_networks[net_i][layer_i], store_prob[net_i][layer_i], mx_context)
 
 </#if>
                 if tic is None:
@@ -497,16 +511,14 @@ class ${tc.fileNameWithoutEnding}:
 
 <#include "saveAttentionImageTrain.ftl">
 
-                    for i in range(num_pus):
-                        predictions = []
-                        for output_name in outputs[i]:
-                            if mx.nd.shape_array(mx.nd.squeeze(output_name)).size > 1:
-                                predictions.append(mx.nd.argmax(output_name, axis=1))
-                            else:
-                                predictions.append(output_name)
+                    predictions = []
+                    for output_name in outputs:
+                        if mx.nd.shape_array(mx.nd.squeeze(output_name)).size > 1:
+                            predictions.append(mx.nd.argmax(output_name, axis=1))
+                        else:
+                            predictions.append(output_name)
 
-                        labels_metric = [[labels[j][i] for j in range(len(labels))] for i in range(num_pus)]
-                        metric.update(preds=predictions, labels=labels_metric[i])
+                    metric.update(preds=predictions, labels=[labels[j] for j in range(len(labels))])
 
                 train_metric_score = metric.get()[1]
             else:
@@ -524,22 +536,22 @@ class ${tc.fileNameWithoutEnding}:
 
 <#include "saveAttentionImageTest.ftl">
 
-                for loss in losses:
-                    global_loss_test += loss.sum().asscalar()
-    
+                for element in lossList:
+                    loss = loss + element
+
+                global_loss_test += loss.sum().asscalar()
+
                 test_batches += 1
 
-                for i in range(num_pus):
-                    predictions = []
-                    for output_name in outputs[i]:
-                        predictions.append(output_name)
+                predictions = []
+                for output_name in outputs:
+                    predictions.append(output_name)
 
-                    labels_metric = [[labels[j][i] for j in range(len(labels))] for i in range(num_pus)]
-                    metric.update(preds=predictions, labels=labels_metric[i])
+                metric.update(preds=predictions, labels=[labels[j] for j in range(len(labels))])
 
             test_metric_score = metric.get()[1]
 
-            global_loss_test /= (test_batches * batch_size)
+            global_loss_test /= (test_batches * single_pu_batch_size)
 
             logging.info("Epoch[%d] Train metric: %f, Test metric: %f, Train loss: %f, Test loss: %f" % (epoch, train_metric_score, test_metric_score, global_loss_train, global_loss_test))
 
@@ -547,10 +559,6 @@ class ${tc.fileNameWithoutEnding}:
                 for i, network in self._networks.items():
                     network.save_parameters(self.parameter_path(i) + '-' + str(epoch).zfill(4) + '.params')
 
-        print("--------------------------------------Speed------------------------------------------")
-        print(avg_speed/n)
-        print("--------------------------------------Speed------------------------------------------")
-    
         for i, network in self._networks.items():
             network.save_parameters(self.parameter_path(i) + '-' + str(num_epoch + begin_epoch + 1).zfill(4) + '.params')
             network.export(self.parameter_path(i) + '_newest', epoch=0)

src/main/resources/templates/gluon/Optimizer/AdamW.ftl

+import os
+import numpy as np
+import mxnet as mx
+from mxnet import nd
+from mxnet.ndarray.contrib import mp_adamw_update, adamw_update, multi_mp_adamw_update, multi_adamw_update
+    
+class AdamW(mx.optimizer.Optimizer):
+    def __init__(self, learning_rate=0.001, beta1=0.9, beta2=0.999, epsilon=1e-8, **kwargs):
+        super(AdamW, self).__init__(learning_rate=learning_rate, **kwargs)
+        self.beta1 = beta1
+        self.beta2 = beta2
+        self.epsilon = epsilon
+        self.aggregate_num = max(1, min(50, int(os.getenv('MXNET_OPTIMIZER_AGGREGATION_SIZE', '4'))))
+
+    def create_state(self, _, weight):
+        return (nd.zeros(weight.shape, weight.context, dtype=weight.dtype), #mean
+                nd.zeros(weight.shape, weight.context, dtype=weight.dtype)) #variance
+
+    def create_state_multi_precision(self, index, weight):
+        weight_master_copy = None
+        if self.multi_precision and weight.dtype == numpy.float16:
+            weight_master_copy = weight.astype(numpy.float32)
+            return (self.create_state(index, weight_master_copy), weight_master_copy)
+        return self.create_state(index, weight)
+
+    def update(self, index, weight, grad, state):
+        self._update_impl(index, weight, grad, state, multi_precision=False)
+
+    def update_multi_precision(self, index, weight, grad, state):
+        use_multi_precision = self.multi_precision and weight[0].dtype == numpy.float16
+        self._update_impl(index, weight, grad, state, multi_precision=use_multi_precision)
+
+    def _update_impl(self, indices, weight, grad, state, multi_precision=False):
+        """update function"""
+        aggregate = self.aggregate_num > 1
+        if not isinstance(indices, (tuple, list)):
+            indices = [indices]
+            weight = [weight]
+            grad = [grad]
+            state = [state]
+        for w_i, g_i in zip(weight, grad):
+            assert(isinstance(w_i, nd.NDArray))
+            assert(isinstance(g_i, nd.NDArray))
+            aggregate = (aggregate and
+                         w_i.stype == 'default' and
+                         g_i.stype == 'default')
+        self._update_count(indices)
+        lrs = self._get_lrs(indices)
+        wds = self._get_wds(indices)
+
+        # pylint: disable=access-member-before-definition
+        if not isinstance(self.rescale_grad, nd.NDArray):
+            self.rescale_grad = nd.full(shape=(1,), val=self.rescale_grad, ctx=weight[0].context)
+        else:
+            self.rescale_grad = self.rescale_grad.as_in_context(weight[0].context)
+
+        kwargs = {'beta1': self.beta1, 'beta2': self.beta2, 'epsilon': self.epsilon,
+                  'rescale_grad': self.rescale_grad}
+        if self.clip_gradient:
+            kwargs['clip_gradient'] = self.clip_gradient
+
+        if aggregate:
+            current_index = 0
+            while current_index < len(indices):
+                sidx = current_index
+                eidx = min(current_index + self.aggregate_num, len(indices))
+                if not multi_precision:
+                    mean, var = list(zip(*state[sidx:eidx]))
+                    multi_adamw_update(weight[sidx:eidx],
+                                       grad[sidx:eidx],
+                                       mean, var,
+                                       out=weight[sidx:eidx],
+                                       size=len(weight[sidx:eidx]),
+                                       lrs=list(np.ones(len(weight[sidx:eidx]))),
+                                       wds=wds[sidx:eidx],
+                                       etas=lrs[sidx:eidx],
+                                       **kwargs)
+                else:
+                    mean_var = list(zip(*state[sidx:eidx]))[0]
+                    tmean_var = list(zip(*mean_var))
+                    mean = tmean_var[0]
+                    var = tmean_var[1]
+                    multi_mp_adamw_update(weight[sidx:eidx],
+                                          grad[sidx:eidx],
+                                          mean, var,
+                                          list(zip(*state[sidx:eidx]))[1],
+                                          out=weight[sidx:eidx],
+                                          size=len(weight[sidx:eidx]),
+                                          lrs=list(np.ones(len(weight[sidx:eidx]))),
+                                          wds=wds[sidx:eidx],
+                                          etas=lrs[sidx:eidx],
+                                          **kwargs)
+                current_index += self.aggregate_num
+        else:
+            for w_i, g_i, s_i, lr, wd in zip(weight, grad, state, lrs, wds):
+                if not multi_precision:
+                    mean, var = s_i
+                    adamw_update(w_i, g_i, mean, var, out=w_i, lr=1, wd=wd, eta=lr, **kwargs)
+                else:
+                    mean, var = s_i[0]
+                    mp_adamw_update(w_i, g_i, mean, var, s_i[1], out=w_i, lr=1, wd=wd, eta=lr, **kwargs)    
\ No newline at end of file

src/main/resources/templates/gluon/elements/LoadNetwork.ftl

@@ -18,13 +18,13 @@
 
 <#if numInputs == "1">
             inputNames = ["data"]      
-            zeroInputs = [nd.zeros(<#if element.element.inputTypes[0].dimensions[0] == 1>(${tc.join(element.element.inputTypes[0].dimensions?reverse, ",")})<#else>(1,${tc.join(element.element.inputTypes[0].dimensions?reverse, ",")})</#if>, ctx=mx_context[0])]
+            zeroInputs = [nd.zeros(<#if element.element.inputTypes[0].dimensions[0] == 1 && tc.cutDimensionsInteger(element.element.inputTypes[0].dimensions)?size != 1>(${tc.join(element.element.inputTypes[0].dimensions?reverse, ",")})<#else>(1,${tc.join(element.element.inputTypes[0].dimensions?reverse, ",")})</#if>, ctx=mx_context[0])]
 <#else>
             inputNames = []
             zeroInputs = []
 <#list element.element.inputTypes as inType>
             inputNames.append("data" + str(${inType?index}))
-            zeroInputs.append(nd.zeros(<#if inType.dimensions[0] == 1>(${tc.join(tc.cutDimensionsInteger(inType.dimensions)?reverse, ",")})<#else>(1,${tc.join(tc.cutDimensionsInteger(inType.dimensions)?reverse, ",")})</#if>, ctx=mx_context[0]))
+            zeroInputs.append(nd.zeros(<#if inType.dimensions[0] == 1 && tc.cutDimensionsInteger(inType.dimensions)?size != 1>(${tc.join(tc.cutDimensionsInteger(inType.dimensions)?reverse, ",")})<#else>(1,${tc.join(tc.cutDimensionsInteger(inType.dimensions)?reverse, ",")})</#if>, ctx=mx_context[0]))
 </#list>
 </#if>
             with warnings.catch_warnings():

src/main/resources/templates/gluon/pythonEpisodicExecuteTrain.ftl

@@ -19,8 +19,6 @@
                                 for gradient_step in range(layer.replay_gradient_steps):
                                     with autograd.record():
 
-                                        nd.waitall()
-
                                         episodic_output = [self._networks[${networkInstruction?index}].episodic_sub_nets[layer_i](*(episodic_data[i]))[0] for i in range(num_pus)]
                                         for i in range(layer_i+1, len(episodic_layers[${networkInstruction?index}])):
                                             episodic_output = self._networks[${networkInstruction?index}].episodic_sub_nets[i](*episodic_output)[0]

src/main/resources/templates/gluon/pythonExecuteTest.ftl

-                    labels = [gluon.utils.split_and_load(batch.label[i], ctx_list=mx_context, even_split=False) for i in range(${tc.architectureOutputs?size?c})]
+                    labels = [gluon.utils.split_and_load(batch.label[i], ctx_list=mx_context, even_split=False)[0] for i in range(${tc.architectureOutputs?size?c})]
 <#list tc.architectureInputs as input_name>
-                    ${input_name} = gluon.utils.split_and_load(batch.data[${input_name?index}], ctx_list=mx_context, even_split=False)
+                    ${input_name} = gluon.utils.split_and_load(batch.data[${input_name?index}], ctx_list=mx_context, even_split=False)[0]
 </#list>
 
 <#if tc.architectureOutputSymbols?size gt 1>