Merge pull request #29 from narendasan/mobilenet

Add operators necessary for mobilenet

.gitignore

@@ -13,4 +13,4 @@ experiments/
 py/build/
 py/tmp/
 py/.eggs
- 
+.vscode/

core/conversion/conversion.cpp

@@ -10,7 +10,7 @@ namespace trtorch {
 namespace core {
 namespace conversion {
 
-// Defined in core/conversion/conversion_blacklist.cpp 
+// Defined in core/conversion/conversion_blacklist.cpp
 bool isNodeConversionBlacklisted(const torch::jit::Node* n);
 
 bool OpSupported(const torch::jit::Node* n) {
@@ -24,8 +24,8 @@ c10::optional<torch::jit::IValue> EvaluateNode(ConversionCtx* ctx, const torch::
     // Also probably a better way to deal with the two error cases;
     TRTORCH_CHECK(level < limit, "Failed to evaluate node: " << *n              \
                            << "Reason: Exceeded evaluation stack limit (limit=" \
-                           << limit << ")"); 
-    
+                           << limit << ")");
+
     LOG_DEBUG(ctx->logger, "Evaluating " << util::node_info(n));
     evaluators::kwargs eval_args;
     for (auto eval_in : n->inputs()) {
@@ -55,7 +55,7 @@ c10::optional<torch::jit::IValue> EvaluateNode(ConversionCtx* ctx, const torch::
     return eval;
 }
 
-bool AddLayer(ConversionCtx* ctx, const torch::jit::Node* n) {
+void AddLayer(ConversionCtx* ctx, const torch::jit::Node* n) {
     LOG_INFO(ctx->logger,
              "Adding Layer " << util::node_info(n) << " (ctx.AddLayer)");
     converters::args node_args;
@@ -87,36 +87,34 @@ bool AddLayer(ConversionCtx* ctx, const torch::jit::Node* n) {
             TRTORCH_THROW_ERROR("Unable to retrieve all node inputs for node: " \
                                 << util::node_info(n) << " (ctx.AddLayer)\nSpecifically failed to retrieve value for input: " \
                                 << *input_node);
-            return false;
         }
-
     }
 
     if (n->inputs().size() != node_args.size()) {
         TRTORCH_THROW_ERROR("Unable to retrieve all node inputs for node: " << *n);
-        return false;
     }
 
-    
+
     auto schema = n->maybeSchema();
     TRTORCH_CHECK(schema, "Unable to get schema for Node " << util::node_info(n) \
                   << " (conversion.AddLayer)");
-    
+
     auto converter = converters::get_node_converter_for(schema);
     TRTORCH_CHECK(converter, "Unable to convert node: " << util::node_info(n) \
                   << " (conversion.AddLayer)\nSchema: " << *schema
                   << "\nConverter for " << schema->name()
                   << " requested, but no such converter was found.\nIf you need a converter for this operator, you can try implementing one yourself\n"
-                  << "or request a converter: https://www.github.com/NVIDIA/TRTorch/issues"); 
-    converter(ctx, n, node_args);
+                  << "or request a converter: https://www.github.com/NVIDIA/TRTorch/issues");
 
-    return true;
+    TRTORCH_CHECK(converter(ctx, n, node_args),
+                  "Converter for " << *schema << " failed to convert node: "
+                  << util::node_info(n) << "please report this error to https://www.github.com/NVIDIA/TRTorch/issues");
 }
 
-bool AddInputs(ConversionCtx* ctx,
+void AddInputs(ConversionCtx* ctx,
                 at::ArrayRef<const torch::jit::Value*> inputs,
                 std::vector<InputRange>& input_dims) {
-    
+
     auto type_lut = torch::jit::script::string_to_type_lut();
     std::vector<const torch::jit::Value*> input_tensors;
     for (auto in : inputs) {
@@ -130,15 +128,15 @@ bool AddInputs(ConversionCtx* ctx,
             input_tensors.push_back(in);
         }
     }
-    
+
     TRTORCH_CHECK(input_tensors.size() == input_dims.size(),
                   "Expected dimension specifications for all input tensors" \
                   << ", but found " << input_tensors.size()                 \
                   << " input tensors and "                                  \
                   << input_dims.size() << "dimension specs (conversion.AddInputs)");
 
     auto profile = ctx->builder->createOptimizationProfile();
-    
+
     for (size_t i = 0; i < input_tensors.size(); i++) {
         auto in = input_tensors[i];
         auto dims = input_dims[i];
@@ -158,20 +156,23 @@ bool AddInputs(ConversionCtx* ctx,
     }
 
     TRTORCH_CHECK(profile->isValid(), "Optimization profile is invalid, please check the input range provided (conversion.AddInputs)");
-    
+
     ctx->cfg->addOptimizationProfile(profile);
-    return true;
 }
 
-bool MarkOutputs(ConversionCtx* ctx, at::ArrayRef<const torch::jit::Value*> outputs) {
+void MarkOutputs(ConversionCtx* ctx, at::ArrayRef<const torch::jit::Value*> outputs) {
     for (auto out : outputs) {
-        ctx->net->markOutput(*(ctx->value_tensor_map[out]));
+        auto it = ctx->value_tensor_map.find(out);
+        // Leaves the potential for unused outputs to be populated with nullptr "safely"
+        TRTORCH_CHECK(it != ctx->value_tensor_map.end() && it->second,
+                      "No corresponding output TRT Tensor found for TorchScript output: " << out->debugName());
+        auto out_tensor = it->second;
+        ctx->net->markOutput(*out_tensor);
         LOG_INFO(ctx->logger,
                  "Marking Output " << out->debugName() << " (ctx.MarkOutput)");
     }
-    return true;
 }
-    
+
 void AddParamsToCtxValueMap(ConversionCtx* ctx, GraphParams& params) {
     for (auto p : params) {
         ctx->evaluated_value_map[p.first] = torch::jit::IValue(p.second.clone());
@@ -191,13 +192,8 @@ void ConvertBlockToNetDef(ConversionCtx* ctx, const torch::jit::Block* b, ExtraI
         bool to_eval = evaluators::shouldEvalAtConversionTime(n);
         bool blacklisted = isNodeConversionBlacklisted(n);
         if (!to_eval && !blacklisted) {
-            if (!AddLayer(ctx, n)) {
-                //TODO: Exception things 
-                LOG_ERROR(ctx->logger,
-                          "Failed to add layer: " << *n \
-                          << " (ctx.AddLayer)");
-                return;
-            }
+            // Should error out if something fails
+            AddLayer(ctx, n);
         } else {
             std::string reason = "";
             if (to_eval) {
@@ -207,7 +203,13 @@ void ConvertBlockToNetDef(ConversionCtx* ctx, const torch::jit::Block* b, ExtraI
                 reason += " (explicitly blacklisted)";
             }
             LOG_DEBUG(ctx->logger,
-                      "Skipping Node: " << (n->kind().toQualString()) << reason);
+                      "Skipping Node: " << util::node_info(n) << reason);
+        }
+    }
+
+    for (const auto n : nodes) {
+        if (converters::node_is_convertable(n)) {
+            ctx->CheckLayerAddition(n);
         }
     }
 
@@ -218,7 +220,7 @@ void ConvertBlockToNetDef(ConversionCtx* ctx, const torch::jit::Block* b, ExtraI
 // Converts a already lowered block (blocks with no sub blocks) to
 // a serialized TensorRT engine that can be deserialized and run
 
-// Probably should consolidate these two functions 
+// Probably should consolidate these two functions
 std::string ConvertBlockToEngine(const torch::jit::Block* b, ExtraInfo build_info, GraphParams& static_params) {
     ConversionCtx ctx(build_info.engine_settings);
     ConvertBlockToNetDef(&ctx, b, build_info, static_params);
@@ -247,7 +249,7 @@ bool VerifyConverterSupportForBlock(const torch::jit::Block* b) {
         for (auto s : unsupported_ops) {
             unsupported_msg << "  -  " << s << std::endl;
         }
-        unsupported_msg << "You can either implement converters for these ops in your application or file a bug" << std::endl;
+        unsupported_msg << "You can either implement converters for these ops in your application or request implementation" << std::endl;
         unsupported_msg <<  "https://www.github.com/nvidia/TRTorch/issues" << std::endl;
         LOG_ERROR(unsupported_msg.str());
     }

core/conversion/conversionctx/ConversionCtx.cpp

@@ -37,11 +37,11 @@ ConversionCtx::ConversionCtx(BuilderSettings build_settings)
     switch(settings.op_precision) {
     case nvinfer1::DataType::kHALF:
         cfg->setFlag(nvinfer1::BuilderFlag::kFP16);
-        input_type = nvinfer1::DataType::kHALF; 
+        input_type = nvinfer1::DataType::kHALF;
         break;
     // case nvinfer1::DataType::kINT8:
     //     cfg->setFlag(nvinfer1::BuilderFlag::kINT8);
-    //         input_type = nvinfer1::DataType::kFLOAT;    
+    //         input_type = nvinfer1::DataType::kFLOAT;
     //     break;
     case nvinfer1::DataType::kFLOAT:
     default:
@@ -80,13 +80,30 @@ ConversionCtx::~ConversionCtx() {
         free(ptr);
     }
 }
-
+
+nvinfer1::ITensor* ConversionCtx::AssociateValueAndTensor(const torch::jit::Value* value, nvinfer1::ITensor* tensor) {
+    tensor->setName(value->debugName().c_str());
+    this->value_tensor_map[value] = tensor;
+    return tensor;
+}
+
 std::string ConversionCtx::SerializeEngine() {
     auto engine = builder->buildEngineWithConfig(*net, *cfg);
     auto serialized_engine = engine->serialize();
     return std::string((const char*)serialized_engine->data(), serialized_engine->size());
 }
 
+bool ConversionCtx::CheckLayerAddition(const torch::jit::Node* n) {
+    for (auto out : n->outputs()) {
+        auto iter = this->value_tensor_map.find(out);
+        if (iter == this->value_tensor_map.end()) {
+            LOG_WARNING("Node " << util::node_info(n) << " output: " << out->debugName() << " does not have a coresponding output, may potentially indicate a defective converter");
+            return false;
+        }
+    }
+    return true;
+}
+
 } // namespace conversion
 } // namespace core
 } // namespace trtorch

core/conversion/conversionctx/ConversionCtx.h

@@ -30,12 +30,15 @@ struct BuilderSettings {
 
     BuilderSettings() = default;
     BuilderSettings(const BuilderSettings& other) = default;
-    friend std::ostream& operator<<(std::ostream& os, const BuilderSettings& s); 
+    friend std::ostream& operator<<(std::ostream& os, const BuilderSettings& s);
 };
-    
+
 struct ConversionCtx {
     ConversionCtx(BuilderSettings settings);
     std::string SerializeEngine();
+    nvinfer1::ITensor* AssociateValueAndTensor(const torch::jit::Value* value, nvinfer1::ITensor* tensor);
+    bool CheckLayerAddition(const torch::jit::Node* n);
+
     ~ConversionCtx();
 
     nvinfer1::IBuilder* builder;
@@ -50,12 +53,12 @@ struct ConversionCtx {
     // is constructed from a PyTorch Tensor it allocates the data here to store a
     // copy of the values
     std::vector<void*> builder_resources;
-    
+
     std::unordered_map<const torch::jit::Value*, nvinfer1::ITensor*> value_tensor_map;
     std::unordered_map<const torch::jit::Value*, torch::jit::IValue> evaluated_value_map;
 };
 
-} // namespace conversion    
+} // namespace conversion
 } // namespace core
 } // namespace trtorch
-    
+

core/conversion/converters/Arg.cpp

@@ -85,9 +85,9 @@ std::string Arg::type_name() const {
     default:
         return "None";
     }
-        
+
 }
-    
+
 const torch::jit::IValue* Arg::IValue() const {
     if (type_ == Type::kIValue) {
         return ptr_.ivalue;
@@ -150,7 +150,7 @@ double Arg::unwrapToDouble(double default_val) {
 
 double Arg::unwrapToDouble() {
     return this->unwrapTo<double>();
-} 
+}
 
 bool Arg::unwrapToBool(bool default_val) {
     return this->unwrapTo<bool>(default_val);
@@ -194,26 +194,41 @@ c10::List<bool> Arg::unwrapToBoolList() {
 
 template<typename T>
 T Arg::unwrapTo(T default_val) {
-    if (isIValue()) {
-        // TODO: implement Tag Checking 
-        return ptr_.ivalue->to<T>();
+    try {
+        return this->unwrapTo<T>();
+    } catch(trtorch::Error& e) {
+        LOG_DEBUG("In arg unwrapping, returning default value provided (" << e.what() << ")");
+        return default_val;
     }
-    LOG_DEBUG("In arg unwrapping, returning default value provided");
-    return default_val;
 }
 
-
 template<typename T>
 T Arg::unwrapTo() {
-    if (isIValue()) {
-        //TODO: Implement Tag checking
-        return ptr_.ivalue->to<T>();
-        //TODO: Exception
-        //LOG_INTERNAL_ERROR("Requested unwrapping of arg IValue assuming it was " << typeid(T).name() << " however type is " << ptr_.ivalue->type());
-
+    TRTORCH_CHECK(isIValue(), "Requested unwrapping of arg assuming it was an IValue, however arg type is " << type_name());
+    auto ivalue = ptr_.ivalue;
+    bool correct_type = false;
+    if (typeid(T) == typeid(double)) {
+        correct_type = ivalue->isDouble();
+    } else if (typeid(T) == typeid(bool)) {
+        correct_type = ivalue->isBool();
+    } else if (typeid(T) == typeid(int64_t)) {
+        correct_type = ivalue->isInt();
+    } else if (typeid(T) == typeid(at::Tensor)) {
+        correct_type = ivalue->isTensor();
+    } else if (typeid(T) == typeid(c10::Scalar)) {
+        correct_type = ivalue->isScalar();
+    } else if (typeid(T) == typeid(c10::List<int64_t>)) {
+        correct_type = ivalue->isIntList();
+    } else if (typeid(T) == typeid(c10::List<double>)) {
+        correct_type = ivalue->isDoubleList();
+    } else if (typeid(T) == typeid(c10::List<bool>)) {
+        correct_type = ivalue->isBoolList();
+    } else {
+        TRTORCH_THROW_ERROR("Requested unwrapping of arg to an unsupported type: " << typeid(T).name());
     }
-    TRTORCH_THROW_ERROR("Requested unwrapping of arg assuming it was an IValue, however arg type is " << type_name());
-    return T();
+
+    TRTORCH_CHECK(correct_type, "Requested unwrapping of arg IValue assuming it was " << typeid(T).name() << " however type is " << *(ptr_.ivalue->type()));
+    return ptr_.ivalue->to<T>();
 }
 
 

core/conversion/converters/BUILD

@@ -16,6 +16,7 @@ cc_library(
         "impl/element_wise.cpp",
         "impl/linear.cpp",
         "impl/pooling.cpp",
+        "impl/reduce.cpp",
         "impl/softmax.cpp",
         "impl/unary.cpp",
     ],

core/conversion/converters/NodeConverterRegistry.cpp

@@ -41,20 +41,20 @@ std::string canonical_schema_string(const torch::jit::FunctionSchema& schema) {
 }
 
 namespace {
-using ConverterLUT = std::unordered_map<torch::jit::Symbol, OpConverter>;
+using ConverterLUT = std::unordered_map<c10::OperatorName, OpConverter>;
 
 class NodeConverterRegistry {
 public:
     bool RegisterConverter(torch::jit::FunctionSchema* signature, OpConverter& converter) {
         LOG_DEBUG("Registering Converter for " << canonical_schema_string(*signature));
-        auto sym = torch::jit::Symbol::fromQualString(signature->name());
-        converter_lut_[sym] = std::move(converter);
+        auto name = signature->operator_name();
+        converter_lut_[name] = std::move(converter);
         return true;
     }
 
     OpConverter GetConverter(const torch::jit::FunctionSchema* signature) {
-        auto sym = torch::jit::Symbol::fromQualString(signature->name());
-        auto iter = converter_lut_.find(sym);
+        auto name = signature->operator_name();
+        auto iter = converter_lut_.find(name);
         if (iter == converter_lut_.end()) {
             LOG_ERROR("Requested converter for " << signature->name() << ", but no such converter was found");
             // ASK: Is there a better way than returning a nullptr?
@@ -66,8 +66,8 @@ class NodeConverterRegistry {
     bool Convertable(const torch::jit::Node* n) {
         auto schema = n->maybeSchema();
         if (schema) {
-            auto sym = torch::jit::Symbol::fromQualString(schema->name());
-            auto iter = converter_lut_.find(sym);
+            auto name = schema->operator_name();
+            auto iter = converter_lut_.find(name);
             if (iter == converter_lut_.end()) {
                return false;
             } else {
@@ -79,7 +79,7 @@ class NodeConverterRegistry {
             return false;
         }
     }
-    
+
 private:
     ConverterLUT converter_lut_;
 };
@@ -111,7 +111,7 @@ OpConverter get_node_converter_for(const torch::jit::FunctionSchema* signature)
 bool node_is_convertable(const torch::jit::Node* n) {
     return get_converter_registry().Convertable(n);
 }
-    
+
 RegisterNodeConversionPatterns&& RegisterNodeConversionPatterns::pattern(ConversionPattern p) && {
     register_node_converter(std::move(p));
     return std::move(*this);