src/modules/optim/clip.cc¶

// Copyright 2026 The Tenzr Project Authors
//
// SPDX-License-Identifier: Apache-2.0
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//      https://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

#include "tenzr/modules/optim/clip.h"

#include <algorithm>
#include <cmath>
#include <vector>

#include "tenzr/common/error.h"

namespace tenzr::optim {

  std::unordered_map<std::string, Tensor> ClipByValue(const std::unordered_map<std::string, Tensor>& grads, double min_val, double max_val) {
    TENZR_CHECK(min_val <= max_val, "min_val must be <= max_val");
    std::unordered_map<std::string, Tensor> clipped;

    for (const auto& [name, g] : grads) {
      auto raw = g.ToVectorAs<double>();
      for (double& v : raw) {
        v = std::clamp(v, min_val, max_val);
      }
      clipped[name] = Tensor::FromVector<double>(raw, g.shape()).Astype(g.Dtype());
    }
    return clipped;
  }

  std::unordered_map<std::string, Tensor> ClipByGlobalNorm(const std::unordered_map<std::string, Tensor>& grads, double max_norm, double* calculated_norm) {
    TENZR_CHECK(max_norm > 0.0, "max_norm must be positive");
    double total_sq_norm = 0.0;

    for (const auto& [name, g] : grads) {
      auto raw = g.ToVectorAs<double>();
      for (double v : raw) {
        total_sq_norm += v * v;
      }
    }

    double global_norm = std::sqrt(total_sq_norm);
    if (calculated_norm != nullptr) {
      *calculated_norm = global_norm;
    }

    std::unordered_map<std::string, Tensor> clipped;
    if (global_norm <= max_norm || global_norm == 0.0) {
      return grads;
    }

    double scale = max_norm / (global_norm + 1.0e-6);
    for (const auto& [name, g] : grads) {
      auto raw = g.ToVectorAs<double>();
      for (double& v : raw) {
        v *= scale;
      }
      clipped[name] = Tensor::FromVector<double>(raw, g.shape()).Astype(g.Dtype());
    }

    return clipped;
  }

}  // namespace tenzr::optim