/*
 * GStreamer gstreamer-classifiertensordecoder
 * Copyright (C) 2025 Collabora Ltd.
 *  @author: Daniel Morin <daniel.morin@dmohub.org>
 *
 * gstclassifiertensordecoder.c
 *
 * This library is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Library General Public
 * License as published by the Free Software Foundation; either
 * version 2 of the License, or (at your option) any later version.
 *
 * This library is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Library General Public License for more details.
 *
 * You should have received a copy of the GNU Library General Public
 * License along with this library; if not, write to the
 * Free Software Foundation, Inc., 51 Franklin St, Fifth Floor,
 * Boston, MA 02110-1301, USA.
 */

/**
 * SECTION:element-classifiertensordecoder.c
 * @short_description: Decode tensors from classification model using a common
 * tensor output format.
 *
 *
 * This element can parse per-buffer inference tensor meta data generated by
 * an upstream inference element.
 *
 * Tensor format must be:
 *   Dims: [batch-size, class_count]
 *   Datatype: float32
 *
 *   Tensor [M,N]
 *   Batch 0 | Class 0 confidence level | ... | Class N confidence level |
 *   ...
 *   Batch M | Class 0 confidence level | ... | Class N confidence level |
 *
 *   In-memory tensor format:
 *
 *   |Batch 0, Class 0 confidence level |
 *   |Batch 0,           ...            |
 *   |Batch 0, Class N confidence level |
 *   |               ...                |
 *   |Batch M, Class 0 confidence level |
 *   |Batch M,           ...            |
 *   |Batch M, Class N confidence level |
 *
 *
 * ## Example launch command:
 * |[
 * gst-launch-1.0 filesrc location=/onnx-models/images/bus.jpg                 \
 *  ! jpegdec                                                                  \
 *  ! videoconvertscale add-borders=1                                          \
 *  ! onnxinference execution-provider=cpu                                     \
 *    model-file=/onnx-models/models/mobilenet_v1.onnx                         \
 *  ! classifiertensordecoder labels-file=labels.txt ! fakesink               \
 * ]| This pipeline create an tensor-decoder for classification model
 *
 */

#ifdef HAVE_CONFI_H
#include "config.h"
#endif

#include "gstclassifiertensordecoder.h"
#include <gst/gst.h>
#include <math.h>
#include <gst/analytics/analytics.h>

const gchar GST_MODEL_STD_IMAGE_CLASSIFICATION[] = "classification-generic-out";

GST_DEBUG_CATEGORY_STATIC (classifier_tensor_decoder_debug);
#define GST_CAT_DEFAULT classifier_tensor_decoder_debug
#define gst_classifier_tensor_decoder_parent_class parent_class

GST_ELEMENT_REGISTER_DEFINE (classifier_tensor_decoder,
    "classifiertensordecoder", GST_RANK_PRIMARY,
    GST_TYPE_CLASSIFIER_TENSOR_DECODER);


/* GstClassifierTensorDecoder properties */
enum
{
  PROP_0,
  PROP_THRESHOLD,
  PROP_LABEL_FILE
};

static const float DEFAULT_THRESHOLD = 0.7f;

static GstStaticPadTemplate gst_classifier_tensor_decoder_src_template =
GST_STATIC_PAD_TEMPLATE ("src",
    GST_PAD_SRC,
    GST_PAD_ALWAYS,
    GST_STATIC_CAPS_ANY);

static GstStaticPadTemplate gst_classifier_tensor_decoder_sink_template =
GST_STATIC_PAD_TEMPLATE ("sink",
    GST_PAD_SINK,
    GST_PAD_ALWAYS,
    GST_STATIC_CAPS_ANY);

static void gst_classifier_tensor_decoder_set_property (GObject * object,
    guint prop_id, const GValue * value, GParamSpec * pspec);
static void gst_classifier_tensor_decoder_get_property (GObject * object,
    guint prop_id, GValue * value, GParamSpec * pspec);

static void gst_classifier_tensor_decoder_finalize (GObject * object);

static GstFlowReturn
gst_classifier_tensor_decoder_transform_ip (GstBaseTransform * trans,
    GstBuffer * buf);

static GstStateChangeReturn
gst_classifier_tensor_decoder_change_state (GstElement * element,
    GstStateChange transition);

#define softmax(len, values, results, max_val)                                \
  gsize i;                                                                    \
  gfloat sum = 0.0;                                                           \
  gfloat value;                                                               \
  g_return_if_fail (values != NULL);                                          \
  g_return_if_fail (result != NULL);                                          \
                                                                              \
  /* Calculate exponential of every value */                                  \
  for (i = 0; i < len; i++) {                                                 \
    value = values[i] / max_val;                                              \
    result[i] = exp (value);                                                  \
    sum += result[i];                                                         \
  }                                                                           \
                                                                              \
  /* Complete softmax */                                                      \
  for (i = 0; i < len; i++) {                                                 \
    result[i] = result[i] / sum;                                              \
  }

static void
softmax_u8 (gsize len, const guint8 * values, gfloat * result)
{
  softmax (len, values, results, 255.0);
}

static void
softmax_f32 (gsize len, const gfloat * values, gfloat * result)
{
  softmax (len, values, results, 1.0);
}

G_DEFINE_TYPE (GstClassifierTensorDecoder, gst_classifier_tensor_decoder,
    GST_TYPE_BASE_TRANSFORM);

static void
gst_classifier_tensor_decoder_class_init (GstClassifierTensorDecoderClass *
    klass)
{
  GObjectClass *gobject_class = (GObjectClass *) klass;
  GstElementClass *element_class = (GstElementClass *) klass;
  GstBaseTransformClass *basetransform_class = (GstBaseTransformClass *) klass;

  GST_DEBUG_CATEGORY_INIT (classifier_tensor_decoder_debug,
      "classifiertensordecoder", 0,
      "Tensor decoder for classification model with common output format");

  gobject_class->set_property = gst_classifier_tensor_decoder_set_property;
  gobject_class->get_property = gst_classifier_tensor_decoder_get_property;
  gobject_class->finalize = gst_classifier_tensor_decoder_finalize;

  g_object_class_install_property (G_OBJECT_CLASS (klass),
      PROP_THRESHOLD,
      g_param_spec_float ("class-confidence-threshold",
          "Class confidence threshold",
          "Classes with a confidence level inferior to this threshold "
          "will be excluded",
          0.0, 1.0, DEFAULT_THRESHOLD,
          (GParamFlags) (G_PARAM_READWRITE | G_PARAM_STATIC_STRINGS)));

  g_object_class_install_property (G_OBJECT_CLASS (klass),
      PROP_LABEL_FILE,
      g_param_spec_string ("labels-file",
          "Class labels file",
          "Path to a file containing class label. COCO format",
          NULL, (G_PARAM_READWRITE | G_PARAM_STATIC_STRINGS)));

  element_class->change_state = gst_classifier_tensor_decoder_change_state;

  gst_element_class_set_static_metadata (element_class,
      "classifiertensordecoder", "Tensordecoder",
      "Decode tensors output from classification model using common format.\n"
      "\tTensor format must be: \n" "\t\tDims: [batch-size, class_count]\n"
      "\t\tDatatype: float32 \n" "\n" "\t\tTensor [M,N]\n"
      "\t\t\tBatch 0   | Class 0 confidence level | ... | Class N-1 confidence level |\n"
      "\t\t\t...\n"
      "\t\t\tBatch M-1 | Class 0 confidence level | ... | Class N-1 confidence level |\n"
      "\t\t\n" "\tIn-memory tensor format:\n" "\n"
      "\t\t|Batch 0, Class 0 confidence level     |\n"
      "\t\t|Batch 0,           ...                |\n"
      "\t\t|Batch 0, Class N-1 confidence level   |\n"
      "\t\t|               ...                    |\n"
      "\t\t|Batch M-1, Class 0 confidence level   |\n"
      "\t\t|Batch M-1,           ...              |\n"
      "\t\t|Batch M-1, Class N-1 confidence level |\n" "\n" " model",
      "Daniel Morin <daniel.morin@collabora.com>");

  gst_element_class_add_pad_template (element_class,
      gst_static_pad_template_get
      (&gst_classifier_tensor_decoder_src_template));
  gst_element_class_add_pad_template (element_class,
      gst_static_pad_template_get
      (&gst_classifier_tensor_decoder_sink_template));

  basetransform_class->transform_ip =
      GST_DEBUG_FUNCPTR (gst_classifier_tensor_decoder_transform_ip);
}

static void
gst_classifier_tensor_decoder_init (GstClassifierTensorDecoder * self)
{
  self->threshold = DEFAULT_THRESHOLD;
  self->labels_file = NULL;
  self->softmax_res = NULL;

  gst_base_transform_set_passthrough (GST_BASE_TRANSFORM (self), FALSE);
}

static void
gst_classifier_tensor_decoder_finalize (GObject * object)
{
  GstClassifierTensorDecoder *self = GST_CLASSIFIER_TENSOR_DECODER (object);

  g_free (self->labels_file);
  G_OBJECT_CLASS (gst_classifier_tensor_decoder_parent_class)->finalize
      (object);
}

static void
gst_classifier_tensor_decoder_set_property (GObject * object, guint prop_id,
    const GValue * value, GParamSpec * pspec)
{
  GstClassifierTensorDecoder *self = GST_CLASSIFIER_TENSOR_DECODER (object);
  static GFileTest filetest = (G_FILE_TEST_EXISTS | G_FILE_TEST_IS_REGULAR);

  switch (prop_id) {
    case PROP_THRESHOLD:
      self->threshold = g_value_get_float (value);
      break;
    case PROP_LABEL_FILE:
      self->labels_file = g_strdup (g_value_get_string (value));

      if (self->labels_file) {
        if (!g_file_test (self->labels_file, filetest)) {
          GST_ERROR_OBJECT (self, "Unable to load %s", self->labels_file);
          g_free (g_steal_pointer (&self->labels_file));
        }
      } else {
        GST_ERROR_OBJECT (self, "Invalid file");
      }
      break;
    default:
      G_OBJECT_WARN_INVALID_PROPERTY_ID (object, prop_id, pspec);
      break;
  }
}

static void
gst_classifier_tensor_decoder_get_property (GObject * object, guint prop_id,
    GValue * value, GParamSpec * pspec)
{
  GstClassifierTensorDecoder *self = GST_CLASSIFIER_TENSOR_DECODER (object);

  switch (prop_id) {
    case PROP_THRESHOLD:
      g_value_set_float (value, self->threshold);
      break;
    case PROP_LABEL_FILE:
      g_value_set_string (value, self->labels_file);
      break;
    default:
      G_OBJECT_WARN_INVALID_PROPERTY_ID (object, prop_id, pspec);
      break;
  }
}

static guint
gst_classifier_tensor_decoder_load_labels (GstClassifierTensorDecoder * self)
{
  gchar *content = NULL;
  gchar **tokens = NULL;
  gsize len;
  GError *err = NULL;
  GQuark val;

  if (self->labels_file == NULL) {
    GST_ERROR_OBJECT (self, "Missing label file");
    return 0;
  }
  if (!g_file_get_contents (self->labels_file, &content, &len, &err)) {
    GST_ERROR_OBJECT (self, "Could not load labels file %s: %s",
        self->labels_file, err->message);
    g_error_free (err);
    return 0;
  }

  if (len == 0) {
    GST_ERROR_OBJECT (self, "Labels file %s is empty", self->labels_file);
    g_free (content);
    return 0;
  }

  tokens = g_strsplit (content, "\n", 0);
  g_free (content);

  if (tokens[0] == NULL) {
    GST_ERROR_OBJECT (self, "Labels file %s has no labels", self->labels_file);
    g_free (content);
    return 0;
  }

  self->class_quark = g_array_new (FALSE, FALSE, sizeof (GQuark));

  for (int i = 0; tokens[i] != NULL && tokens[i][0] != '\0'; i++) {
    val = g_quark_from_string (tokens[i]);
    g_array_append_val (self->class_quark, val);
  }

  self->softmax_res = g_array_sized_new (FALSE, TRUE, sizeof (gfloat),
      self->class_quark->len);

  g_strfreev (tokens);
  return self->class_quark->len;
}

static GstStateChangeReturn
gst_classifier_tensor_decoder_change_state (GstElement * element,
    GstStateChange transition)
{
  GstClassifierTensorDecoder *self = GST_CLASSIFIER_TENSOR_DECODER (element);
  GstStateChangeReturn ret;

  switch (transition) {
    case GST_STATE_CHANGE_NULL_TO_READY:
      if (!gst_classifier_tensor_decoder_load_labels (self)) {
        return GST_STATE_CHANGE_FAILURE;
      }
      break;
    default:
      break;
  }

  ret = GST_ELEMENT_CLASS (parent_class)->change_state (element, transition);

  switch (transition) {
    case GST_STATE_CHANGE_READY_TO_NULL:
      g_array_free (self->class_quark, FALSE);
      g_array_free (self->softmax_res, TRUE);
      break;
    default:
      break;
  }

  return ret;
}

static GstTensorMeta *
gst_classifier_tensor_decoder_get_tensor_meta (GstClassifierTensorDecoder *
    self, GstBuffer * buf)
{
  GstMeta *meta = NULL;
  gpointer iter_state = NULL;

  if (!gst_buffer_get_meta (buf, GST_TENSOR_META_API_TYPE)) {
    GST_DEBUG_OBJECT (self,
        "missing tensor meta from buffer %" GST_PTR_FORMAT, buf);
    return NULL;
  }

  while ((meta = gst_buffer_iterate_meta_filtered (buf, &iter_state,
              GST_TENSOR_META_API_TYPE))) {
    GstTensorMeta *tensor_meta = (GstTensorMeta *) meta;

    if (tensor_meta->num_tensors != 1)
      continue;

    gint index = gst_tensor_meta_get_index_from_id (tensor_meta,
        g_quark_from_static_string (GST_MODEL_STD_IMAGE_CLASSIFICATION));

    if (index == -1)
      continue;

    return tensor_meta;
  }

  return NULL;
}

static GstFlowReturn
gst_classifier_tensor_decoder_decode (GstClassifierTensorDecoder * self,
    GstBuffer * buf, GstAnalyticsRelationMeta * rmeta, GstTensorMeta * tmeta)
{
  GstMapInfo map_info = GST_MAP_INFO_INIT;
  gfloat max = 0.0;
  gfloat *softmax_res = (gfloat *) self->softmax_res->data;
  gsize len;
  GQuark q, qmax;
  gint max_idx = -1;
  const GstTensor *tensor;
  GstAnalyticsClsMtd cls_mtd;
  const gsize DIMS[] = { 1, G_MAXSIZE };

  tensor = gst_tensor_meta_get_typed_tensor (tmeta,
      g_quark_from_static_string (GST_MODEL_STD_IMAGE_CLASSIFICATION),
      GST_TENSOR_DATA_TYPE_FLOAT32, GST_TENSOR_DIM_ORDER_ROW_MAJOR, 1, NULL);
  if (tensor == NULL)
    tensor = gst_tensor_meta_get_typed_tensor (tmeta,
        g_quark_from_static_string (GST_MODEL_STD_IMAGE_CLASSIFICATION),
        GST_TENSOR_DATA_TYPE_FLOAT32, GST_TENSOR_DIM_ORDER_ROW_MAJOR, 2, DIMS);
  if (tensor == NULL)
    tensor = gst_tensor_meta_get_typed_tensor (tmeta,
        g_quark_from_static_string (GST_MODEL_STD_IMAGE_CLASSIFICATION),
        GST_TENSOR_DATA_TYPE_UINT8, GST_TENSOR_DIM_ORDER_ROW_MAJOR, 1, NULL);
  if (tensor == NULL)
    tensor = gst_tensor_meta_get_typed_tensor (tmeta,
        g_quark_from_static_string (GST_MODEL_STD_IMAGE_CLASSIFICATION),
        GST_TENSOR_DATA_TYPE_UINT8, GST_TENSOR_DIM_ORDER_ROW_MAJOR, 2, DIMS);

  if (tensor == NULL) {
    GST_ELEMENT_ERROR (GST_BASE_TRANSFORM (self), STREAM, FAILED,
        (NULL), ("Could not find classification tensor"));
    return GST_FLOW_ERROR;
  }

  len = tensor->dims[tensor->num_dims - 1];

  if (len != self->class_quark->len) {
    GST_WARNING_OBJECT (self, "Labels file has size %zu, but the tensor has"
        " %u entries, it is probably not the right labels file",
        len, self->class_quark->len);
    len = MIN (len, self->class_quark->len);
  }

  if (!gst_buffer_map (tensor->data, &map_info, GST_MAP_READ)) {
    GST_ELEMENT_ERROR (self, STREAM, FAILED, (NULL),
        ("Failed to map tensor data"));
    return GST_FLOW_ERROR;
  }

  GST_TRACE_OBJECT (self, "Tensor shape dims %zu", tensor->num_dims);

  if (gst_debug_category_get_threshold (GST_CAT_DEFAULT) >= GST_LEVEL_TRACE) {
    for (gint i = 0; i < tensor->num_dims; i++) {
      GST_TRACE_OBJECT (self, "Tensor dim %d: %zu", i, tensor->dims[i]);
    }
  }

  switch (tensor->data_type) {
    case GST_TENSOR_DATA_TYPE_FLOAT32:
      softmax_f32 (len, (gfloat *) map_info.data, softmax_res);
      break;
    case GST_TENSOR_DATA_TYPE_UINT8:
      softmax_u8 (len, (guint8 *) map_info.data, softmax_res);
      break;
    default:
      g_return_val_if_reached (GST_FLOW_ERROR);
      break;
  }
  gst_buffer_unmap (tensor->data, &map_info);

  for (gint j = 0; j < len; j++) {
    q = g_array_index (self->class_quark, GQuark, j);
    if (softmax_res[j] > max) {
      max = softmax_res[j];
      max_idx = j;
      qmax = q;
    }
  }

  if (max_idx != -1) {
    gst_analytics_relation_meta_add_one_cls_mtd (rmeta, max, qmax, &cls_mtd);

    GST_LOG_OBJECT (self, "Max class is %d:%s with %f", max_idx,
        g_quark_to_string (qmax), max);
  }

  return GST_FLOW_OK;
}

static GstFlowReturn
gst_classifier_tensor_decoder_transform_ip (GstBaseTransform * trans,
    GstBuffer * buf)
{
  GstClassifierTensorDecoder *self = GST_CLASSIFIER_TENSOR_DECODER (trans);
  GstTensorMeta *tmeta;
  GstAnalyticsRelationMeta *rmeta;

  tmeta = gst_classifier_tensor_decoder_get_tensor_meta (self, buf);
  if (tmeta != NULL) {
    rmeta = gst_buffer_add_analytics_relation_meta (buf);
    g_assert (rmeta != NULL);
  } else {
    GST_WARNING_OBJECT (trans, "missing tensor meta");
    return TRUE;
  }

  return gst_classifier_tensor_decoder_decode (self, buf, rmeta, tmeta);
}