torchcodec 0.7.0__cp313-cp313-win_amd64.whl → 0.8.0__cp313-cp313-win_amd64.whl

torchcodec/_core/SingleStreamDecoder.h

@@ -17,6 +17,7 @@
 #include "src/torchcodec/_core/FFMPEGCommon.h"
 #include "src/torchcodec/_core/Frame.h"
 #include "src/torchcodec/_core/StreamOptions.h"
+#include "src/torchcodec/_core/Transform.h"
 
 namespace facebook::torchcodec {
 
@@ -83,6 +84,7 @@ class SingleStreamDecoder {
 
   void addVideoStream(
       int streamIndex,
+      std::vector<Transform*>& transforms,
       const VideoStreamOptions& videoStreamOptions = VideoStreamOptions(),
       std::optional<FrameMappings> customFrameMappings = std::nullopt);
   void addAudioStream(
@@ -106,7 +108,7 @@ class SingleStreamDecoder {
 
   // Returns frames at the given indices for a given stream as a single stacked
   // Tensor.
-  FrameBatchOutput getFramesAtIndices(const std::vector<int64_t>& frameIndices);
+  FrameBatchOutput getFramesAtIndices(const torch::Tensor& frameIndices);
 
   // Returns frames within a given range. The range is defined by [start, stop).
   // The values retrieved from the range are: [start, start+step,
@@ -121,7 +123,7 @@ class SingleStreamDecoder {
   // seconds=5.999, etc.
   FrameOutput getFramePlayedAt(double seconds);
 
-  FrameBatchOutput getFramesPlayedAt(const std::vector<double>& timestamps);
+  FrameBatchOutput getFramesPlayedAt(const torch::Tensor& timestamps);
 
   // Returns frames within a given pts range. The range is defined by
   // [startSeconds, stopSeconds) with respect to the pts values for frames. The
@@ -226,17 +228,8 @@ class SingleStreamDecoder {
     std::vector<FrameInfo> keyFrames;
     std::vector<FrameInfo> allFrames;
 
-    // TODO since the decoder is single-stream, these should be decoder fields,
-    // not streamInfo fields. And they should be defined right next to
-    // `cursor_`, with joint documentation.
-    int64_t lastDecodedAvFramePts = 0;
-    int64_t lastDecodedAvFrameDuration = 0;
     VideoStreamOptions videoStreamOptions;
     AudioStreamOptions audioStreamOptions;
-
-    // color-conversion fields. Only one of FilterGraphContext and
-    // UniqueSwsContext should be non-null.
-    UniqueSwrContext swrContext;
   };
 
   // --------------------------------------------------------------------------
@@ -318,6 +311,7 @@ class SingleStreamDecoder {
       int streamIndex,
       AVMediaType mediaType,
       const torch::Device& device = torch::kCPU,
+      const std::string_view deviceVariant = "default",
       std::optional<int> ffmpegThreadCount = std::nullopt);
 
   // Returns the "best" stream index for a given media type. The "best" is
@@ -356,16 +350,49 @@ class SingleStreamDecoder {
   const int NO_ACTIVE_STREAM = -2;
   int activeStreamIndex_ = NO_ACTIVE_STREAM;
 
-  bool cursorWasJustSet_ = false;
   // The desired position of the cursor in the stream. We send frames >= this
   // pts to the user when they request a frame.
   int64_t cursor_ = INT64_MIN;
+  bool cursorWasJustSet_ = false;
+  int64_t lastDecodedAvFramePts_ = 0;
+  int64_t lastDecodedAvFrameDuration_ = 0;
+
+  // Audio only. We cache it for performance. The video equivalents live in
+  // deviceInterface_. We store swrContext_ here because we only handle audio
+  // on the CPU.
+  UniqueSwrContext swrContext_;
+
   // Stores various internal decoding stats.
   DecodeStats decodeStats_;
+
   // Stores the AVIOContext for the input buffer.
   std::unique_ptr<AVIOContextHolder> avioContextHolder_;
+
+  // We will receive a vector of transforms upon adding a stream and store it
+  // here. However, we need to know if any of those operations change the
+  // dimensions of the output frame. If they do, we need to figure out what are
+  // the final dimensions of the output frame after ALL transformations. We
+  // figure this out as soon as we receive the transforms. If any of the
+  // transforms change the final output frame dimensions, we store that in
+  // resizedOutputDims_. If resizedOutputDims_ has no value, that means there
+  // are no transforms that change the output frame dimensions.
+  //
+  // The priority order for output frame dimension is:
+  //
+  // 1. resizedOutputDims_; the resize requested by the user always takes
+  //    priority.
+  // 2. The dimemnsions of the actual decoded AVFrame. This can change
+  //    per-decoded frame, and is unknown in SingleStreamDecoder. Only the
+  //    DeviceInterface learns it immediately after decoding a raw frame but
+  //    before the color transformation.
+  // 3. metdataDims_; the dimensions we learned from the metadata.
+  std::vector<std::unique_ptr<Transform>> transforms_;
+  std::optional<FrameDims> resizedOutputDims_;
+  FrameDims metadataDims_;
+
   // Whether or not we have already scanned all streams to update the metadata.
   bool scannedAllStreams_ = false;
+
   // Tracks that we've already been initialized.
   bool initialized_ = false;
 };
@@ -375,6 +402,4 @@ std::ostream& operator<<(
     std::ostream& os,
     const SingleStreamDecoder::DecodeStats& stats);
 
-SingleStreamDecoder::SeekMode seekModeFromString(std::string_view seekMode);
-
 } // namespace facebook::torchcodec

torchcodec/_core/StreamOptions.h

@@ -9,11 +9,11 @@
 #include <torch/types.h>
 #include <optional>
 #include <string>
+#include <string_view>
 
 namespace facebook::torchcodec {
 
 enum ColorConversionLibrary {
-  // TODO: Add an AUTO option later.
   // Use the libavfilter library for color conversion.
   FILTERGRAPH,
   // Use the libswscale library for color conversion.
@@ -28,16 +28,26 @@ struct VideoStreamOptions {
   // utilize all cores. If not set, it will be the default FFMPEG behavior for
   // the given codec.
   std::optional<int> ffmpegThreadCount;
+
   // Currently the dimension order can be either NHWC or NCHW.
   // H=height, W=width, C=channel.
   std::string dimensionOrder = "NCHW";
-  // The output height and width of the frame. If not specified, the output
-  // is the same as the original video.
-  std::optional<int> width;
-  std::optional<int> height;
-  std::optional<ColorConversionLibrary> colorConversionLibrary;
+
+  // By default we have to use filtergraph, as it is more general. We can only
+  // use swscale when we have met strict requirements. See
+  // CpuDeviceInterface::initialze() for the logic.
+  ColorConversionLibrary colorConversionLibrary =
+      ColorConversionLibrary::FILTERGRAPH;
+
   // By default we use CPU for decoding for both C++ and python users.
   torch::Device device = torch::kCPU;
+  // Device variant (e.g., "default", "beta", etc.)
+  std::string_view deviceVariant = "default";
+
+  // Encoding options
+  // TODO-VideoEncoder: Consider adding other optional fields here
+  // (bit rate, gop size, max b frames, preset)
+  std::optional<int> crf;
 };
 
 struct AudioStreamOptions {

torchcodec/_core/Transform.cpp

@@ -0,0 +1,60 @@
+// Copyright (c) Meta Platforms, Inc. and affiliates.
+// All rights reserved.
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include "src/torchcodec/_core/Transform.h"
+#include <torch/types.h>
+#include "src/torchcodec/_core/FFMPEGCommon.h"
+
+namespace facebook::torchcodec {
+
+namespace {
+
+std::string toFilterGraphInterpolation(
+    ResizeTransform::InterpolationMode mode) {
+  switch (mode) {
+    case ResizeTransform::InterpolationMode::BILINEAR:
+      return "bilinear";
+    default:
+      TORCH_CHECK(
+          false,
+          "Unknown interpolation mode: " +
+              std::to_string(static_cast<int>(mode)));
+  }
+}
+
+int toSwsInterpolation(ResizeTransform::InterpolationMode mode) {
+  switch (mode) {
+    case ResizeTransform::InterpolationMode::BILINEAR:
+      return SWS_BILINEAR;
+    default:
+      TORCH_CHECK(
+          false,
+          "Unknown interpolation mode: " +
+              std::to_string(static_cast<int>(mode)));
+  }
+}
+
+} // namespace
+
+std::string ResizeTransform::getFilterGraphCpu() const {
+  return "scale=" + std::to_string(outputDims_.width) + ":" +
+      std::to_string(outputDims_.height) +
+      ":sws_flags=" + toFilterGraphInterpolation(interpolationMode_);
+}
+
+std::optional<FrameDims> ResizeTransform::getOutputFrameDims() const {
+  return outputDims_;
+}
+
+bool ResizeTransform::isResize() const {
+  return true;
+}
+
+int ResizeTransform::getSwsFlags() const {
+  return toSwsInterpolation(interpolationMode_);
+}
+
+} // namespace facebook::torchcodec

torchcodec/_core/Transform.h

@@ -0,0 +1,59 @@
+// Copyright (c) Meta Platforms, Inc. and affiliates.
+// All rights reserved.
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#pragma once
+
+#include <optional>
+#include <string>
+#include "src/torchcodec/_core/Frame.h"
+
+namespace facebook::torchcodec {
+
+class Transform {
+ public:
+  virtual std::string getFilterGraphCpu() const = 0;
+  virtual ~Transform() = default;
+
+  // If the transformation does not change the output frame dimensions, then
+  // there is no need to override this member function. The default
+  // implementation returns an empty optional, indicating that the output frame
+  // has the same dimensions as the input frame.
+  //
+  // If the transformation does change the output frame dimensions, then it
+  // must override this member function and return the output frame dimensions.
+  virtual std::optional<FrameDims> getOutputFrameDims() const {
+    return std::nullopt;
+  }
+
+  // The ResizeTransform is special, because it is the only transform that
+  // swscale can handle.
+  virtual bool isResize() const {
+    return false;
+  }
+};
+
+class ResizeTransform : public Transform {
+ public:
+  enum class InterpolationMode { BILINEAR };
+
+  ResizeTransform(const FrameDims& dims)
+      : outputDims_(dims), interpolationMode_(InterpolationMode::BILINEAR) {}
+
+  ResizeTransform(const FrameDims& dims, InterpolationMode interpolationMode)
+      : outputDims_(dims), interpolationMode_(interpolationMode) {}
+
+  std::string getFilterGraphCpu() const override;
+  std::optional<FrameDims> getOutputFrameDims() const override;
+  bool isResize() const override;
+
+  int getSwsFlags() const;
+
+ private:
+  FrameDims outputDims_;
+  InterpolationMode interpolationMode_;
+};
+
+} // namespace facebook::torchcodec

torchcodec/_core/__init__.py

@@ -25,6 +25,7 @@ from .ops import (
     encode_audio_to_file,
     encode_audio_to_file_like,
     encode_audio_to_tensor,
+    encode_video_to_file,
     get_ffmpeg_library_versions,
     get_frame_at_index,
     get_frame_at_pts,

torchcodec/_core/custom_ops.cpp

@@ -10,6 +10,7 @@
 #include <string>
 #include "c10/core/SymIntArrayRef.h"
 #include "c10/util/Exception.h"
+#include "src/torchcodec/_core/AVIOFileLikeContext.h"
 #include "src/torchcodec/_core/AVIOTensorContext.h"
 #include "src/torchcodec/_core/Encoder.h"
 #include "src/torchcodec/_core/SingleStreamDecoder.h"
@@ -31,15 +32,20 @@ TORCH_LIBRARY(torchcodec_ns, m) {
   m.def("create_from_file(str filename, str? seek_mode=None) -> Tensor");
   m.def(
       "encode_audio_to_file(Tensor samples, int sample_rate, str filename, int? bit_rate=None, int? num_channels=None, int? desired_sample_rate=None) -> ()");
+  m.def(
+      "encode_video_to_file(Tensor frames, int frame_rate, str filename, int? crf=None) -> ()");
   m.def(
       "encode_audio_to_tensor(Tensor samples, int sample_rate, str format, int? bit_rate=None, int? num_channels=None, int? desired_sample_rate=None) -> Tensor");
+  m.def(
+      "_encode_audio_to_file_like(Tensor samples, int sample_rate, str format, int file_like_context, int? bit_rate=None, int? num_channels=None, int? desired_sample_rate=None) -> ()");
   m.def(
       "create_from_tensor(Tensor video_tensor, str? seek_mode=None) -> Tensor");
-  m.def("_convert_to_tensor(int decoder_ptr) -> Tensor");
   m.def(
-      "_add_video_stream(Tensor(a!) decoder, *, int? width=None, int? height=None, int? num_threads=None, str? dimension_order=None, int? stream_index=None, str? device=None, (Tensor, Tensor, Tensor)? custom_frame_mappings=None, str? color_conversion_library=None) -> ()");
+      "_create_from_file_like(int file_like_context, str? seek_mode=None) -> Tensor");
   m.def(
-      "add_video_stream(Tensor(a!) decoder, *, int? width=None, int? height=None, int? num_threads=None, str? dimension_order=None, int? stream_index=None, str? device=None, (Tensor, Tensor, Tensor)? custom_frame_mappings=None) -> ()");
+      "_add_video_stream(Tensor(a!) decoder, *, int? num_threads=None, str? dimension_order=None, int? stream_index=None, str device=\"cpu\", str device_variant=\"default\", str transform_specs=\"\", (Tensor, Tensor, Tensor)? custom_frame_mappings=None, str? color_conversion_library=None) -> ()");
+  m.def(
+      "add_video_stream(Tensor(a!) decoder, *, int? num_threads=None, str? dimension_order=None, int? stream_index=None, str device=\"cpu\", str device_variant=\"default\", str transform_specs=\"\", (Tensor, Tensor, Tensor)? custom_frame_mappings=None) -> ()");
   m.def(
       "add_audio_stream(Tensor(a!) decoder, *, int? stream_index=None, int? sample_rate=None, int? num_channels=None) -> ()");
   m.def("seek_to_pts(Tensor(a!) decoder, float seconds) -> ()");
@@ -49,7 +55,7 @@ TORCH_LIBRARY(torchcodec_ns, m) {
   m.def(
       "get_frame_at_index(Tensor(a!) decoder, *, int frame_index) -> (Tensor, Tensor, Tensor)");
   m.def(
-      "get_frames_at_indices(Tensor(a!) decoder, *, int[] frame_indices) -> (Tensor, Tensor, Tensor)");
+      "get_frames_at_indices(Tensor(a!) decoder, *, Tensor frame_indices) -> (Tensor, Tensor, Tensor)");
   m.def(
       "get_frames_in_range(Tensor(a!) decoder, *, int start, int stop, int? step=None) -> (Tensor, Tensor, Tensor)");
   m.def(
@@ -57,7 +63,7 @@ TORCH_LIBRARY(torchcodec_ns, m) {
   m.def(
       "get_frames_by_pts_in_range_audio(Tensor(a!) decoder, *, float start_seconds, float? stop_seconds) -> (Tensor, Tensor)");
   m.def(
-      "get_frames_by_pts(Tensor(a!) decoder, *, float[] timestamps) -> (Tensor, Tensor, Tensor)");
+      "get_frames_by_pts(Tensor(a!) decoder, *, Tensor timestamps) -> (Tensor, Tensor, Tensor)");
   m.def("_get_key_frame_indices(Tensor(a!) decoder) -> Tensor");
   m.def("get_json_metadata(Tensor(a!) decoder) -> str");
   m.def("get_container_json_metadata(Tensor(a!) decoder) -> str");
@@ -165,6 +171,81 @@ std::string mapToJson(const std::map<std::string, std::string>& metadataMap) {
   return ss.str();
 }
 
+SingleStreamDecoder::SeekMode seekModeFromString(std::string_view seekMode) {
+  if (seekMode == "exact") {
+    return SingleStreamDecoder::SeekMode::exact;
+  } else if (seekMode == "approximate") {
+    return SingleStreamDecoder::SeekMode::approximate;
+  } else if (seekMode == "custom_frame_mappings") {
+    return SingleStreamDecoder::SeekMode::custom_frame_mappings;
+  } else {
+    TORCH_CHECK(false, "Invalid seek mode: " + std::string(seekMode));
+  }
+}
+
+int checkedToPositiveInt(const std::string& str) {
+  int ret = 0;
+  try {
+    ret = std::stoi(str);
+  } catch (const std::invalid_argument&) {
+    TORCH_CHECK(false, "String cannot be converted to an int:" + str);
+  } catch (const std::out_of_range&) {
+    TORCH_CHECK(false, "String would become integer out of range:" + str);
+  }
+  TORCH_CHECK(ret > 0, "String must be a positive integer:" + str);
+  return ret;
+}
+
+// Resize transform specs take the form:
+//
+//   "resize, <height>, <width>"
+//
+// Where "resize" is the string literal and <height> and <width> are positive
+// integers.
+Transform* makeResizeTransform(
+    const std::vector<std::string>& resizeTransformSpec) {
+  TORCH_CHECK(
+      resizeTransformSpec.size() == 3,
+      "resizeTransformSpec must have 3 elements including its name");
+  int height = checkedToPositiveInt(resizeTransformSpec[1]);
+  int width = checkedToPositiveInt(resizeTransformSpec[2]);
+  return new ResizeTransform(FrameDims(height, width));
+}
+
+std::vector<std::string> split(const std::string& str, char delimiter) {
+  std::vector<std::string> tokens;
+  std::string token;
+  std::istringstream tokenStream(str);
+  while (std::getline(tokenStream, token, delimiter)) {
+    tokens.push_back(token);
+  }
+  return tokens;
+}
+
+// The transformSpecsRaw string is always in the format:
+//
+//   "name1, param1, param2, ...; name2, param1, param2, ...; ..."
+//
+// Where "nameX" is the name of the transform, and "paramX" are the parameters.
+std::vector<Transform*> makeTransforms(const std::string& transformSpecsRaw) {
+  std::vector<Transform*> transforms;
+  std::vector<std::string> transformSpecs = split(transformSpecsRaw, ';');
+  for (const std::string& transformSpecRaw : transformSpecs) {
+    std::vector<std::string> transformSpec = split(transformSpecRaw, ',');
+    TORCH_CHECK(
+        transformSpec.size() >= 1,
+        "Invalid transform spec: " + transformSpecRaw);
+
+    auto name = transformSpec[0];
+    if (name == "resize") {
+      transforms.push_back(makeResizeTransform(transformSpec));
+    } else {
+      TORCH_CHECK(false, "Invalid transform name: " + name);
+    }
+  }
+  return transforms;
+}
+
 } // namespace
 
 // ==============================
@@ -203,33 +284,47 @@ at::Tensor create_from_tensor(
     realSeek = seekModeFromString(seek_mode.value());
   }
 
-  auto contextHolder = std::make_unique<AVIOFromTensorContext>(video_tensor);
+  auto avioContextHolder =
+      std::make_unique<AVIOFromTensorContext>(video_tensor);
 
   std::unique_ptr<SingleStreamDecoder> uniqueDecoder =
-      std::make_unique<SingleStreamDecoder>(std::move(contextHolder), realSeek);
+      std::make_unique<SingleStreamDecoder>(
+          std::move(avioContextHolder), realSeek);
   return wrapDecoderPointerToTensor(std::move(uniqueDecoder));
 }
 
-at::Tensor _convert_to_tensor(int64_t decoder_ptr) {
-  auto decoder = reinterpret_cast<SingleStreamDecoder*>(decoder_ptr);
-  std::unique_ptr<SingleStreamDecoder> uniqueDecoder(decoder);
+at::Tensor _create_from_file_like(
+    int64_t file_like_context,
+    std::optional<std::string_view> seek_mode) {
+  auto fileLikeContext =
+      reinterpret_cast<AVIOFileLikeContext*>(file_like_context);
+  TORCH_CHECK(
+      fileLikeContext != nullptr, "file_like_context must be a valid pointer");
+  std::unique_ptr<AVIOFileLikeContext> avioContextHolder(fileLikeContext);
+
+  SingleStreamDecoder::SeekMode realSeek = SingleStreamDecoder::SeekMode::exact;
+  if (seek_mode.has_value()) {
+    realSeek = seekModeFromString(seek_mode.value());
+  }
+
+  std::unique_ptr<SingleStreamDecoder> uniqueDecoder =
+      std::make_unique<SingleStreamDecoder>(
+          std::move(avioContextHolder), realSeek);
   return wrapDecoderPointerToTensor(std::move(uniqueDecoder));
 }
 
 void _add_video_stream(
     at::Tensor& decoder,
-    std::optional<int64_t> width = std::nullopt,
-    std::optional<int64_t> height = std::nullopt,
     std::optional<int64_t> num_threads = std::nullopt,
     std::optional<std::string_view> dimension_order = std::nullopt,
     std::optional<int64_t> stream_index = std::nullopt,
-    std::optional<std::string_view> device = std::nullopt,
+    std::string_view device = "cpu",
+    std::string_view device_variant = "default",
+    std::string_view transform_specs = "",
     std::optional<std::tuple<at::Tensor, at::Tensor, at::Tensor>>
         custom_frame_mappings = std::nullopt,
     std::optional<std::string_view> color_conversion_library = std::nullopt) {
   VideoStreamOptions videoStreamOptions;
-  videoStreamOptions.width = width;
-  videoStreamOptions.height = height;
   videoStreamOptions.ffmpegThreadCount = num_threads;
 
   if (dimension_order.has_value()) {
@@ -253,37 +348,46 @@ void _add_video_stream(
           ". color_conversion_library must be either filtergraph or swscale.");
     }
   }
-  if (device.has_value()) {
-    videoStreamOptions.device = createTorchDevice(std::string(device.value()));
-  }
+
+  validateDeviceInterface(std::string(device), std::string(device_variant));
+
+  videoStreamOptions.device = torch::Device(std::string(device));
+  videoStreamOptions.deviceVariant = device_variant;
+
+  std::vector<Transform*> transforms =
+      makeTransforms(std::string(transform_specs));
+
   std::optional<SingleStreamDecoder::FrameMappings> converted_mappings =
       custom_frame_mappings.has_value()
       ? std::make_optional(makeFrameMappings(custom_frame_mappings.value()))
       : std::nullopt;
   auto videoDecoder = unwrapTensorToGetDecoder(decoder);
   videoDecoder->addVideoStream(
-      stream_index.value_or(-1), videoStreamOptions, converted_mappings);
+      stream_index.value_or(-1),
+      transforms,
+      videoStreamOptions,
+      converted_mappings);
 }
 
 // Add a new video stream at `stream_index` using the provided options.
 void add_video_stream(
     at::Tensor& decoder,
-    std::optional<int64_t> width = std::nullopt,
-    std::optional<int64_t> height = std::nullopt,
     std::optional<int64_t> num_threads = std::nullopt,
     std::optional<std::string_view> dimension_order = std::nullopt,
     std::optional<int64_t> stream_index = std::nullopt,
-    std::optional<std::string_view> device = std::nullopt,
+    std::string_view device = "cpu",
+    std::string_view device_variant = "default",
+    std::string_view transform_specs = "",
     const std::optional<std::tuple<at::Tensor, at::Tensor, at::Tensor>>&
         custom_frame_mappings = std::nullopt) {
   _add_video_stream(
       decoder,
-      width,
-      height,
       num_threads,
       dimension_order,
       stream_index,
       device,
+      device_variant,
+      transform_specs,
       custom_frame_mappings);
 }
 
@@ -344,11 +448,9 @@ OpsFrameOutput get_frame_at_index(at::Tensor& decoder, int64_t frame_index) {
 // Return the frames at given indices for a given stream
 OpsFrameBatchOutput get_frames_at_indices(
     at::Tensor& decoder,
-    at::IntArrayRef frame_indices) {
+    const at::Tensor& frame_indices) {
   auto videoDecoder = unwrapTensorToGetDecoder(decoder);
-  std::vector<int64_t> frameIndicesVec(
-      frame_indices.begin(), frame_indices.end());
-  auto result = videoDecoder->getFramesAtIndices(frameIndicesVec);
+  auto result = videoDecoder->getFramesAtIndices(frame_indices);
   return makeOpsFrameBatchOutput(result);
 }
 
@@ -367,10 +469,9 @@ OpsFrameBatchOutput get_frames_in_range(
 // Return the frames at given ptss for a given stream
 OpsFrameBatchOutput get_frames_by_pts(
     at::Tensor& decoder,
-    at::ArrayRef<double> timestamps) {
+    const at::Tensor& timestamps) {
   auto videoDecoder = unwrapTensorToGetDecoder(decoder);
-  std::vector<double> timestampsVec(timestamps.begin(), timestamps.end());
-  auto result = videoDecoder->getFramesPlayedAt(timestampsVec);
+  auto result = videoDecoder->getFramesPlayedAt(timestamps);
   return makeOpsFrameBatchOutput(result);
 }
 
@@ -397,6 +498,21 @@ OpsAudioFramesOutput get_frames_by_pts_in_range_audio(
   return makeOpsAudioFramesOutput(result);
 }
 
+void encode_video_to_file(
+    const at::Tensor& frames,
+    int64_t frame_rate,
+    std::string_view file_name,
+    std::optional<int64_t> crf = std::nullopt) {
+  VideoStreamOptions videoStreamOptions;
+  videoStreamOptions.crf = crf;
+  VideoEncoder(
+      frames,
+      validateInt64ToInt(frame_rate, "frame_rate"),
+      file_name,
+      videoStreamOptions)
+      .encode();
+}
+
 void encode_audio_to_file(
     const at::Tensor& samples,
     int64_t sample_rate,
@@ -441,6 +557,36 @@ at::Tensor encode_audio_to_tensor(
       .encodeToTensor();
 }
 
+void _encode_audio_to_file_like(
+    const at::Tensor& samples,
+    int64_t sample_rate,
+    std::string_view format,
+    int64_t file_like_context,
+    std::optional<int64_t> bit_rate = std::nullopt,
+    std::optional<int64_t> num_channels = std::nullopt,
+    std::optional<int64_t> desired_sample_rate = std::nullopt) {
+  auto fileLikeContext =
+      reinterpret_cast<AVIOFileLikeContext*>(file_like_context);
+  TORCH_CHECK(
+      fileLikeContext != nullptr, "file_like_context must be a valid pointer");
+  std::unique_ptr<AVIOFileLikeContext> avioContextHolder(fileLikeContext);
+
+  AudioStreamOptions audioStreamOptions;
+  audioStreamOptions.bitRate = validateOptionalInt64ToInt(bit_rate, "bit_rate");
+  audioStreamOptions.numChannels =
+      validateOptionalInt64ToInt(num_channels, "num_channels");
+  audioStreamOptions.sampleRate =
+      validateOptionalInt64ToInt(desired_sample_rate, "desired_sample_rate");
+
+  AudioEncoder encoder(
+      samples,
+      validateInt64ToInt(sample_rate, "sample_rate"),
+      format,
+      std::move(avioContextHolder),
+      audioStreamOptions);
+  encoder.encode();
+}
+
 // For testing only. We need to implement this operation as a core library
 // function because what we're testing is round-tripping pts values as
 // double-precision floating point numbers from C++ to Python and back to C++.
@@ -694,14 +840,16 @@ void scan_all_streams_to_update_metadata(at::Tensor& decoder) {
 TORCH_LIBRARY_IMPL(torchcodec_ns, BackendSelect, m) {
   m.impl("create_from_file", &create_from_file);
   m.impl("create_from_tensor", &create_from_tensor);
-  m.impl("_convert_to_tensor", &_convert_to_tensor);
+  m.impl("_create_from_file_like", &_create_from_file_like);
   m.impl(
       "_get_json_ffmpeg_library_versions", &_get_json_ffmpeg_library_versions);
 }
 
 TORCH_LIBRARY_IMPL(torchcodec_ns, CPU, m) {
   m.impl("encode_audio_to_file", &encode_audio_to_file);
+  m.impl("encode_video_to_file", &encode_video_to_file);
   m.impl("encode_audio_to_tensor", &encode_audio_to_tensor);
+  m.impl("_encode_audio_to_file_like", &_encode_audio_to_file_like);
   m.impl("seek_to_pts", &seek_to_pts);
   m.impl("add_video_stream", &add_video_stream);
   m.impl("_add_video_stream", &_add_video_stream);