Video Encoder

class VideoEncoder

__init__(

coding_structure: CodingStructure,

shared_coolchic_parameter: CoolChicEncoderParameter,

shared_frame_encoder_manager: FrameEncoderManager,

)

A VideoEncoder object is our main object. Its purpose is to encode a video i.e. one I-frame followed by 0 to N inter (P or B) frames.

Parameters:

• coding_structure (CodingStructure) – The coding structure (organization of the different frames) used to encode the video.

• shared_coolchic_parameter (CoolChicEncoderParameter) – Common parameters for all Cool-chic of all frames (e.g. synthesis architecture). Can be overridden later in the encode function to better suits the need of each individual frame.

• shared_frame_encoder_manager (FrameEncoderManager) – Common training parameters for all frames (e.g. max. number of iterations). It can be overridden later in the encode function to better suits the need of each individual frame.

encode(

path_original_sequence: str,

device: Literal['cpu', 'cuda:0'],

workdir: str,

job_duration_min: int = -1,

) → TrainingExitCode

Main training function of a VideoEncoder. Encode all required frames (i.e. as stated in self.coding_structure) of the video located at path_original_sequence. This will fill the dictionary self.all_frame_encoders containing the successively overfitted frame encoders.

There is a series of nested loops to encode the video, following roughly this process:

# Code all frames
for idx_coding_order in range(n_frames):
    # Perform n_loops independent encoding
    for idx_loop in range(n_loops):
        # Find the best initialization
        frame_encoder = warmup(...)
        # Perform the successive training stages
        for training_phase in all_training_phases:
            frame_encoder = train(frame_encoder, training_phase)
        # Training is over, test and save
        results = test(frame_encoder)
        frame_encoder.save()

Parameters:

• path_original_sequence (str) – Absolute path to the original image or video to be compressed.

• device (Literal['cpu', 'cuda:0']) – On which device should the training run

• workdir (str) – Where we’ll save many thing

• job_duration_min (int) – Exit and save the job after this duration is passed. Use -1 to only exit at the end of the entire encoding. Default to -1.

Returns:

Either TrainingExitCode.REQUEUE if job has run for longer than job_duration_min and should be put back into the job queue, or TrainingExitCode.END if the encoding is actually over.

Return type:

TrainingExitCode

get_frame_workdir(workdir: str, frame_display_order: int) → str

Compute the absolute path for the workdir of one frame.

Parameters:

• workdir (str) – Main working directory of the video encoder

• frame_display_order (int) – Display order of the frame

Returns:

Working directory of the frame

Return type:

str

concat_results_file(workdir: str) → None

Look at all the already encoded frames inside workdir and concatenate their result files (workdir/frame_XXX/results_best.tsv) into a single result file workdir/results_best.tsv.

Parameters:

workdir (str) – Working directory of the video encoder

Return type:

None

get_ref_data(

frame: Frame,

) → List[FrameData]

Return a list of the (decoded) reference frames. The decoded data are obtained by recursively inferring the already learned FrameEncoder.

Parameters:

frame (Frame) – The frame whose reference(s) we want.

Returns:

The decoded reference frames.

Return type:

List[FrameData]

get_lmbda_from_depth(depth: float, initial_lmbda: float) → float

Perform the QP offset as follows:

\[\lambda_i = (\frac{3}{2})^{d} \lambda,\]

Parameters:

• depth (float) – The depth \(d\) of the frame in the GOP. See encoder/utils/coding_structure.py for more info

• initial_lmbda (float) – The lmbda of the I frame \(\lambda\).

Returns:

The lambda of the current frame \(\lambda_i\).

Return type:

float

save(save_path: str) → None

Save current VideoEncoder at given path. It contains everything, the CodingStructure, the shared parameters between the different frames as well as all the successive FrameEncoder and their respective FrameEncoderManager.

Parameters:

save_path (str) – Where to save the model

Return type:

None

load_video_encoder(load_path: str) → VideoEncoder

Load a video encoder.

Parameters:

load_path (str) – Absolute path where the VideoEncoder should be loaded.

Returns:

The loaded VideoEncoder

Return type:

VideoEncoder