Apply default values with yaml

app.py

@@ -1,6 +1,7 @@
 import os
 import argparse
 import gradio as gr
+import yaml
 
 from modules.whisper.whisper_factory import WhisperFactory
 from modules.whisper.faster_whisper_inference import FasterWhisperInference
@@ -33,102 +34,119 @@ class App:
             output_dir=os.path.join(self.args.output_dir, "translations")
         )
 
+        default_param_path = os.path.join("configs", "default_parameters.yaml")
+        with open(default_param_path, 'r', encoding='utf-8') as file:
+            self.default_params = yaml.safe_load(file)
+
     def create_whisper_parameters(self):
+        whisper_params = self.default_params["whisper"]
+        vad_params = self.default_params["vad"]
+        diarization_params = self.default_params["diarization"]
+
         with gr.Row():
-            dd_model = gr.Dropdown(choices=self.whisper_inf.available_models, value="large-v2",
+            dd_model = gr.Dropdown(choices=self.whisper_inf.available_models, value=whisper_params["model_size"],
                                    label="Model")
             dd_lang = gr.Dropdown(choices=["Automatic Detection"] + self.whisper_inf.available_langs,
-                                  value="Automatic Detection", label="Language")
+                                  value=whisper_params["lang"], label="Language")
             dd_file_format = gr.Dropdown(choices=["SRT", "WebVTT", "txt"], value="SRT", label="File Format")
         with gr.Row():
-            cb_translate = gr.Checkbox(value=False, label="Translate to English?", interactive=True)
+            cb_translate = gr.Checkbox(value=whisper_params["is_translate"], label="Translate to English?",
+                                       interactive=True)
         with gr.Row():
             cb_timestamp = gr.Checkbox(value=True, label="Add a timestamp to the end of the filename",
                                        interactive=True)
         with gr.Accordion("Advanced Parameters", open=False):
-            nb_beam_size = gr.Number(label="Beam Size", value=5, precision=0, interactive=True,
+            nb_beam_size = gr.Number(label="Beam Size", value=whisper_params["beam_size"], precision=0, interactive=True,
                                      info="Beam size to use for decoding.")
-            nb_log_prob_threshold = gr.Number(label="Log Probability Threshold", value=-1.0, interactive=True,
+            nb_log_prob_threshold = gr.Number(label="Log Probability Threshold", value=whisper_params["log_prob_threshold"], interactive=True,
                                               info="If the average log probability over sampled tokens is below this value, treat as failed.")
-            nb_no_speech_threshold = gr.Number(label="No Speech Threshold", value=0.6, interactive=True,
+            nb_no_speech_threshold = gr.Number(label="No Speech Threshold", value=whisper_params["no_speech_threshold"], interactive=True,
                                                info="If the no speech probability is higher than this value AND the average log probability over sampled tokens is below 'Log Prob Threshold', consider the segment as silent.")
             dd_compute_type = gr.Dropdown(label="Compute Type", choices=self.whisper_inf.available_compute_types,
                                           value=self.whisper_inf.current_compute_type, interactive=True,
                                           info="Select the type of computation to perform.")
-            nb_best_of = gr.Number(label="Best Of", value=5, interactive=True,
+            nb_best_of = gr.Number(label="Best Of", value=whisper_params["best_of"], interactive=True,
                                    info="Number of candidates when sampling with non-zero temperature.")
-            nb_patience = gr.Number(label="Patience", value=1, interactive=True,
+            nb_patience = gr.Number(label="Patience", value=whisper_params["patience"], interactive=True,
                                     info="Beam search patience factor.")
-            cb_condition_on_previous_text = gr.Checkbox(label="Condition On Previous Text", value=True,
+            cb_condition_on_previous_text = gr.Checkbox(label="Condition On Previous Text", value=whisper_params["condition_on_previous_text"],
                                                         interactive=True,
                                                         info="Condition on previous text during decoding.")
-            sld_prompt_reset_on_temperature = gr.Slider(label="Prompt Reset On Temperature", value=0.5,
+            sld_prompt_reset_on_temperature = gr.Slider(label="Prompt Reset On Temperature", value=whisper_params["prompt_reset_on_temperature"],
                                                         minimum=0, maximum=1, step=0.01, interactive=True,
                                                         info="Resets prompt if temperature is above this value."
                                                              " Arg has effect only if 'Condition On Previous Text' is True.")
             tb_initial_prompt = gr.Textbox(label="Initial Prompt", value=None, interactive=True,
                                            info="Initial prompt to use for decoding.")
-            sd_temperature = gr.Slider(label="Temperature", value=0, step=0.01, maximum=1.0, interactive=True,
+            sd_temperature = gr.Slider(label="Temperature", value=whisper_params["temperature"], minimum=0.0,
+                                       step=0.01, maximum=1.0, interactive=True,
                                        info="Temperature for sampling. It can be a tuple of temperatures, which will be successively used upon failures according to either `Compression Ratio Threshold` or `Log Prob Threshold`.")
-            nb_compression_ratio_threshold = gr.Number(label="Compression Ratio Threshold", value=2.4, interactive=True,
+            nb_compression_ratio_threshold = gr.Number(label="Compression Ratio Threshold", value=whisper_params["compression_ratio_threshold"],
+                                                       interactive=True,
                                                        info="If the gzip compression ratio is above this value, treat as failed.")
             with gr.Group(visible=isinstance(self.whisper_inf, FasterWhisperInference)):
-                nb_length_penalty = gr.Number(label="Length Penalty", value=1,
+                nb_length_penalty = gr.Number(label="Length Penalty", value=whisper_params["length_penalty"],
                                               info="Exponential length penalty constant.")
-                nb_repetition_penalty = gr.Number(label="Repetition Penalty", value=1,
+                nb_repetition_penalty = gr.Number(label="Repetition Penalty", value=whisper_params["repetition_penalty"],
                                                   info="Penalty applied to the score of previously generated tokens (set > 1 to penalize).")
-                nb_no_repeat_ngram_size = gr.Number(label="No Repeat N-gram Size", value=0, precision=0,
+                nb_no_repeat_ngram_size = gr.Number(label="No Repeat N-gram Size", value=whisper_params["no_repeat_ngram_size"],
+                                                    precision=0,
                                                     info="Prevent repetitions of n-grams with this size (set 0 to disable).")
-                tb_prefix = gr.Textbox(label="Prefix", value=lambda: None,
+                tb_prefix = gr.Textbox(label="Prefix", value=lambda: whisper_params["prefix"],
                                        info="Optional text to provide as a prefix for the first window.")
-                cb_suppress_blank = gr.Checkbox(label="Suppress Blank", value=True,
+                cb_suppress_blank = gr.Checkbox(label="Suppress Blank", value=whisper_params["suppress_blank"],
                                                 info="Suppress blank outputs at the beginning of the sampling.")
-                tb_suppress_tokens = gr.Textbox(label="Suppress Tokens", value="[-1]",
+                tb_suppress_tokens = gr.Textbox(label="Suppress Tokens", value=whisper_params["suppress_tokens"],
                                                 info="List of token IDs to suppress. -1 will suppress a default set of symbols as defined in the model config.json file.")
-                nb_max_initial_timestamp = gr.Number(label="Max Initial Timestamp", value=1.0,
+                nb_max_initial_timestamp = gr.Number(label="Max Initial Timestamp", value=whisper_params["max_initial_timestamp"],
                                                      info="The initial timestamp cannot be later than this.")
-                cb_word_timestamps = gr.Checkbox(label="Word Timestamps", value=False,
+                cb_word_timestamps = gr.Checkbox(label="Word Timestamps", value=whisper_params["word_timestamps"],
                                                  info="Extract word-level timestamps using the cross-attention pattern and dynamic time warping, and include the timestamps for each word in each segment.")
-                tb_prepend_punctuations = gr.Textbox(label="Prepend Punctuations", value="\"'“¿([{-",
+                tb_prepend_punctuations = gr.Textbox(label="Prepend Punctuations", value=whisper_params["prepend_punctuations"],
                                                      info="If 'Word Timestamps' is True, merge these punctuation symbols with the next word.")
-                tb_append_punctuations = gr.Textbox(label="Append Punctuations", value="\"'.。,，!！?？:：”)]}、",
+                tb_append_punctuations = gr.Textbox(label="Append Punctuations", value=whisper_params["append_punctuations"],
                                                     info="If 'Word Timestamps' is True, merge these punctuation symbols with the previous word.")
-                nb_max_new_tokens = gr.Number(label="Max New Tokens", value=lambda: None, precision=0,
+                nb_max_new_tokens = gr.Number(label="Max New Tokens", value=lambda: whisper_params["max_new_tokens"],
+                                              precision=0,
                                               info="Maximum number of new tokens to generate per-chunk. If not set, the maximum will be set by the default max_length.")
-                nb_chunk_length = gr.Number(label="Chunk Length", value=lambda: None, precision=0,
+                nb_chunk_length = gr.Number(label="Chunk Length", value=lambda: whisper_params["chunk_length"],
+                                            precision=0,
                                             info="The length of audio segments. If it is not None, it will overwrite the default chunk_length of the FeatureExtractor.")
                 nb_hallucination_silence_threshold = gr.Number(label="Hallucination Silence Threshold (sec)",
-                                                               value=lambda: None,
+                                                               value=lambda: whisper_params["hallucination_silence_threshold"],
                                                                info="When 'Word Timestamps' is True, skip silent periods longer than this threshold (in seconds) when a possible hallucination is detected.")
-                tb_hotwords = gr.Textbox(label="Hotwords", value=None,
+                tb_hotwords = gr.Textbox(label="Hotwords", value=lambda: whisper_params["hotwords"],
                                          info="Hotwords/hint phrases to provide the model with. Has no effect if prefix is not None.")
-                nb_language_detection_threshold = gr.Number(label="Language Detection Threshold", value=None,
+                nb_language_detection_threshold = gr.Number(label="Language Detection Threshold", value=lambda: whisper_params["language_detection_threshold"],
                                                             info="If the maximum probability of the language tokens is higher than this value, the language is detected.")
-                nb_language_detection_segments = gr.Number(label="Language Detection Segments", value=1, precision=0,
+                nb_language_detection_segments = gr.Number(label="Language Detection Segments", value=lambda: whisper_params["language_detection_segments"],
+                                                           precision=0,
                                                            info="Number of segments to consider for the language detection.")
             with gr.Group(visible=isinstance(self.whisper_inf, InsanelyFastWhisperInference)):
-                nb_chunk_length_s = gr.Number(label="Chunk Lengths (sec)", value=30, precision=0)
-                nb_batch_size = gr.Number(label="Batch Size", value=24, precision=0)
+                nb_chunk_length_s = gr.Number(label="Chunk Lengths (sec)", value=whisper_params["chunk_length_s"],
+                                              precision=0)
+                nb_batch_size = gr.Number(label="Batch Size", value=whisper_params["batch_size"], precision=0)
 
         with gr.Accordion("VAD", open=False):
-            cb_vad_filter = gr.Checkbox(label="Enable Silero VAD Filter", value=False, interactive=True)
-            sd_threshold = gr.Slider(minimum=0.0, maximum=1.0, step=0.01, label="Speech Threshold", value=0.5,
+            cb_vad_filter = gr.Checkbox(label="Enable Silero VAD Filter", value=vad_params["vad_filter"],
+                                        interactive=True)
+            sd_threshold = gr.Slider(minimum=0.0, maximum=1.0, step=0.01, label="Speech Threshold", value=vad_params["threshold"],
                                      info="Lower it to be more sensitive to small sounds.")
-            nb_min_speech_duration_ms = gr.Number(label="Minimum Speech Duration (ms)", precision=0, value=250,
+            nb_min_speech_duration_ms = gr.Number(label="Minimum Speech Duration (ms)", precision=0, value=vad_params["min_speech_duration_ms"],
                                                   info="Final speech chunks shorter than this time are thrown out")
-            nb_max_speech_duration_s = gr.Number(label="Maximum Speech Duration (s)", value=9999,
+            nb_max_speech_duration_s = gr.Number(label="Maximum Speech Duration (s)", value=vad_params["max_speech_duration_s"],
                                                  info="Maximum duration of speech chunks in \"seconds\". Chunks longer"
                                                         " than this time will be split at the timestamp of the last silence that"
                                                         " lasts more than 100ms (if any), to prevent aggressive cutting.")
-            nb_min_silence_duration_ms = gr.Number(label="Minimum Silence Duration (ms)", precision=0, value=2000,
+            nb_min_silence_duration_ms = gr.Number(label="Minimum Silence Duration (ms)", precision=0, value=vad_params["min_silence_duration_ms"],
                                                    info="In the end of each speech chunk wait for this time"
                                                         " before separating it")
-            nb_speech_pad_ms = gr.Number(label="Speech Padding (ms)", precision=0, value=400,
+            nb_speech_pad_ms = gr.Number(label="Speech Padding (ms)", precision=0, value=vad_params["speech_pad_ms"],
                                          info="Final speech chunks are padded by this time each side")
 
         with gr.Accordion("Diarization", open=False):
-            cb_diarize = gr.Checkbox(label="Enable Diarization")
-            tb_hf_token = gr.Text(label="HuggingFace Token", value="",
+            cb_diarize = gr.Checkbox(label="Enable Diarization", value=diarization_params["is_diarize"])
+            tb_hf_token = gr.Text(label="HuggingFace Token", value=diarization_params["hf_token"],
                                   info="This is only needed the first time you download the model. If you already have models, you don't need to enter. To download the model, you must manually go to \"https://huggingface.co/pyannote/speaker-diarization-3.1\" and agree to their requirement.")
             dd_diarization_device = gr.Dropdown(label="Device",
                                                 choices=self.whisper_inf.diarizer.get_available_device(),