File size: 11,986 Bytes
79fc12a |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 |
import json
import os
import shutil
import sys
from collections import defaultdict
from statistics import mean
import pandas as pd
import requests
from constants import BASE_WHISPERKIT_BENCHMARK_URL
from text_normalizer import text_normalizer
from utils import compute_average_wer, download_dataset
def fetch_evaluation_data(url):
"""
Fetches evaluation data from the given URL.
:param url: The URL to fetch the evaluation data from.
:returns: The evaluation data as a dictionary.
:rauses: sys.exit if the request fails
"""
response = requests.get(url)
if response.status_code == 200:
return json.loads(response.text)
else:
sys.exit(f"Failed to fetch WhisperKit evals: {response.text}")
def process_benchmark_file(file_path, dataset_dfs, device_map, results):
"""
Processes a single benchmark file and updates the results dictionary.
:param file_path: Path to the benchmark JSON file.
:param dataset_dfs: Dictionary of DataFrames containing dataset information.
:param results: Dictionary to store the processed results.
This function reads a benchmark JSON file, extracts relevant information,
and updates the results dictionary with various metrics including WER,
speed, tokens per second, and quality of inference (QoI).
"""
with open(file_path, "r") as file:
test_results = json.load(file)
if len(test_results) == 0:
return
commit_hash_timestamp = file_path.split("/")[-2]
commit_timestamp, commit_hash = commit_hash_timestamp.split("_")
first_test_result = test_results[0]
if first_test_result is None:
return
filename = file_path.split("/")[-1].strip(".json")
device, company, model, dataset_dir, timestamp = filename.split("_")
model = f"{company}_{model}"
if device not in device_map:
return
device = device_map[device]
os_info = first_test_result["staticAttributes"]["os"]
key = (model, device, os_info, commit_timestamp)
dataset_name = dataset_dir
for test_result in test_results:
if test_result is None:
continue
test_info = test_result["testInfo"]
audio_file_name = test_info["audioFile"]
dataset_df = dataset_dfs[dataset_name]
wer_entry = {
"prediction": text_normalizer(test_info["prediction"]),
"reference": text_normalizer(test_info["reference"]),
}
results[key]["timestamp"] = timestamp
results[key]["average_wer"].append(wer_entry)
input_audio_seconds = test_info["timings"]["inputAudioSeconds"]
full_pipeline = test_info["timings"]["fullPipeline"] / 1000
time_elapsed = test_result["latencyStats"]["measurements"]["timeElapsed"]
total_decoding_loops = test_info["timings"]["totalDecodingLoops"]
results[key]["dataset_speed"][dataset_name][
"inputAudioSeconds"
] += input_audio_seconds
results[key]["dataset_speed"][dataset_name]["fullPipeline"] += full_pipeline
results[key]["speed"]["inputAudioSeconds"] += input_audio_seconds
results[key]["speed"]["fullPipeline"] += full_pipeline
results[key]["commit_hash"] = commit_hash
results[key]["commit_timestamp"] = commit_timestamp
results[key]["dataset_tokens_per_second"][dataset_name][
"totalDecodingLoops"
] += total_decoding_loops
results[key]["dataset_tokens_per_second"][dataset_name][
"timeElapsed"
] += time_elapsed
results[key]["tokens_per_second"]["totalDecodingLoops"] += total_decoding_loops
results[key]["tokens_per_second"]["timeElapsed"] += time_elapsed
audio = audio_file_name.split(".")[0]
audio = audio.split("-")[0]
dataset_row = dataset_df.loc[dataset_df["file"].str.contains(audio)].iloc[0]
reference_wer = dataset_row["wer"]
prediction_wer = test_info["wer"]
results[key]["qoi"].append(1 if prediction_wer <= reference_wer * 110 else 0)
def calculate_and_save_performance_results(
performance_results, performance_output_path
):
"""
Calculates final performance metrics and saves them to a JSON file.
:param performance_results: Dictionary containing raw performance data.
:param performance_output_path: Path to save the processed performance results.
This function processes the raw performance data, calculates average metrics,
and writes the final results to a JSON file, with each entry representing
a unique combination of model, device, and OS.
"""
not_supported = []
with open(performance_output_path, "w") as performance_file:
for key, data in performance_results.items():
model, device, os_info, timestamp = key
speed = round(
data["speed"]["inputAudioSeconds"] / data["speed"]["fullPipeline"], 2
)
# if speed < 1.0:
# not_supported.append((model, device, os_info))
# continue
performance_entry = {
"model": model.replace("_", "/"),
"device": device,
"os": os_info.replace("_", " "),
"timestamp": data["timestamp"],
"speed": speed,
"tokens_per_second": round(
data["tokens_per_second"]["totalDecodingLoops"]
/ data["tokens_per_second"]["timeElapsed"],
2,
),
"dataset_speed": {
dataset: round(
speed_info["inputAudioSeconds"] / speed_info["fullPipeline"], 2
)
for dataset, speed_info in data["dataset_speed"].items()
},
"dataset_tokens_per_second": {
dataset: round(
tps_info["totalDecodingLoops"] / tps_info["timeElapsed"], 2
)
for dataset, tps_info in data["dataset_tokens_per_second"].items()
},
"average_wer": compute_average_wer(data["average_wer"]),
"qoi": round(mean(data["qoi"]), 2),
"commit_hash": data["commit_hash"],
"commit_timestamp": data["commit_timestamp"],
}
json.dump(performance_entry, performance_file)
performance_file.write("\n")
return not_supported
def generate_support_matrix(performance_data_path="dashboard_data/performance_data.json", output_file="dashboard_data/support_data.csv"):
"""
Generate a support matrix CSV showing model compatibility across devices and OS versions.
✅: All tests passed
⚠️: Some tests failed
"""
support_matrix = defaultdict(lambda: defaultdict(lambda: {
"os_versions": set(),
"dataset_count": 0
}))
models = set()
devices = set()
# Process performance data
with open(performance_data_path, 'r') as f:
for line in f:
entry = json.loads(line)
model = entry["model"]
device = entry["device"]
os_info = entry["os"]
models.add(model)
devices.add(device)
support_matrix[model][device]["os_versions"].add(os_info)
if "dataset_speed" in entry:
support_matrix[model][device]["dataset_count"] = len(entry["dataset_speed"])
# Create DataFrame with correct headers
df = pd.DataFrame(columns=['', 'Model'] + [f'"{device}"' for device in sorted(devices)])
# Add each model with its data
for model in sorted(models):
row_data = {'': model, 'Model': model}
for device in sorted(devices):
info = support_matrix[model].get(device, {"dataset_count": 0, "os_versions": set()})
os_versions = ', '.join(sorted(info["os_versions"]))
if info["dataset_count"] == 0:
row_data[f'"{device}"'] = "Not Supported"
elif info["dataset_count"] >= 2:
row_data[f'"{device}"'] = f"✅ {os_versions}"
else:
row_data[f'"{device}"'] = f"⚠️ {os_versions}"
df = pd.concat([df, pd.DataFrame([row_data])], ignore_index=True)
# Save to CSV
df.to_csv(output_file, index=False)
def main():
"""
Main function to orchestrate the performance data generation process.
This function performs the following steps:
1. Downloads benchmark data if requested.
2. Fetches evaluation data for various datasets.
3. Processes benchmark files and summary files.
4. Calculates and saves performance and support results.
"""
source_xcresult_repo = "argmaxinc/whisperkit-evals-dataset"
source_xcresult_subfolder = "benchmark_data/"
source_xcresult_directory = f"{source_xcresult_repo}/{source_xcresult_subfolder}"
if len(sys.argv) > 1 and sys.argv[1] == "download":
try:
shutil.rmtree(source_xcresult_repo)
except:
print("Nothing to remove.")
download_dataset(
source_xcresult_repo, source_xcresult_repo, source_xcresult_subfolder
)
datasets = {
"Earnings-22": "https://huggingface.co/datasets/argmaxinc/whisperkit-evals/resolve/main/WhisperOpenAIAPI/openai_whisper-large-v2/earnings22/2024-03-04_13%3A39%3A42_GMT-0800.json",
"LibriSpeech": "https://huggingface.co/datasets/argmaxinc/whisperkit-evals/resolve/main/WhisperOpenAIAPI/openai_whisper-large-v2/librispeech/2024-02-28_18%3A45%3A02_GMT-0800.json?download=true",
"earnings22-10mins": "https://huggingface.co/datasets/argmaxinc/whisperkit-evals/resolve/main/WhisperOpenAIAPI/openai_whisper-large-v2/earnings22/2024-03-04_13%3A39%3A42_GMT-0800.json",
"librispeech-10mins": "https://huggingface.co/datasets/argmaxinc/whisperkit-evals/resolve/main/WhisperOpenAIAPI/openai_whisper-large-v2/librispeech/2024-02-28_18%3A45%3A02_GMT-0800.json?download=true",
"earnings22-12hours": "https://huggingface.co/datasets/argmaxinc/whisperkit-evals/resolve/main/WhisperOpenAIAPI/openai_whisper-large-v2/earnings22/2024-03-04_13%3A39%3A42_GMT-0800.json",
"librispeech": "https://huggingface.co/datasets/argmaxinc/whisperkit-evals/resolve/main/WhisperOpenAIAPI/openai_whisper-large-v2/librispeech/2024-02-28_18%3A45%3A02_GMT-0800.json?download=true",
}
dataset_dfs = {}
for dataset_name, url in datasets.items():
evals = fetch_evaluation_data(url)
dataset_dfs[dataset_name] = pd.json_normalize(evals["results"])
performance_results = defaultdict(
lambda: {
"average_wer": [],
"qoi": [],
"speed": {"inputAudioSeconds": 0, "fullPipeline": 0},
"tokens_per_second": {"totalDecodingLoops": 0, "timeElapsed": 0},
"dataset_speed": defaultdict(
lambda: {"inputAudioSeconds": 0, "fullPipeline": 0}
),
"dataset_tokens_per_second": defaultdict(
lambda: {"totalDecodingLoops": 0, "timeElapsed": 0}
),
"timestamp": None,
"commit_hash": None,
"commit_timestamp": None,
"test_timestamp": None,
}
)
with open("dashboard_data/device_map.json", "r") as f:
device_map = json.load(f)
for subdir, _, files in os.walk(source_xcresult_directory):
for filename in files:
file_path = os.path.join(subdir, filename)
if not filename.endswith(".json"):
continue
else:
process_benchmark_file(file_path, dataset_dfs, device_map, performance_results)
calculate_and_save_performance_results(
performance_results, "dashboard_data/performance_data.json"
)
generate_support_matrix()
if __name__ == "__main__":
main()
|