update readme

README.md

@@ -82,8 +82,8 @@ Training code can be found in [training/](training/) directory (not used for inf
 - **Recall**: 89.97%
 - **F-score**: 91,71%
 - **Environmental Impact**:
-  - Emissions tracked in gCO2eq
-  - Energy consumption tracked in Wh
+  - Emissions tracked in gCO2eq: 5.19
+  - Energy consumption tracked in Wh: 14.07
 - **Mistakes**
   - False positive represents 38.35% of mistakes
   - False negative represents 61.65% of mistakes

requirements.txt

@@ -2,7 +2,6 @@ fastapi>=0.68.0
 uvicorn>=0.15.0
 codecarbon>=2.3.1
 datasets>=2.14.0
-scikit-learn>=1.0.2
 pydantic>=1.10.0
 python-dotenv>=1.0.0
 gradio>=4.0.0

tasks/audio.py

@@ -1,7 +1,6 @@
 from fastapi import APIRouter
 from datetime import datetime
 from datasets import load_dataset
-#from sklearn.metrics import accuracy_score
 import os
 import torch
 
@@ -15,7 +14,7 @@ load_dotenv()
 
 router = APIRouter()
 
-DESCRIPTION = "Chainsaw goes brrr ⇒ GPU goes brrr"
+DESCRIPTION = "ChainsawDetector"
 ROUTE = "/audio"
 
 
@@ -35,6 +34,7 @@ async def evaluate_audio(request: AudioEvaluationRequest):
         "chainsaw": 0,
         "environment": 1
     }
+
     # Load and prepare the dataset
     # Because the dataset is gated, we need to use the HF_TOKEN environment variable to authenticate
     batch_size = 16

tasks/utils/preprocess.py

@@ -4,7 +4,7 @@ from math import floor
 import torch
 from torch.nn.functional import pad
 from torchaudio.transforms import Resample
-#from random import randint
+from random import randint
 
 
 def get_dataloader(dataset, device, batch_size=16, shuffle=True):
@@ -43,9 +43,7 @@ def preprocess(X, newsr, n_fft, win_length, hop_length, gain=0.8, bias=10, power
     return X
 
 
-
 def prepare_batch(samples):
-    #maxlen=60
     newsr = 4000
     n_fft = 2**10 # power of 2
     win_length = 2**10
@@ -66,34 +64,36 @@ def prepare_batch(samples):
     labels = torch.tensor(labels, dtype=float)
     return batch, labels
 
-# def random_mask(sample):
-#     # random rectangular mask
-#     B, H, W = sample.shape
-#     for b in range(B):
-#         for _ in range(randint(3,12)):
-#             w = randint(5, 15)
-#             h = randint(10, 100)
-#             x1 = randint(0, W-w)
-#             y1 = randint(0, H-h)
-#             sample[b, y1:y1+h, x1:x1+w] = 0
-#     return sample
+# Data augmentation
+
+def random_mask(sample):
+    # random rectangular mask
+    B, H, W = sample.shape
+    for b in range(B):
+        for _ in range(randint(3,12)):
+            w = randint(5, 15)
+            h = randint(10, 100)
+            x1 = randint(0, W-w)
+            y1 = randint(0, H-h)
+            sample[b, y1:y1+h, x1:x1+w] = 0
+    return sample
 
-# def timeshift(sample):
-#     padsize = randint(0, 6)
-#     length = sample.size(2)
-#     randpad = torch.zeros((sample.size(0), sample.size(1), padsize), dtype=torch.float32)
-#     sample = torch.cat((randpad, sample), dim=2)
-#     sample = sample[:,:,:length]
-#     return sample
+def timeshift(sample):
+    padsize = randint(0, 6)
+    length = sample.size(2)
+    randpad = torch.zeros((sample.size(0), sample.size(1), padsize), dtype=torch.float32)
+    sample = torch.cat((randpad, sample), dim=2)
+    sample = sample[:,:,:length]
+    return sample
 
-# def add_noise(sample):
-#     #noise = np.random.normal(0, 0.05*sample.max(), sample.shape)
-#     noise = 0.05*sample.max()*torch.randn(sample.shape, dtype=torch.float32)
-#     sample = sample + noise
-#     return sample
+def add_noise(sample):
+    #noise = np.random.normal(0, 0.05*sample.max(), sample.shape)
+    noise = 0.05*sample.max()*torch.randn(sample.shape, dtype=torch.float32)
+    sample = sample + noise
+    return sample
 
-# def augment(sample):
-#     sample = timeshift(sample)
-#     sample = random_mask(sample)
-#     sample = add_noise(sample)
-#     return sample
+def augment(sample):
+    sample = timeshift(sample)
+    sample = random_mask(sample)
+    sample = add_noise(sample)
+    return sample