working FGSM config for ASR + lint (#859)

armory/data/librispeech/librispeech_dev_clean_split.py

@@ -80,11 +80,11 @@ class LibrispeechConfig(tfds.core.BuilderConfig):
     def __init__(self, text_encoder_config=None, **kwargs):
         """Constructs a LibrispeechConfig.
 
-    Args:
-      text_encoder_config: `tfds.features.text.TextEncoderConfig`, configuration
-        for the `tfds.features.text.TextEncoder` used for the text feature.
-      **kwargs: keyword arguments forwarded to super.
-    """
+        Args:
+          text_encoder_config: `tfds.features.text.TextEncoderConfig`, configuration
+            for the `tfds.features.text.TextEncoder` used for the text feature.
+          **kwargs: keyword arguments forwarded to super.
+        """
         name = kwargs.get("name")
         if name is None:
             name = text_encoder_config.name if text_encoder_config else "plain_text"
@@ -107,12 +107,12 @@ def __init__(self, text_encoder_config=None, **kwargs):
 def _make_builder_configs():
     """Make built-in Librispeech BuilderConfigs.
 
-  Uses 3 text encodings (plain_text, subwords with 8k vocab, subwords with 32k
-  vocab).
+    Uses 3 text encodings (plain_text, subwords with 8k vocab, subwords with 32k
+    vocab).
 
-  Returns:
-    `list<tfds.audio.LibrispeechConfig>`
-  """
+    Returns:
+      `list<tfds.audio.LibrispeechConfig>`
+    """
     text_encoder_configs = [
         None,
         tfds.features.text.TextEncoderConfig(

scenario_configs/asr_deepspeech_baseline_fgsm.json

@@ -0,0 +1,72 @@
+{
+    "_description": "Baseline DeepSpeech ASR on LibriSpeech, contributed by MITRE Corporation",
+    "adhoc": {
+        "skip_adversarial": false
+    },
+    "attack": {
+        "knowledge": "white",
+        "kwargs": {
+            "batch_size": 1,
+            "eps": 0.2,
+            "eps_step": 0.1,
+            "minimal": false,
+            "num_random_init": 0,
+            "targeted": true
+        },
+        "module": "art.attacks.evasion",
+        "name": "FastGradientMethod",
+        "targeted": true,
+        "targeted_labels": {
+            "scheme": "string",
+            "value": "TEST STRING"
+        },
+        "use_label": false
+    },
+    "dataset": {
+        "batch_size": 1,
+        "eval_split": "test_clean",
+        "framework": "numpy",
+        "module": "armory.data.datasets",
+        "name": "librispeech",
+        "train_split": "train_clean100"
+    },
+    "defense": null,
+    "metric": {
+        "means": false,
+        "perturbation": "linf",
+        "record_metric_per_sample": true,
+        "task": [
+            "word_error_rate"
+        ]
+    },
+    "model": {
+        "fit": false,
+        "fit_kwargs": {
+            "nb_epochs": 20000
+        },
+        "model_kwargs": {},
+        "module": "armory.baseline_models.pytorch.deep_speech",
+        "name": "get_art_model",
+        "predict_kwargs": {
+            "transcription_output": true
+        },
+        "weights_file": null,
+        "wrapper_kwargs": {
+            "pretrained_model": "librispeech"
+        }
+    },
+    "scenario": {
+        "kwargs": {},
+        "module": "armory.scenarios.audio_asr",
+        "name": "AutomaticSpeechRecognition"
+    },
+    "sysconfig": {
+        "docker_image": "twosixarmory/pytorch-deepspeech:0.12.2-dev",
+        "external_github_repo": "hkakitani/deepspeech.pytorch",
+        "gpus": "all",
+        "local_repo_path": null,
+        "output_dir": null,
+        "output_filename": null,
+        "use_gpu": false
+    }
+}