commit files to HF hub

.gitattributes

@@ -15,3 +15,4 @@
 *.pt filter=lfs diff=lfs merge=lfs -text
 *.pth filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,12 @@
+---
+library_name: superb
+benchmark: superb
+task: sd
+datasets:
+- superb
+tags:
+- speaker-diarization
+- osanseviero/hubert_base
+---
+
+# Fine-tuned s3prl model for SD

args_2021-07-30-22-56-56.yaml

@@ -0,0 +1,26 @@
+auto_resume: false
+backend: nccl
+cache_dir: null
+config: ./downstream/diarization/config.yaml
+device: cuda
+downstream: diarization
+downstream_variant: null
+evaluate_split: test
+expdir: result/downstream/sd-push-to-hub
+expname: sd-push-to-hub
+from_hf_hub: true
+hub: huggingface
+init_ckpt: null
+local_rank: null
+mode: train
+override: config.downstream_expert.loaderrc.train_dir='/data/lewis/superb/Libri2Mix/train',,config.downstream_expert.loaderrc.dev_dir='/data/lewis/superb/Libri2Mix/dev',,config.downstream_expert.loaderrc.test_dir='/data/lewis/superb/Libri2Mix/test',,config.runner.total_steps=2000
+past_exp: null
+seed: 1337
+upstream: osanseviero/hubert_base
+upstream_ckpt: null
+upstream_feature_selection: hidden_states
+upstream_model_config: null
+upstream_model_name: model.pt
+upstream_refresh: false
+upstream_trainable: false
+verbose: false

best-states-dev.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6dd37726e49314bea4ef308eb079de842b4400d803777b6c5c47fb9dda3975d3
+size 31526636

config_2021-07-30-22-56-56.yaml

@@ -0,0 +1,34 @@
+downstream_expert:
+  datarc:
+    chunk_size: 2000
+    frame_shift: 160
+    label_delay: 0
+    num_speakers: 2
+    rate: 16000
+    subsampling: 1
+  loaderrc:
+    dev_dir: /data/lewis/superb/Libri2Mix/dev
+    eval_batchsize: 1
+    num_workers: 8
+    test_dir: /data/lewis/superb/Libri2Mix/test
+    train_batchsize: 8
+    train_dir: /data/lewis/superb/Libri2Mix/train
+  modelrc:
+    hidden_size: 512
+    rnn_layers: 1
+  scorerc:
+    save_predictions: true
+optimizer:
+  lr: 0.0001
+  name: TorchOptim
+  torch_optim_name: Adam
+runner:
+  eval_dataloaders:
+  - dev
+  eval_step: 500
+  gradient_accumulate_steps: 4
+  gradient_clipping: 1
+  log_step: 500
+  max_keep: 1
+  save_step: 500
+  total_steps: 2000

events.out.tfevents.1627678623.vorace

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6fb8837a8aa7b3e84a8af54ef3adab01f3b9a8ea91982bc931f8834f931fc8ea
+size 748

hub_repo

@@ -0,0 +1 @@
+Subproject commit 6adc9cb4f0b6f1dc428b178979c526a94f353494

model.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6dd37726e49314bea4ef308eb079de842b4400d803777b6c5c47fb9dda3975d3
+size 31526636

model.py

@@ -0,0 +1,39 @@
+from s3prl.downstream.runner import Runner
+from typing import Dict
+import torch
+import os
+
+
+class PreTrainedModel(Runner):
+    def __init__(self, path=""):
+        """
+        Initialize downstream model.
+        """
+        ckp_file = os.path.join(path, "model.ckpt")
+        ckp = torch.load(ckp_file, map_location='cpu')
+        ckp["Args"].init_ckpt = ckp_file
+        ckp["Args"].mode = "inference"
+        ckp["Args"].device = "cpu"
+
+        Runner.__init__(self, ckp["Args"], ckp["Config"])
+
+    def __call__(self, inputs)-> Dict[str, str]:
+        """
+        Args:
+            inputs (:obj:`np.array`):
+                The raw waveform of audio received. By default at 16KHz.
+        Return:
+            A :obj:`dict`:. The object should return a dictionary like
+            {"frames": "XXX"} which contains the frames where one, both, or none
+            of the speakers are speaking.
+        """
+        for entry in self.all_entries:
+            entry.model.eval()
+
+        inputs = [torch.FloatTensor(inputs)]
+
+        with torch.no_grad():
+            features = self.upstream.model(inputs)
+            features = self.featurizer.model(inputs, features)
+            preds = self.downstream.model.inference(features, [])
+        return {"frames": preds[0]}