|
|
@@ -0,0 +1,769 @@ |
|
|
|
{ |
|
|
|
"cells": [ |
|
|
|
{ |
|
|
|
"cell_type": "markdown", |
|
|
|
"metadata": { |
|
|
|
"id": "C_jdZ5vHJ4A9" |
|
|
|
}, |
|
|
|
"source": [ |
|
|
|
"# Task description\n", |
|
|
|
"- Classify the speakers of given features.\n", |
|
|
|
"- Main goal: Learn how to use transformer.\n", |
|
|
|
"- Baselines:\n", |
|
|
|
" - Easy: Run sample code and know how to use transformer.\n", |
|
|
|
" - Medium: Know how to adjust parameters of transformer.\n", |
|
|
|
" - Strong: Construct [conformer](https://arxiv.org/abs/2005.08100) which is a variety of transformer. \n", |
|
|
|
" - Boss: Implement [Self-Attention Pooling](https://arxiv.org/pdf/2008.01077v1.pdf) & [Additive Margin Softmax](https://arxiv.org/pdf/1801.05599.pdf) to further boost the performance.\n", |
|
|
|
"\n", |
|
|
|
"- Other links\n", |
|
|
|
" - Kaggle: [link](https://www.kaggle.com/t/ac77388c90204a4c8daebeddd40ff916)\n", |
|
|
|
" - Slide: [link](https://docs.google.com/presentation/d/1HLAj7UUIjZOycDe7DaVLSwJfXVd3bXPOyzSb6Zk3hYU/edit?usp=sharing)\n", |
|
|
|
" - Data: [link](https://github.com/MachineLearningHW/ML_HW4_Dataset)\n", |
|
|
|
"\n", |
|
|
|
"# Download dataset\n", |
|
|
|
"- Data is [here](https://github.com/MachineLearningHW/ML_HW4_Dataset)" |
|
|
|
] |
|
|
|
}, |
|
|
|
{ |
|
|
|
"cell_type": "code", |
|
|
|
"execution_count": null, |
|
|
|
"metadata": { |
|
|
|
"id": "LhLNWB-AK2Z5" |
|
|
|
}, |
|
|
|
"outputs": [], |
|
|
|
"source": [ |
|
|
|
"\"\"\"\n", |
|
|
|
"If the links below become inaccessible, please connect TAs.\n", |
|
|
|
"\"\"\"\n", |
|
|
|
"\n", |
|
|
|
"!wget https://github.com/MachineLearningHW/ML_HW4_Dataset/raw/0.0.1/Dataset.tar.gz.partaa\n", |
|
|
|
"!wget https://github.com/MachineLearningHW/ML_HW4_Dataset/raw/0.0.1/Dataset.tar.gz.partab\n", |
|
|
|
"!wget https://github.com/MachineLearningHW/ML_HW4_Dataset/raw/0.0.1/Dataset.tar.gz.partac\n", |
|
|
|
"!wget https://github.com/MachineLearningHW/ML_HW4_Dataset/raw/0.0.1/Dataset.tar.gz.partad\n", |
|
|
|
"\n", |
|
|
|
"!cat Dataset.tar.gz.parta* > Dataset.tar.gz\n", |
|
|
|
"\n", |
|
|
|
"!tar zxvf Dataset.tar.gz" |
|
|
|
] |
|
|
|
}, |
|
|
|
{ |
|
|
|
"cell_type": "markdown", |
|
|
|
"metadata": { |
|
|
|
"id": "ENWVAUDVJtVY" |
|
|
|
}, |
|
|
|
"source": [ |
|
|
|
"## Fix Random Seed" |
|
|
|
] |
|
|
|
}, |
|
|
|
{ |
|
|
|
"cell_type": "code", |
|
|
|
"execution_count": null, |
|
|
|
"metadata": { |
|
|
|
"id": "E6burzCXIyuA" |
|
|
|
}, |
|
|
|
"outputs": [], |
|
|
|
"source": [ |
|
|
|
"import numpy as np\n", |
|
|
|
"import torch\n", |
|
|
|
"import random\n", |
|
|
|
"\n", |
|
|
|
"def set_seed(seed):\n", |
|
|
|
" np.random.seed(seed)\n", |
|
|
|
" random.seed(seed)\n", |
|
|
|
" torch.manual_seed(seed)\n", |
|
|
|
" if torch.cuda.is_available():\n", |
|
|
|
" torch.cuda.manual_seed(seed)\n", |
|
|
|
" torch.cuda.manual_seed_all(seed)\n", |
|
|
|
" torch.backends.cudnn.benchmark = False\n", |
|
|
|
" torch.backends.cudnn.deterministic = True\n", |
|
|
|
"\n", |
|
|
|
"set_seed(87)" |
|
|
|
] |
|
|
|
}, |
|
|
|
{ |
|
|
|
"cell_type": "markdown", |
|
|
|
"metadata": { |
|
|
|
"id": "k7dVbxW2LASN" |
|
|
|
}, |
|
|
|
"source": [ |
|
|
|
"# Data\n", |
|
|
|
"\n", |
|
|
|
"## Dataset\n", |
|
|
|
"- Original dataset is [Voxceleb2](https://www.robots.ox.ac.uk/~vgg/data/voxceleb/vox2.html).\n", |
|
|
|
"- The [license](https://creativecommons.org/licenses/by/4.0/) and [complete version](https://www.robots.ox.ac.uk/~vgg/data/voxceleb/files/license.txt) of Voxceleb2.\n", |
|
|
|
"- We randomly select 600 speakers from Voxceleb2.\n", |
|
|
|
"- Then preprocess the raw waveforms into mel-spectrograms.\n", |
|
|
|
"\n", |
|
|
|
"- Args:\n", |
|
|
|
" - data_dir: The path to the data directory.\n", |
|
|
|
" - metadata_path: The path to the metadata.\n", |
|
|
|
" - segment_len: The length of audio segment for training. \n", |
|
|
|
"- The architecture of data directory \\\\\n", |
|
|
|
" - data directory \\\\\n", |
|
|
|
" |---- metadata.json \\\\\n", |
|
|
|
" |---- testdata.json \\\\\n", |
|
|
|
" |---- mapping.json \\\\\n", |
|
|
|
" |---- uttr-{random string}.pt \\\\\n", |
|
|
|
"\n", |
|
|
|
"- The information in metadata\n", |
|
|
|
" - \"n_mels\": The dimention of mel-spectrogram.\n", |
|
|
|
" - \"speakers\": A dictionary. \n", |
|
|
|
" - Key: speaker ids.\n", |
|
|
|
" - value: \"feature_path\" and \"mel_len\"\n", |
|
|
|
"\n", |
|
|
|
"\n", |
|
|
|
"For efficiency, we segment the mel-spectrograms into segments in the traing step." |
|
|
|
] |
|
|
|
}, |
|
|
|
{ |
|
|
|
"cell_type": "code", |
|
|
|
"execution_count": null, |
|
|
|
"metadata": { |
|
|
|
"id": "KpuGxl4CI2pr" |
|
|
|
}, |
|
|
|
"outputs": [], |
|
|
|
"source": [ |
|
|
|
"import os\n", |
|
|
|
"import json\n", |
|
|
|
"import torch\n", |
|
|
|
"import random\n", |
|
|
|
"from pathlib import Path\n", |
|
|
|
"from torch.utils.data import Dataset\n", |
|
|
|
"from torch.nn.utils.rnn import pad_sequence\n", |
|
|
|
" \n", |
|
|
|
" \n", |
|
|
|
"class myDataset(Dataset):\n", |
|
|
|
"\tdef __init__(self, data_dir, segment_len=128):\n", |
|
|
|
"\t\tself.data_dir = data_dir\n", |
|
|
|
"\t\tself.segment_len = segment_len\n", |
|
|
|
"\t\n", |
|
|
|
"\t\t# Load the mapping from speaker neme to their corresponding id. \n", |
|
|
|
"\t\tmapping_path = Path(data_dir) / \"mapping.json\"\n", |
|
|
|
"\t\tmapping = json.load(mapping_path.open())\n", |
|
|
|
"\t\tself.speaker2id = mapping[\"speaker2id\"]\n", |
|
|
|
"\t\n", |
|
|
|
"\t\t# Load metadata of training data.\n", |
|
|
|
"\t\tmetadata_path = Path(data_dir) / \"metadata.json\"\n", |
|
|
|
"\t\tmetadata = json.load(open(metadata_path))[\"speakers\"]\n", |
|
|
|
"\t\n", |
|
|
|
"\t\t# Get the total number of speaker.\n", |
|
|
|
"\t\tself.speaker_num = len(metadata.keys())\n", |
|
|
|
"\t\tself.data = []\n", |
|
|
|
"\t\tfor speaker in metadata.keys():\n", |
|
|
|
"\t\t\tfor utterances in metadata[speaker]:\n", |
|
|
|
"\t\t\t\tself.data.append([utterances[\"feature_path\"], self.speaker2id[speaker]])\n", |
|
|
|
" \n", |
|
|
|
"\tdef __len__(self):\n", |
|
|
|
"\t\t\treturn len(self.data)\n", |
|
|
|
" \n", |
|
|
|
"\tdef __getitem__(self, index):\n", |
|
|
|
"\t\tfeat_path, speaker = self.data[index]\n", |
|
|
|
"\t\t# Load preprocessed mel-spectrogram.\n", |
|
|
|
"\t\tmel = torch.load(os.path.join(self.data_dir, feat_path))\n", |
|
|
|
"\n", |
|
|
|
"\t\t# Segmemt mel-spectrogram into \"segment_len\" frames.\n", |
|
|
|
"\t\tif len(mel) > self.segment_len:\n", |
|
|
|
"\t\t\t# Randomly get the starting point of the segment.\n", |
|
|
|
"\t\t\tstart = random.randint(0, len(mel) - self.segment_len)\n", |
|
|
|
"\t\t\t# Get a segment with \"segment_len\" frames.\n", |
|
|
|
"\t\t\tmel = torch.FloatTensor(mel[start:start+self.segment_len])\n", |
|
|
|
"\t\telse:\n", |
|
|
|
"\t\t\tmel = torch.FloatTensor(mel)\n", |
|
|
|
"\t\t# Turn the speaker id into long for computing loss later.\n", |
|
|
|
"\t\tspeaker = torch.FloatTensor([speaker]).long()\n", |
|
|
|
"\t\treturn mel, speaker\n", |
|
|
|
" \n", |
|
|
|
"\tdef get_speaker_number(self):\n", |
|
|
|
"\t\treturn self.speaker_num" |
|
|
|
] |
|
|
|
}, |
|
|
|
{ |
|
|
|
"cell_type": "markdown", |
|
|
|
"metadata": { |
|
|
|
"id": "668hverTMlGN" |
|
|
|
}, |
|
|
|
"source": [ |
|
|
|
"## Dataloader\n", |
|
|
|
"- Split dataset into training dataset(90%) and validation dataset(10%).\n", |
|
|
|
"- Create dataloader to iterate the data." |
|
|
|
] |
|
|
|
}, |
|
|
|
{ |
|
|
|
"cell_type": "code", |
|
|
|
"execution_count": null, |
|
|
|
"metadata": { |
|
|
|
"id": "B7c2gZYoJDRS" |
|
|
|
}, |
|
|
|
"outputs": [], |
|
|
|
"source": [ |
|
|
|
"import torch\n", |
|
|
|
"from torch.utils.data import DataLoader, random_split\n", |
|
|
|
"from torch.nn.utils.rnn import pad_sequence\n", |
|
|
|
"\n", |
|
|
|
"\n", |
|
|
|
"def collate_batch(batch):\n", |
|
|
|
"\t# Process features within a batch.\n", |
|
|
|
"\t\"\"\"Collate a batch of data.\"\"\"\n", |
|
|
|
"\tmel, speaker = zip(*batch)\n", |
|
|
|
"\t# Because we train the model batch by batch, we need to pad the features in the same batch to make their lengths the same.\n", |
|
|
|
"\tmel = pad_sequence(mel, batch_first=True, padding_value=-20) # pad log 10^(-20) which is very small value.\n", |
|
|
|
"\t# mel: (batch size, length, 40)\n", |
|
|
|
"\treturn mel, torch.FloatTensor(speaker).long()\n", |
|
|
|
"\n", |
|
|
|
"\n", |
|
|
|
"def get_dataloader(data_dir, batch_size, n_workers):\n", |
|
|
|
"\t\"\"\"Generate dataloader\"\"\"\n", |
|
|
|
"\tdataset = myDataset(data_dir)\n", |
|
|
|
"\tspeaker_num = dataset.get_speaker_number()\n", |
|
|
|
"\t# Split dataset into training dataset and validation dataset\n", |
|
|
|
"\ttrainlen = int(0.9 * len(dataset))\n", |
|
|
|
"\tlengths = [trainlen, len(dataset) - trainlen]\n", |
|
|
|
"\ttrainset, validset = random_split(dataset, lengths)\n", |
|
|
|
"\n", |
|
|
|
"\ttrain_loader = DataLoader(\n", |
|
|
|
"\t\ttrainset,\n", |
|
|
|
"\t\tbatch_size=batch_size,\n", |
|
|
|
"\t\tshuffle=True,\n", |
|
|
|
"\t\tdrop_last=True,\n", |
|
|
|
"\t\tnum_workers=n_workers,\n", |
|
|
|
"\t\tpin_memory=True,\n", |
|
|
|
"\t\tcollate_fn=collate_batch,\n", |
|
|
|
"\t)\n", |
|
|
|
"\tvalid_loader = DataLoader(\n", |
|
|
|
"\t\tvalidset,\n", |
|
|
|
"\t\tbatch_size=batch_size,\n", |
|
|
|
"\t\tnum_workers=n_workers,\n", |
|
|
|
"\t\tdrop_last=True,\n", |
|
|
|
"\t\tpin_memory=True,\n", |
|
|
|
"\t\tcollate_fn=collate_batch,\n", |
|
|
|
"\t)\n", |
|
|
|
"\n", |
|
|
|
"\treturn train_loader, valid_loader, speaker_num" |
|
|
|
] |
|
|
|
}, |
|
|
|
{ |
|
|
|
"cell_type": "markdown", |
|
|
|
"metadata": { |
|
|
|
"id": "5FOSZYxrMqhc" |
|
|
|
}, |
|
|
|
"source": [ |
|
|
|
"# Model\n", |
|
|
|
"- TransformerEncoderLayer:\n", |
|
|
|
" - Base transformer encoder layer in [Attention Is All You Need](https://arxiv.org/abs/1706.03762)\n", |
|
|
|
" - Parameters:\n", |
|
|
|
" - d_model: the number of expected features of the input (required).\n", |
|
|
|
"\n", |
|
|
|
" - nhead: the number of heads of the multiheadattention models (required).\n", |
|
|
|
"\n", |
|
|
|
" - dim_feedforward: the dimension of the feedforward network model (default=2048).\n", |
|
|
|
"\n", |
|
|
|
" - dropout: the dropout value (default=0.1).\n", |
|
|
|
"\n", |
|
|
|
" - activation: the activation function of intermediate layer, relu or gelu (default=relu).\n", |
|
|
|
"\n", |
|
|
|
"- TransformerEncoder:\n", |
|
|
|
" - TransformerEncoder is a stack of N transformer encoder layers\n", |
|
|
|
" - Parameters:\n", |
|
|
|
" - encoder_layer: an instance of the TransformerEncoderLayer() class (required).\n", |
|
|
|
"\n", |
|
|
|
" - num_layers: the number of sub-encoder-layers in the encoder (required).\n", |
|
|
|
"\n", |
|
|
|
" - norm: the layer normalization component (optional)." |
|
|
|
] |
|
|
|
}, |
|
|
|
{ |
|
|
|
"cell_type": "code", |
|
|
|
"execution_count": null, |
|
|
|
"metadata": { |
|
|
|
"id": "iXZ5B0EKJGs8" |
|
|
|
}, |
|
|
|
"outputs": [], |
|
|
|
"source": [ |
|
|
|
"import torch\n", |
|
|
|
"import torch.nn as nn\n", |
|
|
|
"import torch.nn.functional as F\n", |
|
|
|
"\n", |
|
|
|
"\n", |
|
|
|
"class Classifier(nn.Module):\n", |
|
|
|
"\tdef __init__(self, d_model=80, n_spks=600, dropout=0.1):\n", |
|
|
|
"\t\tsuper().__init__()\n", |
|
|
|
"\t\t# Project the dimension of features from that of input into d_model.\n", |
|
|
|
"\t\tself.prenet = nn.Linear(40, d_model)\n", |
|
|
|
"\t\t# TODO:\n", |
|
|
|
"\t\t# Change Transformer to Conformer.\n", |
|
|
|
"\t\t# https://arxiv.org/abs/2005.08100\n", |
|
|
|
"\t\tself.encoder_layer = nn.TransformerEncoderLayer(\n", |
|
|
|
"\t\t\td_model=d_model, dim_feedforward=256, nhead=2\n", |
|
|
|
"\t\t)\n", |
|
|
|
"\t\t# self.encoder = nn.TransformerEncoder(self.encoder_layer, num_layers=2)\n", |
|
|
|
"\n", |
|
|
|
"\t\t# Project the the dimension of features from d_model into speaker nums.\n", |
|
|
|
"\t\tself.pred_layer = nn.Sequential(\n", |
|
|
|
"\t\t\tnn.Linear(d_model, d_model),\n", |
|
|
|
"\t\t\tnn.ReLU(),\n", |
|
|
|
"\t\t\tnn.Linear(d_model, n_spks),\n", |
|
|
|
"\t\t)\n", |
|
|
|
"\n", |
|
|
|
"\tdef forward(self, mels):\n", |
|
|
|
"\t\t\"\"\"\n", |
|
|
|
"\t\targs:\n", |
|
|
|
"\t\t\tmels: (batch size, length, 40)\n", |
|
|
|
"\t\treturn:\n", |
|
|
|
"\t\t\tout: (batch size, n_spks)\n", |
|
|
|
"\t\t\"\"\"\n", |
|
|
|
"\t\t# out: (batch size, length, d_model)\n", |
|
|
|
"\t\tout = self.prenet(mels)\n", |
|
|
|
"\t\t# out: (length, batch size, d_model)\n", |
|
|
|
"\t\tout = out.permute(1, 0, 2)\n", |
|
|
|
"\t\t# The encoder layer expect features in the shape of (length, batch size, d_model).\n", |
|
|
|
"\t\tout = self.encoder_layer(out)\n", |
|
|
|
"\t\t# out: (batch size, length, d_model)\n", |
|
|
|
"\t\tout = out.transpose(0, 1)\n", |
|
|
|
"\t\t# mean pooling\n", |
|
|
|
"\t\tstats = out.mean(dim=1)\n", |
|
|
|
"\n", |
|
|
|
"\t\t# out: (batch, n_spks)\n", |
|
|
|
"\t\tout = self.pred_layer(stats)\n", |
|
|
|
"\t\treturn out" |
|
|
|
] |
|
|
|
}, |
|
|
|
{ |
|
|
|
"cell_type": "markdown", |
|
|
|
"metadata": { |
|
|
|
"id": "W7yX8JinM5Ly" |
|
|
|
}, |
|
|
|
"source": [ |
|
|
|
"# Learning rate schedule\n", |
|
|
|
"- For transformer architecture, the design of learning rate schedule is different from that of CNN.\n", |
|
|
|
"- Previous works show that the warmup of learning rate is useful for training models with transformer architectures.\n", |
|
|
|
"- The warmup schedule\n", |
|
|
|
" - Set learning rate to 0 in the beginning.\n", |
|
|
|
" - The learning rate increases linearly from 0 to initial learning rate during warmup period." |
|
|
|
] |
|
|
|
}, |
|
|
|
{ |
|
|
|
"cell_type": "code", |
|
|
|
"execution_count": null, |
|
|
|
"metadata": { |
|
|
|
"id": "ykt0N1nVJJi2" |
|
|
|
}, |
|
|
|
"outputs": [], |
|
|
|
"source": [ |
|
|
|
"import math\n", |
|
|
|
"\n", |
|
|
|
"import torch\n", |
|
|
|
"from torch.optim import Optimizer\n", |
|
|
|
"from torch.optim.lr_scheduler import LambdaLR\n", |
|
|
|
"\n", |
|
|
|
"\n", |
|
|
|
"def get_cosine_schedule_with_warmup(\n", |
|
|
|
"\toptimizer: Optimizer,\n", |
|
|
|
"\tnum_warmup_steps: int,\n", |
|
|
|
"\tnum_training_steps: int,\n", |
|
|
|
"\tnum_cycles: float = 0.5,\n", |
|
|
|
"\tlast_epoch: int = -1,\n", |
|
|
|
"):\n", |
|
|
|
"\t\"\"\"\n", |
|
|
|
"\tCreate a schedule with a learning rate that decreases following the values of the cosine function between the\n", |
|
|
|
"\tinitial lr set in the optimizer to 0, after a warmup period during which it increases linearly between 0 and the\n", |
|
|
|
"\tinitial lr set in the optimizer.\n", |
|
|
|
"\n", |
|
|
|
"\tArgs:\n", |
|
|
|
"\t\toptimizer (:class:`~torch.optim.Optimizer`):\n", |
|
|
|
"\t\tThe optimizer for which to schedule the learning rate.\n", |
|
|
|
"\t\tnum_warmup_steps (:obj:`int`):\n", |
|
|
|
"\t\tThe number of steps for the warmup phase.\n", |
|
|
|
"\t\tnum_training_steps (:obj:`int`):\n", |
|
|
|
"\t\tThe total number of training steps.\n", |
|
|
|
"\t\tnum_cycles (:obj:`float`, `optional`, defaults to 0.5):\n", |
|
|
|
"\t\tThe number of waves in the cosine schedule (the defaults is to just decrease from the max value to 0\n", |
|
|
|
"\t\tfollowing a half-cosine).\n", |
|
|
|
"\t\tlast_epoch (:obj:`int`, `optional`, defaults to -1):\n", |
|
|
|
"\t\tThe index of the last epoch when resuming training.\n", |
|
|
|
"\n", |
|
|
|
"\tReturn:\n", |
|
|
|
"\t\t:obj:`torch.optim.lr_scheduler.LambdaLR` with the appropriate schedule.\n", |
|
|
|
"\t\"\"\"\n", |
|
|
|
"\tdef lr_lambda(current_step):\n", |
|
|
|
"\t\t# Warmup\n", |
|
|
|
"\t\tif current_step < num_warmup_steps:\n", |
|
|
|
"\t\t\treturn float(current_step) / float(max(1, num_warmup_steps))\n", |
|
|
|
"\t\t# decadence\n", |
|
|
|
"\t\tprogress = float(current_step - num_warmup_steps) / float(\n", |
|
|
|
"\t\t\tmax(1, num_training_steps - num_warmup_steps)\n", |
|
|
|
"\t\t)\n", |
|
|
|
"\t\treturn max(\n", |
|
|
|
"\t\t\t0.0, 0.5 * (1.0 + math.cos(math.pi * float(num_cycles) * 2.0 * progress))\n", |
|
|
|
"\t\t)\n", |
|
|
|
"\n", |
|
|
|
"\treturn LambdaLR(optimizer, lr_lambda, last_epoch)" |
|
|
|
] |
|
|
|
}, |
|
|
|
{ |
|
|
|
"cell_type": "markdown", |
|
|
|
"metadata": { |
|
|
|
"id": "-LN2XkteM_uH" |
|
|
|
}, |
|
|
|
"source": [ |
|
|
|
"# Model Function\n", |
|
|
|
"- Model forward function." |
|
|
|
] |
|
|
|
}, |
|
|
|
{ |
|
|
|
"cell_type": "code", |
|
|
|
"execution_count": null, |
|
|
|
"metadata": { |
|
|
|
"id": "N-rr8529JMz0" |
|
|
|
}, |
|
|
|
"outputs": [], |
|
|
|
"source": [ |
|
|
|
"import torch\n", |
|
|
|
"\n", |
|
|
|
"\n", |
|
|
|
"def model_fn(batch, model, criterion, device):\n", |
|
|
|
"\t\"\"\"Forward a batch through the model.\"\"\"\n", |
|
|
|
"\n", |
|
|
|
"\tmels, labels = batch\n", |
|
|
|
"\tmels = mels.to(device)\n", |
|
|
|
"\tlabels = labels.to(device)\n", |
|
|
|
"\n", |
|
|
|
"\touts = model(mels)\n", |
|
|
|
"\n", |
|
|
|
"\tloss = criterion(outs, labels)\n", |
|
|
|
"\n", |
|
|
|
"\t# Get the speaker id with highest probability.\n", |
|
|
|
"\tpreds = outs.argmax(1)\n", |
|
|
|
"\t# Compute accuracy.\n", |
|
|
|
"\taccuracy = torch.mean((preds == labels).float())\n", |
|
|
|
"\n", |
|
|
|
"\treturn loss, accuracy" |
|
|
|
] |
|
|
|
}, |
|
|
|
{ |
|
|
|
"cell_type": "markdown", |
|
|
|
"metadata": { |
|
|
|
"id": "cwM_xyOtNCI2" |
|
|
|
}, |
|
|
|
"source": [ |
|
|
|
"# Validate\n", |
|
|
|
"- Calculate accuracy of the validation set." |
|
|
|
] |
|
|
|
}, |
|
|
|
{ |
|
|
|
"cell_type": "code", |
|
|
|
"execution_count": null, |
|
|
|
"metadata": { |
|
|
|
"id": "YAiv6kpdJRTJ" |
|
|
|
}, |
|
|
|
"outputs": [], |
|
|
|
"source": [ |
|
|
|
"from tqdm import tqdm\n", |
|
|
|
"import torch\n", |
|
|
|
"\n", |
|
|
|
"\n", |
|
|
|
"def valid(dataloader, model, criterion, device): \n", |
|
|
|
"\t\"\"\"Validate on validation set.\"\"\"\n", |
|
|
|
"\n", |
|
|
|
"\tmodel.eval()\n", |
|
|
|
"\trunning_loss = 0.0\n", |
|
|
|
"\trunning_accuracy = 0.0\n", |
|
|
|
"\tpbar = tqdm(total=len(dataloader.dataset), ncols=0, desc=\"Valid\", unit=\" uttr\")\n", |
|
|
|
"\n", |
|
|
|
"\tfor i, batch in enumerate(dataloader):\n", |
|
|
|
"\t\twith torch.no_grad():\n", |
|
|
|
"\t\t\tloss, accuracy = model_fn(batch, model, criterion, device)\n", |
|
|
|
"\t\t\trunning_loss += loss.item()\n", |
|
|
|
"\t\t\trunning_accuracy += accuracy.item()\n", |
|
|
|
"\n", |
|
|
|
"\t\tpbar.update(dataloader.batch_size)\n", |
|
|
|
"\t\tpbar.set_postfix(\n", |
|
|
|
"\t\t\tloss=f\"{running_loss / (i+1):.2f}\",\n", |
|
|
|
"\t\t\taccuracy=f\"{running_accuracy / (i+1):.2f}\",\n", |
|
|
|
"\t\t)\n", |
|
|
|
"\n", |
|
|
|
"\tpbar.close()\n", |
|
|
|
"\tmodel.train()\n", |
|
|
|
"\n", |
|
|
|
"\treturn running_accuracy / len(dataloader)" |
|
|
|
] |
|
|
|
}, |
|
|
|
{ |
|
|
|
"cell_type": "markdown", |
|
|
|
"metadata": { |
|
|
|
"id": "g6ne9G-eNEdG" |
|
|
|
}, |
|
|
|
"source": [ |
|
|
|
"# Main function" |
|
|
|
] |
|
|
|
}, |
|
|
|
{ |
|
|
|
"cell_type": "code", |
|
|
|
"execution_count": null, |
|
|
|
"metadata": { |
|
|
|
"id": "Usv9s-CuJSG7" |
|
|
|
}, |
|
|
|
"outputs": [], |
|
|
|
"source": [ |
|
|
|
"from tqdm import tqdm\n", |
|
|
|
"\n", |
|
|
|
"import torch\n", |
|
|
|
"import torch.nn as nn\n", |
|
|
|
"from torch.optim import AdamW\n", |
|
|
|
"from torch.utils.data import DataLoader, random_split\n", |
|
|
|
"\n", |
|
|
|
"\n", |
|
|
|
"def parse_args():\n", |
|
|
|
"\t\"\"\"arguments\"\"\"\n", |
|
|
|
"\tconfig = {\n", |
|
|
|
"\t\t\"data_dir\": \"./Dataset\",\n", |
|
|
|
"\t\t\"save_path\": \"model.ckpt\",\n", |
|
|
|
"\t\t\"batch_size\": 32,\n", |
|
|
|
"\t\t\"n_workers\": 8,\n", |
|
|
|
"\t\t\"valid_steps\": 2000,\n", |
|
|
|
"\t\t\"warmup_steps\": 1000,\n", |
|
|
|
"\t\t\"save_steps\": 10000,\n", |
|
|
|
"\t\t\"total_steps\": 70000,\n", |
|
|
|
"\t}\n", |
|
|
|
"\n", |
|
|
|
"\treturn config\n", |
|
|
|
"\n", |
|
|
|
"\n", |
|
|
|
"def main(\n", |
|
|
|
"\tdata_dir,\n", |
|
|
|
"\tsave_path,\n", |
|
|
|
"\tbatch_size,\n", |
|
|
|
"\tn_workers,\n", |
|
|
|
"\tvalid_steps,\n", |
|
|
|
"\twarmup_steps,\n", |
|
|
|
"\ttotal_steps,\n", |
|
|
|
"\tsave_steps,\n", |
|
|
|
"):\n", |
|
|
|
"\t\"\"\"Main function.\"\"\"\n", |
|
|
|
"\tdevice = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n", |
|
|
|
"\tprint(f\"[Info]: Use {device} now!\")\n", |
|
|
|
"\n", |
|
|
|
"\ttrain_loader, valid_loader, speaker_num = get_dataloader(data_dir, batch_size, n_workers)\n", |
|
|
|
"\ttrain_iterator = iter(train_loader)\n", |
|
|
|
"\tprint(f\"[Info]: Finish loading data!\",flush = True)\n", |
|
|
|
"\n", |
|
|
|
"\tmodel = Classifier(n_spks=speaker_num).to(device)\n", |
|
|
|
"\tcriterion = nn.CrossEntropyLoss()\n", |
|
|
|
"\toptimizer = AdamW(model.parameters(), lr=1e-3)\n", |
|
|
|
"\tscheduler = get_cosine_schedule_with_warmup(optimizer, warmup_steps, total_steps)\n", |
|
|
|
"\tprint(f\"[Info]: Finish creating model!\",flush = True)\n", |
|
|
|
"\n", |
|
|
|
"\tbest_accuracy = -1.0\n", |
|
|
|
"\tbest_state_dict = None\n", |
|
|
|
"\n", |
|
|
|
"\tpbar = tqdm(total=valid_steps, ncols=0, desc=\"Train\", unit=\" step\")\n", |
|
|
|
"\n", |
|
|
|
"\tfor step in range(total_steps):\n", |
|
|
|
"\t\t# Get data\n", |
|
|
|
"\t\ttry:\n", |
|
|
|
"\t\t\tbatch = next(train_iterator)\n", |
|
|
|
"\t\texcept StopIteration:\n", |
|
|
|
"\t\t\ttrain_iterator = iter(train_loader)\n", |
|
|
|
"\t\t\tbatch = next(train_iterator)\n", |
|
|
|
"\n", |
|
|
|
"\t\tloss, accuracy = model_fn(batch, model, criterion, device)\n", |
|
|
|
"\t\tbatch_loss = loss.item()\n", |
|
|
|
"\t\tbatch_accuracy = accuracy.item()\n", |
|
|
|
"\n", |
|
|
|
"\t\t# Updata model\n", |
|
|
|
"\t\tloss.backward()\n", |
|
|
|
"\t\toptimizer.step()\n", |
|
|
|
"\t\tscheduler.step()\n", |
|
|
|
"\t\toptimizer.zero_grad()\n", |
|
|
|
"\n", |
|
|
|
"\t\t# Log\n", |
|
|
|
"\t\tpbar.update()\n", |
|
|
|
"\t\tpbar.set_postfix(\n", |
|
|
|
"\t\t\tloss=f\"{batch_loss:.2f}\",\n", |
|
|
|
"\t\t\taccuracy=f\"{batch_accuracy:.2f}\",\n", |
|
|
|
"\t\t\tstep=step + 1,\n", |
|
|
|
"\t\t)\n", |
|
|
|
"\n", |
|
|
|
"\t\t# Do validation\n", |
|
|
|
"\t\tif (step + 1) % valid_steps == 0:\n", |
|
|
|
"\t\t\tpbar.close()\n", |
|
|
|
"\n", |
|
|
|
"\t\t\tvalid_accuracy = valid(valid_loader, model, criterion, device)\n", |
|
|
|
"\n", |
|
|
|
"\t\t\t# keep the best model\n", |
|
|
|
"\t\t\tif valid_accuracy > best_accuracy:\n", |
|
|
|
"\t\t\t\tbest_accuracy = valid_accuracy\n", |
|
|
|
"\t\t\t\tbest_state_dict = model.state_dict()\n", |
|
|
|
"\n", |
|
|
|
"\t\t\tpbar = tqdm(total=valid_steps, ncols=0, desc=\"Train\", unit=\" step\")\n", |
|
|
|
"\n", |
|
|
|
"\t\t# Save the best model so far.\n", |
|
|
|
"\t\tif (step + 1) % save_steps == 0 and best_state_dict is not None:\n", |
|
|
|
"\t\t\ttorch.save(best_state_dict, save_path)\n", |
|
|
|
"\t\t\tpbar.write(f\"Step {step + 1}, best model saved. (accuracy={best_accuracy:.4f})\")\n", |
|
|
|
"\n", |
|
|
|
"\tpbar.close()\n", |
|
|
|
"\n", |
|
|
|
"\n", |
|
|
|
"if __name__ == \"__main__\":\n", |
|
|
|
"\tmain(**parse_args())" |
|
|
|
] |
|
|
|
}, |
|
|
|
{ |
|
|
|
"cell_type": "markdown", |
|
|
|
"metadata": { |
|
|
|
"id": "NLatBYAhNNMx" |
|
|
|
}, |
|
|
|
"source": [ |
|
|
|
"# Inference\n", |
|
|
|
"\n", |
|
|
|
"## Dataset of inference" |
|
|
|
] |
|
|
|
}, |
|
|
|
{ |
|
|
|
"cell_type": "code", |
|
|
|
"execution_count": null, |
|
|
|
"metadata": { |
|
|
|
"colab": { |
|
|
|
"background_save": true |
|
|
|
}, |
|
|
|
"id": "efS4pCmAJXJH" |
|
|
|
}, |
|
|
|
"outputs": [], |
|
|
|
"source": [ |
|
|
|
"import os\n", |
|
|
|
"import json\n", |
|
|
|
"import torch\n", |
|
|
|
"from pathlib import Path\n", |
|
|
|
"from torch.utils.data import Dataset\n", |
|
|
|
"\n", |
|
|
|
"\n", |
|
|
|
"class InferenceDataset(Dataset):\n", |
|
|
|
"\tdef __init__(self, data_dir):\n", |
|
|
|
"\t\ttestdata_path = Path(data_dir) / \"testdata.json\"\n", |
|
|
|
"\t\tmetadata = json.load(testdata_path.open())\n", |
|
|
|
"\t\tself.data_dir = data_dir\n", |
|
|
|
"\t\tself.data = metadata[\"utterances\"]\n", |
|
|
|
"\n", |
|
|
|
"\tdef __len__(self):\n", |
|
|
|
"\t\treturn len(self.data)\n", |
|
|
|
"\n", |
|
|
|
"\tdef __getitem__(self, index):\n", |
|
|
|
"\t\tutterance = self.data[index]\n", |
|
|
|
"\t\tfeat_path = utterance[\"feature_path\"]\n", |
|
|
|
"\t\tmel = torch.load(os.path.join(self.data_dir, feat_path))\n", |
|
|
|
"\n", |
|
|
|
"\t\treturn feat_path, mel\n", |
|
|
|
"\n", |
|
|
|
"\n", |
|
|
|
"def inference_collate_batch(batch):\n", |
|
|
|
"\t\"\"\"Collate a batch of data.\"\"\"\n", |
|
|
|
"\tfeat_paths, mels = zip(*batch)\n", |
|
|
|
"\n", |
|
|
|
"\treturn feat_paths, torch.stack(mels)" |
|
|
|
] |
|
|
|
}, |
|
|
|
{ |
|
|
|
"cell_type": "markdown", |
|
|
|
"metadata": { |
|
|
|
"id": "tl0WnYwxNK_S" |
|
|
|
}, |
|
|
|
"source": [ |
|
|
|
"## Main funcrion of Inference" |
|
|
|
] |
|
|
|
}, |
|
|
|
{ |
|
|
|
"cell_type": "code", |
|
|
|
"execution_count": null, |
|
|
|
"metadata": { |
|
|
|
"colab": { |
|
|
|
"background_save": true |
|
|
|
}, |
|
|
|
"id": "i8SAbuXEJb2A" |
|
|
|
}, |
|
|
|
"outputs": [], |
|
|
|
"source": [ |
|
|
|
"import json\n", |
|
|
|
"import csv\n", |
|
|
|
"from pathlib import Path\n", |
|
|
|
"from tqdm.notebook import tqdm\n", |
|
|
|
"\n", |
|
|
|
"import torch\n", |
|
|
|
"from torch.utils.data import DataLoader\n", |
|
|
|
"\n", |
|
|
|
"def parse_args():\n", |
|
|
|
"\t\"\"\"arguments\"\"\"\n", |
|
|
|
"\tconfig = {\n", |
|
|
|
"\t\t\"data_dir\": \"./Dataset\",\n", |
|
|
|
"\t\t\"model_path\": \"./model.ckpt\",\n", |
|
|
|
"\t\t\"output_path\": \"./output.csv\",\n", |
|
|
|
"\t}\n", |
|
|
|
"\n", |
|
|
|
"\treturn config\n", |
|
|
|
"\n", |
|
|
|
"\n", |
|
|
|
"def main(\n", |
|
|
|
"\tdata_dir,\n", |
|
|
|
"\tmodel_path,\n", |
|
|
|
"\toutput_path,\n", |
|
|
|
"):\n", |
|
|
|
"\t\"\"\"Main function.\"\"\"\n", |
|
|
|
"\tdevice = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n", |
|
|
|
"\tprint(f\"[Info]: Use {device} now!\")\n", |
|
|
|
"\n", |
|
|
|
"\tmapping_path = Path(data_dir) / \"mapping.json\"\n", |
|
|
|
"\tmapping = json.load(mapping_path.open())\n", |
|
|
|
"\n", |
|
|
|
"\tdataset = InferenceDataset(data_dir)\n", |
|
|
|
"\tdataloader = DataLoader(\n", |
|
|
|
"\t\tdataset,\n", |
|
|
|
"\t\tbatch_size=1,\n", |
|
|
|
"\t\tshuffle=False,\n", |
|
|
|
"\t\tdrop_last=False,\n", |
|
|
|
"\t\tnum_workers=8,\n", |
|
|
|
"\t\tcollate_fn=inference_collate_batch,\n", |
|
|
|
"\t)\n", |
|
|
|
"\tprint(f\"[Info]: Finish loading data!\",flush = True)\n", |
|
|
|
"\n", |
|
|
|
"\tspeaker_num = len(mapping[\"id2speaker\"])\n", |
|
|
|
"\tmodel = Classifier(n_spks=speaker_num).to(device)\n", |
|
|
|
"\tmodel.load_state_dict(torch.load(model_path))\n", |
|
|
|
"\tmodel.eval()\n", |
|
|
|
"\tprint(f\"[Info]: Finish creating model!\",flush = True)\n", |
|
|
|
"\n", |
|
|
|
"\tresults = [[\"Id\", \"Category\"]]\n", |
|
|
|
"\tfor feat_paths, mels in tqdm(dataloader):\n", |
|
|
|
"\t\twith torch.no_grad():\n", |
|
|
|
"\t\t\tmels = mels.to(device)\n", |
|
|
|
"\t\t\touts = model(mels)\n", |
|
|
|
"\t\t\tpreds = outs.argmax(1).cpu().numpy()\n", |
|
|
|
"\t\t\tfor feat_path, pred in zip(feat_paths, preds):\n", |
|
|
|
"\t\t\t\tresults.append([feat_path, mapping[\"id2speaker\"][str(pred)]])\n", |
|
|
|
"\n", |
|
|
|
"\twith open(output_path, 'w', newline='') as csvfile:\n", |
|
|
|
"\t\twriter = csv.writer(csvfile)\n", |
|
|
|
"\t\twriter.writerows(results)\n", |
|
|
|
"\n", |
|
|
|
"\n", |
|
|
|
"if __name__ == \"__main__\":\n", |
|
|
|
"\tmain(**parse_args())" |
|
|
|
] |
|
|
|
} |
|
|
|
], |
|
|
|
"metadata": { |
|
|
|
"accelerator": "GPU", |
|
|
|
"colab": { |
|
|
|
"collapsed_sections": [], |
|
|
|
"name": "hw04.ipynb", |
|
|
|
"provenance": [] |
|
|
|
}, |
|
|
|
"kernelspec": { |
|
|
|
"display_name": "Python 3", |
|
|
|
"name": "python3" |
|
|
|
}, |
|
|
|
"language_info": { |
|
|
|
"name": "python" |
|
|
|
} |
|
|
|
}, |
|
|
|
"nbformat": 4, |
|
|
|
"nbformat_minor": 0 |
|
|
|
} |