Yak-hbdx's picture
uploaded TransfoRNA repo
0b11a42 verified
raw
history blame
24.7 kB
import logging
import math
import os
import random
from pathlib import Path
from random import randint
import numpy as np
import pandas as pd
import torch
from hydra._internal.utils import _locate
from hydra.utils import instantiate
from omegaconf import DictConfig, OmegaConf
from scipy.stats import entropy
from sklearn.model_selection import train_test_split
from sklearn.utils.class_weight import (compute_class_weight,
compute_sample_weight)
from skorch.dataset import Dataset
from skorch.helper import predefined_split
from ..callbacks.metrics import get_callbacks
from ..score.score import infer_from_model
from .energy import *
from .file import load
logger = logging.getLogger(__name__)
def update_config_with_inference_params(config:DictConfig,mc_or_sc:str='sub_class',trained_on:str = 'id',path_to_models:str = 'models/tcga/') -> DictConfig:
inference_config = config.copy()
model = config['model_name']
model = "-".join([word.capitalize() for word in model.split("-")])
transforna_folder = "TransfoRNA_ID"
if trained_on == "full":
transforna_folder = "TransfoRNA_FULL"
inference_config['inference_settings']["model_path"] = f'{path_to_models}{transforna_folder}/{mc_or_sc}/{model}/ckpt/model_params_tcga.pt'
inference_config["inference"] = True
inference_config["log_logits"] = False
inference_config = DictConfig(inference_config)
#train and model config should be fetched from teh inference model
train_cfg_path = get_hp_setting(inference_config, "train_config")
model_cfg_path = get_hp_setting(inference_config, "model_config")
train_config = instantiate(train_cfg_path)
model_config = instantiate(model_cfg_path)
# prepare configs as structured dicts
train_config = OmegaConf.structured(train_config)
model_config = OmegaConf.structured(model_config)
# update model config with the name of the model
model_config["model_input"] = inference_config["model_name"]
inference_config = OmegaConf.merge({"train_config": train_config, "model_config": model_config}, inference_config)
return inference_config
def update_config_with_dataset_params_benchmark(train_data_df,configs):
'''
After tokenizing the dataset, some features in the config needs to be updated as they will be used
later by sub modules
'''
# set feedforward input dimension and vocab size
#ss_tokens_id and tokens_id are the same
configs["model_config"].second_input_token_len = train_data_df["second_input"].shape[1]
configs["model_config"].tokens_len = train_data_df["tokens_id"].shape[1]
#set batch per epoch (number of batches). This will be used later by both the criterion and the LR
configs["train_config"].batch_per_epoch = train_data_df["tokens_id"].shape[0]/configs["train_config"].batch_size
return
def update_config_with_dataset_params_tcga(dataset_class,all_data_df,configs):
configs["model_config"].ff_input_dim = all_data_df['second_input'].shape[1]
configs["model_config"].vocab_size = len(dataset_class.seq_tokens_ids_dict.keys())
configs["model_config"].second_input_vocab_size = len(dataset_class.second_input_tokens_ids_dict.keys())
configs["model_config"].tokens_len = dataset_class.tokens_len
configs["model_config"].second_input_token_len = dataset_class.tokens_len
if configs["model_name"] == "seq-seq":
configs["model_config"].tokens_len = math.ceil(dataset_class.tokens_len/2)
configs["model_config"].second_input_token_len = math.ceil(dataset_class.tokens_len/2)
def update_dataclass_inference(cfg,dataset_class):
seq_token_dict,ss_token_dict = get_tokenization_dicts(cfg)
dataset_class.seq_tokens_ids_dict = seq_token_dict
dataset_class.second_input_tokens_ids_dict = ss_token_dict
dataset_class.tokens_len =cfg["model_config"].tokens_len
dataset_class.max_length = get_hp_setting(cfg,'max_length')
dataset_class.min_length = get_hp_setting(cfg,'min_length')
return dataset_class
def set_seed_and_device(seed:int = 0,device_no:int=0):
# set seed
torch.backends.cudnn.deterministic = True
random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
np.random.seed(seed)
torch.cuda.set_device(device_no)
#CUDA_LAUNCH_BLOCKING=1 #for debugging
def sync_skorch_with_config(skorch_cfg: DictConfig,cfg:DictConfig):
'''
skorch config contains duplicate params to the train and model configs
values of skorch config should be populated by those in the trian and
model config
'''
#populate skorch params with params in train or model config if exists
for key in skorch_cfg:
if key in cfg["train_config"]:
skorch_cfg[key] = cfg["train_config"][key]
if key in cfg["model_config"]:
skorch_cfg[key] = cfg["model_config"][key]
return
def instantiate_predictor(skorch_cfg: DictConfig,cfg:DictConfig,path: str=None):
# convert config to omegaconf container
predictor_config = OmegaConf.to_container(skorch_cfg)
# Patch model device argument from the run config:
if "device" in predictor_config:
predictor_config["device"] = skorch_cfg["device"]
for key, val in predictor_config.items():
try:
predictor_config[key] = _locate(val)
except:
continue
#add callbacks to list of params
predictor_config["callbacks"] = get_callbacks(path,cfg)
#save callbacks as instantiate changes the lrcallback from tuple to list,
#then skorch's instantiate_callback throws an error
callbacks_list = predictor_config["callbacks"]
predictor_config["callbacks"] = "disable"
#remove model from the cfg otherwise intantiate will throw an error as
#models' scoring doesnt recieve input params
predictor_config["module__main_config"] = \
{key:cfg[key] for key in cfg if key not in ["model"]}
#in case of tcga task, remove dataset at it its already instantiated
if 'dataset' in predictor_config['module__main_config']:
del predictor_config['module__main_config']['dataset']
#set train split to false in skorch model
if not cfg['train_split']:
predictor_config['train_split'] = False
net = instantiate(predictor_config)
#restore callback and instantiate it
net.callbacks = callbacks_list
net.task = cfg['task']
net.initialize_callbacks()
#prevents double initialization
net.initialized_=True
return net
def get_fused_seqs(seqs,num_sequences:int=1,max_len:int=30):
'''
fuse num_sequences sequences from seqs
'''
fused_seqs = []
while len(fused_seqs) < num_sequences:
#get two random sequences
seq1 = random.choice(seqs)[:max_len]
seq2 = random.choice(seqs)[:max_len]
#select indeex to tuncate seq1 at between 60 to 70% of its length
idx = random.randint(math.floor(len(seq1)*0.3),math.floor(len(seq1)*0.7))
len_to_be_added_from_seq2 = len(seq1)-idx
#truncate seq1 at idx
seq1 = seq1[:idx]
#get the rest from the beg of seq2
seq2 = seq2[:len_to_be_added_from_seq2]
#fuse seq1 and seq2
fused_seq = seq1+seq2
if fused_seq not in fused_seqs and fused_seq not in seqs:
fused_seqs.append(fused_seq)
return fused_seqs
def revert_seq_tokenization(tokenized_seqs,configs):
window = configs["model_config"].window
if configs["model_config"].tokenizer != "overlap":
logger.error("Sequences are not reverse tokenized")
return tokenized_seqs
#currently only overlap tokenizer can be reverted
seqs_concat = []
for seq in tokenized_seqs.values:
seqs_concat.append(''.join(seq[seq!='pad'])[::window]+seq[seq!='pad'][-1][window-1])
return pd.DataFrame(seqs_concat,columns=["Sequences"])
def introduce_mismatches(seq, n_mismatches):
seq = list(seq)
for i in range(n_mismatches):
rand_nt = randint(0,len(seq)-1)
seq[rand_nt] = ['A','G','C','T'][randint(0,3)]
return ''.join(seq)
def prepare_split(split_data_df,configs):
'''
This function returns tokens, token ids and labels for a given dataframes' split.
It also moves tokens and labels to device
'''
model_input_cols = ['tokens_id','second_input','seqs_length']
#token_ids
split_data = torch.tensor(
np.array(split_data_df[model_input_cols].values, dtype=float),
dtype=torch.float,
)
split_weights = torch.tensor(compute_sample_weight('balanced',split_data_df['Labels']))
split_data = torch.cat([split_data,split_weights[:,None]],dim=1)
#tokens (chars)
split_rna_seq = revert_seq_tokenization(split_data_df["tokens"],configs)
#labels
split_labels = torch.tensor(
np.array(split_data_df["Labels"], dtype=int),
dtype=torch.long,
)
return split_data, split_rna_seq, split_labels
def prepare_model_inference(cfg,path):
# instantiate skorch model
net = instantiate_predictor(cfg["model"]["skorch_model"], cfg,path)
net.initialize()
logger.info(f"Model loaded from {cfg['inference_settings']['model_path']}")
net.load_params(f_params=f'{cfg["inference_settings"]["model_path"]}')
net.labels_mapping_dict = dict(zip(cfg["model_config"].class_mappings,list(np.arange(cfg["model_config"].num_classes))))
#save embeddings
if cfg['log_embedds']:
net.save_embedding=True
net.gene_embedds = []
net.second_input_embedds = []
return net
def prepare_data_benchmark(tokenizer,test_ad, configs):
"""
This function recieves anddata and prepares the anndata in a format suitable for training
It also set default parameters in the config that cannot be known until preprocessing step
is done.
all_data_df is heirarchical pandas dataframe, so can be accessed [AA,AT,..,AC ]
"""
###get tokenized train set
train_data_df = tokenizer.get_tokenized_data()
### update config with data specific params
update_config_with_dataset_params_benchmark(train_data_df,configs)
###tokenize test set
test_data_df = tokenize_set(tokenizer,test_ad.var)
### get tokens(on device), seqs and labels(on device)
train_data, train_rna_seq, train_labels = prepare_split(train_data_df,configs)
test_data, test_rna_seq, test_labels = prepare_split(test_data_df,configs)
class_weights = compute_class_weight(class_weight='balanced',classes=np.unique(train_labels.flatten()),y=train_labels.flatten().numpy())
#omegaconfig does not support float64 as datatype so conversion to str is done
# and reconversion is done in criterion
configs['model_config'].class_weights = [str(x) for x in list(class_weights)]
if configs["train_split"]:
#stratify train to get valid
train_data,valid_data,train_labels,valid_labels = stratify(train_data,train_labels,configs["valid_size"])
valid_ds = Dataset(valid_data,valid_labels)
valid_ds=predefined_split(valid_ds)
else:
valid_ds = None
all_data= {"train_data":train_data,
"valid_ds":valid_ds,
"test_data":test_data,
"train_rna_seq":train_rna_seq,
"test_rna_seq":test_rna_seq,
"train_labels_numeric":train_labels,
"test_labels_numeric":test_labels}
if configs["task"] == "premirna":
generalization_test_set = get_add_test_set(tokenizer,\
dataset_path=configs["train_config"].datset_path_additional_testset)
#get all vocab from both test and train set
configs["model_config"].vocab_size = len(tokenizer.seq_tokens_ids_dict.keys())
configs["model_config"].second_input_vocab_size = len(tokenizer.second_input_tokens_ids_dict.keys())
configs["model_config"].tokens_mapping_dict = tokenizer.seq_tokens_ids_dict
if configs["task"] == "premirna":
generalization_test_data = []
for test_df in generalization_test_set:
#no need to read the labels as they are all one
test_data_extra, _, _ = prepare_split(test_df,configs)
generalization_test_data.append(test_data_extra)
all_data["additional_testset"] = generalization_test_data
#get inference dataset
# if do inference and inference datasert path exists
get_inference_data(configs,tokenizer,all_data)
return all_data
def prepare_inference_results_benchmarck(net,cfg,predicted_labels,logits,all_data):
iterables = [["Sequences"], np.arange(1, dtype=int)]
index = pd.MultiIndex.from_product(iterables, names=["type of data", "indices"])
rna_seqs_df = pd.DataFrame(columns=index, data=np.vstack(all_data["infere_rna_seq"]["Sequences"].values))
iterables = [["Logits"], list(net.labels_mapping_dict.keys())]
index = pd.MultiIndex.from_product(iterables, names=["type of data", "indices"])
logits_df = pd.DataFrame(columns=index, data=np.array(logits))
#add Labels,entropy to df
all_data["infere_rna_seq"]["Labels",'0'] = predicted_labels
all_data["infere_rna_seq"].set_index("Sequences",inplace=True)
#log logits if required
if cfg["log_logits"]:
seq_logits_df = logits_df.join(rna_seqs_df).set_index(("Sequences",0))
all_data["infere_rna_seq"] = all_data["infere_rna_seq"].join(seq_logits_df)
else:
all_data["infere_rna_seq"].columns = ['Labels']
return
def prepare_inference_results_tcga(cfg,predicted_labels,logits,all_data,max_len):
logits_clf = load('/'.join(cfg["inference_settings"]["model_path"].split('/')[:-2])\
+'/analysis/logits_model_coef.yaml')
threshold = round(logits_clf['Threshold'],2)
iterables = [["Sequences"], np.arange(1, dtype=int)]
index = pd.MultiIndex.from_product(iterables, names=["type of data", "indices"])
rna_seqs_df = pd.DataFrame(columns=index, data=np.vstack(all_data["infere_rna_seq"]["Sequences"].values))
iterables = [["Logits"], cfg['model_config'].class_mappings]
index = pd.MultiIndex.from_product(iterables, names=["type of data", "indices"])
logits_df = pd.DataFrame(columns=index, data=np.array(logits))
#add Labels,novelty to df
all_data["infere_rna_seq"]["Net-Label"] = predicted_labels
all_data["infere_rna_seq"]["Is Familiar?"] = entropy(logits,axis=1) <= threshold
all_data["infere_rna_seq"].set_index("Sequences",inplace=True)
#log logits if required
if cfg["log_logits"]:
seq_logits_df = logits_df.join(rna_seqs_df).set_index(("Sequences",0))
all_data["infere_rna_seq"] = all_data["infere_rna_seq"].join(seq_logits_df)
all_data["infere_rna_seq"].index.name = f'Sequences, Max Length={max_len}'
return
def prepare_inference_data(cfg,infer_pd,dataset_class):
#tokenize sequences
infere_data_df = tokenize_set(dataset_class,infer_pd,inference=True)
infere_data,infere_rna_seq,_ = prepare_split(infere_data_df,cfg)
all_data = {}
all_data["infere_data"] = infere_data
all_data["infere_rna_seq"] = infere_rna_seq
return all_data
def get_inference_data(configs,dataset_class,all_data):
if configs["inference"]==True and configs["inference_settings"]["sequences_path"] is not None:
inference_file = configs["inference_settings"]["sequences_path"]
inference_path = Path(__file__).parent.parent.parent.absolute() / f"{inference_file}"
infer_data = load(inference_path)
#check if infer_data has secondary structure
if "Secondary" not in infer_data:
infer_data['Secondary'] = dataset_class.get_secondary_structure(infer_data["Sequences"])
if "Labels" not in infer_data:
infer_data["Labels"] = [0]*len(infer_data["Sequences"].values)
dataset_class.seqs_dot_bracket_labels = infer_data
dataset_class.min_length = 0
dataset_class.limit_seqs_to_range(logger)
infere_data_df = dataset_class.get_tokenized_data(inference=True)
infere_data,infere_rna_seq,_ = prepare_split(infere_data_df,configs)
all_data["infere_data"] = infere_data
all_data["infere_rna_seq"] = infere_rna_seq
def get_add_test_set(dataset_class,dataset_path):
all_added_test_set = []
#get paths of all files in mirbase and mirgene
paths_mirbase = dataset_path+"mirbase/"
files_mirbase = os.listdir(paths_mirbase)
for f_idx,_ in enumerate(files_mirbase):
files_mirbase[f_idx] = paths_mirbase+files_mirbase[f_idx]
paths_mirgene = dataset_path + "mirgene/"
files_mirgene = os.listdir(paths_mirgene)
for f_idx,_ in enumerate(files_mirgene):
files_mirgene[f_idx] = paths_mirgene+files_mirgene[f_idx]
files = files_mirbase+files_mirgene
for f in files:
#tokenize test set
test_pd = load(f)
test_pd = test_pd.drop(columns='Unnamed: 0')
test_pd["Sequences"] = test_pd["Sequences"].astype(object)
test_pd["Secondary"] = test_pd["Secondary"].astype(object)
#convert dataframe to anndata
test_pd["Labels"] = 1
dataset_class.seqs_dot_bracket_labels = test_pd
dataset_class.limit_seqs_to_range()
all_added_test_set.append(dataset_class.get_tokenized_data())
return all_added_test_set
def get_tokenization_dicts(cfg):
tokenization_path='/'.join(cfg['inference_settings']['model_path'].split('/')[:-2])
seq_token_dict = load(tokenization_path+'/seq_tokens_ids_dict')
ss_token_dict = load(tokenization_path+'/second_input_tokens_ids_dict')
return seq_token_dict,ss_token_dict
def get_hp_setting(cfg,hp_param):
model_parent_path=Path('/'.join(cfg['inference_settings']['model_path'].split('/')[:-2]))
hp_settings = load(model_parent_path/'meta/hp_settings.yaml')
#hp_param could be in hp_settings .keyes or in a key of a key
hp_val = hp_settings.get(hp_param)
if not hp_val:
for key in hp_settings.keys():
try:
hp_val = hp_settings[key].get(hp_param)
except:
pass
if hp_val != None:
break
if hp_val == None:
raise ValueError(f"hp_param {hp_param} not found in hp_settings")
return hp_val
def get_model(cfg,path):
cfg["model_config"] = get_hp_setting(cfg,'model_config')
sync_skorch_with_config(cfg["model"]["skorch_model"],cfg)
cfg['model_config']['model_input'] = cfg['model_name']
net = prepare_model_inference(cfg,path)
return cfg,net
def stratify(train_data,train_labels,valid_size):
return train_test_split(train_data, train_labels,
stratify=train_labels,
test_size=valid_size)
def tokenize_set(dataset_class,test_pd,inference:bool=False):
#reassign the sequences to test
dataset_class.seqs_dot_bracket_labels = test_pd
#prevent sequences with len < min lenght from being deleted
dataset_class.limit_seqs_to_range()
return dataset_class.get_tokenized_data(inference)
def add_original_seqs_to_predictions(short_to_long_df,pred_df):
short_to_long_df.set_index('Sequences',inplace=True)
pred_df = pd.merge(pred_df,short_to_long_df[['Trimmed','Original_Sequence']],right_index=True,left_index=True,how='left')
#filter repeated indexes
pred_df = pred_df[~pred_df.index.duplicated(keep='first')]
return pred_df
def add_ss_and_labels(infer_data):
#check if infer_data has secondary structure
if "Secondary" not in infer_data:
infer_data["Secondary"] = fold_sequences(infer_data["Sequences"].tolist())['structure_37'].values
if "Labels" not in infer_data:
infer_data["Labels"] = [0]*len(infer_data["Sequences"].values)
return infer_data
def chunkstring_overlap(string, window):
return (
string[0 + i : window + i] for i in range(0, len(string) - window + 1, 1)
)
def create_short_seqs_from_long(df,max_len):
long_seqs = df['Sequences'][df['Sequences'].str.len()>max_len].values
short_seqs_pd = df[df['Sequences'].str.len()<=max_len]
feature_tokens_gen = list(
chunkstring_overlap(feature, max_len)
for feature in long_seqs
)
original_seqs = []
shortened_seqs = []
for i,feature_tokens in enumerate(feature_tokens_gen):
curr_trimmed_seqs = [feature for feature in feature_tokens]
shortened_seqs.extend(curr_trimmed_seqs)
original_seqs.extend([long_seqs[i]]*len(curr_trimmed_seqs))
short_to_long_dict = dict(zip(shortened_seqs,original_seqs))
shortened_df = pd.DataFrame(data=shortened_seqs,columns=['Sequences'])
df = shortened_df.append(short_seqs_pd).reset_index(drop=True)
#add a column in df to indicate if the sequence was trimmed and another column to indicate the original sequence
df['Trimmed'] = False
df.loc[shortened_df.index,'Trimmed'] = True
df['Original_Sequence'] = df['Sequences']
df.loc[shortened_df.index,'Original_Sequence'] = df.loc[shortened_df.index,'Sequences'].map(short_to_long_dict)
return df
def infer_from_pd(cfg,net,infer_pd,DataClass,attention_flag:bool=False):
try:
max_len = net.module_.transformer_layers.pos_encoder.pe.shape[1]+1
except:
max_len = 30#for baseline models
if cfg['model_name'] == 'seq-seq':
max_len = max_len*2 - 1
if len(infer_pd['Sequences'][infer_pd['Sequences'].str.len()>max_len].values)>0:
infer_pd = create_short_seqs_from_long(infer_pd,max_len)
infer_pd = add_ss_and_labels(infer_pd)
if cfg['model_name'] == 'seq-seq':
cfg['model_config']['tokens_len'] *=2
cfg['model_config']['second_input_token_len'] *=2
#create dataclass to tokenize infer sequences
dataset_class = DataClass(infer_pd,cfg)
#update datasetclass with tokenization dicts and tokens_len
dataset_class = update_dataclass_inference(cfg,dataset_class)
#tokenize sequences
all_data = prepare_inference_data(cfg,infer_pd,dataset_class)
#inference on custom data
predicted_labels,logits,attn_scores_first_list,attn_scores_second_list = infer_from_model(net,all_data["infere_data"])
if attention_flag:
#in case of baseline or seq models
if not attn_scores_second_list:
attn_scores_second_list = attn_scores_first_list
attn_scores_first = np.array(attn_scores_first_list)
seq_lengths = all_data['infere_rna_seq']['Sequences'].str.len().values
#get attention scores for each sequence
attn_scores_list = [attn_scores_first[i,:seq_lengths[i],:seq_lengths[i]].flatten().tolist() for i in range(len(seq_lengths))]
attn_scores_first_df = pd.DataFrame(data = {'attention_first':attn_scores_list})
attn_scores_first_df.index = all_data['infere_rna_seq']['Sequences'].values
attn_scores_second = np.array(attn_scores_second_list)
attn_scores_list = [attn_scores_second[i,:seq_lengths[i],:seq_lengths[i]].flatten().tolist() for i in range(len(seq_lengths))]
attn_scores_second_df = pd.DataFrame(data = {'attention_second':attn_scores_list})
attn_scores_second_df.index = all_data['infere_rna_seq']['Sequences'].values
attn_scores_df = attn_scores_first_df.join(attn_scores_second_df)
attn_scores_df['Secondary'] = infer_pd["Secondary"].values
else:
attn_scores_df = None
gene_embedds_df = None
#net.gene_embedds is a list of tensors. convert them to a numpy array
if cfg['log_embedds']:
gene_embedds = np.vstack(net.gene_embedds)
if cfg['model_name'] not in ['baseline']:
second_input_embedds = np.vstack(net.second_input_embedds)
gene_embedds = np.concatenate((gene_embedds,second_input_embedds),axis=1)
gene_embedds_df = pd.DataFrame(data=gene_embedds)
gene_embedds_df.index = all_data['infere_rna_seq']['Sequences'].values
gene_embedds_df.columns = ['gene_embedds_'+str(i) for i in range(gene_embedds_df.shape[1])]
return predicted_labels,logits,gene_embedds_df,attn_scores_df,all_data,max_len,net,infer_pd
def log_embedds(cfg,net,seqs_df):
gene_embedds = np.vstack(net.gene_embedds)
if not cfg['model_name'] in ['seq','baseline']:
second_input_embedds = np.vstack(net.second_input_embedds)
gene_embedds = np.concatenate((gene_embedds,second_input_embedds),axis=1)
return seqs_df.join(pd.DataFrame(data=gene_embedds))