def __init__(self,
rawdata_dir,
max_docs_per_item=50,
gold_sterr=0.5,
n_items=0):
self.df = pd.read_csv(rawdata_dir,
sep=" ",
error_bad_lines=False,
names=["topic_item", "na", "doc", "gold"])
self.df = self.df[self.df.na == 0]
self.df = self.trunc_docs(max_docs_per_item)
topics = self.df.topic_item.unique()
n_items = max(n_items, len(topics))
extradfs = []
for i in range(n_items - len(topics)):
topic = np.random.choice(topics)
tdf = self.df[self.df.topic_item == topic]
new_topic = topic * (10 + i)
newdf = tdf.copy(deep=True)
newdf.topic_item = new_topic
newdf.gold = np.random.permutation(newdf.gold.values)
extradfs.append(newdf)
if len(extradfs):
self.df = pd.concat([self.df] + extradfs)
self.df.gold = self.df.gold + np.random.normal(0, gold_sterr,
len(self.df.gold))
self.topic_lookup = utils.make_categorical(self.df, "topic_item")
self.doc_lookup = utils.make_categorical(self.df, "doc")
def rank_docs(data):
return data.sort_values("gold", ascending=False).doc.values
self.gold = self.df.groupby("topic_item").apply(rank_docs)
def setup(self,
annodf,
golddf,
c_anno_uid=None,
c_anno_item=None,
c_anno_label=None,
c_gold_item=None,
c_gold_label=None):
renamey = lambda y: self.label_colname if "label" in y else self.item_colname if "item" in y else self.uid_colname if "uid" in y else "_"
localargs = locals()
colrename = {
localargs[k]: renamey(k)
for k in localargs if "c_" in k and localargs[k] is not None
}
self.annodf = annodf[[
c_anno_uid or self.uid_colname, c_anno_item or self.item_colname,
c_anno_label or self.label_colname
]]
self.annodf = self.annodf.dropna().copy().rename(columns=colrename)[[
self.uid_colname, self.item_colname, self.label_colname
]]
uiddict = utils.make_categorical(self.annodf, "uid")
itemdict = utils.make_categorical(self.annodf, "item")
golddf = golddf[[
c_gold_item or self.item_colname, c_gold_label
or self.label_colname
]]
golddf = golddf.rename(columns=colrename)[[
self.item_colname, self.label_colname
]]
golddf = utils.translate_categorical(golddf, self.item_colname,
itemdict)
self.golddict = golddf.set_index("item").to_dict()[self.label_colname]
self.produce_stan_data()
def setup(self):
userIDs = []
itemIDs = []
labels = []
golds = []
hmmcrowds = []
majorityvotes = []
for row in self.rawdf.iterrows():
itemID = row[1]["docid"]
data = row[1]["Participants"]
gold = self.golddf[self.golddf["docid"] ==
itemID]["Participants"].values[0]
gold = gold.get("MedicalStudent")
if gold is None:
continue
agg = self.aggdf[self.aggdf["docid"] ==
itemID]["Participants"].values[0]
for userID, label in data.items():
userIDs.append(userID)
itemIDs.append(itemID)
labels.append(label2tvr(label, default=[]))
golds.append(label2tvr(gold))
hmmcrowds.append(agg["HMMCrowd"])
majorityvotes.append(agg["MajorityVote"])
df = pd.DataFrame({
"uid": userIDs,
"itemID": itemIDs,
"label": labels,
"gold": golds,
"HMMCrowd": hmmcrowds,
"MajorityVote": majorityvotes
})
df = df.sort_values("itemID")
userIdDict = utils.make_categorical(df, "uid")
itemIdDict = utils.make_categorical(df, "itemID")
anno_df = df.copy()
super().setup(anno_df, anno_df, c_gold_label="gold")
mv_labels = {
k: label2tvr(v)
for k, v in dict(
df.groupby("itemID").first()["MajorityVote"].dropna()).items()
}
hmm_labels = {
k: label2tvr(v)
for k, v in dict(
df.groupby("itemID").first()["HMMCrowd"].dropna()).items()
}
self.register_baseline("Tokenwise MV", mv_labels)
self.register_baseline("Crowd-HMM", hmm_labels)
def setup(self, annodf, golddf=None, c_anno_uid=None, c_anno_item=None, c_anno_label=None, c_gold_item=None, c_gold_label=None, merge_index=None):
renamey = lambda y: self.label_colname if "label" in y else self.item_colname if "item" in y else self.uid_colname if "uid" in y else y
localargs = locals()
colrename = {localargs[k]:renamey(k) for k in localargs if "c_" in k and localargs[k] is not None}
self.annodf = annodf[[c_anno_uid or self.uid_colname, c_anno_item or self.item_colname, c_anno_label or self.label_colname]]
self.annodf = self.annodf.rename(columns=colrename)[[self.uid_colname, self.item_colname, self.label_colname]]
if merge_index is not None:
self.merge_index_colname = merge_index
self.annodf[merge_index] = annodf[merge_index]
self.annodf = self.annodf.dropna().copy()
self.uiddict = utils.make_categorical(self.annodf, self.uid_colname)
self.itemdict = utils.make_categorical(self.annodf, self.item_colname)
if golddf is not None:
golddf = golddf[[c_gold_item or self.item_colname, c_gold_label or self.label_colname]]
golddf = golddf.rename(columns=colrename)[[self.item_colname, self.label_colname]]
golddf = utils.translate_categorical(golddf, self.item_colname, self.itemdict)
self.golddict = golddf.set_index(self.item_colname).to_dict()[self.label_colname]
self.golddict = {k: v for k, v in self.golddict.items() if v is not None}
self.produce_stan_data()
def __init__(self,
rawdata_dir='data/coco/person_keypoints_train2017.json',
max_items=500,
minlabelsperitem=4):
with open(rawdata_dir) as f:
dataset = json.load(f)
self.category_id_skeletons = {
c["id"]: np.array(c["skeleton"]) - 1
for c in iter(dataset["categories"])
}
img_label = {}
for dataset_annotation in iter(dataset["annotations"]):
v = img_label.setdefault(dataset_annotation["image_id"], [])
v.append(dataset_annotation)
img_label_minlen = {
k: v
for k, v in img_label.items() if len(v) >= minlabelsperitem
}
i = 0
rows = []
item = []
annotation = []
category = []
for dataset_annotations in iter(img_label_minlen.values()):
for dataset_annotation in dataset_annotations:
kp = np.reshape(dataset_annotation["keypoints"], (-1, 3))
kp = kp[kp[:, 2] > -90][:, :2]
if len(kp) == 0:
continue
item.append(dataset_annotation["image_id"])
annotation.append(kp)
category.append(dataset_annotation["category_id"])
i += 1
if i > max_items:
break
kp_df = pd.DataFrame({
"item": item,
"gold": annotation,
"category": category
})
self.df = kp_df.groupby("item")["gold"].apply(list).reset_index()
self.itemdict = utils.make_categorical(self.df, "item")
def __init__(self, rawdata_dir, max_items=10000):
self.df = pd.read_csv(rawdata_dir, error_bad_lines=False, header=None, sep=" ", names=["img", "x", "y", "w", "h"])
self.df = self.df[:max_items]
self.df["goldcoords"] = self.df.apply(lambda row: [row["x"], row["y"], row["x"] + row["w"], row["y"] + row["h"]], axis=1)
self.img_lookup = utils.make_categorical(self.df, "img")
self.gold = self.df.set_index("img")["goldcoords"]