-
Notifications
You must be signed in to change notification settings - Fork 3
/
api.py
222 lines (177 loc) · 7.73 KB
/
api.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
from flask import Flask, request, send_from_directory
import os
import argparse
from tqdm import tqdm
from flask import jsonify
import csv
import numpy
from bling_fire_tokenizer import BlingFireTokenizer
import yaml
app = Flask(__name__, static_url_path='')
app.config['SEND_FILE_MAX_AGE_DEFAULT'] = 0
#
# data loading & prep
#
with open(os.environ.get("RUN_CONFIG"), 'r') as ymlfile:
yaml_cfg = yaml.load(ymlfile)
runs = yaml_cfg["runs"]
max_doc_char_length = 100_000
def load_qrels(path):
with open(path,'r') as f:
qids_to_relevant_passageids = {}
for l in f:
try:
l = l.strip().split()
qid = l[0]
if l[3] != "0":
if qid not in qids_to_relevant_passageids:
qids_to_relevant_passageids[qid] = []
qids_to_relevant_passageids[qid].append(l[2].strip())
except:
raise IOError('\"%s\" is not valid format' % l)
return qids_to_relevant_passageids
qrels = []
clusters = []
collection = []
queries = []
queries_with_stats = []
secondary_model = []
secondary_qd = []
collection_cache = {}
queries_cache = {}
for run in runs:
qrels.append(load_qrels(run["qrels"]))
with open(run["cluster-stats"],"r") as csv_file:
cluster_csv = csv.DictReader(csv_file)
_clusters = {}
for row in cluster_csv:
_clusters[row["cluster"]] = dict(row)
_clusters[row["cluster"]]["queries"] = []
with open(run["queries"],"r") as csv_file:
query_csv = csv.DictReader(csv_file)
_queries = {}
_queries_with_stats = {}
for row in query_csv:
_clusters[row["cluster"]]["queries"].append(dict(row))
_queries[row["qid"]] = row["text"]
_queries_with_stats[row["qid"]] = dict(row)
queries.append(_queries)
queries_with_stats.append(_queries_with_stats)
clusters.append(_clusters)
if run["collection"] in collection_cache:
collection.append(collection_cache[run["collection"]])
else:
_collection = {} # int id -> full line dictionary
with open(run["collection"],"r",encoding="utf8") as collection_file:
for line in tqdm(collection_file):
ls = line.split("\t") # id<\t>text ....
_id = ls[0]
_collection[_id] = ls[1].rstrip()[:max_doc_char_length]
collection_cache[run["collection"]]= _collection
collection.append(_collection)
secondary = numpy.load(run["secondary-output"])
secondary_model.append(secondary.get("model_data")[()])
secondary_qd.append(secondary.get("qd_data")[()])
if run["run-info"]["score_type"]=="tk":
run["run-info"]["model_weights_log_len_mix"] = secondary.get("model_data")[()]["dense_comb_weight"][0].tolist()
import gc
gc.collect()
#
# api endpoints
#
@app.route('/dist/<path:path>')
def send_dist(path):
return send_from_directory('dist', path)
@app.route("/")
def main():
return send_from_directory('', 'index.html')
@app.route("/run-info")
def run_info():
return jsonify(runs=[r["run-info"] for r in runs])
@app.route("/evaluated-queries/<run>")
def all_queries(run):
return jsonify(clusters=clusters[int(run)])
@app.route("/query/<run>/<qid>")
def query(qid,run):
run = int(run)
documents = []
for doc in secondary_qd[run][qid]:
documents.append(get_document_info(runs[run]["run-info"]["score_type"],qid,doc,secondary_qd[run][qid][doc],run))
return jsonify(documents=documents)
#
# helper methods
#
tokenizer = BlingFireTokenizer()
def analyze_weighted_param_1D(name,values, param_weight,bias=None,last_x=5):
#print(name, ": value * weight + bias")
rolling_sum = 0
rolling_sum_after_x = 0
kernels = {}
after_x = len(values) - last_x
for i,val in enumerate(values):
param = param_weight[i]
if i < after_x:
kernels[i] = (float(val),float(param))
#print("["+str(i)+"]", str(val) + " * "+str(param) + " = "+ str(val*param))
rolling_sum += val*param
if i >= after_x:
rolling_sum_after_x += val*param
#if bias != None:
#print("Sum:",rolling_sum + bias)
#print("Sum(>="+str(after_x)+")",rolling_sum_after_x + bias)
#else:
#print("Sum:",rolling_sum)
#print("Sum(>="+str(after_x)+")",rolling_sum_after_x)
#print("-----------")
if bias != None:
rolling_sum = rolling_sum + bias
rolling_sum_after_x = rolling_sum_after_x + bias
return (kernels, float(rolling_sum),float(rolling_sum_after_x))
def get_document_info(score_type,qid,did,secondary_info,run):
document_info = {"id":float(did),"score":float(secondary_info["score"]),"judged_relevant": did in qrels[run][qid]}
if score_type == "tk":
document_info["val_log"] = analyze_weighted_param_1D("log-kernels",secondary_info["per_kernel"],secondary_model[run]["dense_weight"][0],last_x=runs[run]["run-info"]["rest-kernels-last"])
document_info["val_len"] = analyze_weighted_param_1D("len-norm-kernels",secondary_info["per_kernel_mean"],secondary_model[run]["dense_mean_weight"][0],last_x=runs[run]["run-info"]["rest-kernels-last"])
if score_type == "knrm":
document_info["val_log"] = analyze_weighted_param_1D("log-kernels",secondary_info["per_kernel"],secondary_model[run]["kernel_weight"][0],last_x=runs[run]["run-info"]["rest-kernels-last"])
document_info["tokenized_query"] = tokenizer.tokenize(queries[run][qid])
document_info["tokenized_document"] = tokenizer.tokenize(collection[run][did])
#matches = []
matches_per_kernel = []
matches_per_kernel_strongest = []
original_mm = numpy.transpose(secondary_info["cosine_matrix_masked"][:len(document_info["tokenized_query"]),:len(document_info["tokenized_document"])]).astype('float64')
kernel_transformed = numpy.exp(- pow(numpy.expand_dims(original_mm,2) - numpy.array(runs[run]["run-info"]["kernels_mus"]), 2) / (2 * pow(0.1, 2)))
kernel_transformed_max_query_per_kernel = numpy.max(kernel_transformed,axis=1)
#for t,token in enumerate(document_info["tokenized_document"]):
# #largest_sim = secondary_info["cosine_matrix_masked"][max_query_id_per_doc[t]][t]
#
# kernel_results = [0]*len(runs["run-info"]["kernels_mus"])
# #matches_per_doc = []
# for i,m in enumerate(runs["run-info"]["kernels_mus"]):
# for q in range(secondary_info["cosine_matrix_masked"].shape[0]):
# kernel_results[i] = float(max(kernel_results[i],(kernel_transformed[q][t][i])))
# #matches_per_doc.append(float(secondary_info["cosine_matrix_masked"][q][t]))
#
# #matches.append(matches_per_doc)
# matches_per_kernel.append(kernel_results)
#
# strongest_kernel = numpy.argmax(numpy.array(kernel_results),axis=0).tolist()
# matches_per_kernel_strongest.append(strongest_kernel)
#print(secondary_info["cosine_matrix_masked"].dtype)
#print(original_mm.dtype)
#print(kernel_transformed.shape)
#print(kernel_transformed.dtype)
#print(original_mm)
#print(numpy.around(original_mm,3).dtype)
#print(numpy.around(original_mm,3).tolist())
#print(numpy.around(kernel_transformed,3).dtype)
document_info["matches"] = numpy.around(original_mm,3).tolist()
document_info["matches_per_kernel"] = numpy.around(kernel_transformed,3).tolist()
document_info["matches_per_kernel_max"] = numpy.around(kernel_transformed_max_query_per_kernel,3).tolist()
#for q in range(len(document_info["tokenized_query"])):
# mq = []
# for d in range(len(document_info["tokenized_document"])):
# mq.append(float(secondary_info["cosine_matrix_masked"][q][d]))
# matches.append(mq)
#document_info["matches"] = matches
return document_info