def main(_log, _config):
p = _config
modelname = file2name[p['modelfn']]
mod_model = importlib.import_module('models.%s' % p['modelfn'])
model_cls = getattr(mod_model, modelname)
model_params_raw = {k: v for k, v in p.items() if k in model_cls.params or k == 'modelfn'}
list_of_model_params = get_model_param(model_params_raw)
expids = list()
for model_params in list_of_model_params:
model = model_cls(model_params, rnd_seed=p['seed'])
expid = model.params_to_string(model_params, True)
expids.append(expid)
raw_expid = model.params_to_string(model_params_raw, True)
for train_years in train_test_years:
for i in range(len(train_test_years[train_years])):
test_year, val_year = train_test_years[train_years][i], train_test_years[train_years][1 - i]
train_years='wt09_10'
test_year='wt11'
val_year='wt11'
pred_dirs, val_dirs = list(), list()
for expid in expids:
pred_dir = '%s/train_%s/%s/predict_per_epoch/test_%s/%s' % (p['parentdir'], train_years, p['expname'], test_year, expid)
val_dir = '%s/train_%s/%s/predict_per_epoch/test_%s/%s' % (p['parentdir'], train_years, p['expname'], val_year, expid)
if not os.path.isdir(pred_dir) or not os.path.isdir(val_dir):
warnings.warn('No such dir {0}/{1}'.format(pred_dir, val_dir), RuntimeWarning)
continue
pred_dirs.append(pred_dir)
val_dirs.append(val_dir)
output_file='%s/train_%s/%s/evaluations/statdocpair/%s_v-%s_t-%s/%s'%(p['outdir'], train_years,\
p['expname'], '-'.join(train_years.split('_')), val_year[2:], test_year[2:], raw_expid)
try:
if not os.path.isdir(os.path.dirname(output_file)):
os.makedirs(os.path.dirname(output_file))
except OSError as e:
pass
_log.info('evaluate {0} on {1} based on val {2} \
over docpairs benchmark and output to {3}'.format(expid, test_year, val_year, output_file))
test_qids = year_qids[test_year]
qrelf = get_qrelf(qrelfdir, test_year)
qid_cwid_label = read_qrel(qrelf, test_qids, include_spam=False)
year_pkey_docpairs = create_docpairs(qid_cwid_label, test_qids, qid_year)
best_pred_dir, argmax_epoch, argmax_run, argmax_ndcg, argmax_err = get_epoch_from_val(pred_dirs, val_dirs)
qid_cwid_invrank, _, runid = read_run(os.path.join(best_pred_dir, argmax_run))
pkey_qid_acc = eval_docpair_predaccuracy(qid_cwid_invrank, year_pkey_docpairs, test_year)
sored_data = sorted(pkey_qid_acc)
print("pkey_qid_acc : ",sored_data)
'''
def pred(_log, _config):
p = _config
modelname = file2name[p['modelfn']]
mod_model = importlib.import_module('models.%s' % p['modelfn'])
model_cls = getattr(mod_model, modelname)
model_params = {k: v for k, v in p.items() if k in model_cls.params or k == 'modelfn'}
model = model_cls(model_params, rnd_seed=p['seed'])
expid = model.params_to_string(model_params)
outdir_plot=trunc_dir('%s/train_%s/%s/predict_per_epoch/test_%s' % (p['parentdir'], p['train_years'],
p['expname'], p['test_year']))
outdir_run=trunc_dir('%s/%s'%(outdir_plot, expid))
tmp_dir=trunc_dir(os.path.join(outdir_run,'tmp'))
weight_dir=trunc_dir('%s/train_%s/%s/model_weight/%s' % (p['parentdir'], p['train_years'],p['expname'], expid))
detail_outdir=trunc_dir('%s/train_%s/%s/model_detail/' % (p['parentdir'], p['train_years'], p['expname']))
assert os.path.isdir(weight_dir), "weight_dir " + weight_dir + " does not exist. Make sure you trained the model."
assert os.path.isdir(detail_dir), "detail_dir " + detail_dir + " does not exist. Make sure you trained the model."
if len(os.listdir(weight_dir)) < 1:
raise SoftFailure('weight dir empty')
try:
if not os.path.isdir(outdir_run):
os.makedirs(outdir_run)
os.makedirs(tmp_dir)
except OSError:
pass
_log.info('Processing {0}'.format(outdir_run))
###################
label2tlabel={4:2,3:2,2:2,1:1,0:0,-2:0}
topk4eval=20
NGRAM_NFILTER, N_GRAMS = get_ngram_nfilter(p['winlen'], p['qproximity'], p['maxqlen'], p['xfilters'])
_log.info('process {0} and output to {1}'.format(weight_dir, outdir_run))
_log.info('{0} {1} {2} {3} {4}'.format(p['distill'], 'NGRAM_NFILTER', NGRAM_NFILTER, 'N_GRAMS', N_GRAMS))
# prepare train data
qids = get_train_qids(p['test_year'])
qrelf = get_qrelf(qrelfdir, p['test_year'])
qid_cwid_label = read_qrel(qrelf, qids, include_spam=False)
test_qids =[qid for qid in qids if qid in qid_cwid_label]
_log.info('%s test_num %d '%(p['test_year'], len(test_qids)))
f_ndcg=dict()
f_epochs = set()
# sort weights by time and only use the first weights for each epoch
# (in case there are duplicate weights from a failed/re-run train)
for f in sorted(os.listdir(weight_dir),
key=lambda x: os.path.getctime(os.path.join(weight_dir, x))):
if f.split('.')[-1] != 'h5':
continue
cols = f.split('.')[0].split('_')
if len(cols) == 4:
nb_epoch, loss, n_batch, n_samples = int(cols[0]), int(cols[1]), int(cols[2]), int(cols[3])
if nb_epoch <= p['epochs'] and nb_epoch not in f_epochs:
f_epochs.add(nb_epoch)
f_ndcg[f]=(nb_epoch, loss, n_batch, n_samples)
finished_epochs = {}
for fn in sorted(os.listdir(outdir_run),
key=lambda x: os.path.getctime(os.path.join(outdir_run, x))):
if fn.endswith(".run"):
fields = fn[:-4].split("_") # trim .run
assert len(fields) == 5
epoch, loss = int(fields[0]), int(fields[4])
ndcg, mapv, err = float(fields[1]), float(fields[2]), float(fields[3])
#assert epoch not in finished_epochs
if epoch in finished_epochs:
_log.error("TODO two weights exist for same epoch")
finished_epochs[epoch] = (epoch, err, ndcg, mapv, loss)
_log.info('skipping finished epochs: {0}'.format(finished_epochs))
def model_pred(NGRAM_NFILTER, weight_file, test_data, test_docids, test_qids):
dump_modelplot(model.build(), detail_outdir + 'predplot_' + expid)
model_predict = model.build_from_dump(weight_file)
qid_cwid_pred = pred_label(model_predict, test_data, test_docids, test_qids)
return qid_cwid_pred
test_doc_vec, test_docids, test_qids=load_test_data(qids, rawdoc_mat_dir, qid_cwid_label, N_GRAMS, p)
epoch_err_ndcg_loss=list()
_log.info('start {0} {1} {2}'.format(expid, p['train_years'], p['test_year']))
for f in sorted(f_ndcg, key=lambda x:f_ndcg[x][0]):
nb_epoch, loss, n_batch, n_samples = f_ndcg[f]
if nb_epoch in finished_epochs:
epoch_err_ndcg_loss.append(finished_epochs[nb_epoch])
continue
weight_file = os.path.join(weight_dir, f)
qid_cwid_pred = model_pred(NGRAM_NFILTER, weight_file, test_doc_vec, test_docids, test_qids)
ndcg20, err20, mapv = eval_run(_log, qid_cwid_pred, expid, perlf, treceval, tmp_dir, topk4eval, qrelf)
loss = int(loss)
out_name = '%d_%0.4f_%0.4f_%0.4f_%d.run' % (nb_epoch, ndcg20, mapv, err20, loss)
epoch_err_ndcg_loss.append((nb_epoch, err20, ndcg20, mapv, loss))
print_run(qid_cwid_pred, outdir_run, out_name, expid)
_log.info('finished {0}'.format(f))
_log.info('finish {0} {1} {2}'.format(expid, p['train_years'], p['test_year']))
plot_curve(epoch_err_ndcg_loss, outdir_plot, expid, p)
if max(f_epochs) < p['epochs'] - 3:
raise SoftFailure("prediction finished, but not all epochs are available yet. last epoch found: %s" % max(f_epochs))
def main(_log, _config):
p = _config
modelname = file2name[p['modelfn']]
mod_model = importlib.import_module('models.%s' % p['modelfn'])
# load the model to be employed, say from models/pacrr.py
model_cls = getattr(mod_model, modelname)
model_params = {k: v for k, v in p.items() if k in model_cls.params or k == 'modelfn'}
model = model_cls(model_params, rnd_seed=p['seed'])
# create a expid based on the configured parameters
expid = model.params_to_string(model_params)
# the model files
outdir='%s/train_%s/%s/model_weight/%s'%(p['parentdir'], p['train_years'], p['expname'], expid)
# the plots for the model, the training loss etc..
detail_outdir='%s/train_%s/%s/model_detail/'%(p['parentdir'], p['train_years'], p['expname'])
if not os.path.isdir(detail_outdir + 'outs'):
print(detail_outdir + 'outs')
os.makedirs(detail_outdir + 'outs')
_log.info('Input parameters: {0}'.format(p))
label2tlabel={4:2,3:2,2:2,1:1,0:0,-2:0}
sample_label_prob=dict()
_log.info('{0} {1} {2}'.format(p['expname'], p['train_years'], sample_label_prob))
NGRAM_NFILTER, N_GRAMS = get_ngram_nfilter(p['winlen'], p['qproximity'], p['maxqlen'], p['xfilters'])
_log.info('process and output to %s'%outdir)
_log.info('{0} {1} {2} {3} {4}'.format(p['distill'], 'NGRAM_NFILTER', NGRAM_NFILTER, 'N_GRAMS', N_GRAMS))
if os.path.exists(outdir) and len(os.listdir(outdir)) == p['epochs']:
_log.info("outdir already seems to be full... exiting early")
return
# prepare train data
qids = get_train_qids(p['train_years'])
qrelf = get_qrelf(qrelfdir, p['train_years'])
qid_cwid_label = read_qrel(qrelf, qids, include_spam=False)
train_qids =[qid for qid in qids if qid in qid_cwid_label]
_log.info('%s train_num %d '%(p['train_years'], len(train_qids)))
def plot_curve_loss(epoch_train_loss, outdir, name, plot_id, series):
epochs, losses = zip(*list(enumerate(epoch_train_loss)))
argmin_loss_epoch = np.argmin(epoch_train_loss)
fig = plt.figure()
plt.plot(epochs, losses, 'k:')
plt.ylabel('Training Loss')
plt.tick_params('y')
plt.xlabel('epoches')
plt.title('loss:%d %.3f'%(argmin_loss_epoch, epoch_train_loss[argmin_loss_epoch]))
fig.savefig(trunc_dir(outdir) + trunc_file(name + '_' + plot_id + '.pdf'), format='pdf')
plt.close()
# dump model plot
built_model = model.build()
model.build_predict() # run build_predict to verify it's working
dump_modelplot(built_model, detail_outdir + 'model_' + expid)
# callback function, dump the model and compute ndcg/map
dump_weight = DumpWeight(outdir, batch_size=p['batch'], nb_sample=p['nsamples'])
# keras 2 steps per epoch is number of batches per epoch, not number of samples per epoch
steps_per_epoch = np.int(p['nsamples'] / p['batch'])
# the generator for training data
train_data_generator=\
load_train_data_generator(qids, rawdoc_mat_dir, qid_cwid_label, N_GRAMS, p,\
label2tlabel=label2tlabel, sample_label_prob=sample_label_prob)
history = built_model.fit_generator(train_data_generator, steps_per_epoch=steps_per_epoch, epochs=p['epochs'],
verbose=0, callbacks=[dump_weight], max_q_size=15, workers=1, pickle_safe=False)
epoch_train_loss = history.history['loss']
# plot the training loss for debugging
plot_curve_loss(epoch_train_loss, detail_outdir, 'train_', expid, ['loss'])
historyfile = detail_outdir + 'hist_' + expid + '.history'
with open(detail_outdir + 'hist_' + expid + '.p', 'wb') as handle:
pickle.dump(history.history, handle, protocol=pickle.HIGHEST_PROTOCOL)
def main(_log, _config):
p = _config
modelname = file2name[p['modelfn']]
mod_model = importlib.import_module('models.%s' % p['modelfn'])
model_cls = getattr(mod_model, modelname)
model_params_raw = {
k: v
for k, v in p.items() if k in model_cls.params or k == 'modelfn'
}
list_of_model_params = get_model_param(model_params_raw)
expids = list()
for model_params in list_of_model_params:
model = model_cls(model_params, rnd_seed=p['seed'])
expid = model.params_to_string(model_params, True)
expids.append(expid)
raw_expid = model.params_to_string(model_params_raw, True)
for train_years in train_test_years:
for i in range(len(train_test_years[train_years])):
test_year, val_year = train_test_years[train_years][
i], train_test_years[train_years][1 - i]
pred_dirs, val_dirs = list(), list()
for expid in expids:
pred_dir = '%s/train_%s/%s/predict_per_epoch/test_%s/%s' % (
p['parentdir'], train_years, p['expname'], test_year,
expid)
val_dir = '%s/train_%s/%s/predict_per_epoch/test_%s/%s' % (
p['parentdir'], train_years, p['expname'], val_year, expid)
if not os.path.isdir(pred_dir) or not os.path.isdir(val_dir):
warnings.warn(
'No such dir {0}/{1}'.format(pred_dir, val_dir),
RuntimeWarning)
continue
pred_dirs.append(pred_dir)
val_dirs.append(val_dir)
output_ql='%s/train_%s/%s/evaluations/rerank-ql/%s_v-%s_t-%s/%s'%(p['outdir'], train_years,\
p['expname'], '-'.join(train_years.split('_')), val_year[2:], test_year[2:], raw_expid)
output_rrall='%s/train_%s/%s/evaluations/rerank-all/%s_v-%s_t-%s/%s'%(p['outdir'], train_years,\
p['expname'], '-'.join(train_years.split('_')), val_year[2:], test_year[2:], raw_expid)
reranked_run_dir='%s/train_%s/%s/reranking/trecrun/v%s-t%s_%s'%(p['outdir'],train_years,\
p['expname'],val_year, test_year,raw_expid)
reranked_metric_dir='%s/train_%s/%s/reranking/ndcgerr/v%s-t%s_%s'%(p['outdir'],train_years,\
p['expname'],val_year, test_year,raw_expid)
try:
if not os.path.isdir(reranked_run_dir):
os.makedirs(reranked_run_dir)
except OSError as e:
print(e)
try:
if not os.path.isdir(reranked_metric_dir):
os.makedirs(reranked_metric_dir)
except OSError as e:
print(e)
try:
if not os.path.isdir(os.path.dirname(output_ql)):
os.makedirs(os.path.dirname(output_ql))
except OSError as e:
print(e)
try:
if not os.path.isdir(os.path.dirname(output_rrall)):
os.makedirs(os.path.dirname(output_rrall))
except OSError as e:
print(e)
_log.info('evaluate {0} on {1} based on val {2} \
over docpairs benchmark and output to {3}, {4}'.format(
expid, test_year, val_year, output_ql, output_rrall))
trec_run_dir = '{0}/{1}'.format(trec_run_basedir, test_year)
eval_trecrun_dir = '{0}/{1}'.format(eval_trec_run_basedir,
test_year)
test_qids = year_qids[test_year]
qrelf = get_qrelf(qrelfdir, test_year)
qid_cwid_label = read_qrel(qrelf, test_qids, include_spam=False)
best_pred_dir, argmax_epoch, argmax_run, argmax_ndcg, argmax_err = get_epoch_from_val(
pred_dirs, val_dirs)
# create re-rank
qid_cwid_pred, _, pred_expid = read_run(
os.path.join(best_pred_dir, argmax_run))
sysn_qid_cwid_pred = trec_run_predscore(_log, trec_run_dir,
qid_cwid_pred)
print_rerun(sysn_qid_cwid_pred, reranked_run_dir, pred_expid,
val_year, test_year)
# eval re-rank
for runfile in os.listdir(reranked_run_dir):
outfile = '%s/%s.ndcg_err' % (reranked_metric_dir,
runfile[:-3])
with open(outfile, 'w') as outf:
subprocess.Popen(
[perlf, qrelf,
'%s/%s' % (reranked_run_dir, runfile)],
stdout=outf)
_log.info(
'finished {0} train on {1}, val on {2}, test on {3}'.
format(runfile, train_years, val_year, test_year))
# read in eval and generate results
trecrun_qid_ndcgerr = read_eval_res(eval_trecrun_dir)
while True:
rr_trecrun_qid_ndcgerr = read_eval_res(reranked_metric_dir)
_log.error('mismatched #run {0} != {1}'.format(
len(trecrun_qid_ndcgerr), len(rr_trecrun_qid_ndcgerr)))
if len(trecrun_qid_ndcgerr) == len(rr_trecrun_qid_ndcgerr):
break
# latency for subprocess.Popen
time.sleep(2)
# orig_rank, orig_score, qidscores, rr_rank, rr_score, qidscores
orig_rr_ndcg_rank, orig_rr_err_rank = get_rank(
rr_trecrun_qid_ndcgerr, trecrun_qid_ndcgerr)
if test_year in fold_names:
# query likelihood benchmark
cols = [
'QL-Variants', 'Measures', 'TREC', 'Trec-Rank', 'Rerank',
'Rerank-Rank', 'Comparison', 'p-value'
]
tabledict = dict()
measure_ind = {'ERR': 1, 'nDCG': 0}
for j, col in enumerate(cols):
tabledict[col] = list()
for method in ['cwindri']:
for measure in ['ERR', 'nDCG']:
if j == 0:
tabledict[col].append(method)
elif j == 1:
tabledict[col].append(measure)
# original trec score
elif j == 2:
if measure == 'ERR':
tabledict[col].append(
'%.3f' % orig_rr_err_rank[method][1])
elif measure == 'nDCG':
tabledict[col].append(
'%.3f' % orig_rr_ndcg_rank[method][1])
# original trec rank
elif j == 3:
if measure == 'ERR':
tabledict[col].append(
orig_rr_err_rank[method][0])
elif measure == 'nDCG':
tabledict[col].append(
orig_rr_ndcg_rank[method][0])
# reranked score
elif j == 4:
if measure == 'ERR':
tabledict[col].append(
'%.3f' % orig_rr_err_rank[method][4])
elif measure == 'nDCG':
tabledict[col].append(
'%.3f' % orig_rr_ndcg_rank[method][4])
# reranked rank
elif j == 5:
if measure == 'ERR':
tabledict[col].append(
orig_rr_err_rank[method][3])
elif measure == 'nDCG':
tabledict[col].append(
orig_rr_ndcg_rank[method][3])
# comparison: (r-t)/t * 100 %
elif j == 6:
if measure == 'ERR':
comp = (orig_rr_err_rank[method][4] -
orig_rr_err_rank[method][1]
) / orig_rr_err_rank[method][1]
tabledict[col].append('%.0f%%' %
(comp * 100))
elif measure == 'nDCG':
comp = (orig_rr_ndcg_rank[method][4] -
orig_rr_ndcg_rank[method][1]
) / orig_rr_ndcg_rank[method][1]
tabledict[col].append('%.0f%%' %
(comp * 100))
# comparison: p-value
elif j == 7:
if measure == 'ERR':
_, p_err_diff = ttest_rel(
orig_rr_err_rank[method][2],
orig_rr_err_rank[method][5])
tabledict[col].append('%.3f' %
(p_err_diff))
elif measure == 'nDCG':
_, p_ndcg_diff = ttest_rel(
orig_rr_ndcg_rank[method][2],
orig_rr_ndcg_rank[method][5])
tabledict[col].append('%.3f' %
(p_ndcg_diff))
dftable = pd.DataFrame(tabledict, columns=cols, index=None)
_log.info('\n' + dftable.to_string())
dftable.to_csv(output_ql + '.csv',
float_format='%.3f',
header=True,
index=False,
sep=',',
mode='w')
_log.info('finished ql benchmark {0} {1} {2} {3}'.format(
expid, train_years, val_year, test_year))
# re-rank all benchmark
def comparison(orig_rr_rank):
count = 0
percents = list()
for r in orig_rr_rank:
orig_rank, orig_score, orig_scores, rr_rank, rr_score, rr_scores = orig_rr_rank[
r]
if rr_rank < orig_rank:
count += 1
# compute micro avg
qid_chg = (rr_score - orig_score) / orig_score
percents.append(qid_chg)
return count, np.mean(percents), np.median(percents)
cols = ['Measures', '#Total Runs', '#Improved', 'Avg', 'Median']
orig_rr_ranks = [orig_rr_ndcg_rank, orig_rr_err_rank]
tabledict = list()
for i, measure in enumerate(['nDCG', 'ERR']):
tabledict.append(dict())
count, avg_chg, median_chg = comparison(orig_rr_ranks[i])
for j, col in enumerate(cols):
if j == 0:
tabledict[i][col] = measure
elif j == 1:
tabledict[i][col] = len(orig_rr_ranks[i])
elif j == 2:
tabledict[i][col] = count
elif j == 3:
tabledict[i][col] = '%.0f%%' % (avg_chg * 100)
elif j == 4:
tabledict[i][col] = '%.0f%%' % (median_chg * 100)
dftable = pd.DataFrame(tabledict, columns=cols, index=None)
_log.info('\n' + dftable.to_string())
dftable.to_csv(output_rrall + '.csv',
float_format='%.3f',
header=True,
index=False,
sep=',',
mode='w')
_log.info('finished rerank all benchmark {0} {1} {2} {3}'.format(
expid, train_years, val_year, test_year))
# rank vs percentage of change
def rank_improve(orig_rr_rank):
oscore_percent = list()
for r in orig_rr_rank:
orig_rank, orig_score, orig_scores, rr_rank, rr_score, rr_scores = orig_rr_rank[
r]
percent = (rr_score - orig_score) / orig_score
oscore_percent.append((orig_score, percent))
return [
p for s, p in sorted(
oscore_percent, key=lambda s_p: s_p[0], reverse=True)
]
def plot_curve(ranks, ndcg_ps, err_ps, outfilename):
fig, ax = plt.subplots()
rects1 = ax.scatter(ranks,
ndcg_ps,
s=25,
c='b',
marker="^",
lw=0)
rects2 = ax.scatter(ranks,
err_ps,
s=25,
c='r',
marker="o",
lw=0)
vals = ax.get_yticks()
ax.set_yticklabels(['{:3.0f}%'.format(x * 100) for x in vals])
ax.set_xlabel(
'Rank of runs from TREC sorted by corresponding measures')
ax.set_ylabel('Relative improvement based on Err/nDCG')
ax.legend((rects1, rects2), ('Improvements based on nDCG',
'Improvements based on ERR'))
plt.grid(b=False, linestyle='--')
fig.savefig(outfilename + '.pdf', format='pdf')
plt.close()
ndcg_ps = rank_improve(orig_rr_ndcg_rank)
err_ps = rank_improve(orig_rr_err_rank)
ranks = range(1, len(ndcg_ps) + 1)
plot_curve(ranks, ndcg_ps, err_ps, output_rrall)
def main(_log, _config):
p = _config
modelname = file2name[p['modelfn']]
mod_model = importlib.import_module('models.%s' % p['modelfn'])
model_cls = getattr(mod_model, modelname)
model_params_raw = {
k: v
for k, v in p.items() if k in model_cls.params or k == 'modelfn'
}
list_of_model_params = get_model_param(model_params_raw)
expids = list()
for model_params in list_of_model_params:
model = model_cls(model_params, rnd_seed=p['seed'])
expid = model.params_to_string(model_params, True)
expids.append(expid)
raw_expid = model.params_to_string(model_params_raw, True)
for train_years in train_test_years:
for i in range(len(train_test_years[train_years])):
test_year, val_year = train_test_years[train_years][
i], train_test_years[train_years][1 - i]
pred_dirs, val_dirs = list(), list()
for expid in expids:
default_dir = lambda year: '%s/train_%s/%s/predict_per_epoch/test_%s/%s' % \
(p['parentdir'], train_years, p['expname'], year, expid)
pred_dir = default_dir(test_year)
while not os.path.isdir(pred_dir):
print pred_dir, 'is not a valid pred_dir'
pred_dir = raw_input('Enter new pred_dir: ')
val_dir = default_dir(val_year)
while not os.path.isdir(val_dir):
print val_dir, 'is not a valid val_dir'
val_dir = raw_input('Enter new val_dir: ')
pred_dirs.append(pred_dir)
val_dirs.append(val_dir)
# we want to avoid making the output file too long, so we truncate it to the PATH_MAX
output_file_long='%s/train_%s/%s/evaluations/statdocpair/%s_v-%s_t-%s/%s'%(p['outdir'], train_years,\
p['expname'], '-'.join(train_years.split('_')), val_year[2:], test_year[2:], raw_expid)
output_file = output_file_long[:PATH_MAX]
if not os.path.isdir(os.path.dirname(output_file)):
os.makedirs(os.path.dirname(output_file))
_log.info('evaluate {0} on {1} based on val {2} \
over docpairs benchmark and output to {3}'.format(
expid, test_year, val_year, output_file))
test_qids = year_qids[test_year]
qrelf = get_qrelf(qrelfdir, test_year)
qid_cwid_label = read_qrel(qrelf, test_qids, include_spam=False)
year_pkey_docpairs = create_docpairs(qid_cwid_label, test_qids,
qid_year)
best_pred_dir, argmax_epoch, argmax_run, argmax_ndcg, argmax_err = get_epoch_from_val(
pred_dirs, val_dirs)
qid_cwid_invrank, _, runid = read_run(
os.path.join(best_pred_dir, argmax_run))
pkey_qid_acc = eval_docpair_predaccuracy(qid_cwid_invrank,
year_pkey_docpairs,
test_year)
dftable = df(pkey_qid_acc,
index=sorted(list(qid_cwid_invrank.keys())) + [0, -1])
_log.info('\n' + dftable.to_string())
dftable.to_csv(output_file + '.csv',
float_format='%.3f',
header=True,
index=True,
sep=',',
mode='w')
_log.info('finished {0} {1} {2} {3}'.format(
expid, train_years, val_year, test_year))