Open maymay1982 opened 4 years ago
@maymay1982 The input files should be RESULTS_FILE after evaluation, "python model_main.py --eval --eval_output=RESULTS_FILE --model_path=CHECKPOINT_FILE". It should work after your getting those files. "
@braveheartwithlove Thanks for your reply. But when I ran python model_main.py --eval --eval_output=cqcc_result --model_path=./models/model_logical_cqcc_100_32_5e-05/epoch_99.pth, I got such error:
File "model_main.py", line 192, in
@braveheartwithlove I changed the type of input file:python model_main.py --eval --eval_output=cqcc_result --model_path=cache_train_LA_cqcc.npy
Then, got following error:
Traceback (most recent call last):
File "/mnt/publicStoreA/lishaomei/asvspoof2019-master/model_main.py", line 192, in
@braveheartwithlove I solved it by add "--features cqcc" when I called model_main.py --eval --eval_output=cqcc_result --model_path=cache_train_LA_cqcc.npy --features cqcc. Thanks a lot.
@maymay1982 That is great!
@maymay1982 by the way, could you please share here your final performance using cqcc features? My EER on dev data set is pretty good. 0.7%, but the EER on eval dataset is ~17%, which is too high in my setup. Thanks
@maymay1982 Thank you for helping us. @braveheartwithlove We also want to note three things: First, initialization plays a role in model performance, so use a different random seed might help. Secondly, the best EER is given by a fusion model, so using other varients of this network might help. Finally, you can also play with the number of CCs you extract and compare their performance.
@braveheartwithlove I'm sorry that I had left the lab yesterday when I got your message. The EER of cqcc on dev dataset is 0.55, and I don't know how to test it on eval dataset. If convenient, would you please send a readme to my email: lishaomei_may@126.com? Thanks a lot.
@braveheartwithlove I solved it by add "--features cqcc" when I called model_main.py --eval --eval_output=cqcc_result --model_path=cache_train_LA_cqcc.npy --features cqcc. Thanks a lot.
Hi. The AttributeError still persists for me even after I added --features cqcc.
May I know which version of PyTorch you are using?
@braveheartwithlove In my experiments using cqcc features, the EER on dev data set is 0.55%, and the EER on eval dataset is 12.1%, which is the worst among the three features.
@maymay1982 It is similar to my simulation. In my test, Spect gives me the best performance so far.
@ddave25 I'm sorry, the right command is: python model_main.py --eval --eval_output=cqcc_result --model_path=model_logical_cqcc_100_32_5e-05.pth --features cqcc。And the version of my PyTorch is 0.4.
@maymay1982 Thank you. That does work without errors.
After run model_main.py, I got many .npy files and .pth files. But when I put these .npy files as input to run fuse_result.py, I got the following error: python fuse_result.py --input cache_train_LA_spect.npy cache_train_LA_mfcc.npy cache_train_LA_cqcc.npy --output=cache_train_LA.npy Processing input files : ['cache_train_LA_spect.npy', 'cache_train_LA_mfcc.npy', 'cachetrain LA_cqcc.npy'] Traceback (most recent call last): File "fuse_result.py", line 35, in
fuse_result = fuse(args.input)
File "fuse_result.py", line 21, in fuse
frames = [read_frame(f) for f in file_list]
File "fuse_result.py", line 21, in
frames = [read_frame(f) for f in file_list]
File "fuse_result.py", line 10, in read_frame
data_np = np.genfromtxt(fname, dtype=str)
File "/home/lishaomei/anaconda3/lib/python3.7/site-packages/numpy/lib/npyio.py", line 1761, i n genfromtxt
first_line = _decode_line(next(fhd), encoding)
File "/home/lishaomei/anaconda3/lib/python3.7/codecs.py", line 322, in decode
(result, consumed) = self._buffer_decode(data, self.errors, final)
UnicodeDecodeError: 'utf-8' codec can't decode byte 0x80 in position 0: invalid start byte
The same error happened when I put these .pth files as input to run fuse_result.py.