Error in Colab Example

[msfasha]

When running the following command, the error presented below is raised:

Beam search is the default generation method on Turjuman
!turjuman_translate --text "As US reaches one million COVID deaths, how are Americans coping?"

IndexError: too many indices for tensor of dimension 2

/usr/local/bin/turjuman_translate:8 in <module>                              │
│                                                                              │
│   5 from turjuman_cli.translate import translate_cli                         │
│   6 if __name__ == '__main__':                                               │
│   7 │   sys.argv[0] = re.sub(r'(-script\.pyw|\.exe)?$', '', sys.argv[0])     │
│ ❱ 8 │   sys.exit(translate_cli())                                            │
│   9                                                                          │
│                                                                              │
│ /usr/local/lib/python3.9/dist-packages/turjuman_cli/translate.py:76 in       │
│ translate_cli                                                                │
│                                                                              │
│   73 │                                                                       │
│   74 │   torj = turjuman(logger, args.cache_dir)                             │
│   75 │   if input_source=="text":                                            │
│ ❱ 76 │   │   torj.translate_from_text (args.text, args.search_method, args.s │
│   77 │   elif input_source=="file":                                          │
│   78 │   │   torj.translate_from_file (args.input_file, args.search_method,  │
│   79                                                                         │
│                                                                              │
│ /usr/local/lib/python3.9/dist-packages/turjuman/turjuman.py:93 in            │
│ translate_from_text                                                          │
│                                                                              │
│    90 │   │   outputs = self.translate(sources, search_method, seq_length, m │
│    91 │   │                                                                  │
│    92 │   │   if max_outputs==1:                                             │
│ ❱  93 │   │   │   targets = outputs['target'][0]                             │
│    94 │   │   else:                                                          │
│    95 │   │   │   targets = outputs[str(max_outputs)+'_targets'][0]          │
│    96 │   │   if type(targets) == list:     

[obada-jaras]

same

[obada-jaras]

Use this Colab example. Note that in this example, the output is the token ids. To solve this problem, you have to follow this comment.

Also, I faced another error and solved it using these two lines of code:

import torch
torch.set_default_tensor_type('torch.cuda.FloatTensor')

I think you have to use this only if you select the runtime to GPU.

---

Here's my final code that runs properly:

import torch
torch.set_default_tensor_type('torch.cuda.FloatTensor')

(2) Translate using beam search (default)

beam_options = {"search_method":"beam", "seq_length": 1024, "num_beams":5, "no_repeat_ngram_size":2, "max_outputs":3}
target = torj.translate("As US reaches one million COVID deaths, how are Americans coping?",**beam_options)
result = torj.tokenizer.batch_decode(target, skip_special_tokens=True)
print (result)

(3) Translate using greedy search

greedy_options = {"search_method":"greedy", "seq_length": 1024}
target = torj.translate("As US reaches one million COVID deaths, how are Americans coping?",**greedy_options)
result = torj.tokenizer.batch_decode(target, skip_special_tokens=True)
print (result)

(4) Translate using sampling search

sampling_options = {"search_method":"sampling", "seq_length": 1024, "max_outputs":3, "top_p":0.95, "top_k":50}
target = torj.translate("As US reaches one million COVID deaths, how are Americans coping?",**sampling_options)
result = torj.tokenizer.batch_decode(target, skip_special_tokens=True)
print (result)