langs=["ru", "en"], src_vocab_size=1000, tgt_vocab_size=1000, d_model=4, encoder_layers=1, decoder_layers=1, encoder_ffn_dim=4, decoder_ffn_dim=4, encoder_attention_heads=1, decoder_attention_heads=1, ) tiny_model = FSMTForConditionalGeneration(config) print(f"num of params {tiny_model.num_parameters()}") # Test batch = tokenizer.prepare_seq2seq_batch(["Making tiny model"]) outputs = tiny_model(**batch, return_dict=True) print("test output:", len(outputs.logits[0])) # Save tiny_model.half() # makes it smaller tiny_model.save_pretrained(mname_tiny) tokenizer.save_pretrained(mname_tiny) print(f"Generated {mname_tiny}") # Upload # transformers-cli upload tiny-wmt19-en-ru
merges_file=merges_file, ) config = FSMTConfig( langs=['ru', 'en'], src_vocab_size=1000, tgt_vocab_size=1000, d_model=4, encoder_layers=1, decoder_layers=1, encoder_ffn_dim=4, decoder_ffn_dim=4, encoder_attention_heads=1, decoder_attention_heads=1, ) tiny_model = FSMTForConditionalGeneration(config) print(f"num of params {tiny_model.num_parameters()}") # Test batch = tokenizer.prepare_seq2seq_batch(["Making tiny model"], return_tensors="pt") outputs = tiny_model(**batch) print("test output:", len(outputs.logits[0])) # Save tiny_model.half() # makes it smaller tiny_model.save_pretrained(mname_tiny) tokenizer.save_pretrained(mname_tiny) print(f"Generated {mname_tiny}") # Upload # transformers-cli upload tiny-wmt19-en-ru