description: Evaluate VAE LM on Wiki2 Dataset auth: # which virtual cluster you belong to (msrlabs, etc.). Everyone has access to "pnrsy". vc: resrchprojvc6 # physical cluster to use (cam, gcr, rr1) or Azure clusters (eu1, eu2, etc.) # cluster: rr2, eu2, eu1 et1 cluster: rr1 # docker environment (vm) in which your job will run. we provide "generic" dockers # with the main deep learning toolkit installed (PyTorch, TF, Chainer, etc.) docker: # image: philly/jobs/custom/generic-docker:py27 # registry: phillyregistry.azurecr.io image: chunyl/pytorch-transformers:v0 registry: index.docker.io storage: _default: #use_phillyfs: True storage_account_name: textae container_name: bigtextae mount_path: /mnt/_default code: # local directory of the code. this will be uploaded to the server. # $CONFIG_DIR is expanded to the directory of this config file code_upload: False remote_dir: code/ local_dir: $CONFIG_DIR/code #data: # data upload is not required for this example #data_upload: False search: job_template: name: vq_{experiment_name:s}_{bs_option:.0f} sku: G4 # G4 # G1 command: - pip install --user --editable . - python examples/big_ae/run_lm_vae_pretraining.py --output_dir ../output/local_lm_vae_wiki2_bert_gpt --encoder_model_type bert --encoder_model_name_or_path bert-base-uncased --decoder_model_type gpt2 --decoder_model_name_or_path gpt2 --train_data_file ../data/datasets/wikitext-2/train.txt --do_eval --eval_data_file ../data/datasets/wikitext-2/valid.txt --overwrite_output_dir --per_gpu_train_batch_size {bs_option} max_trials: 20 type: grid params: - name: bs_option spec: discrete values: [1] # [top,bottom]