Merge pull request #371 from YaoYinYing/main

re-PR for pretrained ESM weights issue.

scripts/download_weights.sh

@@ -0,0 +1,60 @@
+#! /bin/sh
+# Usage: bash <repo>/esm/scripts/download_weights.sh /path/to/weights/
+# re run the command to continue downloading incomplete files.
+echo $0
+run_dir=$(readlink -f $(dirname $0)) ;
+
+if [[ "$(which aria2c)" == "" ]];then
+  echo aria2c is required for downloading!
+  exit 1
+fi
+
+for _sh_exec in /bin/sh bash zsh;do
+  # only bash and zsh are tested.
+  _sh_exec=$(which ${_sh_exec})
+  if [[ "$(${_sh_exec} --version)" =~ "bash" || "$(${_sh_exec} --version)" =~ "zsh" ]];then
+    echo $(which ${_sh_exec});
+    break
+  fi
+done
+
+model_pth=$1
+if [[ "$model_pth" == "" ]];then
+  model_pth=$PWD
+fi
+
+mkdir -p $model_pth/checkpoints
+
+pushd $model_pth/checkpoints
+cat $run_dir/../README.md |grep -e '^|' |grep -e 'fair-esm/models' |tr -d '|' | \
+  awk 'BEGIN{print "set -e"};
+  {
+      # read urls starts with https
+      split($0,arr,"https");
+      url="https"arr[2];
+
+      # remove blank spaces after url string
+      split(url,url_arr," ")
+      url=url_arr[1]
+
+      # guessing regression pt url
+      url_regression=url;
+      sub("models","regression",url_regression);
+      sub(".pt","-contact-regression.pt",url_regression);
+
+      # downloading weight
+      url_basename_idx=split(url,url_arr,"/")
+      url_basename=url_arr[url_basename_idx]
+      print "if [[ ! -f "url_basename" || -f "url_basename".aria2 ]];then echo Download not complete: "url_basename";aria2c -x 10 "url";else echo Download complete: "url_basename";fi"
+
+      # downloading regression, if not existing and through an error, we just ignore it.
+      url_regression_basename_idx=split(url_regression,url_arr,"/")
+      url_regression_basename=url_arr[url_regression_basename_idx]
+      print "if [[ ! -f "url_regression_basename" || -f "url_regression_basename".aria2 ]];then echo Download not complete: "url_regression_basename";aria2c -x 10 "url_regression" 2>/dev/null || echo Never mind. "url_regression_basename" may not exist. ;else echo Download complete: "url_regression_basename";fi"
+
+      }' |$_sh_exec
+
+
+echo "Your model directory is located at \`$(readlink -f ${model_pth})\`."
+
+popd

scripts/esmfold_inference.py

@@ -1,3 +1,6 @@
+
+from pathlib import Path
+import sys,os
 import argparse
 import logging
 import sys
@@ -83,6 +86,9 @@ def create_batched_sequence_datasest(
     parser.add_argument(
         "-o", "--pdb", help="Path to output PDB directory", type=Path, required=True
     )
+    parser.add_argument(
+        "-m", "--model-dir", help="Parent path to Pretrained ESM data directory. ", type=Path, default=None
+    )
     parser.add_argument(
         "--num-recycles",
         type=int,
@@ -121,7 +127,15 @@ def create_batched_sequence_datasest(
     logger.info(f"Loaded {len(all_sequences)} sequences from {args.fasta}")
 
     logger.info("Loading model")
+
+    # Use pre-downloaded ESM weights from model_pth.
+    if args.model_dir is not None:
+        # if pretrained model path is available
+        torch.hub.set_dir(args.model_dir)
+
     model = esm.pretrained.esmfold_v1()
+
+
     model = model.eval()
     model.set_chunk_size(args.chunk_size)